Near-Real-Time Detection of Illegal Deforestation and Artisanal Gold Mining in the Peruvian Amazon Using Multi-Source Satellite Imagery and Deep Learning

Illegal deforestation and artisanal gold mining in the Peruvian Amazon represent two of the most acute and accelerating environmental crises of the current decade. The Madre de Dios department alone has lost an estimated 95,750 hectares of primary forest to small-scale mining since 2009, with the rate of loss accelerating in recent years. Traditional monitoring approaches — ground surveys, periodic satellite audits, and annual deforestation statistics — cannot keep pace with an activity that can clear several hectares within 48 hours.

This paper presents the architecture, methods, and early operational results of the Canopy Vision Lab monitoring system: an end-to-end pipeline that ingests multi-source satellite imagery (Sentinel-2 L2A, PlanetScope, Landsat 8/9, and Sentinel-1 SAR), applies a cascade of deep learning models for land-cover segmentation and multi-temporal change detection, enriches candidate alerts with geospatial context, and delivers actionable notifications to partner organizations within a median of 6.4 hours of the satellite acquisition event.

Our primary segmentation model — a SegFormer-B3 architecture finetuned on approximately 42,000 labeled tiles across three Peruvian departments — achieves a mean Intersection over Union (mIoU) of 0.81 across five disturbance classes: intact forest, secondary growth, recent clearing, mine pit, and tailings/pond. A downstream Siamese change-detection model achieves a precision of 0.93. During the 2025 calendar year the system delivered 14,210 alerts to eleven partner organizations. This paper documents the full technical pipeline, dataset construction methodology, model architecture, evaluation framework, and operational deployment experience.