text
stringlengths 1
6.27k
| id
int64 0
3.07M
| raw_id
stringlengths 2
9
| shard_id
int64 0
0
| num_shards
int64 16
16
|
|---|---|---|---|---|
stage, respectively. L2, P2, E2 and Z2 represent flag leaves, anther, embryo sac and leaf sheath at pre-flowering stage, respectively. L3, P3 and Z3 represent flag leaves, grain and leaf sheath at three days after flowering, respectively Co-expression network analysis of differentially expressed genes in different tissues by weighted gene coexpression network analysis (WGCNA) WGCNA, which is a systems biology tool, was used to understand the relationships and networks in a set of genes. In this study, WGCNA was constructed using RNA-seq data, and 28 WGCNA modules were identified (Fig. 7a, Additional file 20: Table S12). The gene numbers in these modules were ranged from 30 (MEwhite module) to 9294 (MEgrey module). Interestingly, the turquoise and grey modules consist of 67.72% of the genes in the network analysis. We found that some modules showed correlation with the different tissues in F 1 and parents (Fig. 7b), for example, MEbrown module in leaf, MEblue module in mature anther, MEred module in mature embryo sac, and MEturquoise and MEgrey module in three tissues (anther, embryo sac and grain), which indicated that these modules may play putatively important roles in tetraploid rice leaf and reproductive organs. Furthermore, a total of 1335 genes were involved in the MEbrown module, and GO enrichment analysis showed significant enriched terms that were related to photosynthesis light harvesting, light reaction, carboxylic acid metabolic process and chlorophyll metabolic process. These results indicated that brown module genes may play an important role in the photosynthesis in tetraploid rice. In total, 9291 genes were involved in the | 3,069,000 | 149445636 | 0 | 16 |
MEturquoise module, and GO analysis revealed significant terms associated with DNA repair, carbohydrate metabolic process and transport, which indicated that MEturquoise module plays an important role in the fertility and yield of tetraploid rice (Additional file 21: Table S13). Association between DNA sequence variations and differentially expressed genes in F 1 hybrid compared to parents A total of 100,395,344 and 150,936,032 clean reads were obtained in T449 and H1 by using genome re-sequencing, respectively. Approximately 98.16% (T449) and 96.11% (H1) of clean reads were mapped onto the Nipponbare reference genome, and the reads coverage depths were 35× and 49× in T449 and H1, respectively (Additional file 22: Table S14). A total of 912,892 SNPs and 195,976 InDels were detected between T449 and H1 by using the two filter conditions (coverage ≥10 and ≤ 100, and removal of heterozygous SNPs and InDels). We found that about 5% of SNPs and 6% of InDels were detected in intergenic regions, and 60% SNPs and 65% InDels were identified in up or down regulatory regions, which might be related to the differentially expressed genes (Additional file 23: Table S17). These results were nearly consistent with the GO enrichment of DEG FPU . Expression patterns of saccharides metabolism and starch synthase related genes in the hybrid compared to parents The GO and KEGG analyses of DEG FPU showed that there were significant differences for carbohydrate metabolic process in nine tissues between F 1 and its parents, and DEG FPU were involved in sucrose synthase, cell wall invertase, 6-phosphofructokinase, and hexokinase. Many | 3,069,001 | 149445636 | 0 | 16 |
saccharide metabolic genes were up-regulated in the F 1 compared to its parents in L1, P1 and P3 (Additional file 26: Table S18). Interestingly, the saccharide transporters were up-regulated in the F 1 compared to its parents in L1, P1 and P3, and these results were consistent with saccharide metabolic genes (Additional file 26: Table S18). The two saccharide transporter genes (LOC_Os02g10800 and LOC_Os03g07480) were also up-regulated in F 1 compared to both parents in P3 (Additional file 26: Table S18). In addition, the invertase (OsINV3 and OsINV4), sucrose synthase (OsSUS3 and OsSUS4), hexokinase gene (OsHXK6), starch branching enzyme (OsBEIIb) and two starch synthase genes (OsSSIIIa and wx) displayed higher levels of expressions in F 1 than parents in the grains (three days after flowering) (Fig. 8a). Moreover, the promoter regions of OsSUS3, OsBEIIb and OsSSIIIa also exhibited differences between maternal and paternal rice lines by re-sequencing (Fig. 8b). High heterosis and the frequency of bivalents in F 1 hybrid harboring double neutral genes In the current study, the paternal line was high fertility neo-tetraploid rice and maternal line was autotetraploid rice harboring double neutral genes. The hybrid displayed stronger heterosis for yield and yield-related traits, such as filled grains per plant, total grains per plant, grain yield per plant and seed setting. These results were consistent with the previous studies, where autotetraploid hybrids exhibited high heterosis for filled grains per panicle, grain yield per plant and seed setting (Shahid et al. 2012;Guo et al. 2017). In addition, the pollen and embryo sac fertilities were investigated | 3,069,002 | 149445636 | 0 | 16 |
to evaluate the fertility of hybrid and its parents. Our results showed that the embryo sac fertility of hybrid and its parents was higher than 89%, and pollen fertility of hybrid and paternal line was high, while pollen fertility of maternal line was low. These results revealed that pollen fertility has a greater impact on seed setting than embryo sac fertility in autotetraploid rice hybrid and its parents. It is well known that meiosis process has a great effect on plant reproductive development, and chromosome behavior and configuration play an important role in the plant meiosis and directly correlated with pollen fertility. The way of quadrivalent separation depends on chromosome configuration. Many quadrivalent Fig. 6 Gene expression levels of 47 genes and predicted protein-protein interaction network. a, The distribution of 47 genes exhibited up-regulation in F 1 hybrid and H1 compared to T449 and down-regulation in T449 compared to E249 during meiosis stage, b, Predicted protein-protein interaction network of differently expressed genes (black), meiosis-specific (blue) and meiosis-related (red) genes. T represents T449, F represents hybrid, and H represents H1 chromosomes were found in Triticum monococcum during diakinesis and metaphase I (Kim and Kuspira 1993), while chain, ring and frying pan shapes were three main types of chromosome configurations in autotetraploid rice (Luan et al. 2007;He et al. 2011a). Ring, chain, frying pan, "X" and "OK" shapes were observed in the present study, and we have drawn these shape models based on the observation. The ring shape quadrivalent was found more frequently compared to other quadrivalent types | 3,069,003 | 149445636 | 0 | 16 |
in the present study. Our results were in agreement with the previous studies, who also observed ring shape quadrivalent in autotetraploid rice (Luan et al. 2007;He et al. 2011b). He et al. (2011b) revealed that high frequency of bivalent was related to high pollen fertility and seed setting in autotetraploid rice hybrid. We also observed higher frequency of bivalents in the autotetraploid hybrid than parents at diakinesis and metaphase I. Chromosome behavior had a direct relationship with pollen fertility and seed setting in autotetraploid rice hybrids and interspecific hybrid between Brachiaria ruziziensis and B. brizantha (Adamowski et al. 2008;He et al. 2011b;Guo et al. 2016). Here, the frequency of normal chromosome behavior was significantly higher in the hybrid with high fertility than maternal line with double neutral genes. It is worth to mention that the frequency of abnormal chromosome behavior is much higher during anaphase II Fig. 7 WGCNA based gene expression matrix between hybrid and parents. a Hierarchical cluster tree showing co-expression modules identified by WGCNA. Each leaf in the tree represents one gene. The major tree branches constitute 28 modules labeled with different colors. b Modulesample relationship. Each row corresponds to a module. Each column corresponds to a sample. The color of each cell at the row-column intersection indicates the correlation coefficient between the module and the sample than other stages in this study. Here, many types of abnormal chromosome behaviors, including asynchronous meiocytes, abnormal spindles and straggling chromosomes were observed at anaphase II (Additional file 5: Table S3). Of these abnormalities, asynchronous meiocytes | 3,069,004 | 149445636 | 0 | 16 |
was the highest, and the frequencies of asynchronous meiocytes were 66.9%, 65.3% and 56.0% in T449, F 1 and H1, respectively. We inferred that most of asynchronous meiocytes could develop normal tetrad according to their morphology, which might be the major reason for lower frequency of the normal cells at anaphase II than other phases. The expression patterns of meiosis and meiosis-related genes promote high fertility in F 1 hybrid A number of meiosis-related and meiosis-specific genes were detected in rice (Fujita et al. 2010;Tang et al. 2010;Deveshwar et al. 2011;Yant et al. 2013;Luo et al. 2014;Wright et al. 2015). A total of 55 meiosis-related or meiosis-stage-specific genes were found to be down-regulated, which increased pollen sterility loci interactions in autotetraploid rice hybrids (Wu et al. 2015). In the present study, the meiotic stages were determined by the floret length and according to the observation by 4′, 6-diamidino-2-phenylindole (DAPI) staining (He et al. 2011a). The DAPI staining could clearly distinguish between meiotic stages and other pollen development stages. For the proper understanding of meiosis-related and meiosis-specific genes between hybrid and parents, we dissected anthers from floret for RNA-seq analysis. A total of four meiosis-related and 26 meiosis-stage specific genes were identified, which were found to be up-regulated in hybrid and paternal line compared to maternal line. Interestingly, of the four meiosis-related and 26 meiosis-specific genes, two meiosis-related and 19 meiosis-specific genes were also found to be down-regulated in autotetraploid rice compared to diploid progenitors (Chen et al. 2018a). For example, DPW gene encodes a fatty acyl | 3,069,005 | 149445636 | 0 | 16 |
ACP reductase, and was found to be essential for anther cuticle, pollen wall and pollen sporopollenin biosynthesis (Shi et al. 2011), and maternal and paternal lines showed upstream and intron variations in DPW gene. OsACOS12, which is an acyl-CoA synthetase, is essential for sporopollenin synthesis in rice (Li et al. 2016a;Yang et al. 2017), and maternal and paternal lines displayed downstream, upstream, intron and non-synonymous variations in OsACOS12. PDA1 encodes an ABC transporter (OsABCG15) and required for the transport of lipidic precursors for anther cuticle and pollen exine development (Zhao et al. 2015). CYP703A3, cytochrome P450 hydroxylase, is involved in the tapetum degeneration retardation, a known pollen exine formation (Yang et al. 2014). The meiosis-related gene (LOC_Os12g24420) encoded cyclin-dependent kinase, which is homolog to CDGK1 in Arabidopsis, and CDKG1 protein kinase is crucial for synapsis and recombination in Arabidopsis during meiosis (Zheng et al. 2014), and we observed changes in the downstream and intron region of CDKG1 in maternal and paternal lines. These results suggested that the expression profiles of important meiosis-related or meiosis-specific genes have a significant effect on the fertility of polyploidy hybrid rice. Dominance, non-additive and yield-related genes/QTLs contribute to heterosis The DEG FP were divided into five basic groups according to their expression profiles, including over-dominance (HBP), under-dominance (LBP), dominance (CHP and CLP), and mid-parent (BBP) . In our data, the dominance expression was the most prevalent class among DEG FP (46.34-77.71%). Similarly, the dominance expression patterns were found to be the most abundant among DEG FP in wheat and rice hybrids | 3,069,006 | 149445636 | 0 | 16 |
by RNA-seq analysis (Zhai et al. 2013;Liu et al. 2018). These results indicated that dominance expression have a great effect on the performance of hybrids. According to gene expression levels of hybrid and its parents, the gene expression profiles of F 1 could be divided into two types, the first type is called as additive expression, which is contributed by each allele from its parents in a hybrid, and another is non-additive expression that differed from the mid-parent value (MPV) (Wei et al. 2009). In the previous study, whether or not a transcript shows non-additive expression is most likely to be affected by the contributions of cisand trans-acting element of a gene (Zhang et al. 2008). In this study, NAGs only accounted for 2.1-6.5% of the total expressed genes, but 57.3-72.4% of DEG FPU were NAGs in each tissue. These results were consistent with diploid rice (Wei et al. 2009), wheat ) and maize heterosis (Swanson-Wagner et al. 2006), who also detected NAGs in F 1 . The previous studies have shown that NAGs play vital roles in heterosis (Zhang et al. 2008;Liu et al. 2018), and NAGs were associated with circadian rhythm, flowering time, and panicle branching in rice (Li et al. 2016b). Overall less number of NAGs detected in this study, but major portion of DEG FPU was constituted of NAGs. Therefore, we speculated that NAGs play important roles in high F 1 heterosis of polyploid rice. The potential relationships between differently expressed genes and QTLs have been proposed in many yield-related QTL regions | 3,069,007 | 149445636 | 0 | 16 |
using RNA-seq (Zhai et al. 2013;Chen et al. 2018b). A recent study showed that nine genes (Hd3a, TAC1, Ghd8, Sd-1, NAL1, Hd1, GW6a, IPA1 and DEP1) have a major impact on heterosis in diploid hybrid rice (Huang et al. 2016). In our study, DEP1, which is related to rice panicle and located in SKPNB (spikelet number, AQBK037), was found to be up-regulated in grains (three days after flowering) of F 1 hybrid compared to maternal line and also up-regulated in leaf of hybrid than parents during meiosis stage. GW6a, which involved in grain-weight and located in TSDWT (1000-seed weight, AQEB012), revealed much higher expression in anthers of hybrid than parents during meiosis. The semi-dwarf gene (Sd-1), which is involved in biosynthesis of gibberellin and located in GRYLD (grain yield, AQQ005), exhibited higher levels of expressions in leaf sheath (three day after flowering) of hybrid than parents. Hd3a, which is related to rice flowering and regional adaptation, and located in FGRNB (filled grain number, AQCF008), was up-regulated in leaf (before flowering) of hybrid compared to parents. These results showed that these candidate genes in QTL regions may contribute to heterosis in autotetraploid hybrid rice. Saccharides metabolism and starch synthase related genes play an important role in heterosis Carbohydrate metabolism plays an essential role in the plant growth and development. RNA-seq analysis showed that processes of carbohydrate metabolism are related to heterosis in rice (Wei et al. 2009;Zhai et al. 2013). In addition, our research group reported that abnormal distribution of saccharides and saccharides-related genes may influence pollen | 3,069,008 | 149445636 | 0 | 16 |
fertility and cause decrease in the yield of autotetraploid rice (Chen et al. 2018a). Here, the DEG FPU was significantly enriched in the carbohydrate metabolism process between hybrid and its parents in nine tissues sampled across different development stages. The grain filling stage is an important part of growth and development in rice, and a large quantity of carbohydrates are synthesized and transported into the grain at this stage, particularly the embryo starts exponential growth at three days after flowering that shows dual rhythmicity (Itoh et al. 2005). The stage of three days after flowering is important stage for grain development, so we focused on saccharides metabolism in the grains three days after flowering. The carbohydrates supply from the leaves to pollen or grain involves sucrose transport and degradation, monosaccharides formation and transport, and starch generation (Ruan 2012). There are two types of enzymes that catalyze the sucrose degradation in plants, one is sucrose synthase (SUS) and the other is invertase (Ruan et al. 2010). Here, the invertase (OsINV3 and OsINV4) and sucrose synthase (OsSUS3 and OsSUS4) were found to be up-regulated in the grains of hybrid compared to parents. After sucrose degradation, the resulting hexoses undergo phosphorylation by hexokinase for starch synthesis. Subsequently, hexokinase plays important role in hexose signaling and sensing (Cho et al. 2009;Kim et al. 2016). The hexokinase gene (OsHXK6) was up-regulated in grains (three after days flowering) of the hybrid compared to its parents. The starch synthase, starch debranching and starch branching enzyme have a great influence on the starch generation | 3,069,009 | 149445636 | 0 | 16 |
and metabolism (Zeeman et al. 2010). The starch branching enzyme (OsBEIIb) and two starch synthase (OsSSIIIa and wx) genes were up-regulated in the grains of hybrid compared to parents. In addition, we detected differences between maternal and paternal rice lines in the promoter regions of OsSUS3, OsBEIIb and OsSSIIIa by re-sequencing. Consequently, the genetic effects of OsSUS3, OsBEIIb and OsSSIIIa may cause allelic heterozygosity in promoter regions of hybrid. Sucrose and monosaccharide transporters are important proteins for the translocation of saccharides from source to sink organs (Ruan et al. 2010). OsSUT1 primarily play a role in the transport pathway (Scofield et al. 2007). In our study, sucrose transporter (OsSUT1) displayed much higher expression patterns in the grain (three after days flowering) of hybrid than parents. OsBT1, which encodes putatively ADP-glucose transporter and localizes in the amyloplast envelope membrane, plays a crucial role in starch synthesis (Cakir et al. 2016;Li et al. 2017a). We found that OsBT1 was up-regulated in the grain (three after days flowering) of hybrid compared to parents. Transcriptome profiling showed high expression levels of saccharides metabolism and starch synthase related genes in the hybrid, which might be an indication of enhanced source for sink tissues and resulted in high yield of hybrid. In our previous studies, we found that the double neutral genes can overcome the hybrid sterility in autotetraploid rice , and detected specific differentially expressed genes associated with fertility and heterosis in neo-tetraploid rice by RNA-seq (Guo et al. 2017). Here, an autotetraploid rice line (T449), harboring Sa-n and Sb-n double | 3,069,010 | 149445636 | 0 | 16 |
neutral genes for pollen sterility loci, was used to generate the hybrids by crossing with neo-tetraploid rice, and investigated the heterosis and fertility by cytological and RNA-seq methods. We further want to understand the role of double neutral genes in heterosis and fertility of neo-tetraploid rice. Therefore, we observed chromosome behavior and gene expression patterns during important growth stages. The results showed that seed setting of F 1 hybrid improved with the increase in number of bivalents, and many important genes, including meiosis-related and meiosis-specific genes and saccharides metabolism and starch synthase related genes, exhibited heterosis specific expression patterns in polyploid rice during different development stages. Conclusions In this study, we observed the chromosome behavior and configuration in hybrid and its parents, and found higher frequency of bivalent and normal chromosome behavior in hybrid than parents, which promoted high fertility (heterosis) in the hybrid harboring double neutral genes. Furthermore, we systematically investigated the global transcriptome of hybrid and its parents by RNA-seq. We obtained a large number of DEG FPU , and detected substantial candidate genes, including meiosis-related and meiosis-specific genes, saccharides metabolism and starch synthase related genes, which were up-regulated in hybrid having improved fertility and yield. Our results provided new resource for polyploid rice breeding and exploring of these candidate genes will provide valuable information for revealing molecular mechanisms of heterosis in polyploidy rice. Rice material An autotetraploid rice line, DN18-4x (T449), harboring Sa-n and Sb-n double neutral genes for pollen sterility loci, was used to generate the hybrids by crossing with five neo-tetraploid | 3,069,011 | 149445636 | 0 | 16 |
