DETECTING ILLEGAL LOGGING USING DEEP LEARNING ON SENTINEL-1 SAR IMAGERY

Illegal logging in tropical peat swamp forests represents a significant threat to global climate stability, by contributing to around 10-15% of worldwide greenhouse emissions and degrading vital carbon storage ecosystems.This study addresses the necessity of automated detection of illegal logging activities in the Mawas Conservation Area of Central Kalimantan, Indonesia, by developing a deep learning model trained on Synthetic Aperture Radar (SAR) imagery from Sentinel-1. The model structure involves a combination of unet and deeplabv3 architecture with efficientnet-B4 as the encoder backbone, enhanced by Spatial and Channel Squeeze & Excitation (SCSE) attention mechanisms for improved feature extraction. The model was trained on 690 SAR images, captured from March 2015 to December 2016. The deep learning model shows promising results with an F1-Score of 66% and an iou of 49%. The overall accuracy is high at 89.55% and a precision is 67.41%. These results demonstrate the potential of deep learning for monitoring illegal logging in data-sparse tropical forest regions.