mpg-ranch/drone-lsr

Light Stable Representations Dataset

Dataset Description

This dataset contains aerial orthomosaic tiles captured at three different times of day (10:00, 12:00, and 15:00). The dataset is organized into three configurations: default (raw images + canopy height), dinov2_base (DINOv2 embeddings), and dinov3_sat (DINOv3 embeddings). All configurations share consistent train/test splits with matching tile identifiers for cross-referencing. The dataset is designed for training vision encoders that maintain consistent feature representations despite changes in illumination, with applications in remote sensing and environmental monitoring.

Dataset Configurations

The dataset is organized into three configurations, each serving different research needs:

Configuration: default

Raw imagery and environmental data for direct analysis:

Feature	Type	Shape	Description
idx	string	-	Tile identifier in format {ROW}_{COL} for geographic referencing
image_t0	Image	1024×1024×3	Morning capture at 10:00 AM (time=1000)
image_t1	Image	1024×1024×3	Noon capture at 12:00 PM (time=1200)
image_t2	Image	1024×1024×3	Afternoon capture at 3:00 PM (time=1500)
canopy_height	int32	[1024, 1024]	Canopy height grid in centimeters from canopy height model

Configuration: dinov2_base

Pre-computed DINOv2 Base (ViT-B/14) embeddings:

Feature	Type	Shape	Description
idx	string	-	Tile identifier matching other configurations
cls_t0	float32	[768]	DINOv2 CLS token (global features) for morning image
cls_t1	float32	[768]	DINOv2 CLS token (global features) for noon image
cls_t2	float32	[768]	DINOv2 CLS token (global features) for afternoon image
patch_t0	float32	[256, 768]	DINOv2 patch tokens (16×16 spatial grid) for morning image
patch_t1	float32	[256, 768]	DINOv2 patch tokens (16×16 spatial grid) for noon image
patch_t2	float32	[256, 768]	DINOv2 patch tokens (16×16 spatial grid) for afternoon image

Configuration: dinov3_sat

Pre-computed DINOv3 Large (ViT-L/16) embeddings with satellite pretraining:

Feature	Type	Shape	Description
idx	string	-	Tile identifier matching other configurations
cls_t0	float32	[1024]	DINOv3 CLS token (global features) for morning image
cls_t1	float32	[1024]	DINOv3 CLS token (global features) for noon image
cls_t2	float32	[1024]	DINOv3 CLS token (global features) for afternoon image
patch_t0	float32	[196, 1024]	DINOv3 patch tokens (14×14 spatial grid) for morning image
patch_t1	float32	[196, 1024]	DINOv3 patch tokens (14×14 spatial grid) for noon image
patch_t2	float32	[196, 1024]	DINOv3 patch tokens (14×14 spatial grid) for afternoon image

Notes:

• Canopy height values represent centimeters above ground; missing data is encoded as -2147483648
• All configurations use consistent 80%/20% train/test splits with matching idx values
• Patch tokens represent spatial features in different grid resolutions: 16×16 (DINOv2) vs 14×14 (DINOv3)

Usage Example

from datasets import load_dataset

# Load specific configurations
dataset_default = load_dataset("mpg-ranch/drone-lsr", "default")
dataset_dinov2 = load_dataset("mpg-ranch/drone-lsr", "dinov2_base")
dataset_dinov3 = load_dataset("mpg-ranch/drone-lsr", "dinov3_sat")

# Access raw imagery and canopy height
sample_default = dataset_default['train'][0]
morning_image = sample_default['image_t0']      # RGB image
noon_image = sample_default['image_t1']         # RGB image
afternoon_image = sample_default['image_t2']    # RGB image
canopy_height = sample_default['canopy_height'] # Height grid in cm
tile_id = sample_default['idx']                 # Geographic identifier

# Access DINOv2 embeddings (same tile via matching idx)
sample_dinov2 = dataset_dinov2['train'][0]
dinov2_cls_morning = sample_dinov2['cls_t0']     # Global features (768-dim)
dinov2_patches_morning = sample_dinov2['patch_t0'] # Spatial features (256×768)

# Access DINOv3 embeddings (same tile via matching idx)
sample_dinov3 = dataset_dinov3['train'][0]
dinov3_cls_morning = sample_dinov3['cls_t0']     # Global features (1024-dim)
dinov3_patches_morning = sample_dinov3['patch_t0'] # Spatial features (196×1024)

# Verify consistent tile identifiers across configurations
assert sample_default['idx'] == sample_dinov2['idx'] == sample_dinov3['idx']

# Access test sets for evaluation
test_default = dataset_default['test'][0]
test_dinov2 = dataset_dinov2['test'][0]
test_dinov3 = dataset_dinov3['test'][0]

Pre-computed Embeddings

The dataset includes pre-computed embeddings from two state-of-the-art vision transformers:

DINOv2 Base (facebook/dinov2-base)

• Architecture: Vision Transformer Base with 14×14 patch size
• CLS Tokens: 768-dimensional global feature vectors capturing scene-level representations
• Patch Tokens: 256×768 arrays (16×16 spatial grid) encoding local features
• Training: Self-supervised learning on natural images

DINOv3 Large (facebook/dinov3-vitl16-pretrain-sat493m)

• Architecture: Vision Transformer Large with 16×16 patch size
• CLS Tokens: 1024-dimensional global feature vectors capturing scene-level representations
• Patch Tokens: 196×1024 arrays (14×14 spatial grid) encoding local features
• Training: Self-supervised learning with satellite imagery pretraining

Purpose: Enable efficient training and analysis without requiring on-the-fly feature extraction, while providing comparison between natural image and satellite-pretrained models.

Dataset Information

• Location: Lower Partridge Alley, MPG Ranch, Montana, USA
• Survey Date: November 7, 2024
• Coverage: 620 complete tile sets (80% train / 20% test split via seeded random sampling)
• Resolution: 1024×1024 pixels at 1.2cm ground resolution
• Total Size: ~6.4GB of image data plus embeddings
• Quality Control: Tiles with transient objects, such as vehicles, were excluded from the dataset. RGB imagery and canopy rasters are removed together to keep modalities aligned.

Use Cases

This dataset is intended for:

• Developing vision encoders robust to lighting variations
• Representation stability research in computer vision
• Time-invariant feature learning
• Remote sensing applications requiring lighting robustness
• Comparative analysis of illumination effects on vision model features

Citation

If you use this dataset in your research, please cite:

@dataset{mpg_ranch_light_stable_semantics_2024,
  title={Light Stable Representations Dataset},
  author={Kyle Doherty and Erik Samose and Max Gurinas and Brandon Trabucco and Ruslan Salakhutdinov},
  year={2024},
  month={November},
  url={https://huggingface.co/datasets/mpg-ranch/drone-lsr},
  publisher={Hugging Face},
  note={Aerial orthomosaic tiles with DINOv2 and DINOv3 embeddings for light-stable representation vision encoder training},
  location={MPG Ranch, Montana, USA},
  survey_date={2024-11-07},
  organization={MPG Ranch}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Attribution Requirements:

• You must give appropriate credit to MPG Ranch
• Provide a link to the license
• Indicate if changes were made to the dataset