rice lines, including Huaduo 1 (H1), Huaduo 2 (H2), Huaduo 3 (H3), Huaduo 4 (H4) and Huaduo 8 (H8). All the materials were planted at the experimental farm of South China Agricultural University (SCAU) under natural conditions, and management practices followed the recommendations for the area. Investigation of agronomic traits and data analysis Agronomic traits, including plant height, effective number of panicles per plant, grain length and width, 1000-grain weight, filled grains per plant, total grains per plant, grain yield per plant and seed setting, were investigated. The standard for investigating these agronomic traits was according to the protocols of People's Republic of China for the registration of a new plant variety DUS (Distinctness, Uniformity and Stability) test guidelines of rice (Guidelines for the DUS test in China, 2012) Guo et al. 2017). The mid-parent heterosis (MPH) and high-parent heterosis (HPH) were estimated by the following formula: MPH = (F 1 − MP)/MP × 100%, and HPH = (F 1 − HP)/HP × 100%, where F 1 related to the performance of hybrid, HP was defined as the performance of the best parent, and MP was defined as an average performance of two parents . Cytological observation The chromosome configuration and behaviors were observed according to Wu et al. (2014). The inflorescences of F 1 and its parents lines were collected from the shoots of rice plants with 0 to 4 cm between their flag leaf cushion and the second to last leaf cushion, and fixed in Carnoy's solution (ethanol: acetic acid = 3:1) for 24 | 3,069,012 | 149445636 | 0 | 16 |
h, and the samples were stored in 70% ethanol at 4°C after washing two times. The anther was removed from the floret and placed in a small drop of 1% acetocarmine on a glass slide. The glass slide was covered with a slide cover after 2-3 min, and observed under a microscope (Motic BA200). The pollen fertility was observed according to Shahid et al. (2013b). The normal and abnormal pollens were observed by staining with 1% I 2 -KI under a microscope (Motic BA200). The whole mount eosin B confocal laser scanning microscopy (WE-CLSM) was used to investigate the embryo sac fertility in F 1 and its parents according to Li et al. (2017b) with minor modifications. The ovary was dissected from the floret, and was hydrated in 70%, 50%, 30%, 10% ethanol and distilled water for 30 min each. Then, the samples were dehydrated in 10%, 30%, 50%, 70%, 90% and 100% ethanol for 30 min after eosin B staining for 12 h. Finally, the samples were shifted into a mixture solution (ethanol and methyl salicylate = 1:1) for 2 h, and then keep in pure methyl salicylate and observed under a laser scanning confocal microscope (Leica SPE). RNA-seq experiments and data analysis All samples were collected during meiosis, pre-flowering, and three days after flowering. The meiosis stage is a crucial event for the sexual reproduction of eukaryotes to form haploid spores and gametes ). The pre-flowering stage is an important stage for pollen and embryo sac fertility. The carbohydrates are synthesized and transported into | 3,069,013 | 149445636 | 0 | 16 |
the grains in large quantity during grain filling stage (Itoh et al. 2005). The nutrients produced by leaves are transported to other organs through the leaf sheath, so the leaf sheath has an important role in the transportation of energy. Flag leaf is one of the most important photosynthetic organs in rice and has an important impact on crop yield and quality. In addition, pollen and embryo sac have a significant impact on rice fertility and yield (Shahid et al. 2013b;Wu et al. 2015;Li et al. 2017b). Hence, we collected the nine tissues during these development stages from hybrid and its parents in three biological replicates, including anthers (P1) and flag leaves (L1) at meiosis stage, and flag leaves (L2), leaf sheath (Z2), anther (P2) and embryo sac (E2) at pre-flowering stage, and flag leaves (L3), leaf sheath (Z3) and grains (P3) at three days after flowering (Additional file 4: Figure S2). All tissues of hybrid and its parents were harvested in three biological replicates and immediately kept at − 80°C for RNA extraction. The total RNA was extracted according to the manual instructions of the TRIzol Reagent (Life technologies, California, USA). The library was prepared according to the vendor's recommended protocol. The RNA-seq was performed on the Illumina HiSeq 4000 sequencing platform (LC Sceiences, USA). Using the Illumina paired-end RNA-seq approach, we sequenced the transcriptome that generated millions of paired-end reads. Low quality reads, including reads containing sequencing adaptors, reads containing sequencing primer and nucleotides with quality score lower than 20, were removed. The mapped | 3,069,014 | 149445636 | 0 | 16 |
reads from each sample were assembled using StringTie, and all transcriptome samples were mixed to reconstruct a comprehensive transcriptome by perl scripts. After the generation of transcriptome, the StringTie and Ballgown were used to evaluate the gene expression levels. String-Tie was used to perform expression level for mRNAs by calculating FPKM (fragments per kilobase of transcript per million fragments mapped reads), and false discovery rate (FDR) was used to determine the threshold of the P-value in multiple tests. The Venny software was used to identify the overlapped differentially expressed genes in different samples (http:// bioinfogp.cnb.csic.es/tools/venny/index.html). Hierarchical analysis was carried out for all genes using Cluster 3.0 software after normalization. Transcription factor analysis was done according to transcription factor data (Jin et al. 2017). Gene Ontology (GO) enrichment analysis was employed for functional categorization by using AgriGO tool (http://systemsbiology.cau.edu.cn/agriGOv2/). Expression patterns of differentially expressed genes (DEGs) The expression patterns of DEGs were defined according to Liu et al. (2018). We defined the gene expression in F 1 as EF 1 , and genes expression of both parental lines as ET449 and EH1.We defined the average value of both parental lines as MPV (mid-parental value). If the F 1 was significantly (FDR ≤ 0.05 and fold change ≥2) different from MPV, we defined these genes as non-additive genes (NAGs), if there was non-significant difference between F 1 and MPV, these genes were defined as additive genes. Classification of DEG FP was performed according to the expression of EF1 relative to ET449 and EH1. ">" and "<" represents statistically | 3,069,015 | 149445636 | 0 | 16 |
higher or lower, and "=" represents statistically similar. If EF1 > ET449 > EH1, or EF1 > EH1 > ET449, then expression patterns of these genes were considered as higher than both parents (HBP); if EF1 = ET449 > EH1, or EF1 = EH1 > ET449, then expression patterns of these genes were considered as close to higher parent (CHP); if ET449 > EF1 > EH1, or EH1 > EF1 > ET449, then expression patterns of these genes were considered as between two parents (BBP); if EF1 = ET449 < EH1, or EF1 = EH1 < ET449, then expression patterns of these genes were considered as close to lower parent (CLP); if EF1 < ET449 = EH1, or EF1 < ET449 < EH1, or EF1 < EH1 < ET449, these genes were considered as lower than both parents (LBP). Real-time qRT-PCR analysis A total of 12 DEGs were randomly selected for validation of RNA-Seq data by qRT-PCR. The gene-specific primers were designed using Primer Premier 5.0 software, and checked in the NCBI (National Center for Biotechnology Information) for specific primers (Additional file 27: Table S19). Total RNA was taken from sequenced samples, and the first-strand cDNA was synthesized using the Transcriptor cDNA Synth. Kit 1 (Roche) according to the manufacturer's instructions. The qRT-PCR reaction procedure was 30 s at 95°C, with 40 cycles of 95°C denaturation for 10 s and 60°C annealing and extension for 30 s, and performed on the Light-cycler480 system (Roche). The genes relative expression levels were calculated using the 2-ΔΔCt method (Livak | 3,069,016 | 149445636 | 0 | 16 |
& Schmittgen 2001). All qRT-PCR reactions were performed in triplicate. Mapping DEG FPU to rice QTLs and weighted gene coexpression network analysis (WGCNA) Rice QTL data with physical positions on the MSU Rice Genome Annotation Project Release 6.1 were acquired from Gramene (Youens-Clark et al. 2011). The DEG FPU were mapped onto 1019 yield related QTLs and 26 yield-related traits using gene coordinates from the MSU Rice Genome Annotation Project. The gene co-expression networks were used WGCNA package in R (Langfelder & Horvath 2008). To reduce noise, genes with total FPKM < 5 in 81 samples were removed. The modules were obtained using the automatic network construction with default settings. | 3,069,017 | 149445636 | 0 | 16 |
Estimating the Above-Ground Biomass in Miombo Savanna Woodlands (Mozambique, East Africa) Using L-Band Synthetic Aperture Radar Data The quantification of forest above-ground biomass (AGB) is important for such broader applications as decision making, forest management, carbon (C) stock change assessment and scientific applications, such as C cycle modeling. However, there is a great uncertainty related to the estimation of forest AGB, especially in the tropics. The main goal of this study was to test a combination of field data and Advanced Land Observing Satellite (ALOS) Phased Array L-band Synthetic Aperture Radar (PALSAR) backscatter intensity data to reduce the uncertainty in the estimation of forest AGB in the Miombo savanna woodlands of Mozambique (East Africa). A machine learning algorithm, based on bagging stochastic gradient boosting (BagSGB), was used to model forest AGB as a function of ALOS PALSAR Fine Beam Dual (FBD) backscatter intensity metrics. The application of this method resulted in a coefficient of correlation (R) between observed and predicted (10-fold cross-validation) forest AGB values of 0.95 and a root mean square error of 5.03 Mg·ha−1. However, as a consequence of using bootstrap samples in combination with a cross validation procedure, some bias may have been introduced, and the reported cross validation statistics could be overoptimistic. Therefore and as a consequence of the BagSGB model, a measure of prediction variability (coefficient of variation) on a pixel-by-pixel basis was also produced, with values ranging from 10 to 119% (mean = 25%) across the study area. It provides additional and complementary information regarding the spatial distribution of the | 3,069,018 | 18039205 | 0 | 16 |
error resulting from the application of the fitted model to new observations. Introduction Forests play an important role in the global carbon (C) cycle, and their relation to anthropogenic and climate changes have been recognized in the literature (e.g., [1,2]).Also, the importance of accurately reporting the C content (biomass) of forested lands over time has been acknowledged by several studies (e.g., [3,4]) and is a requirement of international conventions (e.g., the United Nations Framework Convention on Climate Change, UNFCCC).Such information is critical, as it forms the basis of reporting to mechanisms developed under the UNFCCC, such as the initiatives focusing on Clean Development Mechanisms (CDM) and the voluntary post-Kyoto Protocol policy mechanism on Reducing Emissions from Deforestation and forest Degradation (REDD) in developing countries [5], an economic instrument to provide incentives for reducing emissions from the forest Reference Emissions Levels/Reference Levels (RELs/RLs) benchmarks, while providing co-benefits in terms of biodiversity and livelihoods [6,7], and a core issue under the ongoing climate negotiations [8]. Sensors onboard orbital platforms provide the only means of observing the Earth from a global and systematic perspective and, particularly, of assessing its different components, namely, land-use and land-cover change (LULC), forest monitoring and C stocks [9,10].Mapping and understanding the spatial distribution of forest above-ground biomass (AGB) using remote-sensing methods is an important and challenging task [11][12][13].These maps can be used to monitor forests (deforestation, regrowth and degradation processes), to estimate and model greenhouse gas emissions and the effects of conservation actions, sustainable management and enhancement of C stocks [4,14].However, most of the | 3,069,019 | 18039205 | 0 | 16 |
attempts to estimate forest AGB are approximations relying on a combination of land cover type and corresponding mean C values derived from field surveys, instead of spatially explicit biomass maps (e.g., [15]). A few key active sensors onboard orbital and aerial platforms are providing useable information for forest AGB estimation, which could support both countries and industry in their endeavor of setting benchmarks and assess their performance in implementing a series of forest-related activities.Relationships have been established between forest AGB and (a) the backscattering coefficients of Synthetic Aperture Radar (SAR) data (different frequencies and polarizations, e.g., [16]) and (b) the vertical/horizontal distribution of Light Detection and Ranging (LIDAR) returns; (e.g., [17]).The common link between these sensors is that they offer information on the three-dimensional distribution of plant elements because of the penetrative capability [15].Other approaches have utilized tree or canopy height, from LIDAR, SAR interferometry (InSAR) and SAR polarimetric interferometry (PolInSAR), as a surrogate for biomass estimation (e.g., [18][19][20]).However, few approaches for consistent and reliable retrieval of forest AGB are available, as most suffer, for example, from saturation of the signal above certain levels, insufficient coverage, inconclusive or non-repeatable relationships, because of seasonal variation in leaf cover or moisture content and a lack of in situ data to support their calibration/validation (e.g., [15]).Consequently, there is a great uncertainty related to the estimation of forest aboveground biomass, especially in the tropical and sub-tropical regions (e.g., [21]).However, in this study, we bring forth an example of a human modified low carbon ecosystem in southern Africa dominated by Miombo | 3,069,020 | 18039205 | 0 | 16 |
woodlands, the largest savanna in the world [22,23].This ecosystem is strongly influenced by anthropogenic fires and supports the livelihoods of over 100 million people, while at the same time, it is also greatly threatened by desertification processes, deforestation, degradation of land and water resources and loss of biodiversity. The main goal of this study was to test a combination of field data and Japan's Advanced Land Observing Satellite (ALOS) Phased Array L-band Synthetic Aperture Radar (PALSAR) backscatter intensity data to reduce the uncertainty in the estimation of forest AGB in the Miombo savanna woodlands of Mozambique (East Africa).The penetrative capability of L-band synthetic aperture radar (SAR) data and the resulting interaction with vegetation structure is the main motivation for using these data for modeling forest AGB with a view to propose an alternative methodological approach that could be applied for the assessment of national, regional or local forest C stocks with a reduced uncertainty in the reported estimates. The paper is structured as follows.Section 2 provides the framework and the scientific background of the research, including a brief description of the study area.Section 3 describes the field data collection, as well as the remote sensing data used in this study.Section 4 includes a description of the machine learning method used to model forest aboveground biomass as a function of ALOS PALSAR data.The results and corresponding discussion are presented in Section 5, and the study is concluded in Section 6. Background and Study Area The study area is located in the central region of Mozambique, Zambézia province, | 3,069,021 | 18039205 | 0 | 16 |
district of Lugela, near the village of Mocuba (Figure 1) (16°04′S-17°16′S, 36°33′E-37°28′E).The climate is sub-humid, according to the Thorntwaite climate classification system.The dry season extends from March to November, the average annual precipitation ranges from 850 mm to 1,300 mm and the mean temperature from 20 °C to 27 °C. The region is mostly covered by Miombo forests, the most extensive tropical savanna woodland formation of Africa that extends across some of the world's poorest countries [22] and directly supports the livelihoods of the local populations.The main economic activity in the Lugela district is agriculture, which accounts for around 82% of the economically active population, followed by hunting activities, wood collection (for construction timber, charcoal production and domestic fuel), collection of medicinal plants, palm wine extraction, sand extraction and fishing.During the past year, an energy company has been making some efforts to understand and support the livelihoods of local populations (who are still recovering from decades of war) in the context of a suitability assessment for the production of biofuel in the region.For that and as a first step, lacking infrastructures were built, namely roads and a bridge.A detailed assessment was carried out to determine the impact of a plantation with Jatropha curcas L. (a non-edible plant whose oil-rich seeds can be processed into biodiesel).The best location for the installment of the plantation in a ~10,000 ha parcel of land was decided, complying with the sustainability criteria listed in the Directive 2009/28/EC of the European Parliament [24].Under this Directive, the energy from biofuel cannot be derived | 3,069,022 | 18039205 | 0 | 16 |
from raw material obtained from land with a high carbon stock.Therefore, all continuously forested areas, i.e., land spanning more than one hectare with trees higher than five meters and a canopy cover of more than 30% could not be used as plantation areas, and evidence of the low carbon stock of the area with a canopy cover between 10% and 30% before conversion must be provided, thus contributing to the sustainable management (social and environmental) of this area of interest and, ultimately, contributing to the company's carbon emissions reduction targets as a whole.The ~10,000 ha parcel of land was therefore divided into different classes of land use and tree canopy cover (see Section 3.1 for details) to assess the feasibility of the plantations and biofuel production in this area.This study focused on the areas with a tree canopy cover below the 50% threshold, corresponding to 16.5% of the total area. Field Measurements Sampling was designed to accurately account for the total biomass C stocks in the selected carbon pools and was stratified using ancillary data provided from on-screen classification of very high spatial resolution satellite data (WorldView 2) images into four strata with different tree crown cover (0-10, 10-30, 30-40, 40-50%).A systematic grid of points with a random origin was created covering the Miombo area with a tree canopy cover between 10% and 50% in the ~10,000 ha parcel of land and used as a basis for plot location.A total of 51 circular plots with a 20-m radius were randomly selected over the grid and measured | 3,069,023 | 18039205 | 0 | 16 |
