Detection of degraded forests in Guinea, West Africa, using convolutional neural networks and Sentinel-2 time series

Forest degradation is the alteration of forest biomass, structure or services without the conversion to another land cover. Unlike deforestation, forest degradation is subtle and less visible, but it often leads to deforestation eventually. In this study we conducted a comprehensive analysis of degraded forest detection in the Guinea forest region using remote sensing techniques. Our aim was to explore the use of Sentinel-2 satellite imagery in detecting and monitoring forest degradation in Guinea, West Africa, where selective logging is the primary degradation process observed. Consequently, degraded forests exhibit fewer large trees than intact forests, resulting in discontinuities in the canopy structure. This study consists in a comparative analysis between the contextual Random Forest (RF) algorithm previously introduced, three convolutional neural network (CNN) models (U-Net, SegNet, ResNet-UNet), and the photo-interpreted (PI) method, with all model results undergoing independent validation by external Guinean photo-interpreters. The CNN and RF models were trained using subsets of the maps obtained by the PI method. The results show that the CNN U-Net model is the most adequate method, with an 94% agreement with the photo-interpreted map in the Ziama massif for the year 2021 unused for the training. All models were also tested over the Mount Nimba area, which was not included in the training dataset. Again, the U-Net model surpassed all other models with an overall agreement above 91%, and an accuracy of 91.5% as established during a second validation exercise carried out by independent photo-interpreters following the widely used Verified Carbon Standard validation methodology. These results underscore the robustness and efficiency of the U-Net model in accurately identifying degraded forests across diverse areas with similar typology of degraded forests. Altogether, the results show that the method is transferable and applicable across different years and among the different Guinean forest regions, such as the Ziama, Diécké, and Nimba massifs. Based on the superior performance and robustness demonstrated by the U-Net model, we selected it to replace the previous photo-interpretation-based method for forest class updates in the land cover map produced for the Guinean ministry of agriculture.