README.md

---
pretty_name: Geolayers
language: en
language_creators:
  - found
license: cc-by-4.0
multilinguality: monolingual
size_categories:
  - 10K<n<100K
task_categories:
  - image-classification
  - image-segmentation
source_datasets:
  - SustainBench
  - USAVars
  - BigEarthNetv2.0
  - EnviroAtlas
homepage: https://huggingface.co/datasets/arjunrao2000/geolayers
repository: https://huggingface.co/datasets/arjunrao2000/geolayers
download_size: 25570000000
tags:
  - climate
  - remote-sensing

# 1) Index all JPG previews under huggingface_preview/
data_files:
  - "huggingface_preview/**/*.jpg"

# 2) Also index metadata.parquet so DuckDB can read it
  - "metadata.csv"

# 3) Tell HF that metadata.parquet is your preview table,
#    and explicitly name which columns are images.

preview:
  path: metadata.csv
  images:
    - rgb
    - osm
    - dem
    - mask
configs:
  - config_name: benchmark
    data_files:
      - split: sustainbench
        path: metadata.csv
---


# Geolayers-Data
<img src="osm_usavars.png" alt="Sample Geographic Inputs with the USAVars Dataset" width="800"/>

 -->
This dataset card contains usage instructions and metadata for all data-products released with our paper:  
*Using Multiple Input Modalities can Improve Data-Efficiency and O.O.D. Generalization for ML with Satellite Imagery.* We release 3 modified versions of 3 benchmark datasets spanning land-cover segmentation, tree-cover regression, and multi-label land-cover classification tasks. These datasets are augmented with auxiliary, geographic inputs. A full list of contributed data products is shown in the table below.

<table>
  <thead>
    <tr>
      <th>Dataset</th>
      <th>Task Description</th>
      <th>Multispectral Input</th>
      <th>Model</th>
      <th>Additional Data Layers</th>
      <th colspan="2">Dataset Size</th>
      <th>OOD Test Set Present?</th>
    </tr>
    <tr>
      <th></th>
      <th></th>
      <th></th>
      <th></th>
      <th></th>
      <th>Compressed</th>
      <th>Uncompressed</th>
      <th></th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td><a href="https://arxiv.org/abs/2111.04724">SustainBench</a></td>
      <td>Farmland boundary delineation</td>
      <td>Sentinel-2 RGB</td>
      <td>U-Net</td>
      <td>OSM rasters, EU-DEM</td>
      <td>1.76 GB</td>
      <td>1.78 GB</td>
      <td>✗</td>
    </tr>
    <tr>
      <td><a href="https://arxiv.org/abs/2202.14000">EnviroAtlas</a></td>
      <td>Land-cover segmentation</td>
      <td>NAIP RGB + NIR</td>
      <td>FCN</td>
      <td><a href="https://arxiv.org/abs/2202.14000">Prior</a>, OSM rasters</td>
      <td>N/A</td>
      <td>N/A</td>
      <td>✓</td>
    </tr>
    <tr>
      <td><a href="https://bigearth.net/static/documents/Description_BigEarthNet_v2.pdf">BigEarthNet v2.0</a></td>
      <td>Land-cover classification</td>
      <td>Sentinel-2 (10 bands)</td>
      <td>ViT</td>
      <td><a href="https://arxiv.org/abs/2311.17179">SatCLIP</a> embeddings</td>
      <td>120 GB (raw), 91 GB (H5)</td>
      <td>205 GB (raw), 259 GB (H5) </td>
      <td>✓</td>
    </tr>
    <tr>
      <td><a href="https://arxiv.org/abs/2010.08168">USAVars</a></td>
      <td>Tree-cover regression</td>
      <td>NAIP RGB + NIR</td>
      <td>ResNet-50</td>
      <td>OSM rasters</td>
      <td> 23.56 GB </td>
      <td> 167 GB</td>
      <td>✗</td>
    </tr>
  </tbody>
</table>


## Usage Instructions
* Download the `.h5.gz` files in `data/<source dataset name>`. Our source datasets include SustainBench, USAVars, and BigEarthNet2.0. Each dataset with the augmented geographic inputs is detailed in [this section 📦](#geolayersused) 
* You may use pigz (https://linux.die.net/man/1/pigz) to decompress the archive. This is especially recommended for USAVars' train-split, which is 117 GB when uncompressed. This can be done with `pigz -d <.h5.gz>`
* Datasets with auxiliary geographic inputs can be read with H5PY.
* If you plan on using the original datasets, please cite our paper "Using Multiple Input Modalities can Improve Data-Efficiency and O.O.D. Generalization for ML with Satellite Imagery". BibTex can be found at the bottom of this README. 
### Usage Instructions for the BigEarthNetv2.0 dataset (Clasen et al. (2025))
  We detail usage instructions to train on the BigEarthNetv2.0 dataset as a separate section due to the unique fusion mechanism we use for this input modality. Our training code is a distinct Github repository for fusing and training a ViT on the BigEarthNet2.0 dataset using an auxiliary SatCLIP token fused with `TOKEN-FUSE`. [Link to Github Repository](https://github.com/UCBoulder/GeoViTTokenFusion). This repository also uses torchgeo and timm as *submodules*. 