during a field campaign that took place in July and August 2011.At the plot level, the following data were recorded: geographic coordinates (with global positioning system, GPS), physiographic location, dominant aspect and slope.At the stand level, data recorded included the tree crown cover, litter layer, down dead wood, sampling of live non-tree vegetation (herbaceous plants and shrubs) and soil organic carbon, up to 30 cm.At tree level, species were identified, tree vitality classified and diameter at breast height (DBH) measured for all trees with a DBH greater than 5 cm. Several C pools were estimated, namely the live above-and below-ground tree biomass and soil biomass.Root biomass (i.e., BGB) was not measured in the field, but assessed through a root-to-shoot ratio (R:S = 0.42) reported by Ryan et al. [25].As this study focuses only on the estimation of forest AGB, all the other pools measured in the field (down dead wood, litter, soil) were not included in the following sections. (1) The allometric equation developed by Ryan et al. [25] for estimating the C content of the AGB of each tree is given by Equation (1): where B stem is the tree AGB (kg•C), DBH the diameter at breast height (cm) and ln the natural logarithm.C content was estimated by using a conversion factor of 0.47 [25].The coefficient of determination (R 2 ) of this fitted equation was 0.93 and the root mean square error was 0.52 ln(kg C) (n = 29). (2) Chidumayo [26] developed relationships that relate AGB as a function of DBH (in Williams | 3,069,024 | 18039205 | 0 | 16 |
et al. [29]) and are presented in Equations ( 2) and (3): where AGB is in kg and DBH in cm. (3) Separate equations for the estimation of AGB as a function of climate and primarily the mean monthly evapotranspiration (ET) and rainfall (R) were developed by Chave et al. [27], with these being wet (ET > R in less than one month per year), moist (ET > R more than one month and less than five months per year) and dry (ET > R more than five months per year) forests.These criteria basically defined the extent of the growing period (when R is greater than ET) as a proportion of the year and on a monthly basis.Thus, the same criteria can be rearranged in terms of growing period (GP) as (i) wet (GP > 11/12), (ii) moist (11/12 > GP > 7/12) and (iii) dry (GP < 7/12).Using the GP dataset produced by Silva et al. [14], the forests of the study area could be considered as belonging to the dry category.Therefore, Equation 4 was used: where AGB is in kg, ρ is the wood specific density (oven-dried on a green volume basis, g•cm −3 ) and DBH is in cm.(4) Brown et al. [28] developed another equation (Equation ( 5)) for the dry forest life zone: 34.4703 8.0671 0.6589 (5) where AGB is in kg and DBH is in cm.The model has a R 2 = 0.67 and a mean square error of 0.02208 (n = 32). In the end, tree AGB were computed | 3,069,025 | 18039205 | 0 | 16 |
as the arithmetic mean of the values derived from these four equations. ALOS PALSAR The Japan Aerospace Exploration Agency (JAXA) launched ALOS on 24 January 2006, placing the satellite in a polar, sun synchronous orbit at ~700 km and ensuring a 46-day repeat cycle.PALSAR is one of the instruments onboard ALOS, which is an enhanced version of the Japanese Earth Resources Satellite (JERS-1) SAR instrument [30].PALSAR has a center frequency of 1,270 MHz (23.6 cm, i.e., L-band) and a chirp bandwidth of 14 MHz and 28 MHz.The instrument operated in five different observation modes, (i) Fine Beam Single (FBS), (ii) Fine Beam Dual (FBD), (iii) Polarimetric (PLR), (iv) ScanSAR and (v) Direct Transmission (DT).However, data acquired in the FBD mode was used, with these providing HH/HV or VV/VH polarizations (14 MHz bandwidth).Whilst 18 alternative off-nadir viewing angles were available (from 9.9° to 50.8°),only that acquired at 34.3° (HH and HV) polarization) were selected by JAXA and used with these acquired in ascending node with a 70 km swath width.Rosenqvist et al. [30] provide a complete reference and description of the ALOS PALSAR sensor. The European Space Agency (ESA) catalogue was used to select the two ALOS PALSAR FBD (HH and HV polarizations) scenes required to cover the entire study area.The data were acquired on 21 June 2010, with an off-nadir angle of 34.3° (ascending orbits) (Figure 1), and were provided by ESA in level 1.1 (single look complex, SLC).The SARScape software (version 4.3.001)produced by Sarmap SA (http://www.sarmap.ch) was used for all ALOS PALSAR processing, which followed | 3,069,026 | 18039205 | 0 | 16 |
standard SAR processing (e.g., [31][32][33][34]).Prior to geocoding, the SLC data were converted to multi-look intensity (MLI) format.To obtain approximately square pixels in ground range coordinates, a multi-look factor of 1 in range and 5 in azimuth was used.The resulting MLI data was geocoded to a ground resolution of 15 m (both in range and azimuth).No filtering for speckle noise reduction was performed.The geocoding of the MLI data, which refers to a transformation from the slant-range/azimuth geometry to map projection geometry, was performed to obtain geocoded terrain corrected (GTC) images.GTC requires a digital elevation model (DEM), which was applied to the scenes acquired over the study area.A 90 m DEM retrieved from the Shuttle Radar Topography Mission (SRTM) over the study area was used.This DEM (version 4) was provided by the International Centre for Tropical Agriculture (CIAT) and obtained from the European Commission (EC) Joint Research Center (JRC).To obtain GTC images, a backward solution is usually implemented, which considers an input DEM, which is used to convert the positions of the backscatter elements into slant range image coordinates.The transformation of the three-dimensional object coordinates-given in a cartographic reference system-into the two-dimensional row and column coordinates of the slant range image, is performed by rigorously applying the Range and Doppler equations [32].In case of precise satellite orbits, which is the case of ALOS, the geocoding process is run in a fully automatic way, and pixel accuracy can be achieved.Please refer to Meier et al. [32] for a more comprehensive description of the GTC method used in SARScape. Radiometric | 3,069,027 | 18039205 | 0 | 16 |
calibration was carried out by following the radar equation law and involved corrections for the scattering area, the antenna gain pattern and the range spread loss [35,36].A DEM is required to properly determine all required geometric parameters of the radar equation, so the calibration is performed during the data geocoding step, where the required parameters were already calculated.The calibrated value is a normalized dimensionless number (linear units, m 2 •m −2 ), and the corresponding value in the dB scale was generated by applying 10log 10 of the linear value and was generated as gamma nought (γ°).Even after a rigorous radiometric calibration, backscattering coefficient variations are clearly identifiable in range direction and in presence of topography, which requires a radiometric normalization.These variations are an intrinsic property of each imaged object and, thus, might be compensated, but it may not be corrected in absolute terms.In this study, a cosine correction method was applied, which was based on a modified cosine model [31] and applied to the backscattering coefficient to compensate for range variations.After the two scenes had been geocoded, they were combined to create a mosaic.Figure 1c shows the mosaic of the two ALOS PALSAR FBD geocoded scenes over the study area. Extraction of ALOS PALSAR FBD Data at Field Plot Locations Each plot measured in the field had a 20 m radius.Although the coordinates of each plot center were collected with GPS, there are always positional errors, especially when differential corrections are unavailable (errors up to 8-10 m are common).Although pixel accuracy can be achieved with | 3,069,028 | 18039205 | 0 | 16 |
ALOS PALSAR data geocoding (15 m ground pixel spacing), some error will always be present.Therefore, to compensate for these two sources of position errors, a buffer around each plot center with a 50 m radius was created.The buffer was selected as anything larger would have impinged on areas that were not homogeneous in terms of the plot tree canopy cover.All ALOS PALSAR pixels inside each 50 m buffer around each plot center were extracted, with several metrics computed (mean, minimum, maximum and standard deviation) and used to establish relationships with the AGB at the plot level.As the original ALOS PALSAR FBD mosaic had a 15 m spatial resolution and the buffer around each plot center was set to 50 m, then the extracted values per plot were those located approximately on a 6 × 6 pixel window size centered on each plot center, thus extracting the data in a 90 × 90 m window. To assess the amount of speckle of the processed ALOS PALSAR FBD MLI data (15 m ground resolution), the equivalent number of looks (ENL) (Equation ( 6)) (e.g., [33,34]) was estimated over a set of 30 Miombo homogeneous regions of interest spread over the study area: where µ and σ 2 are the mean and variance of the backscatter intensity values (original scale).The estimated ENL mean values were 5.26 and 4.91 for the HH and HV polarizations, respectively.These values are approximately the same as the number of looks used to produce the MLI data, i.e., 1 look in range and 5 look | 3,069,029 | 18039205 | 0 | 16 |
in azimuth, thus resulting in a 5 look image, meaning that the amount of speckle anticipated during multilooking is comparatively the same as that estimated by Equation ( 6).However, as mentioned above, the ALOS PALSAR FBD data extracted per plot were equivalent to extracting data located on a 6 × 6 pixel window size centered on each plot center.Therefore, the amount of speckle present in the ALOS PALSAR FBD data used to model AGB should be that of the ALOS PALSAR FBD data averaged to a 90 m spatial resolution.The ENL corresponding to this ALOS PALSAR FBD data were estimated, using the same set of 30 homogeneous areas and the mean values were 48.86 and 37.82, for the HH and HV polarizations, respectively.It shows that the amount of speckle was substantially reduced by spatially averaging the ALOS PALSAR FBD data. Contribution of Different ALOS PALSAR Polarizations and Metrics Several scatterplots were produced to display and evaluate the strength of the relationship between ALOS PALSAR backscatter intensity (γ°) and AGB data estimated from field data.ALOS PALSAR backscatter intensity (γ°) for the HH and HV polarizations were displayed against AGB data.As the ALOS PALSAR data were extracted on a 50 m radius of each plot center, then several metrics were computed, namely, (i) mean, (ii) minimum, (iii) maximum and (iv) standard deviation.Also, the coefficient of correlation (R) between the ALOS PALSAR metrics and the AGB was computed, either using parametric (Pearson's R) and non-parametric (e.g., Spearman's rank R) approaches [37]. Regression with Stochastic Gradient Boosting (SGB) and Bagging | 3,069,030 | 18039205 | 0 | 16 |
SGB (BagSGB) Traditionally, the attempt of explaining a given variable (dependent variable) as a function of one or more variables (predictor or independent variable(s)) has relied on parametric statistical models, such as linear (simple or multiple) regression models.These models have several assumptions, including normal distribution of errors and variables, as well as homoscedasticity (e.g., [38]).However, in the past decades and largely because of problems presented by large arrays of data, several other methods and algorithms have been developed, which are commonly referred to as machine learning methods.These methods have a large scope of application, from prediction (e.g., [39,40]) to classification (e.g., [41,42]) problems and can have different formulations (e.g., neural nets, ensembles of trees, support vector machines).A comprehensive review of these methods can be found in Hastie et al. [38]. SGB [43,44] originated from the decision tree theory (e.g., [38,45]).Decision tree theory relies on partitions of the space of all possible predictor variables.Starting with the whole predictor space (at the root of the tree), the space is successively split using a series of rules such that, in the end, each terminal node of the tree is assigned to the most probable response class (classification trees) or the mean response in that node (regression trees) [45,46].Some advantages over traditional parametric classification methods include the non-parametric nature of the classifier, quantification of variable importance, disclosure of non-linear and hierarchical relationships between predictor variables, and acceptance of missing values [47,48].However, classification and regression trees are sensitive to small perturbations in the training data, which may originate large changes in | 3,069,031 | 18039205 | 0 | 16 |
the resulting outputs [49].Therefore, these unstable methods can have their accuracy improved with perturbing and combining techniques.These generate multiple perturbed versions of the classifier (a.k.a., ensemble or committee) and combines them into a single predictor [50].These methods can be divided into two types: those that adaptively change the distribution of the training set based on the performance of previous classifiers (e.g., boosting) and those that do not (e.g., bagging) [51]. SGB [43,44] combines both the advantages of bagging and boosting and can be used in regression and classification problems (e.g., [52][53][54][55]).It typically uses a base learner (in our case, binary decision trees) and constructs additive regression (or classification) models by sequentially fitting the chosen base learner to current "pseudo"-residuals by least squares at each iteration using a random fraction of the training data without replacement [44].This process has been shown to substantially improve the prediction accuracy and execution speed, making the approach resilient to overfitting [44].Furthermore, Suen et al. [56] demonstrated that building and combining (by averaging in the case of regression) several SGB models on samples randomly drawn with replacement from the original training dataset (bootstrap sample) performed significantly better than a unique SGB model and concluded that it was accomplished by variance reduction.Therefore, two approaches were followed, (i) generate a SGB model from the original training dataset and (ii) generate several SGB models fitted to bootstrap samples (with replacement) of the original training dataset that were then combined by averaging.This allowed us to compare SGB against bagging SGB (BagSGB). Model fitting under SGB has | 3,069,032 | 18039205 | 0 | 16 |
a number of options that were tested to select the best model.This was done by developing in-house R code [57] to implement a loop using Ridgeway's R gbm package [58] and Elith et al. [40] R code.These options include: (i) distribution (Gaussian, i.e., the loss function, whose measure of deviance to be minimized is the mean squared error), (ii) bagging fraction (i.e., the random fraction of the training data that is randomly selected to build each decision tree; 0.5, 0.6, 0.7, 0.8 and 0.9), (iii) tree complexity (i.e., the number of nodes in each decision tree; 1, 2, 3, 4, 5, 7 and 9) and (iv) shrinkage rate (i.e., controlling the learning rate of the algorithm; 0.01, 0.005, 0.0025, 0.001 and 0.0005) [58].Therefore, the selection of the best SGB model is the result of evaluating 175 candidate individual SGB models (5 bagging fraction values × 7 tree complexity parameters × 5 shrinkage rate values).Elith et al. [40] referred to the fact that small shrinkage rates result in a smaller contribution of each tree and, therefore, are more directed at building a model that will provide a more consistent estimate of the dependent variable; as a rule of thumb, they indicate that, as a minimum, models should be fitted with 1,000 trees.Accordingly, to select the best SGB model out of the various candidates, we selected the model that had the lowest 10-fold cross-validated deviance (i.e., lowest mean squared error), given that a minimum of 1,000 trees were used to generate that model.This procedure was carried out for | 3,069,033 | 18039205 | 0 | 16 |
each bootstrap sample of the original training dataset when generating the BagSGB model.We used 25 bootstrap samples to build a BagSGB model, as Breiman [49] suggests that a higher number of replicates tend not to produce a significant test set error reduction.The final model (bagging SGB or BagSGB) was built by averaging the predictions from the 25 selected SGB models fitted to the bootstrap samples.Also, as each predicted observation is the result of averaging over a set of 25 bootstrap samples, it allows also building a measure of prediction variability.The coefficient of variation (CV) (e.g., [37]), calculated as the standard deviation of a predicted observation divided by the corresponding mean value, was used as a measure of assessing the uncertainty associated with each prediction. AGB was chosen as the dependent variable, and the mean, minimum, maximum and standard deviation of the ALOS PALSAR HH and HV backscatter intensity values extracted for each plot (50 meters buffer around each plot center) were selected as independent variables.The number of pixels that were used to compute those metrics ranged from 28 to 37, the variation being dependent on the plot location regarding the 15 m ALOS PALSAR data. The SGB and BagSGB models were compared using a traditional bias and variance decomposition of the root mean square error (RMSE) (Equations (7-9)).A 10-fold cross validation approach was followed, as the number of observations was not large enough to evaluate model performance with an independent subset (e.g., [38]). where RMSE is the model root mean square error, e i the error | 3,069,034 | 18039205 | 0 | 16 |
(difference between the observed and predicted values) of the i th observation, σ 2 the error variance, b the error bias and n the number of observations; where ē is the mean error; 1 (9) C Stocks and Comparison with Published Biomass Maps Mean AGB and C stocks per tree canopy cover classes under 50% were estimated, as well as the total AGB and C stock in the Miombo forests of the broader study area (~10,000 ha).The default carbon fraction of dry matter of 0.47 obtained from the Intergovernmental Panel on Climate Change [59] was used to convert Miombo forest AGB to C content.This default value is consistent with the dry mass fraction determined by Ryan et al. [25] with subsamples from both the trunk and branches of 19 trees from a Miombo region in central Mozambique. A comparison with published biomass maps, especially when derived from distinct data sources, is important, namely to compare and assess differences in the AGB and C stocks estimates in a same ecosystem and region.For that, two published maps were selected: Baccini et al. [60] and Saatchi et al. [12].Baccini et al. [60] mapped AGB across tropical Africa (1 km spatial resolution) using an ensemble of regression tree-based models (random forests) that relate in situ measurements with data acquired by the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard NASA's Terra and Aqua satellites between 2000 and 2003; a cross-validation approach showed that the model explained 82% of the variance in AGB, corresponding to a RMSE of 50.5 Mg•ha −1 (for AGB | 3,069,035 | 18039205 | 0 | 16 |
