GhanaMining-Sat: A Multi-Site Benchmark Dataset for Artisanal Mining Detection in Ghana

A labeled satellite imagery dataset for detecting artisanal small-scale mining sites in Ghana using deep learning. The dataset contains 300 manually labeled patches (256×256 pixels) extracted from Sentinel-2 satellite imagery across three major mining regions: Prestea, Tarkwa, and Obuasi. Key Features:- 300 high-quality labeled patches (244 mining, 56 non-mining)- Multi-site coverage (3 geographically diverse locations)- Manual labeling with false-color validation methodology- Stratified 80/20 train/test splits with reproducible seed (seed=42)- Baseline models: ResNet-50 (93.33% accuracy) and EfficientNet-B0 (93.33% accuracy)- Cross-site generalization analysis (75.7% average on unseen locations)- Complete documentation (README, METADATA, LICENSE, CITATION) Note: The detection methodology identifies mining sites based on satellite imagery characteristics. Legal status determination requires cross-referencing with mining permit databases. Study sites are documented galamsey (illegal mining) hotspots. This dataset provides a community resource for developing automated mining detection systems and supports research in land use monitoring and environmental assessment across West Africa.