Jan. 10, 2025 The first version of SmallMinesDS is available on HuggingFace

The gradual expansion of unregularized artisanal small-scale gold mining (ASGM) fuels environmental degradation and poses a risk to miners and mining communities. To enforce sustainable mining, support reclamation initiatives, and pave the way for understudying the impacts of mining, we present SmallMinesDS, a dataset for mapping artisanal small-scale gold mining from multi-sensor satellite images. The initial version of the dataset covers five districts in Southwestern Ghana in two time periods. The first version of SmallMinesDS covers five administrative districts in Southwestern Ghana with a ground area of about 3200 sq.km. SmallMinesDS includes satellite images from optical and radar sensors in the dry season (January) for two years (2016 and 2022). The area covered is ecologically diverse, hosting rich biodiversity and serving as a hub for multiple land-use activities, including crop cultivation, forestry, industrial and small-scale gold mining. This multifaceted landscape underscores the importance of reliable mapping and monitoring of ASGM activities.

• SmallMinesDS contain 4270 patches; 2175 each for 2016 and 2022

• Each patch has an input shape of 13 x 128 x 128 and a corresponding mask of 1 x 128 x 128

• The open-access data can be downloaded at Hugging Face

• Conda (for managing environments)
• NVIDIA GPU (for efficient computation)

Two separate virtual environments are required:

1. Terratorch Environment (for TerraTorch dependencies)
2. SAM2 Environment (for Segment Anything Model v2 dependencies)

These environments are necessary as they rely on different PyTorch versions.

conda create -n terratorch python=3.11
conda activate terratorch
pip install -r requirements.txt
conda deactivate

conda create -n sam2 python=3.11
conda activate sam2
chmod +x install_sam2.sh
bash install_sam2.sh
mv scripts/ft-sam2.py sam2/
conda deactivate

Fine-tune Prithvi-2 using our dataset.

conda activate terratorch
python scripts/train-prithvi-v2-300.py

conda activate terratorch
python scripts/train-prithvi-v2-600.py

To compare with Prithvi-2, we train ResNet50 from scratch using six spectral bands: Blue, Green, Red, Narrow NIR, SWIR, and SWIR 2.

conda activate terratorch
python scripts/train-resnet50-6bands.py

Segment Anything Model v2 is a foundational model designed for promptable visual segmentation in RGB images and videos. We fine-tune it using the RGB channels of our dataset.

conda activate sam2
python sam2/ft-sam2.py

For a fair comparison in the RGB domain, we fine-tune a ResNet50 model pretrained on ImageNet and compare its performance with SAM2.

conda activate terratorch
python scripts/ft-resnet50.py

If you use the dataset or supporting code in your research, please cite SmallMinesDS as :

@ARTICLE{10982207,
  author={Ofori-Ampofo, Stella and Zappacosta, Antony and Kuzu, Rıdvan Salih and Schauer, Peter and Willberg, Martin and Zhu, Xiao Xiang},
  journal={IEEE Geoscience and Remote Sensing Letters}, 
  title={SmallMinesDS: A Multi-Modal Dataset for Mapping Artisanal and Small-Scale Gold Mines}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Data mining;Foundation models;Gold;Satellite images;Sentinel-1;Optical sensors;Laser radar;Biological system modeling;Spaceborne radar;Radar imaging;Earth Observation;Semantic Segmentation;Mining;Machine Learning;Foundation Models},
  doi={10.1109/LGRS.2025.3566356}}