values between 0 and 454 Mg•ha −1 ).Saatchi et al. [12] mapped forest C stocks, AGB plus below-ground biomass (BGB), in the pan-tropical belt (~1 km spatial resolution) around the year 2000 using a data fusion algorithm based on the maximum entropy approach (MaxEnt) to model field-based measurements as a function of data acquired by the Geoscience Laser Altimeter System (GLAS) onboard the Ice, Cloud and land Elevation Satellite (ICESat) and extrapolating to the landscape using data from other optical and microwave sensors; the prediction variability ranged from ±6% to ±53% on a pixel basis, but when scaling-up to project-and country-scales, the errors decreased to ±5% and ±1%, respectively. The uncertainty at the study area level (U P ) was estimated using the error propagation approach (Equation ( 10)) (e.g., [61]), which is the same equation used to assess overall uncertainty in REDD projects and national greenhouse gas inventories [59]: where AGB i and U i are the forest AGB and uncertainty at the ith pixel, respectively, and N the number of pixels in the area being assessed. Forest AGB from Field Data As mentioned in section 3.1, the four selected allometric equations were used to produce an estimate of each tree AGB and were subsequently averaged to produce an estimate of AGB of each measured tree.Table 1 depicts the average AGB C stock per tree canopy cover class. Table 1.Average above-ground biomass (AGB) (tC•ha −1 ) values per tree canopy cover class (standard error of the mean in parentheses).A factor of 0.47 was used to | 3,069,036 | 18039205 | 0 | 16 |
convert from biomass to C content [25,59]. Contribution of ALOS PALSAR Polarizations and Metrics The relationship between ALOS PALSAR backscatter intensity data (HH and HV polarizations) and forest AGB data is shown in Figure 2 BagSGB Modeling The final BagSGB model was the result of combining the 25 SGB models built from the corresponding bootstrap samples of the original training dataset.The forest AGB values predicted from the 10-fold cross-validation of each of the 25 models were aggregated (averaged) and the corresponding statistics calculated (Table 3).The comparison between observed and cross-validation predicted forest AGB values is displayed in Figure 3(a).Each cross-validation predicted forest AGB value was the result of averaging between 16 and 36 times, as each bootstrap sample was built with replacement.A close correspondence (R = 0.95) between observed and cross-validation predicted forest AGB values was identified, although the AGB above 40 Mg•ha −1 was slightly underestimated.However, there is no evidence that it is a trend consistent with L-band SAR saturation, as only four plots have AGB values higher than that value.As a comparison, the same combination of parameters (i.e., distribution, bag fraction, tree complexity and learning complexity) was tested to build the best SGB model fitted to the original training dataset.This model was based on 4,700 trees, resulting from a bag fraction of 0.8, a learning rate of 0.0005 and a tree complexity of 2. The 10-fold cross-validation results performed less well than those resulting from the application of an ensemble of SGB models (i.e., a BagSGB model).The cross-validation RMSE of this model was 12.04 | 3,069,037 | 18039205 | 0 | 16 |
Mg•ha −1 (5.03 Mg•ha −1 in the BagSGB model), the variance 144.81 (24.96 in the BagSGB model), the error bias −0.32 Mg•ha −1 (0.58 Mg•ha −1 in the BagSGB model) and the linear coefficient of correlation (R) between observed and cross-validation predicted AGB values was 0.48 (0.95 in the BagSGB model).Figure 3(b) shows the scatterplot between observed and cross-validation predicted forest AGB values for this SGB model.The error bias is approximately the same; therefore, the RMSE decrease (from 12.04 to 5.03 Mg•ha −1 ) was a consequence of variance reduction from 144.81 in the SGB model to 24.96 in the BagSGB model.This is in agreement with the results of Suen et al. [56] already mentioned in Section 4.2.On the basis of its better performance in the cross-validation assessment, the BagSGB was applied to the entire study area.The ALOS PALSAR FBD backscatter intensity data used to build the model were extracted in a 50 m buffer around each plot center ,with this being equivalent to a spatial resolution of approximately 100 m.Therefore, to create the forest AGB map of the study area, the original ALOS PALSAR FBD backscatter intensity data at 15 m spatial resolution was necessarily aggregated to a 90 m spatial resolution, and the minimum, maximum, mean and standard deviation values were computed.Also and for displaying purposed, the continuous AGB map was transformed into a five-class map (Figure 6).As mentioned in Section 4.2, the type of algorithm used allowed the production of a map displaying the prediction variability in each pixel, by computing the coefficient | 3,069,038 | 18039205 | 0 | 16 |
of variation (%) on a pixel-by-pixel basis and for the same reasons was converted into a five-class map (Figure 7).As mentioned above, although the 10-fold cross validation procedure that was implemented may have resulted in overoptimistic values of model assessment, the prediction variability at the pixel level provides additional and complementary information regarding the spatial distribution of the model error that could be used to assess its applicability to new observations.Figure 7. Forest AGB uncertainty classes map of the study area (outlined) obtained with the coefficient of variation (%) resulting from the application of the fitted BagSGB model.The minimum and maximum values presented in the legend are for the area encompassing the mosaic of the two ALOS PALSAR scenes used.In the study area (~10,000 ha), the minimum and maximum forest AGB coefficient of variation values were 10% and 119%, respectively. C Stocks and Comparison with Published Biomass Maps Mean forest AGB and derived C stocks, as well as uncertainty per tree canopy cover class in the study area (~10,000 ha) were calculated on the basis of the forest AGB map produced in this study in the same manner as in the work of Baccini et al. [60] and Saatchi et al. [12].Differences in the time frame of each study (2008 for this study, 2000-2003 for Baccini et al. [60] and early 2000s for Saatchi et al. [12]) were not considered nor addressed here.The forest AGB carbon stock obtained for the entire study area from this study (143,444 Mg•C) is very different from the values obtained with | 3,069,039 | 18039205 | 0 | 16 |
data from Baccini et al. [60] (498,570 Mg•C) and Saatchi et al. [12] (405,986 Mg•C).The mean AGB obtained with the data from this study is 30.6 Mg•ha −1 , while in Baccini et al. [60], it is 106.3Mg•ha −1 and in Saatchi et al. [12], 86.6 Mg•ha −1 .Mitchard et al. [63] on their evaluation of the data published by Baccini et al. [60] used an independent dataset of 1,154 field measurements, obtained in 16 African countries and concluded that large errors were associated with this AGB map; more significantly for our study, there was large underestimation in forests with higher AGB and overestimation in woodland savannas, resulting in a RMSE of 145 Mg•ha −1 .However, Saatchi et al. [12] also reported higher values than those reported here.As mentioned in Section 3.1, in our study, only the area with a tree canopy cover between 10% and 50% was sampled for forest AGB estimation, and therefore, it is not unexpected that the derived C stock and mean AGB are lower than those reported in studies using field data spanning the entire tree canopy cover range.Also, it is important to note that the Miombo forests are diverse in terms of tree canopy cover and, therefore, C content.For example, Glenday [64] reported an average forest AGB value of approximately 92 Mg•ha −1 for Brachystegia sp.dominated dry forests in Kenya.Malimbwi et al. [65], using field data collected in the Miombo woodlands of Tanzania, reported an average forest AGB value around 18 Mg•ha −1 .Ryan et al. [25] estimated an average | 3,069,040 | 18039205 | 0 | 16 |
forest AGB value of 45 Mg•ha −1 for a Miombo woodland in Mozambique.Shirima et al. [66] measured several plots in Miombo forests of Tanzania and estimated a mean forest AGB value of approximately 48 Mg•ha −1 .The study of Ryan et al. [25] was carried out in areas of Miombo formations similar to those addressed in the present study and the estimated mean AGB was 45 Mg•ha −1 ; a value that compares with the 31 Mg•ha −1 estimated during this study.The difference can probably be related to the 10-50% threshold of the measured Miombo in this study. The estimated forest AGB uncertainty at the study area level was 0.21% (coefficient of variation) in this study and 2.82% (error) in Saatchi et al. [12].The values of uncertainty on a pixel basis ranged from 10-119% (mean = 25%) and 21-31% (mean = 27%) in this study and in Saatchi et al. [12] respectively.Although higher on a pixel level, the values decrease substantially when the measure of uncertainty is obtained at a study area level.This is also the case in Saatchi et al. [12]. Furthermore, the mean forest AGB and the total C stocks were estimated for the subset of the study area with a tree canopy cover between 10% and 50% (1,157 ha), which was the area sampled during the field campaign that took place in the 2011 dry season.The estimates are depicted in Table 4 by tree canopy cover classes.Similarly as for the results presented in Table 4, the estimates of mean AGB and total AGB | 3,069,041 | 18039205 | 0 | 16 |
C stock per tree canopy cover class derived from this study are substantially lower than those estimated in Baccini et al. [60] and Saatchi et al. [12].Table 4. Mean AGB, Total AGB, AGB C stock and uncertainty per tree canopy cover class in the study area with tree canopy cover between 10% and 50% (1,157 ha), derived from this study (BagSGB model), Baccini et al. [60] and Saatchi et al. [12] ( 1 the uncertainty refers to the coefficient of variation (%) (this study) and error (%) [12]; n.a., not available.).Romijn [67] concluded that the land use change associated with the introduction of Jatropha curcas L. in Miombo savanna woodlands will only have a positive feedback (i.e., atmospheric carbon sequestration) when introduced in wastelands or severely degraded lands; however, he acknowledges that the data used in his study have a high degree of uncertainty, especially due to substantial regional and local variations in soil, biomass and climate characteristics.Nevertheless, for industry and for the purpose of complying with the objectives of environmental sustainability in biofuel plantations, it is important to guarantee that the C content of the areas to be converted is sufficiently low at the start, so that the biofuel produced actually corresponds to savings in carbon emissions when compared to fossil fuels, and thus, it is effectively contributing to the countries' and companies' renewable energy targets.Given uncertainties in carbon estimations, good practice demands that a conservative approach be applied.Thus, an accurate enough, expedient and cost effective method for spatially explicit carbon quantification whose performance risk | 3,069,042 | 18039205 | 0 | 16 |
is systematic underestimation may constitute a helpful planning tool. Reference When dealing with international funds for performance based payments to developing countries in the context of REDD, the use of the conservativeness principle for estimating emissions from deforestation and forest degradation is always required.In such cases, establishing a reference emission level using a conservative carbon stock is mandatory.Ryan et al. [68] in their study in areas of Miombo savanna woodland located in central Mozambique concluded that deforestation activities were not particularly occurring in high biomass areas.Therefore, by using estimates of forest AGB exclusively from data retrieved from plots that had a tree canopy cover between 10% and 50%, we could actually be encompassing the areas more prone to deforestation and, therefore, provide more conservative estimates of forest AGB for REDD projects.Additionally, the methodology used in this study could be easily adapted to produce spatial conservative estimates of C stocks in the AGB pool.The methodology uses n (in this case n = 25) SGB models and then averages them to estimate the AGB at the pixel level.Therefore, a conservative estimate of AGB at a given pixel could be obtained by using a given percentile (lower than 50%) of the distribution of possible forest AGB values for that pixel.Alternatively, the coefficient of variation that is produced, also on a pixel-by-pixel basis, could be used to discount the value of the estimated forest AGB. Conclusions Consistent forest carbon monitoring methods are required at various levels, namely for developing countries who wish to address international conventions and to access carbon-based | 3,069,043 | 18039205 | 0 | 16 |
financial mechanisms associated with climate change mitigation.A method for reducing the uncertainty in the estimation of forest above-ground biomass (AGB) in Miombo savanna woodlands in southeast Africa (Zambézia province, Mozambique) has been presented.The advancement of this study relied on the use of a machine learning algorithm to establish a relationship between in situ forest AGB and L-band Synthetic Aperture Radar (SAR) backscatter intensity (gamma nought, γ°) data obtained from the Phased Array L-band SAR (PALSAR) sensor onboard the Advanced Land Observing Satellite (ALOS).This algorithm, bagging stochastic gradient boosting (BagSGB), is unique, as it allows also the production of spatial explicit estimates of prediction variability and an indication of the importance of each predictor variable.Estimates of forest AGB with a root mean square error (RMSE) of 5.03 Mg•ha −1 based on 10-fold cross validation were produced with this modeling approach.Also, the coefficient of correlation (R) between observed and predicted (from 10-fold cross validation) forest AGB values was 0.95.The variable contributing the most to this model was the mean backscatter intensity for the HH polarization, which was explained by the low tree canopy cover characterizing Miombo savanna woodlands, thus invoking scattering mechanisms associated with this polarization (e.g., trunk-ground scattering).Furthermore, it was recognized that the optimistic overall validation results (RMSE and R) might be a consequence of the 10-fold cross-validation procedure, especially when dealing with bootstrap samples that were drawn with replacement.Nevertheless, this algorithm was unique in producing estimates of prediction variability (coefficient of variation) on a pixel-by-pixel basis.These estimates ranged from 10 to 119% across the study area, | 3,069,044 | 18039205 | 0 | 16 |
with a mean value of 25%.This map of prediction variability (Figure 7) is a useful instrument to assess how well the model is predicting new observations.One of the reasons for the observed disagreement between the mean forest AGB values and total forest AGB carbon (C) stocks generated from this study and those resulting from the two available forest AGB maps (i.e., [12,60]) could be related to the fact that only the forest areas with tree canopy cover between 10% and 50% were sampled for the collection of in situ data.Therefore, subsequent work will rely on sampling the Miombo areas with tree canopy cover greater than 50%, which will allow a better characterization of the Miombo savanna woodlands in the region and more in situ observations to produce an updated version of the BagSGB model. Figure 1 . Figure 1.(a) Location of Mozambique in Africa; (b) Mozambique provinces and location of the study area in the Zambezia province; (c) mosaic of the two ALOS PALSAR Fine Beam Dual (FBD) scenes (HH polarization 90 m mosaic) and limits (in white) of the study area.ALOS PALSAR FBD zoom over the ~10,000 ha study area; (d) HH polarization; (e) HV polarization. . The mean, minimum, maximum and standard deviation were computed based on the ALOS PALSAR backscatter intensity data extracted over a 50 m radius Figure values the tra values dashed Table 2 . Parametric Pearson's coefficient of correlation (R) and non-parametric Spearman's rank coefficient of correlation (R rank ) between forest AGB and several metrics derived from the ALOS | 3,069,045 | 18039205 | 0 | 16 |
PALSAR FBD (HH and HV polarizations) backscatter intensity data (* significant at a significance level of 0.05). | 3,069,046 | 18039205 | 0 | 16 |
Unsupervised Domain Adaptation for Point Cloud Semantic Segmentation via Graph Matching Unsupervised domain adaptation for point cloud semantic segmentation has attracted great attention due to its effectiveness in learning with unlabeled data. Most of existing methods use global-level feature alignment to transfer the knowledge from the source domain to the target domain, which may cause the semantic ambiguity of the feature space. In this paper, we propose a graph-based framework to explore the local-level feature alignment between the two domains, which can reserve semantic discrimination during adaptation. Specifically, in order to extract local-level features, we first dynamically construct local feature graphs on both domains and build a memory bank with the graphs from the source domain. In particular, we use optimal transport to generate the graph matching pairs. Then, based on the assignment matrix, we can align the feature distributions between the two domains with the graph-based local feature loss. Furthermore, we consider the correlation between the features of different categories and formulate a category-guided contrastive loss to guide the segmentation model to learn discriminative features on the target domain. Extensive experiments on different synthetic-to-real and real-to-real domain adaptation scenarios demonstrate that our method can achieve state-of-the-art performance. I. INTRODUCTION Deep learning methods [1], [2] for point cloud semantic segmentation have shown dramatic success in recent years. However, most of these methods focus on fully supervised learning for point cloud segmentation with a large number of manually annotated labels. Although there are several public datasets providing large amounts of annotation data, it is difficult to directly | 3,069,047 | 251442662 | 0 | 16 |
apply the model trained on a labeled source domain to another unlabeled target domain. The reason lies in that the data collected by different 3D sensors have a huge discrepancy in appearance and sparsity, which results in the domain shift problem. Therefore, how to generalize a well-trained model to another unlabeled domain is a challenging but valuable problem in point cloud semantic segmentation. Unsupervised domain adaptation can alleviate the domain shift problem by transferring the knowledge from the labeled source domain to the unlabeled target domain. Recent advances on unsupervised point cloud domain adaptation tasks mainly focus on reducing the domain gap between the inputs. For example, Yi et al. [3] build a point cloud completion network with sequences of point clouds to bridge the domain gap between LiDAR sensors with different beams. ePointDA [4] and SqueezeSegV2 [5] use auxiliary rendering networks to render dropout noises or intensity on the synthetic dataset, which translate the point clouds from the source domain similar to the target domain. Furthermore, these methods use a series of feature alignment methods to increase the consistency of feature distributions, such as higher-order moment matching [6], and geodesic correlation alignment [7]. However, these methods mainly consider the overall distributions of two domains to form the global-level feature alignment, which ignores the local geometric differences between the domains. In this paper, we propose a domain adaptation framework for unsupervised point cloud segmentation with the locallevel feature alignment. Compared with the global-level feature alignment, our framework can focus on the correlation between the similar local | 3,069,048 | 251442662 | 0 | 16 |
