Detecting Mining Activities in Odisha with Sentinel-1 and Sentinel-2 via Self-Supervised Fine-Tuning

Remote sensing offers a powerful solution for monitoring mining activities in Odisha, India (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$81.5^{\circ}-87.5^{\circ} \mathrm{E}$</tex>, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$17.5^{\circ}-22.5^{\circ} \mathrm{N}$</tex>), where traditional manual surveys and India's Mining Surveillance System (MSS) are hindered by high costs and limited scalability. This study introduces a multimodal approach that integrates Synthetic Aperture Radar (SAR) data from Sentinel-1 (VV, VH polarization bands) and optical imagery from Sentinel-2 (B2, B3, B4, B8, B11, B12 bands), processed through Google Earth Engine to generate multispectral image patches. These patches are enriched with derived indices such as Normalized Difference Vegetation Index (NDVI), Bare Soil Index (BSI), Modified Normalized Difference Water Index (MNDWI), and VH/VV ratio to enhance feature discrimination. An enhanced autoencoder architecture, incorporating a Convolutional Block Attention Module (CBAM) for improved focus on relevant features, is pre-trained on a large set of unlabeled data, followed by fine-tuning a classifier on labeled samples to distinguish mining from nonmining areas. The model achieves a test accuracy of 95.38 %, with a weighted precision of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 4 2 \%}$</tex>, recall of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5. 3 8 \%}$</tex>, and F1 score of 95.40 %, as validated on a held-out test set. This methodology outperforms conventional manual techniques and MSS in cost-effectiveness and scalability, providing a robust framework for large-scale, automated mining monitoring across diverse geographies.