README.md

---
license: cc-by-nc-4.0
---


# 🌲 AGBD: A Global-scale Biomass Dataset 🌳

Authors: Ghjulia Sialelli ([[email protected]](mailto:[email protected])), Torben Peters, Jan Wegner, Konrad Schindler


## 🆕 Updates
* The dataset was last modified on **Feb. 26th, 2025**
* See the [changelog](changelog.md) for more information about what was updated!

## 🚀 Quickstart

To get started quickly with this dataset, use the following code snippet:

```python
# Install the datasets library if you haven't already
!pip install datasets

# Import necessary modules
from datasets import load_dataset

# Load the dataset
dataset = load_dataset('prs-eth/AGBD', trust_remote_code=True, streaming=True)["train"]  # Options: "train", "validation", "test"

# Iterate over the dataset
for sample in dataset:
  features, label = sample['input'], sample['label']

```

This code will load the dataset as an `IterableDataset`. You can find more information on how to work with `IterableDataset` objects in the [Hugging Face documentation](https://huggingface.co/docs/datasets/access#iterabledataset).

---


## 📊 Dataset Overview

Each sample in the dataset contains a **pair of pre-cropped images** along with their corresponding **biomass labels**. For additional resources, including links to the preprocessed uncropped data, please visit the [project page on GitHub](https://github.com/ghjuliasialelli/AGBD/).

### ⚙️ Load Dataset Options

The `load_dataset()` function provides the following configuration options:

- **`norm_strat`** (str) : `{'pct', 'mean_std', 'none'}` (default = `'pct'`)  
  The strategy to apply to process the input features. Valid options are: `'pct'`, which applies min-max scaling with the 1st and 99th percentiles of the data; `'mean_std'` which applies Z-score normalization; and `'none'`, which returns the un-processed data.

- **`encode_strat`** (str) : `{'sin_cos', 'onehot', 'cat2vec', 'none'}` (default = `'sin_cos'`)
  The encoding strategy to apply to the land classification (LC) data. Valid options are: `'onehot'`, one-hot encoding; `'sin_cos'`, sine-cosine encoding; `'cat2vec'`, cat2vec transformation based on embeddings pre-computed on the train set.

- **`input_features`** (dict)
  The features to be included in the data, the default values being:
    ```
    {'S2_bands': ['B01', 'B02', 'B03', 'B04', 'B05', 'B06', 'B07', 'B08', 'B8A', 'B09','B11', 'B12'], 
     'S2_dates' : False, 'lat_lon': True, 'GEDI_dates': False, 'ALOS': True, 'CH': True, 'LC': True, 
     'DEM': True, 'topo': False}
    ```

- **`additional_features`** (list) (default = `[]`)  
  A list of additional features the dataset should include. *Refer to the [documentation below](#add-feat-anchor) for more details.* Possible values are:
  ```
  ['s2_num_days', 'gedi_num_days', 'lat', 'lon', 'agbd_se', 'elev_lowes', 'leaf_off_f', 'pft_class', 'region_cla', 'rh98', 'sensitivity', 'solar_elev', 'urban_prop']
  ```
  This metadata can later be accessed as such:
  ```
  from datasets import load_dataset
  
  dataset = load_dataset('AGBD.py',trust_remote_code=True,streaming=True)
  for sample in dataset['train']:
    lat = sample['lat']
    break
  ```

- **`patch_size`** (int) (default =`15`)  
  The size of the returned patch (in pixels). The maximum value is **25 pixels**, which corresponds to **250 meters**.