structures of point from the two domains, so that the reliable feature alignment can be performed to guide the discriminative semantic feature learning of the target domain. Specifically, through the farthest point sampling, we select a set of centroid points and construct the dynamic local feature graph for each centroid point to capture its local geometry information. Then, in order to enrich the graph of the source domain, we construct a feature graph memory bank to store the generated source-domain feature graphs during the training phase. After that, inspired by the point cloud matching [8], we adopt the optimal-transport cost to measure the graph similarities between the memory bank and target domain, so that a reliable assignment matrix can be obtained to guide the knowledge transferring from the source domain to the target domain. Particularly, in order to further extract the discriminative target-domain feature, we consider the category-wise correlation between the source domain and the target domain, and exploit the contrastive learning to increase category-level discrimination of target graphs. Such category-guided contrastive loss can effectively help cluster and distinguish the feature-graph distributions of different categories. Extensive experiments demonstrate the effectiveness of our framework, where we not only focus on the synthetic-to-real domain adaptation scenarios (vKITTI to SemanticPOSS), but also pay attention to the indoor (S3DIS to ScanNet) and the outdoor (SemanticKITTI to nuScenes) real-to-real domain adaptation scenarios. Our contributions can be summarized as follows: • We propose a novel graph-based framework for locallevel feature alignment for unsupervised domain adaptive point cloud semantic segmentation. • We construct | 3,069,049 | 251442662 | 0 | 16 |
feature graphs to capture the local geometry information of point clouds and use a local feature loss based on an assignment matrix for the alignment of feature graphs. • We develop a category-guided contrastive loss to guide the segmentation model to learn the discriminative features on the target domain. II. RELATED WORK Point Cloud Semantic Segmentation. Recent progress on point cloud semantic segmentation is mainly divided into several categories according to different representations of data. Volumetric-based methods require a preprocessing stage to voxelize the original point cloud. SparseConvNet [9] proposes a submanifold sparse convolution network to deal with spatially-sparse voxel data. MinkowskiNet [10] creates Minkowski space on sparse representation data and proposes a powerful 4-dimensional convolutional neural network to deal with 3D videos. Projection-based methods need to project the point cloud into an image before feeding data into the network. SqueezeSegV2 [5] uses a context aggregation module to improve the robustness to dropout noise on projected 2D LiDAR image. SqueezeSegV3 [11] proposes an efficient spatially-adaptive convolution to deal with the discrepancy of data distribution of different LiDAR image locations. Point-based methods directly use unordered point clouds for semantic segmentation. However, due to the heavy computation, most of the methods first split the point cloud into blocks before training and inferring. PointNet [1] uses multi-layer perceptron and a mini-network (T-Net) to extract features from unordered point clouds. In order to strengthen the local information for point-level segmentation, GACNet [2] and PointWeb [12] use different attention modules to dynamically assign weight to local features. In this work, we | 3,069,050 | 251442662 | 0 | 16 |
leverage PointWeb as our segmentation network because of its efficiency in processing unordered point clouds. Unsupervised Domain Adaptation. Unsupervised domain adaptation (UDA) aims to train the model in the labeled source domain and generalize the knowledge to the target domain through the unsupervised methods. Recent advances on domain adaptation for 3D point cloud mainly study aligning the distributions by input-level and feature-level alignment. Saleh et al. [13] use CycleGAN [14] to translate the synthetic bird's eye view point cloud image to the real point cloud image for domain adaptive vehicle detection. ePointDA [4] use a dropout noise rendering network to achieve uniformity of data distribution between domains and adopt a higher-order moment matching loss for featurelevel alignment. Yi et al. [3] use a completion network to complete the point cloud with sequences data, so that they can recover the 3D surfaces from different LiDAR data and transfer knowledge between different LiDAR sensors. However, the input-level adaptation methods lead to extra challenges and training costs due to the variable geometric structures in different domains. Besides, recent works on 2D UDA [15], [16], [17] are also quite applicable to 3D UDA, where they use the different losses to decrease the domain shift problem, e.g. maximum squares loss, entropy loss, and adversarial loss. Furthermore, self-training is also an effective technique for UDA. [18] proposes a self-supervised task for target domain to learn its useful representations. ST3D [19] proposes a quality-aware triplet memory bank to generate high-quality 3D detection pseudo labels for selftraining. xMUDA [20] and DsCML [21] propose crossmodal | 3,069,051 | 251442662 | 0 | 16 |
constraint to retain the advantages of 2D images and 3D point clouds for domain adaptation. In this paper, considering the input-level methods cannot handle complex domain adaptation scenarios, we develop a general uni-modal 3D UDA framework with feature-level alignment. III. OUR METHOD A. Overview In unsupervised domain adaptive point cloud semantic segmentation, we are able to access the source domain points. Given data X s , Y s and X t , our goal is to train a model which can precisely categorize the point of target data into one of the common semantic categories in the source data, and to alleviate the performance drop problem caused by the domain gap at the same time. As illustrated in Fig.1, we first construct the local feature distributions of the source and target domains with the proposed dynamic feature graphs. Then, by building a source-domain feature graph memory bank, we employ graph matching to obtain graph pairs between the graphs in the memory bank and the target graphs. Finally, with the obtained matching association, we utilize the designed losses for local-level feature alignment. B. Dynamic Feature Graph Different from the global-level feature alignment methods [6], [7], [15], [16], [17] roughly aligning two domains, we consider the differences of local neighborhood context in the target domain for the fine alignment. The main idea of our method is to use the learned dynamic local feature graphs to capture the multi-level features in different neighborhoods of point clouds. Then, based on the local-graph similarity, the correlation between local neighborhoods from two | 3,069,052 | 251442662 | 0 | 16 |
domains can be viewed as the knowledge transferring from the labeled source domain to the unlabeled target domain. We leverage PointWeb [12] as our semantic segmentation backbone, which contains a classifier and a feature extractor with an encoder and a decoder. We use the feature extractor to extract the multi-level features and then build dynamic feature graphs on the sampled centroid points by the feature similarity. Specifically, given a point cloud sample x ∈ R N ×3 with N points, we first extract its local-context feature using the feature extractor. Then, we select N/64 centroid points using the farthest point sampling for three iterations, where the centroid points are then as the kernel points of graphs for feature aggregation at each level. In detail, for each kernel point at different levels, we gather its k-nearest neighbors in the feature space (k is different at each level). Thereby, a feature graph can be constructed by setting the k-NN features as its vertices {v k1 j=1 is the different values in k-NN. We obtain N/64 dynamically updated local feature graphs to represent the local neighborhood context of the given point cloud, which can be formulated as, where N c denotes the number of N/64 centroid points, and f kj i represents the feature embedding containing the vertice and the edge information. As a result, given a source sample and a target sample, the generated graphs are represented as C. Graph-based Local Feature Alignment In this section, based on the constructed dynamic graphs above, we aim to find the | 3,069,053 | 251442662 | 0 | 16 |
intrinsic correlation between the source domain and the target domain. We use the graphs from the source domain to guide the model to extract semantic discriminative features on the target domain. However, at the training stage, the sample category in a batch is limited and their graph patterns tend to present significant structure differences, which may potentially introduce the alignment bias. To address the issue above, we build up a feature graph memory bank G b and store a graph g s i into the bank according to the corresponding category of the centroid point. Therefore, benefiting from such memory bank mechanism, we can sufficiently mine the rich source information from it for reliable target-domain feature learning. The memory bank provides the same capacity B for each category of graphs. Once the number of graphs exceeds the capacity B of the corresponding category, we will update the memory bank by replacing the oldest graphs with the new ones. Given the graphs G t = {g t j } Nc j=1 from target domain, we consider finding the most similar graph from G b to each graph in G t for feature alignment. In particular, we use optimal transport for graph matching. Specifically, the total transport cost of optimal transport is used to measure the similarity between two graphs, and the assignment matrix A ∈ R K×K is used to find the point-level correspondences for K nodes in graphs. In the graph matching formulation, we first compute the distance matrix D ∈ R K×K , where the | 3,069,054 | 251442662 | 0 | 16 |
element D (m,n) indicates the distance between the point m in one graph and the point n in the other graph. Here, we use the squared Euclidean distance in the feature space to measure the pairwise distance between points in graphs, where the points are composed of the corresponding features with edge and vertice information. Once we obtain the distance matrix D, we apply the Sinkhorn algorithm [22] to obtain the final assignment matrix A and the total transport cost through solving the optimal transport problem. In this way, we can measure the relevance between each target graph with all the graphs in the memory bank. As a result, we can find the most similar graph g b (a,l) ∈ G b for the target graph g t j according to the sorting result of the transport cost for knowledge transferring, where a indicates the category and l indicates the index in the memory bank. Based on the generated graph pairs, we formulate a local feature loss based on the assignment matrix for the locallevel feature alignment. Given a target graph g t j , we first select the most similar graph g b (a,l) from the memory bank. At the same time, we are able to access the corresponding assignment matrix A j ∈ R K×K , which can decode the point-level corresponding feature assignment between two graphs. In detail, for each point in target graph g t j , we can obtain the corresponding transport weights for every point in graph g b (a,l) | 3,069,055 | 251442662 | 0 | 16 |
from the assignment matrix A j . Then, we perform a weighted sum of g b (a,l) ∈ R K×D to guide the learning of g t j ∈ R K×D , where D is the number of feature channels. The key point is that the local neighborhood areas with similar semantic contexts need to have similar feature distributions. In this way, we can effectively align the indiscriminate feature distributions of the unlabeled target domain to the source domain. Therefore, we propose the following assignment matrix based local feature loss for feature graph learning in the target domain: Owing to a variety of feature graphs from different categories in our memory bank, we can further exploit the contrastive learning for more discriminative target-domain feature learning. Here, we select the category a of the matched graph g b (a,l) as the positive category and the other categories as the negative categories. In order to obtain the representative features of each category, all graphs in the memory bank are used for calculating the feature representations. Although we have achieved the assignment matrix for each positive or negative pair, it is meaningless for pointlevel adaptation on unmatched pairs. Therefore, we use the mean of all features in the graphs with the same category to represent the feature representation of the corresponding category. It is worth noting that our graph is composed of multi-level features, so we calculate the mean features of different levels separately and then concatenate them as the final feature representation. The positive and negative features for | 3,069,056 | 251442662 | 0 | 16 |
graph g t j can be formulated as, where C is the number of categories and B is the capacity size of each category. The indicator function I returns 1 if the condition is satisfied or returns 0 if unsatisfied. We use Θ to represent the mean and concatenation operators. Then, with the generated positive and negative features, we formulate the following contrastive loss for increasing the intra-category compactness and inter-category separability between the target graph g t j and the graphs in G b . where α is the margin of the contrastive loss and f t j is the mean feature of graph g t j . Therefore, with the proposed local feature loss and the contrastive loss, we consider the feature alignment from two complementary perspectives, which can significantly reduce the domain discrepancy in feature space. D. Domain Adaptation Scheme For the unsupervised domain adaptation, the core challenge is how to learn the discriminative target features without labels. First of all, for the source domain, we use the standard cross-entropy loss for supervised training: where y s ∈ R N ×C is the semantic labels for N points with C semantic categories andŷ s is the outputs from the model. In addition, in our framework, in order to identify the relationship of local features between the source domain and the target domain, we construct dynamic feature graphs and the generated graph pairs based on the graph matching to find correspondences between the two domains. With our developed local feature loss based on the | 3,069,057 | 251442662 | 0 | 16 |
assignment matrix and the category-guided contrastive loss, we can effectively align the local features between the two domains. The overall loss can be formulated as: where λ 1 and λ 2 are hyperparameters balancing the proposed losses with semantic segmentation loss. Furthermore, our framework can be extended into a twostage method in a self-training manner, where we follow Jaritz et al. [20] to use a pseudo-label training strategy. We first use our framework to train a model with the loss in Eq. 7, where the source data X s , Y s and the target data X t are available. Then we fix the parameters of the model and generate pseudo labelsŶ s for target data. After that, the supervised semantic segmentation loss with pseudo labels is used on the target domain. A. Datasets vKITTI to SemanticPOSS. The synthetic dataset vKITTI [23] contains 6 sequences of outdoor scenes in urban settings, where the point cloud are generated from the synthetic 2D depth images. The SemanticPOSS [24] dataset was obtained in dynamic driving scenarios. It is composed of 6 sequences of scenes with a total of 2988 LiDAR scans. Therefore, there is a large gap in data distribution between the vKITTI and the real-world SemanticPOSS. For the domain adaptation scenario from vKITTI to SemanticPOSS, we select 6 semantic categories for domain adaptation: plants, building, road, traffic sign, pole, and car. The point clouds are sampled into blocks of 15m×15m, and each block contains 4096 points. S3DIS to ScanNet. The S3DIS [25] dataset is an indoor point cloud | 3,069,058 | 251442662 | 0 | 16 |
dataset containing 6 areas with 271 rooms and the ScanNet [26] dataset contains 1513 indoor point cloud scenes annotated. For the domain adaptation scenario from S3DIS to ScanNet, we use 8 semantic categories for domain adaptation: floor, wall, window, door, table, chair, sofa, and bookshelf. Due to the sparsity and scene incompleteness of ScanNet, there is a huge domain gap between the datasets. We divide the point clouds into blocks of size 1.5m×1.5m, and each block contains 8192 points. SemanticKITTI to nuScenes. The SemanticKITTI [27] dataset and the nuScenes [28] dataset are real-world datasets. However, the SemanticKITTI dataset is obtained by the 64beam LiDAR scanner, while the nuScenes dataset is obtained by the 32-beam LiDAR scanner. Thus, there is a large gap of data sparsity in the SemanticKITTI-to-nuScenes domain adaptation scenario. We focus on the 10 categories for domain adaptation: car, bicycle, motorcycle, truck, other vehicle, pedestrian, drivable, sidewalk, terrain, and vegetation. The point clouds are sampled into blocks of 10m×10m, and each block contains 4096 points. B. Implementation Details We use the official PyTorch implementation for PointWeb as our segmentation backbone. Stochastic Gradient Descent (SGD) optimizer is selected for training with the momentum 0.9 and the weight decay 0.0001, respectively. Also, we apply the weight decay to the learning rate, where the drop factor is 0.1 and the step size is 30. The initial learning rate for indoor and outdoor scenarios are 0.05 and 0.005. The capacity size B of the memory bank is set to 16. The parameters λ 1 and λ 2 are | 3,069,059 | 251442662 | 0 | 16 |
set to 1.0 and 0.1. The margin α of the contrastive loss is set to 0.4. The values in k-NN for different levels are set to 1, 4, 16 and 64. To train and test our model, we use a single TITAN RTX GPU and the batch size is set to 4. C. Performance Comparison We report the performance of point cloud semantic segmentation with mean Intersection-over-Union (mIoU). Tab.I shows the quantitative results of the comparison between our method with other domain adaptation methods. As shown in the table, our method achieves the highest performance, which shows the effective domain transferability of our method. Specifically, the Supervised means the model of PointWeb is trained on the target domain with semantic labels. The Source Only means the model trained with the source domain and directly tested at the target domain. Due to the domain gap, the performance of Source Only has a significant drop compared to the Supervised, which shows the necessity of unsupervised domain adaptation. In order to verify the effectiveness of our method, we compare our method with a series of general unsupervised domain adaptation methods: MinEnt [16], MaxSquare [15], and ADDA [17]. For a fair comparison, these methods are reproduced with the same setting in our framework, where the hyperparameters are adjusted to obtain the best performance on all domain adaptation scenarios. The PL is the same Pseudo-Label training strategy in the [20] with our framework, which is a two-stage method with extra training cost. Furthermore, we introduce the geodesic correlation alignment used in | 3,069,060 | 251442662 | 0 | 16 |
