FOREST SEMANTIC SEGMENTATION BASED ON DEEP LEARNING USING SENTINEL-2 IMAGES

Abstract

Forests are invaluable for maintaining biodiversity, watersheds, rainfall levels, bioclimatic stability, carbon sequestration and climate change mitigation, and the sustainability of large-scale climate regimes. In other words, forests provide a wide range of ecosystem services and livelihoods for the people and play a critical role in influencing global atmospheric cycles. Providing sustainable, reliable, and accurate information on forest cover change is essential for an holistic forest management, efficient use of resources, neutralizing the effects of global warming and better monitoring of deforestation activities. Within the scope of this study, it is aimed to perform semantic segmentation of 5 different tree species (larch, red pine, yellow pine, oak, spruce) from Sentinel-2 satellite images. For this purpose, the regions where these tree species are densely populated in Turkey (Marmara, Aegean, Eastern Black Sea) were selected as pilot regions. A unique data set was created using the data of the selected pilot regions. As a result of the study, it was possible to determine the forest types temporally for the selected classes with more than 90% Intersection over Union score for all classes. The developed deep learning model with the created forest data set can be implemented to the other forests areas with same species in other parts of the world.