---
### 🖼️ Features Details

Each sample consists of a varying number of channels, based on the `input_features` and `encode_strat` options passed to the `load_dataset()` function. The channels are organized as follows:

| Feature | Channels | Included by default?| Description |
| --- | --- | --- | --- |
| **Sentinel-2 bands** | `B01, B02, B03, B04, B05, B06, B07, B08, B8A, B09, B11, B12` | Y | Sentinel-2 bands, in Surface Reflectance values |
| **Sentinel-2 dates** | `s2_num_days, s2_doy_cos, s2_doy_sin` | N | Date of acquisition of the S2 image (in number of days wrt the beginning of the GEDI mission); sine-cosine encoding of the day of year (DOY).|
| **Geographical coordinates** | `lat_cos, lat_sin, lon_cos, lon_sin` | Y | Sine-cosine encoding of the latitude and longitude.|
| **GEDI dates** | `gedi_num_days, gedi_doy_cos, gedi_doy_sin` | N | Date of acquisition of the GEDI footprint (in number of days wrt the beginning of the GEDI mission); sine-cosine encoding of the day of year (DOY).|
| **ALOS PALSAR-2 bands** | `HH,HV` | Y | ALOS PALSAR-2 bands, gamma-naught values in dB.|
| **Canopy Height** | `ch, ch_std`| Y | Canopy height from Lang et al. and associated standard deviation. |
| **Land Cover Information** | `lc_encoding*, lc_prob`| Y | Encoding of the land class, and classification probability (as a percentage between 0 and 1).|
| **Topography** | `slope, aspect_cos, aspect_sin` | N | Slope (percentage between 0 and 1); sine-cosine encoded aspect of the slope.|
| **Digital Elevation Model (DEM)** | `dem` | Y | Elevation (in meters).|

This corresponds to the following value for `input_features` : 
    ```
    {'S2_bands': ['B01', 'B02', 'B03', 'B04', 'B05', 'B06', 'B07', 'B08', 'B8A', 'B09','B11', 'B12'], 
     'S2_dates' : False, 'lat_lon': True, 'GEDI_dates': False, 'ALOS': True, 'CH': True, 'LC': True, 
     'DEM': True, 'topo': False}
    ```

Regarding `lc_encoding*`, the number of channels follows this convention:
- `sin_cos` (default) : 2 channels
- `cat2vec` : 5 channels
- `onehot` : 14 channels
- `none` : 1 channel

Should you get stuck, you can debug the number of channels using the `compute_num_features()` function in [AGBD.py](AGBD.py).

In summary, the channels are structured as follows:

```plaintext
(Sentinel-2 bands) | (Sentinel-2 dates) | (Geographical coordinates) | (GEDI dates) | (ALOS PALSAR-2 bands) | (Canopy Height) | (Land Cover Information) | (Topography) | DEM
```

---
### ➕ Additional Features
<a name="add-feat-anchor"></a>

You can include a list of additional features from the options below in your dataset configuration:
- **`"agbd_se"` - AGBD Standard Error**: The uncertainty estimate associated with the aboveground biomass density prediction for each GEDI footprint.
- **`"elev_lowes"` - Elevation**: The height above sea level at the location of the GEDI footprint.
- **`"leaf_off_f"` - Leaf-Off Flag**: Indicates whether the measurement was taken during the leaf-off season, which can impact canopy structure data.
- **`"pft_class"` - Plant Functional Type (PFT) Class**: Categorization of the vegetation type (e.g., deciduous broadleaf, evergreen needleleaf).
- **`"region_cla"` - Region Class**: The geographical area where the footprint is located (e.g., North America, South Asia).
- **`"rh98"` - RH98 (Relative Height at 98%)**: The height at which 98% of the returned laser energy is reflected, a key measure of canopy height.
- **`"sensitivity"` - Sensitivity**: The proportion of laser pulse energy reflected back to the sensor, providing insight into vegetation density and structure.
- **`"solar_elev"` - Solar Elevation**: The angle of the sun above the horizon at the time of measurement, which can affect data quality.
- **`"urban_prop"` - Urban Proportion**: The percentage of the footprint area that is urbanized, helping to filter or adjust biomass estimates in mixed landscapes.
- **`"gedi_num_days"` - Date of GEDI Footprints**: The specific date on which each GEDI footprint was captured, adding temporal context to the measurements.
- **`"s2_num_days"` - Date of Sentinel-2 Image**: The specific date on which each Sentinel-2 image was captured, ensuring temporal alignment with GEDI data.
- **`"lat"` - Latitude**: Latitude of the central pixel.
- **`"lon"` - Longitude**: Longitude of the central pixel.