[5] into our segmentation framework to construct additional comparison method 3DGCA. It can be observed from the Tab. I that although the above methods can alleviate the domain discrepancy, they are not efficient for point cloud semantic segmentation. Especially with the global-level feature alignment methods, the model produces the confused semantic information in the feature space. In the S3DIS to ScanNet scenario, these methods even produce negative effects on domain adaptation. Furthermore, we compare our method with the 3D unsupervised domain adaptation methods: SqueezeSegV2 (SQSGV2) [5] and Complete & Label (C&L) [3]. Because these methods use different point cloud semantic segmentation backbones, we directly use the results of SQSGV2 and C&L reported in [3] for comparison. Since SQSGV2 requires spherical projection and C&L requires sequences of point cloud for the completion network, it is limited to reproduce in the vKITTI to SemanticPOSS and S3DIS to ScanNet domain adaptation scenarios. As shown in the Tab. I, compared with the above methods, our method achieves state-of-the-art performance on three domain adaptation scenarios. D. Ablation Studies and Analysis In order to further verify the effects of each module of our method and the effectiveness of the proposed assignment matrix based local feature loss, we conduct ablation studies on the vKITTI to SemanticPOSS scenario. As shown in Tab. II, we first report the performance improvement brought by each proposed loss, where the quantitative results of each loss can show its effective domain transferability. Particularly, the integration of the two losses can benefit the overall domain adaptation framework, and further improve | 3,069,061 | 251442662 | 0 | 16 |
the domain adaptation performance. It can be observed that our framework can benefit from a simple pseudo-label training strategy with additional 3.0% improvement, and they play a complementary role in unsupervised domain adaptation. Secondly, in order to verify that the feature distributions of different categories has been separated, we draw t-SNE [29] visualization with the features from the target domain to show qualitative results. As shown in Fig. 3, our proposed method can effectively enhance the discrimination of features from the target domain. Thirdly, we conduct the ablation study without using the assignment matrix named L loc w/o A. In this case, we use the mean features mentioned in Sec.III-C to represent the local feature graphs, and directly select the nearest neighbor from the memory bank to find the relationship between the graphs from the source domain and the target domain. Instead of using the assignment matrix for the alignment, we use the mean features to directly align the two graph features. As shown in Tab. II, the proposed assignment matrix based local feature loss can achieve a better performance. At last, we show the visualization of point cloud semantic segmentation results to qualitatively illustrate the effectiveness of our method. It can be clearly observed in Fig. 2, compared with the Source Only, only a few noise predictions are produced in our method, which shows the proposed framework can effectively alleviate the domain gap problem and significantly improve the segmentation performance. V. CONCLUSION In this paper, we proposed an unsupervised domain adaptive point cloud semantic segmentation | 3,069,062 | 251442662 | 0 | 16 |
framework based on feature graph matching. With the proposed assignment matrix based local feature loss and category-guided contrastive loss, we can align the local-level feature distributions of the source domain and the target domain more accurately in a meticulous way and guide the segmentation model to learn discriminative features on the target domain. Extensive experiments on different synthetic-to-real and real-to-real domain adaptation scenarios have demonstrated the superiority of our method. | 3,069,063 | 251442662 | 0 | 16 |
Improving a Bimanual Motor Skill Through Unimanual Training When we learn a bimanual motor skill (e.g., rowing a boat), we often break it down into unimanual practices (e.g., a rowing drill with the left or right arm). Such unimanual practice is thought to be useful for learning bimanual motor skills efficiently because the learner can concentrate on learning to perform a simpler component. However, it is not so straightforward to assume that unimanual training (UT) improves bimanual performance. We have previously demonstrated that motor memories for reaching movements consist of three different parts: unimanual-specific, bimanual-specific, and overlapping parts. According to this scheme, UT appears to be less effective, as its training effect is only partially transferred to the same limb for bimanual movement. In the present study, counter-intuitively, we demonstrate that, even after the bimanual skill is almost fully learned by means of bimanual training (BT), additional UT could further improve bimanual skill. We hypothesized that this effect occurs because UT increases the memory content in the overlapping part, which might contribute to an increase in the memory for bimanual movement. To test this hypothesis, we examined whether the UT performed after sufficient BT could improve the bimanual performance. Participants practiced performing bimanual reaching movements (BM) in the presence of a novel force-field imposed only on their left arm. As an index for the motor performance, we used the error-clamp method (i.e., after-effect of the left arm) to evaluate the force output to compensate for the force-field during the reaching movement. After sufficient BT, the training | 3,069,064 | 17386851 | 0 | 16 |
effect reached a plateau. However, UT performed subsequently improved the bimanual performance significantly. In contrast, when the same amount of BT was continued, the bimanual performance remained unchanged, highlighting the beneficial effect of UT on bimanual performance. Considering memory structure, we also expected that BT could improve unimanual performance, which was confirmed by another experiment. These results provide a new interpretation of why UT was useful for improving a bimanual skill, and propose a practical strategy for enhancing performance by performing training in various contexts. INTRODUCTION When we try to learn a complicated motor skill, we often break it down into its simpler fundamental skills (Part practice: Schmidt and Wrisberg, 2007;Schmidt and Lee, 2011). One of the representative examples is bimanual skills. In the case of rowing a boat, for example, it is a common practice to pull an oar with each arm separately before rowing with both arms together (McArthur, 1997). Practically, this type of training is beneficial, because a single yet complicated unimanual skill can be trained first, before performing the same action bimanually (Schmidt and Wrisberg, 2007). However, it is not as straightforward as assuming that unimanual training (UT) improves bimanual performance. We have demonstrated that the adaptation of reaching movements to a novel force-field environment is only partially transferred to the same arm movement when the movement of the opposite arm is absent (i.e., a unimanual reaching movement: UM) or present (i.e., a bimanual reaching movement: BM; Nozaki et al., 2006;Nozaki and Scott, 2009;Kadota et al., 2014). From this observation, we proposed | 3,069,065 | 17386851 | 0 | 16 |
that the motor memories for identical movements are partially segregated: UM-specific, BM-specific, and overlapping parts ( Figure 1A). Such a memory structure would explain why motor adaptation is only partially transferred between unimanual and bimanual movement. Recent studies have shown that the motor memories for an identical movement can be flexibly switched according to different behavioral contexts, such as how the opposite arm is moving (Howard et al., 2010;Yokoi et al., 2011Yokoi et al., , 2014, whether the movement is performed discretely or rhythmically (Ikegami et al., 2010;Howard et al., 2011), and what kind of movement followed afterwards (Howard et al., 2015). This memory structure suggests that UM training is not particularly effective for improving BM skills, because its training effect would be transferred only partially to BM. Counterintuitively, however, we speculated that UM training could improve BM performance based on this memory structure. Consider the case when the adaptation of the left arm to a certain level of force-field is achieved by BM training. The adaptation effect is stored in the BM memory (i.e., BM-specific and overlapping parts; Figure 1A). Once the training effect is virtually saturated, the amount of BM memory cannot be increased any further by performing additional BM training trials ( Figure 1B). It should be noted that Figure 1B does not mean that the memory is full, but that the amount of BM memory has reached the upper limit that can be learned through the training. For example, if the force-field level is further increased, the movement error resulting from the | 3,069,066 | 17386851 | 0 | 16 |
increase can increase the amount of BM memory. However, this is not the only way to increase the amount of BM memory. Even if the forcefield level is maintained, the amount of BM memory can be also improved by performing UM training, as described below. Since the total amount stored in UM memory (i.e., UM-specific and overlapping components) is not yet saturated, performing UM could lead to movement error. This movement error could increase the total amount of UM memory ( Figure 1C). Since this increment is accompanied by an increase in memory in FIGURE 1 | Possible beneficial effect of unimanual training (UT) on bimanual performance, based on partially segregated motor memories. (A) Motor memories for identical reaching movements consist of three distinct parts: unimanual-specific, bimanual-specific, and overlapping parts, as we reported previously (Nozaki et al., 2006;Nozaki and Scott, 2009). (B) Bimanual training (BT) effects should be stored in the bimanual-related parts (i.e., bimanual-specific and overlapping part). Once the memory content reached a plateau, it is not possible to increase the memory content by performing additional BT. (C) However, performing additional UT could enhance bimanual performance by adding training effects to the overlapping part of the memory structure. Broken circles in (B,C) represent the training contexts. the overlapping part, the total amount of motor BM memory (i.e., BM-specific and overlapping parts) is further increased (Figure 1C), which can lead to improvement in BM performance. In other words, the UM training may bring a breakthrough effect over that achieved by BM training. In order to test | 3,069,067 | 17386851 | 0 | 16 |
this hypothesis, we used an experimental paradigm of motor adaptation to a novel force-field during a reaching movement to investigate whether performing UM training after BM training could improve BM performance more than continuing to perform only BM training could. Considering that the memory structure is almost symmetrical (Figure 1A), we predicted that the same effect should be also observed for the UM performance as when additional BM training was performed, and tested this hypothesis. Participants Fifty-two right-handed participants (35 male, 17 female, age: 19-52 years) were recruited in the following experiments. The participants had no reported cognitive, motor, or neurological disorders. We obtained written informed consent from all participants prior to commencing experiments. The Ethics Committee of the University of Tokyo reviewed and approved the experimental protocol that was in accordance with the Declaration of Helsinki. Apparatus and Motor Tasks The participants performed horizontal UM and BM holding robotic handles (KINARM End-Point lab, Bkin Technologies, Kingston, ON, Canada; Figure 2A). They sat in an adjustable chair to which their back was strapped. Their wrists were constrained by braces so that unnecessary wrist movements did not occur. They were instructed to move white cursors (diameter: 10 mm) representing the hand positions, from start positions (diameter: 14 mm) to targets (diameter: 14 mm) on a horizontal display. They could not directly see the movement of their arms. The start positions were located at approximately 15 cm in front of their body and the distance between the positions for both arms was 15 cm. The targets were located | 3,069,068 | 17386851 | 0 | 16 |
10 cm straight ahead of the start positions. To begin each trial, the participants were required to move and maintain the cursor at the start positions for 1 s. A green target appeared and then turned magenta after additional 1-1.5 s holding times, indicating the ''go'' cue. The participants were instructed to perform the UM with the left arm or to perform BM toward the targets, as straight as possible, and to focus on the left hand movements even during the BM condition. Warning messages were presented immediately below the start positions if the peak movement velocity was below (''slow'') or above (''fast'') a range (340-460 mm/s). After completion of each trial, the handles were automatically moved back to the start positions without the participant's efforts. For the training, the participants performed reaching movements (UM or BM) under the presence of a velocitydependent curl force-field (imposed only on the left arm) as f = Bν (Shadmehr and Mussa-Ivaldi, 1994; Figure 2B), where f = [f x ; f y ] (N) is the force to the left handle, ν = [ν x ; ν y ] (m/s) is the left handle velocity, B = [0, −b; b, 0] is the viscosity matrix, and b (N/[m/s]) is the viscosity. Thus, in this study, the left hand movement was trained by the force-field for both UM and BM. Half of the participants adapted to a clockwise forcefield (b = −15) and the other participants adapted to a counterclockwise force-field (b = −15). The right hand movements during BM | 3,069,069 | 17386851 | 0 | 16 |
were not perturbed throughout the experiment. To evaluate the effect of training of the left arm (i.e., the level of adaptation), we used the ''error-clamp method'' with which the movement trajectory of the left handle was constrained to a straight path from the start position to the target by a virtual force-channel (Scheidt et al., 2000; Figure 2B). The forcechannel was created by a virtual spring (6000 N/m) and dumper (100 N/[m/s]) in the perpendicular direction to the straight path of movement. This method enabled us to measure lateral force output exerted against the channel (i.e., the after-effect) directly. Error-clamp trials were not applied to the right arm throughout the experiments. Experimental Groups and Flows There were four experimental groups, determined according to the training protocols assigned ( Figure 2C). In one of the groups (Bimanual Training (BT), n = 12), after 10 BM trials in the null force-field condition, participants performed BM training for four consecutive sets. In each set, participants performed 64 BM training trials in the training period ( Figure 2C). This training period was followed by a testing period in which 10 BM training, 10 BM error-clamp trials, and 10 UM errorclamp trials were pseudo-randomly interleaved, to evaluate the level of motor adaptation for BM and UM. Since such breaks would decay the level of adaptation to the force-field acquired by the previous training set, the participants performed 10 training trials used in the previous set (in this case, BM training) at the beginning of each set (the 2nd, 3rd, and 4th | 3,069,070 | 17386851 | 0 | 16 |
set), to recover the adaptation level again to the level before the break. To test the hypothesis that UM training enhanced BM performance, in another experimental group (Bimanual-Unimanual Training (BUT) group; n = 12), the 3rd training set was replaced by UM training trials ( Figure 2C). That is, the testing period in the BUT group included 10 UM training trials, and 10 BM and 10 UM error-clamp trials. We compared the BM performance at the end of the 3rd set between the BT and BUT groups. If the BM performance was greater for the BUT group than for the BT group, we could conclude that UM training had a beneficial effect on BM performance. We also examined whether BM training enhanced UM performance by testing the UT and UBT group (n = 14 for each group). The UT group performed four consecutive UM training sets, while the UBT group trained with BM training in the 3rd set ( Figure 2C). The settings of the training period, testing period, and short break were identical to those of the BT and BUT groups. Data Analysis The handle positions and forces were sampled at 1000 Hz. The data regarding the handle positions were smoothed by a fourth-order Butterworth filter with a cutoff frequency of 10 Hz. To quantify kinematic errors in the training conditions, we measured the lateral deviation at the peak velocity of the handle, from the straight line between the start positions to the targets. A lateral deviation in the force-field direction was defined as positive. | 3,069,071 | 17386851 | 0 | 16 |
In test trials, we evaluated the after-effect for the adaptation level as f pν /ν pν , where the f pν was the lateral force against the force-channel evaluated at the peak velocity ν pν . The In the BT group, four BM training sets (bars with red hatched lines) were performed (each set consisted of 64 training trials), and in the test period at the end of each set (gray-shaded area), the training effects on BM and UM performance were evaluated by pseudo-randomly interleaving BM and UM test (10 error-clamp trials for each movement). At the beginning of each set, the participants performed 10 training trials in the same training context as used in the previous set, so that the training effect could be recovered after a short break (2-3 min). In the BUT group, the 3rd training set was replaced by UM training. The procedures were identical for the UT and UBT groups. f pν in the opposite direction of the force-field was defined as positive. A value of 15 (i.e., |b|) indicates that the adaptation level was 100%, whereas a value of 0 indicates no adaptation. To examine how performing training trials with different movement patterns in the 3rd set changed the adaptation levels, we performed a 2-way repeated measures ANOVA (Groups and Sets) for each of the movement types (i.e., the UM and BM test trials). If a significant interaction and a simple main effect for groups were obtained, we continued to perform multiple comparisons, with Bonferroni correction as post hoc tests. | 3,069,072 | 17386851 | 0 | 16 |
The statistically significant threshold was set at P < 0.05. Figure 3 shows the lateral deviations of both hands at the peak velocities in training trials for the BT group ( Figure 3A) and the BUT group ( Figure 3B). The kinematic errors of the left hand (the left panel of Figures 3A,B) were produced by the force-field in the initial few trials, but gradually reduced by the end of the 1st set. Note that, since the force-field was imposed only on the left hand, the right hand movements were not affected by it (the right panel of Figures 3A,B). In the 3rd set of the BUT group ( Figure 3B), a slight increase in error was observed, because the training context switched from BM to UM. This result was consistent with a previous finding that the training effect of BM was only partially transferred to UM (Nozaki et al., 2006;Nozaki and Scott, 2009;Kadota et al., 2014). RESULTS At the end of each set, we interleaved 10 BM and 10 UM error-clamp trials to evaluate the level of adaption by the force output (i.e., the after-effect). During this training period, the movement error (i.e., lateral deviation) slightly increased in the training trials during the test period (Figures 3A,B), possibly due to the memory decay caused by the error-clamp trials. We examined whether the adaptation level was maintained throughout the test period. If the motor memory gradually decreased along with memory decay, the lateral deviation of the hand during the test period would also gradually increase. However, | 3,069,073 | 17386851 | 0 | 16 |