We use the original dataset [BigEarthNetv2.0](https://bigearth.net/) dataset which is processed with spatially-buffered train-test splits. We release two **processed** versions of the datasets introduced in Casen et al. (2025)
The first version is stored in directory `data/bigearthnet/raw/`. This dataset, although called `raw` is a pre-processed version of the raw BigEarthNetv2.0 dataset. We follow instructions listed on [this repository](https://git.tu-berlin.de/rsim/reben-training-scripts/-/tree/main?ref_type=heads#data). Steps performed:
  1. We download the raw `BigEarthNet-S2.tar.zst` Sentinel-2 BigEarthNet dataset.
  2. We extract and process the raw S2 tiles to a LMDB 'Lightning' Database. This allows for faster reads during training. We use the rico-hdl tool [here](https://github.com/kai-tub/rico-hdl) to accomplish this.
  3. We download reference maps and sentinel-2 tile metadata with snow and cloud cover rasters
  4. This final dataset is compressed into several chunks and stored in `data/bigearthnet/raw/bigearth.tar.gz.part-a<x>`. Each chunk is 5G large. There are 24 total chunks.

To uncompress and re-assemble the compressed files in `data/bigearthnet/raw/`, download all the parts and run:
```
cat bigearthnet.tar.gz.part-* \
  | pigz -dc \
  | tar -xpf -
```

Note that if this version of the dataset is used, SatCLIP embeddings would need to be re-computed on-the-fly. To use this dataset with the pre-computed SatCLIP embeddings, refer to the note below.

#### 💡 Do you want to try your own input fusion mechanism with BigEarthNetv2.0? 
The second version of the BigEarthNetv2.0 dataset is stored in `data/bigearthnet/`. These datasets are stored as 3 H5PY datasets (`.h5`) for each split in the dataset. 
This version of the processed dataset comes with (i) raw location co-ordinates, and (ii) pre-computed SatCLIP embeddings (L=10, ResNet50 image encoder backbone). 
You may access these embeddings and location metadata with keys `location` and `satclip_embedding`. 

### Usage Instructions for the SustainBench Farmland Boundary Delineation Dataset (Yeh et al. (2021))
1. Unzip the archive in `data/sustainbench-field-boundary-delineation` with `unzip sustainbench.zip`
2. You should see a directory structure as follows:
```
dataset_release/
├── id_augmented_test_split_with_osm_new.h5.gz.zip
├── id_augmented_train_split_with_osm_new.h5.gz.zip
├── id_augmented_val_split_with_osm_new.h5.gz.zip
├── raw_id_augmented_test_split_with_osm_new.h5.gz.zip
├── raw_id_augmented_train_split_with_osm_new.h5.gz.zip
└── raw_id_augmented_val_split_with_osm_new.h5.gz.zip
```
3. Unzip all files using `unzip` and `pigz -d <path to .h5.gz file>`
There are two versions of data released: Datasets that begin with `id_augmented` refer to the version of the SustainBench farmland boundary delineation dataset with the OSM and DEM rasters pre-processed to RGB space following the application of the Gaussian Blur. Datasets that begin with `raw_id_augmented` contain the RGB imagery with 19 categorical rasters for OSM, and 1 raster for the DEM geographic input.

## 📦 <a name="geolayersused"></a> Datasets & Georeferenced Auxiliary Layers
### SustainBench – Farmland Boundary Delineation
* **Optical input:** Sentinel-2 RGB patches (224×224 px, 10 m GSD) covering French cropland in 2017; ≈ 1.6 k training images.  
* **Auxiliary layers (all geo-aligned):**  
  * 19-channel OpenStreetMap (OSM) raster stack (roads, waterways, buildings, biome classes, …)  
  * EU-DEM (20 m GSD, down-sampled to 10 m)  
* **Why:** OSM + DEM give an 8 % Dice boost when labels are scarce; gains appear once the training set drops below ≈ 700 images.

---

### EnviroAtlas – Land-Cover Segmentation
* **Optical input:** NAIP 4-band RGB-NIR aerial imagery at 1 m resolution.  
* **Auxiliary layers:**  
  * OSM rasters (roads, waterbodies, waterways)  
  * **Prior** raster – a hand-crafted fusion of NLCD land-cover and OSM layers (PROC-STACK)  
* **Splits:** Train = Pittsburgh; OOD validation/test = Austin & Durham. Auxiliary layers raise OOD overall accuracy by ~4 pp without extra fine-tuning.

---

### BigEarthNet v2.0 – Multi-Label Land-Cover Classification
* **Optical input:** 10-band Sentinel-2 tile pairs; ≈ 550 k patch/label pairs over 19 classes.  
* **Auxiliary layer:**  
  * **SatCLIP** location embedding (256-D), one per image center, injected as an extra ViT token (TOKEN-FUSE).  
* **Splits:** Grid-based; val/test tiles lie outside the training footprint (spatial OOD by design). SatCLIP token lifts macro-F1 by ~3 pp across *all* subset sizes.

---

### USAVars – Tree-Cover Regression
* **Optical input:** NAIP RGB-NIR images (1 km² tiles); ≈ 100 k samples with tree-cover % labels.  
* **Auxiliary layers:**  
  * Extended OSM raster stack (roads, buildings, land-use, biome classes, …)  
* **Notes:** Stacking the OSM raster boosts R² by 0.16 in the low-data regime (< 250 images); DEM is provided raw for flexibility.



Citation:

```
@inproceedings{
  rao2025using,
  title={Using Multiple Input Modalities can Improve Data-Efficiency and O.O.D. Generalization for {ML} with Satellite Imagery},
  author={Arjun Rao and Esther Rolf},
  booktitle={TerraBytes - ICML 2025 workshop},
  year={2025},
  url={https://openreview.net/forum?id=p5nSQMPUyo}
}
```
---