FAMM: a lightweight multispectral deep learning framework for operational remote sensing

Abstract

Monitoring rapid land-use change requires detection pipelines capable of operating at the pace of newly acquired imagery, rather than static classifiers evaluated solely on static accuracy metrics. In West Africa, artisanal and small-scale mining (ASM) represents one of the region’s fastest-growing extractive activities, presenting key challenges for timely regulatory oversight. We present the Fast-Track ASM Monitor (FAMM), an end-to-end machine learning pipeline for high-cadence ASM detection deployed in south-western Ghana with a design framework scalable across broader mining frontiers. The system applies a dual cloud-probability and Normalised Difference Vegetation Index (NDVI) masking rule with a two-stage focal gap-fill in Google Earth Engine to construct cloud-screened Sentinel-2 composites, subsequently classifying 256×256 pixel patches using a MobileNetV3-Large backbone adapted for ten-band input via two-stage fine-tuning and inverse-frequency class weighting. On a held-out 2023 test partition, the model achieves 91.9% accuracy and an ASM F1-score of 0.93, outperforming a Random Forest baseline trained on identical per-band spectral statistics (87.9% accuracy). This baseline comparison confirms that the convolutional model’s primary operational advantage lies in maintaining a lower false-positive rate, which is critical for preventing resource waste in enforcement dispatch. Applied to independent 2025 imagery across the same locations, the model maintains high ASM recall (98.0%) alongside an accuracy of 88.0%, highlighting performance dynamics under two years of natural temporal landscape evolution. Gradient-based sensitivity analysis demonstrates that the model leverages broad multi-spectral information while drawing primary discriminative signal from visible wavebands (accounting for 69% of combined sensitivity). Within gazetted forest reserves, the system demonstrates a false-positive rate (15.1%) closely aligned with the model’s overall 2023 test rate (14.5%), alongside a reduced ASM recall of 65.0%, establishing initial performance bounds for dense-canopy monitoring. With 4, 205, 602 parameters, the model completes single-patch inference in 55.04 ms on a standard CPU thread, demonstrating high computational efficiency for scalable, high-cadence spatial monitoring.