Figures 3C,D indicate there was no such systematic data trend in lateral deviation during the test period (10 training trials were performed for each set). Figure 4 indicates how the adaptation level changed with sets for the BT group ( Figure 4A) and the BUT group ( Figure 4B). A 2-way repeated measures ANOVA applied to the adaptation level of BM revealed significant interactions between the BT and BUT groups (Groups and Sets: F (1,3) = 5.39, p = 2.22 × 10 −3 , Figures 4A,B). There was no significant simple main effect of group for 2nd (F (1,22) = 1.80, p = 0.19) and 3rd sets (F (1,22) = 1.73, p = 0.20), but a significant simple main effect of sets was observed for both the BT group (F (3,22) = 5.58, p = 5.30 × 10 −3 ) and the BUT group (F (3,22) = 17.37, p = 5.20 × 10 −6 ). Multiple comparisons with Bonferroni corrections indicated that the adaptation level remained unchanged after the 2nd set for the BT group (2nd and 3rd set: t (11) = 0.552, p > 1, corrected), implying that The changes in the adaptation levels from 2nd to 3rd sets. * Indicates statistically significant differences in the levels between the 2nd and 3rd sets only for the BUT group (p < 0.05, Bonferroni correction). (D) A significant correlation was found between improvement in UM performance and that in BM performance. All error bars indicate SEM. the adaptation level had almost reached a plateau ( Figure 4C). | 3,069,074 | 17386851 | 0 | 16 |
In contrast, for the BUT group, the adaptation level was significantly increased from the 2nd to the 3rd set (t (11) = 3.81, p = 0.017, corrected), indicating the beneficial effect of UM training on BM performance ( Figure 4C). The adaptation level appears to decrease from the 3rd to the 4th set, but it was not statistically significant (t (11) = 2.34, p = 0.24, corrected). Of added note, since BT was conducted throughout all four sets, the UM adaptation level quantified by UM error-clamp trials was smaller than that for BM (Figures 4A,B), which was consistent with our previous finding of partial learning transfer from BM to UM skills (Nozaki et al., 2006;Nozaki and Scott, 2009). In the BUT group, the training in the 3rd set was performed unimanually, which increased the UM adaptation level from the 2nd to the 3rd set. A 2-way repeated measures ANOVA applied to the UM adaptation level revealed significant interactions between the BT and BUT groups (Groups and Sets: F (1,3) = 22.64, p = 3.43 × 10 −10 , Figures 4A,B). There was a significant simple main effect of group, not for the 2nd set (F (1,22) = 2.69, p = 0.12), but rather for the 3rd set (F (1,22) = 17.70, p = 3.64 × 10 −4 ). A significant simple main effect of sets was observed only for the BUT group (BUT group: F (3,22) = 33.42, p = 2.27 × 10 −8 ; BT group: F (3,22) = 0.78, p = 0.52). Multiple comparisons | 3,069,075 | 17386851 | 0 | 16 |
with Bonferroni corrections indicated that the adaptation level remained unchanged after the 2nd set for the BT group (2nd and 3rd set: t (11) = 1.74, p = 0.66, corrected; Figure 4A). In contrast, for the BUT group, the UM adaptation level significantly increased from the 2nd to the 3rd set (t (11) = 5.67, p = 8.73 × 10 −4 , corrected). Figure 4D indicates that there was a significant correlation between changes in BM and UM performance from the 2nd to the 3rd set (R 2 = 0.3600, t (10) = 2.372, p = 0.039), suggesting that the increase in the UM adaptation level (UM performance) might contribute to the increase in the BM adaptation level (BM performance), possibly through an increase in memory in the overlapping part of the motor memory structure ( Figure 1C). Figures 5, 6 indicate the results for the UT and UBT groups. Essentially, the results were similar to those observed for the BT and BUT groups. The kinematic errors of the left hand (the left panel of Figures 5A,B) were produced by the forcefield in the initial few trials, but gradually reduced by the end of the 1st set. In the 3rd set of the UBT group (Figure 5B), a slight increase in error was observed, because the training context switched from UM to BM. Figures 5C,D indicate that there was no systematic trend in the lateral deviation of the hand during the test period, indicating that gradual memory decay during the test period did not occur. As | 3,069,076 | 17386851 | 0 | 16 |
for the level of motor adaptation, a 2-way repeated measures ANOVA revealed significant interactions for the adaptation levels of UM between the UT and UBT groups (Groups and Sets: F (1,3) = 2.94, p = 0.038, Figures 6A,B). There was a significant simple main effect of group, not for the 2nd set (F (1,26) = 0.371, p = 0.55), but rather for the 3rd set (F (1,26) = 6.98, p = 0.013), and a simple main effect of sets was also significant for both groups (UT group: F (3,26) = 7.46, p = 9.25 × 10 −4 ; UBT group: F (3,26) = 16.21, p = 3.83 × 10 −8 ). Multiple comparisons with Bonferroni corrections revealed that the adaptation level in the UT group remained unchanged after the 2nd set (2nd and 3rd set: t (13) = 1.36, p > 1, corrected; 3rd and 4th set: t (13) = 1.51, p = 0.93, corrected; Figure 6C), indicating that the adaptation level had virtually reached a plateau after the first two sets. However, the adaptation levels significantly improved from the 2nd to the 3rd set only for the UBT group (t (13) = 3.23, p = 0.039, Adaptation levels were calculated by the ratio of the after-effect (the generated force to the virtual force-channel) to the peak velocity for the UT group (A) and the UBT group (B). (C) Changes in the adaptation levels from the 2nd to 3rd sets. * Indicates the statistically significant differences in the levels between the 2nd and 3rd sets only | 3,069,077 | 17386851 | 0 | 16 |
for the UBT group (p < 0.05, Bonferroni correction). (D) A significant correlation was found between improvement in BM performance and that in UM performance. All error bars indicate SEM. corrected; Figure 6C). Thus, performing BM training after UM training contributed to improvement in the UM performance more than did continuance of UM training. The adaptation level of UM appeared to decrease from the 3rd to the 4th set, but this was not statistically significant (t (13) = 2.02, p = 0.39, corrected). Consistent with the result in the BT group, the BM adaptation level in the UT group was consistently smaller than the UM adaptation level (Figures 6A,B). In the UBT group, the training in the 3rd set was performed bimanually, which increased the BM adaptation level from the 2nd to the 3rd set. Indeed, a 2-way repeated measures ANOVA applied to the BM adaptation level revealed significant interactions between the UT and UBT groups (Groups and Sets: F (1,3) = 21.42, p = 3.21 × 10 −10 , Figures 6A,B). There was a significant simple main effect of group, not for the 2nd set (F (1,26) = 0.81, p = 0.37), but rather for the 3rd set (F (1,26) = 22.59, p = 6.45 × 10 −5 ). A significant simple main effect of sets was observed only for the UBT group (UBT group: F (3,26) = 45.55, p = 1.72 × 10 −10 ; UT group: F (3,26) = 0.14, p = 0.93). Multiple comparisons with Bonferroni corrections indicated that the BM adaptation | 3,069,078 | 17386851 | 0 | 16 |
level remained unchanged after the 2nd set for the UT group (2nd and 3rd set: t (13) = 0.087, p > 1, corrected), but for the UBT group, the adaptation level significantly increased from the 2nd to the 3rd set (t (13) = 6.95, p = 6.08 × 10 −5 , corrected). Figure 6D indicates that there was a significant correlation between the changes in UM performance from the 2nd to the 3rd set and those in BM performance (R 2 = 0.688, t (12) = 5.149, p = 2.415 × 10 −4 ), again indicating that improvement in UM performance was likely to result from the increase in BM performance. Partially Overlapping Memory Structure of UM and BM According to the partial motor learning transfer between UM and BM, we previously proposed the multicompartment memory model, which consisted of three parts: UM-specific, BM-specific, and overlapping parts ( Figure 1A; Nozaki et al., 2006;Nozaki and Scott, 2009). Consistent with these previous observations, the present study also provided evidence of partial learning transfer from UM to BM, and vice versa. For example, in the 2nd set for the BT and BUT groups, the ratio of UM adaptation level to BM adaptation level was 73.3 ± 3.3% (BT group) and 70.0 ± 4.6% (BUT group; Figures 4A,B). Similarly, in the 2nd set for the UT and UBT groups, the ratio of BM adaptation to UM adaptation level was 52.1 ± 6.1% (UT group) and 58.5 ± 6.6% (UBT) group (Figures 6A,B). Previous, single-unit recording studies using nonhuman primates | 3,069,079 | 17386851 | 0 | 16 |
have demonstrated that the MI consists of neurons that respond specifically to either UM or BM (Donchin et al., 1998(Donchin et al., , 2002. For example, in the report by Donchin et al. (2002), among 187 primary motor cortex (MI) neurons recorded from two monkeys, 21 and 38 neurons were active for UM or BM, respectively, and 128 neurons were active for both types of movements. In other words, partially different neural populations in MI are involved in identical reaching actions between UM and BM. It is also well known that the neurons of MI change their activity patterns to adapt to a novel dynamic environment (Li et al., 2001;Arce et al., 2010). Consistent with these findings, the corticospinal excitability in humans while performing wrist movement, evaluated by transcranial magnetic stimulation to MI, also changed after the adaptation to a force-field (Kadota et al., 2014), indicating a significant role of MI for motor adaptation. Thus, partially different neuronal populations could be recruited for motor adaptation involving UM and BM, leading to development of a partially segregated memory structure. Improvement of BM Performance by Additional UM Training The structure of partially segregated motor memories has been shown to enable identical reaching movements to adapt to conflicting force-fields, depending on whether the opposite arm is stationary (UM) or moving together (BM), which might contribute to flexible motor control under the presence of mechanical influence caused by movement of the opposite arm (Yokoi et al., 2011(Yokoi et al., , 2014. However, this structure limits the transfer of the training | 3,069,080 | 17386851 | 0 | 16 |
effect from UM to BM. Although the UM practice is widely used and is considered useful for acquiring bimanual skills, this limiting factor of the training effect needs to be considered. In contrast to the intuitive idea of a limiting training effect, however, we speculated that, due to the memory structure, the UM training could improve the BM performance. This is because even after BM performance reaches a plateau, UM training could enhance BM performance by increasing the motor memory level in the overlapping part of the memory structure ( Figure 1C). Consistent with this speculation, we have demonstrated that interleaving UM training enhanced BM performance from the 2nd to the 3rd set (BUT group) more than did continuation of only BM training for the same number of trials (BT group; Figures 3A,B). We assumed that the beneficial effect of UM training was caused by an increment of the motor memory stored in the overlapping part of the memory structure ( Figure 1C). This assumption was verified by the observation that the degree of improvement of BM was significantly correlated with that of UM induced by the interleaved UM training ( Figure 4D). An alternative explanation would be accounted for by the structure of the test period. For example, the test period of the 2nd set for the BUT group consisted of BM and UM errorclamp trials interleaved with BM training trials. Due to the decay of motor memory resulting from the error-clamp trials, the participants experienced relatively large movement errors during the BM training trials | 3,069,081 | 17386851 | 0 | 16 |
(Figures 3C,D), but no movement error during the BM error-clamp trials. Repeatedly experiencing two different errors for the same BM movements during the test period enabled the participants to detect the contextual change more easily, which might decrease the expression of motor memory (Vaswani and Shadmehr, 2013). In contrast, this decrease in BM memory might not occur in the test period of the 3rd set for the BUT group, because the participants performed only BM error-clamp trials. In other words, the BM motor memory was suppressed only when the BM error-clamp trials were interleaved with the BM training trials. However, this concept was unlikely to explain why the degree of improvement of BM was significantly correlated with that of UM induced by the interleaved UM training ( Figure 4D). Furthermore, if this concept was correct, the movement error (i.e., lateral deviation of the handle) during the test period should have gradually increased within each set as the opportunity to experience two levels of movement error increased. However, we did not observe such a data trend in lateral deviation within each set (Figures 3C,D). Thus, our interpretation that the increase in BM memory resulted from UM training is more likely. However, it should be noted that BM training in the 4th set did not improve UM performance (Figure 4B). This seems inconsistent with the beneficial effect of BM training on UM performance. However, as the BM performance had almost fully developed, virtually reaching a plateau by the end of the 3rd set, performing BM training in the | 3,069,082 | 17386851 | 0 | 16 |
4th set could not add additional training effect to the overlapping memory structure. It should also be noted that the BM adaptation level appeared to become smaller from the 3rd to the 4th set, although the decrease was not significant. However, the BM adaptation level in the 4th set was not significantly different from that in the 2nd set, indicating that the beneficial effect of performing additional UM training on BM performance could be temporary. Future study is necessary to determine the duration of beneficial effect. It should be noted that the nondominant left arm was used as the trained arm in the present study. Thus, one would wonder if the current idea is applicable when the dominant right arm is used as the trained arm. If the structure of the motor memory ( Figure 1A) is different between right and left arms, the beneficial effect of UM training on BM performance should also change. The presence of different memory structures is possible, considering that sensorimotor areas in the dominant hemisphere have shown greater influence over the nondominant hemisphere in both functional magnetic resonance imaging (Hayashi et al., 2008;Diedrichsen et al., 2013) and electrophysiological studies (Netz et al., 1995;Oda and Moritani, 1995;Ziemann and Hallett, 2001;Duque et al., 2007). In our previous study (Yokoi et al., 2014), we had righthanded participants adapt the forward movement of one arm (left or right) to a velocity-dependent curl force-field, while moving the opposite arm in the forward direction, and examined how this adaptation effect was influenced when the movement direction | 3,069,083 | 17386851 | 0 | 16 |
of the opposite arm was changed from the trained direction. We found that the influence on the left hand was greater when the movement direction of the right arm was changed, compared with the influence on the right arm from the left arm, indicating that the motor memory of the nondominant left arm is more strongly influenced by the movement pattern of the dominant right arm than vice versa. If the motor memory of the right arm is relatively independent of the movement of the left arm, as implied by the previous observation described above (Yokoi et al., 2014), the overlap between UM and BM for motor memory of the right arm should be greater than the overlap for motor memory of the left arm. Greater overlap should decrease the beneficial effect of UM training on BM performance, because it is obvious that the benefit is lost when the motor memories completely overlap. However, no previous work has systematically examined the laterality of the motor memory structure. Future work is necessary to address this issue and the resultant beneficial effect of UM training on BM performance for the dominant right arm. Influence of Memory Decay We also need to consider how the memory stored in the BMspecific part changed during the UM training in the 3rd set of BUT group. Since the BUT group performed only UM training during the 3rd set, the memory stored in the BM-specific part of the memory structure was not updated during the 3rd period. This should result in time-dependent memory | 3,069,084 | 17386851 | 0 | 16 |
decay. If the amount of memory decay in the BM-specific part was greater than the increment in the overlapping part induced by the UM training, the total amount of motor memory for BM (i.e., stored in BM-specific and overlapping parts) would decrease. We investigated our data from this point of view. In the BUT group, after the 2nd set, the after-effect of BM and UM was 9.9 and 6.8, respectively (Figure 4B), indicating that the amount of memory in the overlapping part was 6.8 and that in the BM-specific part was 3.1 (= 9.9-6.8; Figure 7A). In the model we previously proposed (Nozaki and Scott, 2009), the memory stored in the BM-specific part while performing UM was assumed to decay with a constant of 0.99 with every UM trial. According to this assumption, the memory in the BMspecific part should decrease to 53% (= 0.99 64 ) during the UM training trials in the 3rd set in the BUT group. Thus, the amount of memory stored in the BM-specific part can be estimated as 3.1 × 0.53 = 1.6 ( Figure 7B). On the other hand, after the UM training in the 3rd set of the BUT group, the after-effect of BM and UM was 12 and 10.2, respectively (Figure 4). Assuming that the memory in the BM-specific part is 1.6, as calculated above, the amount of memory in the overlapping part should be 10.4 (= 12 − 1.6; Figure 7B). However, this is not possible, as the UM after-effect indicates that the memory in the | 3,069,085 | 17386851 | 0 | 16 |
overlapping part cannot exceed 10.2 ( Figure 7B). Therefore, if the memory in the BM-specific part decays by a constant of 0.99 with every trial, the increase in the motor memory in the overlapping part by UM training cannot compensate for the decrement of the memory in the BM-specific part. How then is the beneficial effect explained? Recently, it has been reported that the decay of motor memory is contextdependent (Ingram et al., 2013). In this study, after adapting reaching movements in the backward direction to the velocitydependent force-field, participants performed 30 error-clamp reaching trials in a backward direction (i.e., in the same context) or in a forward direction (i.e., in a different context). During the error clamp trials, the adaptation level of reaching movement in backward direction decayed, but the amount of decay was smaller after 30 reaching movements in the forward direction than in the backward direction. Therefore, they concluded that the decay of motor memory was smaller when performing movements that were different from the training movement. We assumed that this context-dependent decay effect was also likely to occur during UM and BM. More specifically, when the participants performed UM training, the motor memory in the BM-specific part decayed only slightly. When the retention constant was assumed to be 0.9983, as reported by Ingram et al. (2013), the amount of memory stored in the BM-specific part became 3.1 × 0.9983 64 = 2.8 by the end of UM training ( Figure 7C). The memory in the overlapping part can be estimated to be | 3,069,086 | 17386851 | 0 | 16 |
