Mapping Selective Logging in the Amazon with Artificial Intelligence and Sentinel-2

Abstract

The Amazon forest, the largest tropical forest in the world and marked by its rapid change in forest cover, has suffered from intense anthropogenic phenomena such as deforestation and forest degradation, this one caused mainly by fires and selective logging. This study explores a U-NET model to accurately identify selective logging infrastructure (roads, skid trails, storage yards) using Sentinel-2 imagery. Our goal is to improve the SIMEX (System for Monitoring Timber Harvesting) in the Brazilian Amazon, reducing the human workload and increasing the system's accuracy. Data from 780 SIMEX registration polygons (2021–2022) were used, with stratified sampling creating a training data set. The U-NET model, optimized with specific hyperparameters and data augmentation, analyzed six spectral bands (two-year RGB). We achieved an F1 score of ~81% with high precision (73.7%) and recall (90.31%) on the test set, indicating strong performance and generalization. Our model excels at accurately predicting logging infrastructure and potential damage to forest canopies. It provides detailed detection of roads and stockyards, offering a comprehensive view compared to models that generalize explored areas. This refined approach increases its usefulness for forest conservation and management efforts.