Forest Change Monitoring in Siberian Forest: Deep Learning Approach

Abstract

The Siberian forest plays a vital role in the planet’s ecosystem, accounting for approximately 20% of the world’s forests. In today’s rapidly changing world, it requires careful and continuous monitoring to keep deforestation under control and help slow down global warming. This study explores the training and use of the Deep Learning methods, specifically a modified U-Net to detect deforestation areas focussing on forest felling, using bitemporal Sentinel-2 satellite imagery. Two experimental configurations were tested: one using images from a single season and the other using multi-seasonal data. The results demonstrate that the single-season model outperformed the multi-season approach, achieving higher accuracy in detecting deforested areas while minimising false positives and overall good in deforestation detection. On the second hand, detection of deforestation in multiseason conditions did not show that good performance but still showed averagely good performance in detecting large clear deforested areas. The findings demonstrate the potential of DL approaches for deforestation detection, particularly in regions lacking alternative data collection methodologies. However, the study also highlights the crucial role of high-quality ground truth data in ensuring accurate model assessment. Addressing inconsistencies in data annotations through improved pre-processing and post-processing techniques could further enhance the reliability of DL-based deforestation detection. This research contributes to the advancement of Deep Learning methods for the detection and monitoring of environmental changes in vast and less explored areas, such as the Siberian forest.