Multispectral U-Net for Open-Pit Mine Detection in Tropical Indonesia Using Sentinel-2 Satellite Imagery

The open-pit mining activities in Indonesia contribute significantly to the economic growth of the country. Nonetheless, the environmental damage caused by these mining operations is severe. It is essential to monitor these sites on a large scale. In the tropics, dense vegetation often misleads conventional optical detection techniques used for remote sensing. This study implements a physics-informed Multispectral U-Net with Sentinel-2 for automating the task of mine boundaries location. We conduct a systematic assessment of eight spectral configurations ranging from baseline RGB to a 10-band stack containing NIR, SWIR and Red-Edge. Experiments were repeated across multiple random seeds and compared to state-of-the-art architectures (Attention U-Net, DeepLabV3+, SegFormer). The 10-band configuration reached the highest performance (Mean IoU:86.78%, F1-Score:90.11%) by combining SWIR for identifying ores and NIR/Red-Edge for strict masking of vegetation, beating all baselines. External validation affirmed strong cross-continental transferability on a Colombian coal mine (mIoU: 82.75%) while coastal nickel deposits revealed domain shift constraints requiring transfer learning to address performance degradation in >20% of the cases. This multispectral framework, with a ~410 ms inference time, presents a scalable, statistically validated methodology for operational regulatory monitoring in complex tropical landscapes.