Semantic Deep Learning for Open-Pit Mining Detection Using High Resolution SAR Data

Indonesia is rich in mineral resources primarily extracted through open-pit mining. However, these mining activities significantly impact the Earth's surface, necessitating effective monitoring to mitigate environmental consequences through advanced digital technologies. Remote sensing technology has proven effective for monitoring open-pit mining operations. This study utilises high-resolution Synthetic Aperture Radar (SAR) data from the TerraSAR-X, integrated with deep learning, to identify and monitor open-pit mining activities in the South Bangka Regency. This region, known for its extensive tin mining operations, faces substantial ecological challenges due to the disruptive effects of open-pit mining on ecosystems and landscapes. Employing the U-Net model with a ResNet-34 backbone, we processed TerraSAR-X imagery to classify land cover types into open-pit mining areas, vegetation, water bodies, and settlements. The model achieved a classification accuracy of 71%, successfully identifying 2,346.9 hectares as mining areas. However, some classification challenges were encountered, particularly in areas with complex terrain and low backscatter shadow effects. Misclassifications, particularly between mining areas and water bodies due to similar backscatter values classification results showing “box” errors, highlight the necessity for further model refinement. Future research should focus on integrating multispectral data and additional deep learning features to improve classification accuracy. The findings underscore the potential of SAR and artificial intelligence (AI) to enhance sustainable mining practices and support environmental management strategies, reinforcing the need for ongoing advancements in remote sensing technologies for more effective monitoring and regulation of mining activities.