9.2 (= 12-2.8). This value indicates that the memory stored in the overlapping part increased by 2.4 (from 6.8 to 9.2). Considering the UM after-effects after the UM training (i.e., 10.2), the memory in the UM-specific part should be 1.0 (= 10.2 − 9.2). Thus, the UM training increased memory in the UM-specific part and the overlapping part from the 2nd to the 3rd set by 1.0 and 2.4, respectively, indicating that the beneficial effect of UM training on BM performance can be explained by FIGURE 7 | Estimation of how the content for each memory part changed from the 2nd to the 3rd training set for the BUT group. (A) From the data after the 2nd BM training for the BUT group (the adaptation level was 9.9 and 6.8 for BM and UM, respectively; see Figure 4B), the content for each memory part can be estimated as 0 (UM-specific part), 6.8 (overlapping part), and 3.3 (BM-specific part). (B) During the subsequent 3rd UM training, the memory in the BM-specific part was not updated, but should decay with time. If the retention constant was assumed to be 0.99, according to our previous work (Nozaki and Scott, 2009), then the memory content for the BM-specific part should decrease to 1.6. Using the data from after the 3rd UM training (the adaptation level was 12.0 and 10.2 for BM and UM, respectively; see Figure 4B), the memory content could be estimated at −0.2 (UM-specific part) and 10.4 (overlapping part). However, this does not seem possible, because this indicates | 3,069,087 | 17386851 | 0 | 16 |
that the memory content in the UM-specific part decreased to a negative value with UM training. (C) In contrast, if the retention constant was greater, as proposed by Ingram et al. (2013), we could estimate that a reasonable amount of memory content was to be found in all three parts. See the main text for the details of estimation. assuming that the memory in BM-specific part remains stable during UM training. The beneficial effect of BM training on UM performance can be explained in the same way. Practical Implications The issue of how effectively and rapidly motor skills can be acquired (or reacquired) is important for athletes, musicians, and patients with motor dysfunction. For bimanual motor skills, one of the strategies commonly adopted is to break down a whole motor skill into simpler unimanual skills, and then train each hand (or arm) separately (part practice; McArthur, 1997;Burke, 2002;Finch, 2004;Schmidt and Wrisberg, 2007;Schmidt and Lee, 2011). Intuitively, this type of practice should be beneficial for acquiring motor skills by each hand (or arm) before doing it bimanually, but this had not been sufficiently investigated until we had previously demonstrated that the training effect of UM practice was only partially transferred to the same movement performed bimanually (Nozaki et al., 2006). The results of this study seemed to imply that the effect of UM training on BM performance was limited, but the present study has demonstrated that this is not the case: we show that BM performance can be facilitated by performing UM training. As shown in Figures | 3,069,088 | 17386851 | 0 | 16 |
4B,C, even if the BM performance had reached a plateau, subsequently performed UM training could improve the BM performance, possibly through an increase in the motor memory content of the overlapping part of the memory structure. This result may provide a novel insight into the reason for the efficacy of UM training in improving BM performance. Thus far, we have emphasized the beneficial effect of UM training on BM performance; however, based on the consideration of motor memory structure for UM and BM (Figure 1A), we also expected a beneficial effect of BM training on UM performance. Indeed, we verified that BM training after sufficient UM training may also facilitate UM performance. The beneficial effect of BM on UM performance has been reported for the bilateral training adopted for rehabilitation in stroke patients (Cunningham et al., 2002;Cauraugh and Summers, 2005;Choo et al., 2015). The effect of bilateral training on the paretic arm is thought to be because simultaneous movement of the nonparetic limb can facilitate movement of the paretic limb through the interlimb coupling observed in healthy individuals (Swinnen, 2002). Our study has provided further insights into why BM training is beneficial for improving UM performance. To our knowledge, except for the bilateral training described above, no previous studies have shown the beneficial effect of BM training on UM performance. Considering that sports using only one arm (e.g., tennis, throwing a ball, etc.) are ubiquitous, it would be of practical interest to investigate how training with moving the opposite arm could improve UM motor skills. These | 3,069,089 | 17386851 | 0 | 16 |
findings also indicate that the beneficial effect is not related to the notion that UM is a part of BM, but is related to the overlapping memory structure. Recent studies have shown that the motor memories for identical (reaching) movements are changed according to different behavioral contexts, such as the movement pattern of the opposite arm (Yokoi et al., 2011(Yokoi et al., , 2014, discrete vs. rhythmic movements (Ikegami et al., 2010;Howard et al., 2011), or the movement pattern of the lead-in or follow-through movements (Howard et al., 2015). It should be noted that motor memories for different contexts are not completely distinct, but overlap partially. The presence of an overlapping memory structure suggests that performing training in different behavioral contexts rather than in the original context might help to facilitate motor skill development because it might increase the memory content in the overlapping part of the memory structure. This concept is consistent with the classical idea that variable training is more effective for acquiring motor skills (McCracken and Stelmach, 1977;Shea and Kohl, 1991). However, this concept was based on the schema theory (Schmidt, 1975), according to which variable training could contribute to developing a generalized motor program. However, our study provides another mechanism by which variable practice can have a beneficial effect. Finally, since the present study focused only on the short-term training effect, it is not clear how long the facilitated training effect would last. Future studies are necessary to examine the longterm effect of performing training in different behavioral contexts. AUTHOR CONTRIBUTIONS TH | 3,069,090 | 17386851 | 0 | 16 |
performed the experiments and analyzed the data. TH and DN designed the study and wrote the manuscript. | 3,069,091 | 17386851 | 0 | 16 |
Time-synchronized immune-guided SBRT partial bulky tumor irradiation targeting hypoxic segment while sparing the peritumoral immune microenvironment Background A novel unconventional SBRT-based PArtial Tumor irradiation targeting HYpoxic clonogenic cells (SBRT-PATHY) for induction of the tumoricidal bystander (BE) and abscopal effects (AE) was developed by translating our preclinical findings to a clinic in 2016. In order to further improve BE/AE response rate, SBRT-PATHY was upgraded in 2018 by the sparing of peritumoral immune microenvironment as a new OAR, defined by its own dose-constraints. Considering the anti-tumor immune response homeostatic fluctuation, which is cyclically suppressed and incited (“switched off and on”), we synchronized SBRT-PATHY with its most excitable phase, in order to overcome tumor tolerance locally and systemically. The aim of this study, therefore, was to report on the initial results of our latest innovation aimed to further improve BE/AE response rate by testing the effectiveness of the time-synchronized immune-guided SBRT-PATHY. Materials and methods In order to serially map the homeostatic anti-tumor immune response-fluctuations, High Sensitive C-Reactive Protein (HS-CRP), Lactate Dehydrogenase (LDH) and Lymphocyte/Monocyte Ratio (LMR) were analyzed using high-order polynomial trend analysis as surrogate of immune system response. After the biomarker data analysis detected the immune fluctuations and related idiosyncratic immune cycle periodicity, we determined the “most favourable” and “least favourable” treatment time-positions in the immune cycle. In order to evaluate the impact of an idiosyncratic immune cycle on treatment outcomes, our first consecutive four patients were treated on the “most favourable” while the remaining four on the “least favourable” day. Results The median follow-up was 11.8 months. | 3,069,092 | 208625875 | 0 | 16 |
The biomarker data analysis showed periodic immune response fluctuations of regular frequency. The “right” synchronization of SBRT-PATHY with the “most favorable day” of anti-tumor immune response was accompanied with improved clinical outcomes in terms of BE/AE-response rate. Conclusion We believe the right synchronization of radiotherapy with the homeostatically oscillating immune response may improve the probability of inducing BE/AE. Present study has been retrospectively registered on 18th of October 2019 by the ethic committee for Austrian region „Kärnten “in Klagenfurt (AUT), under study number A 37/19. Introduction Clinical exploitation of the bystander (BE) and abscopal effects (AE) was an objective of our long-standing translational oncology research aimed to overcome the outcome-limiting factors related to the unresectable bulky tumors. Despite the developments in oncological therapy BE/AE remain still rare phenomena [1]. In order to improve the therapeutic-ratio by exploiting BE/AE an unconventional partial tumor irradiation targeting the hypoxic segment was developed in 2016 in our institute [2]. Our preclinical findings indicated that the hypoxic in respect to normoxic tumor cells, if selectively irradiated as inductor of BE/AE, show higher potential for the generation of BE/AE [3]. The subsequent translation of these findings to a clinic led to the introduction of a novel SBRT-based PArtial Tumor irradiation targeting HYpoxic clonogenic cells (SBRT-PATHY) showing promising BE/AE-response rates [2,4]. Recently, the Italian group confirmed efficacy of SBRT-PATHY in their initial experience [5]. Considering the immune-mediated nature of BE/AE and in order to further improve BE/AE response rate, SBRT-PATHY was upgraded in 2018 by the sparing of peritumoral immune microenvironment as a | 3,069,093 | 208625875 | 0 | 16 |
new OAR, defined by its own dose-constraints [4,6]. Our concept implied that for successful therapeutic immune modulation, the entire tumor volume may not need to be irradiated but only a part of the tumor. This should initiate antigen shedding, increase effector T cell activation and lead to favorable alterations in radiation-spared peritumoral immune environment [7]. Currently, some studies have described an association between the radiation-induced lymphopenia with poor oncologic outcome, indicating that radiotherapy using large volumes and multiple daily fractions can lead to immunosuppression [8,9]. On the other side, some studies have shown potential therapeutic benefits by eventual ablation of regulatory ("suppressor") T cells with limited (single-dose) systemic therapies [10][11][12] given "at the right time" in order to selectively ablate those suppressor T cells while sparing the effector T cells. Thus, suggesting that the accurate timing of limited therapy may play a major role in treatment efficacy. Following the recent reports of the anti-tumor immune response oscillating over several days [13][14][15][16], we hypothesized the following: by monitoring before the treatment immune-specific biomarkers as the surrogates of homeostatically fluctuating immune response, which is cyclically suppressed and incited ("switched off and on"), it would be possible to determine a periodicity of immune response and, based on that, to synchronize SBRT-PATHY with its most excitable phase, in order to overcome tumor tolerance locally and systemically. The objective of this study, therefore, was to report on the initial results of our latest innovation aimed to further improve BE/AE response rate by testing the effectiveness of the time-synchronized immune-guided SBRT-PATHY. Materials | 3,069,094 | 208625875 | 0 | 16 |
and methods Timing of SBRT-PATHY with respect to an underlying fluctuating anti-tumor immune response Two weeks prior to initiation of SBRT-PATHY, serial (7x) bloods were taken from each patient every second day and assayed for serum biomarkers such as High Sensitive C-Reactive Protein (HS-CRP), Lactate Dehydrogenase (LDH) and also Lymphocyte/Monocyte Ratio (LMR). Data from the assays were analysed to define cyclical fluctuations using high-order polynomial trend analysis. In order to determine each patient's idiosyncratic immune cycle periodicity, the trend/periodicity analysis was performed to generate a standard sine wave of similar periodicity (over several cycles) and visually overlayed and aligned (peak & trough) on the generated polynomial trend graph. This alignment was done in Fig. 2 Immune oscillation monitoring and treatment profile: This figure describes the graphical representation of serial monitoring data over 2 weeks of two example patients in left and right columns. Each graph panel compares the polynomial cyclical relationship and periodicity of biomarkers Hs-CRP/ LDH, Hs-CRP/LMR and LMR/LDH against a standard sine-wave (thick white line). This white sinusoidal line projects forward in time in attempt to identify the putative "most favourable" or "least favourable" dates to treat (marked as red boxes) order to project forward in time to the designated/putative treatment date(s) (Figs. 1 and 2). After the biomarker data analysis detected the immune fluctuations and related idiosyncratic immune cycle periodicity, we determined the "most favourable" and "least favourable" treatment time-positions in the immune cycle. As the biomarkers we used are known to rise and fall periodically over several days with the initiation and | 3,069,095 | 208625875 | 0 | 16 |
then homeostatic termination of the immune response [13,17], we defined a pre-trough time-position as the "most favourable" treatment phase in the immune cycle, while a pre-peak region as the "least favourable" day for SBRT-PATHY. In order to evaluate the impact of an idiosyncratic immune cycle on treatment outcomes, our first consecutive four patients were treated on the "most favourable" while the remaining four on the "least favourable" day. SBRT-PATHY target definition and radiotherapy technique have been previously described in detail [7]. We used a combination of CT and 18F-FDG-PET to define the hypovascularized (contrast-hypo-enhanced) and hypometabolic (SUV max < 3) tumor region representing the "hypoxic" segment between the centralnecrotic and the remaining peripheral-vascularized tumor segments, which was then irradiated with 10Gy × 3 to the 70%-isodose line (Fig. 3). Patients Eight patients with symptomatic, unresectable bulky tumors were prospectively treated between November 2017 and July 2019 with time-synchronized immuneguided SBRT-PATHY. All patients presented either distant metastases or regional metastatic lymph nodes that were not irradiated but followed for AE-induction. The treated patients' main characteristics are summarized in Table 1. Two of eight patients previously received chemotherapy and immunotherapy, and both developed disease progression prior to SBRT-PATHY. No patient received any systemic treatment concomitant with SBRT-PATHY or before the second follow up after it. Two patients which previously received systemic treatment, continued with their treatment after 2 months post-SBRT-PATHY. The response evaluation was performed following the RECIST criteria at 1 month after the treatment by using CT and/or PET-CT, followed by repeated scans at month 2 and | 3,069,096 | 208625875 | 0 | 16 |
then every 3 months. Toxicity was evaluated using the CTCAE Criteria v5.0. All procedures performed in the present study were in accordance with the ethical standards. All the patients signed the informed consent. Present study has been registered by the local ethic committee under study number A 37/19. Idiosyncratic immune cycle periodicity The biomarker data analysis showed immune response fluctuations (Fig. 2) which were synchronized, following similar, regular frequency. The mean immune cycle Clinical outcomes The median follow-up was 11.8 months (range: 4-22). At the time of analysis, one patient (treated at "most favourable" time-position) had died 22 months after SBRT-PATHY because of causes other than cancer. A significant BE (defined as a 30% or greater regression of partially treated bulky tumor) was observed in all four patients treated at "most favourable" time-position (including those two previously being in progression under systemic treatment) with an average tumor shrinkage of 100% (three complete responses (CR), one 80% tumor regression), while among those treated at "least favourable" time-position in two patients with an average tumor shrinkage of 35% (one CR, one 50% tumor regression, two stable bulky tumors reduced for 25%). Significant AE was observed in three patients treated at "most favourable" time-position (in lung and lymph node metastases, and in primary breast cancer) (Fig. 4), while in one among those treated at "least favourable" time-position (in lymph node metastases). Two patients with CR (one from each treatment group) were submitted two months after SBRT-PATHY to surgery, which confirmed pathologic CR at the level of the partially treated | 3,069,097 | 208625875 | 0 | 16 |
bulky tumor but also regional unirradiated lymph node metastases in both patients. Six out of eight patients were free from progression among which all four which were treated at "most favourable" day. In all these four patients was also achieved the symptom relief while among those treated at "least favourable" day it was observed in three patients. Four patients (two from each treatment group) experienced fatigue grade 1. No patient reported any late toxicity. Discussion In addition to the partial tumor irradiation, sparing the loco-regional immune tumor microenvironment, the timing of radiotherapy in relation to the different phases of immune response could be the critical "missing link". Recent evidences suggested that timing of therapy may influence clinical outcomes via immune modulation of the underlying immune response-suppression rather than direct tumor effects [18,19]. In order to serially map the homeostatic immune response-fluctuations, we used Hs-CRP, LDH and LMR as surrogate of immune system interactions [15,18,19]. Since CRP synchronously rises and falls with initiation and termination of the immune response, we determined the start of the cycle as the "most favourable" day of the immune cycle "to release tumor antigen" by SBRT-PATHY, while the first day(s) of CRP fall as the "least favourable" (Fig. 2). This observations were accompanied with clinical outcomes suggesting significant BE/AE with the "most favorable day" approach. The effectiveness in terms of BE/AE response rate of our novel concept could be explained by preserving the pre-existing/endogenous immune signaling in the nonirradiated tumor segment to be modulated by a sufficient threshold of cellular debris/antigen flow | 3,069,098 | 208625875 | 0 | 16 |
by SBRT-PATHY-induced cell damage. This could be seen as analogous to "radio-vaccination" event of manipulating the immunologic homeostatic balance of responsiveness and tolerance via endogenous inflammatory signals. In particular, in order to successfully modulate/disturb the homeostatic tumor immune-suppression, the immune system needs to be preserved as a real OAR. Conclusions To our knowledge, this is the first evidence of a prospectively collected time-synchronized immune-guided , as well as one additional lymph node metastasis with a maximum diameter 2 × 2 cm that was not irradiated (red arrow, lower image left). A dramatic regression of the partially treated bulky lesion and also of unirradiated smaller lymph node metastasis was observed 4 weeks later (green arrows, upper and lower images right) radiotherapy in treating unresectable bulky tumor patients. We believe the right synchronization of radiotherapy with the homeostatically oscillating immune response may improve the probability of inducing BE/AE. Larger prospective trial on time-synchronized immuneguided SBRT-PATHY is ongoing at our institute. | 3,069,099 | 208625875 | 0 | 16 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.