Detection of Degraded Forests in Guinea, West Africa, Using Convolutional Neural Networks and Sentinel-2 Time Series

This study explores forest degradation detection in Guinea using remote sensing, focusing on the Ziama Massif, a UNESCO Biosphere Reserve and using Sentinel-2 satellite imagery. We compared three convolutional neural network models (U-Net, ResNet-UNet, SegNet) against the photointerpreted method, with all approaches undergoing independent validation by external Guinean photointerpreters. The U-Net model, trained on 2015-2020 data, showed 94% accuracy in Ziama and over 91% accuracy in Mount Nimba, which wasn't in the training set. Altogether, the results show that the method is transferable and applicable across different years and Guinean forest regions like Nimba. This demonstrates U-Net's robustness in identifying degraded forests, leading to its adoption for forest class updates in the GAEZ project. This research advances tropical forest monitoring methods, offering crucial insights for environmental stakeholders.