A Novel Integrated ML Approach Utilizing Radar & Satellite Imagery for Selective Logging Detection (Papers Track)

Saraswathy Amjith (MIT); Joshua Fan (Cornell University)

Paper PDF Slides PDF Poster File Cite
Forests

Abstract

Illegal selective logging is devastating to the world’s biodiversity and forests, and many countries face challenges in monitoring these forests. In the Amazon, 94\% of deforestation is attributed to illegal activities, primarily conducted through selective logging—a practice that has received far less research attention than clear-cut deforestation. This study introduces a novel machine learning approach that enhances the detection of selective logging by integrating Sentinel-1 Synthetic Aperture Radar (SAR) data with Sentinel-2 optical imagery, moving beyond conventional methods that rely on a single data source. Requisite data was collected through Google Earth Engine for Sentinel-1 and Sentinel-2 and used to create three new datasets of 134,606 samples (5x other studies). These datasets were then used to train and evaluate three different classification architectures: Convolutional Neural Networks (CNN), Random Forest, and Gradient Boosted Trees. The CNN model trained on both modalities outperformed CNNs trained on individual Sentinel datasets, achieving an accuracy of 95.08\%, compared to 57.51% for Sentinel-1 and 91.95% for Sentinel-2. The Gradient Boosted Trees model, while slightly less accurate with a 94.39% accuracy and a 94.03 F1 score, required significantly less computational power. Our findings demonstrate that the integration of optical and radar data substantially improves the prediction of selective logging activities, achieving up to a 7.08% improvement over previous selective logging detection research and offering a scalable solution for forestry monitoring in resource-limited settings.