SENTINEL-1 TIME-SERIES ANALYSIS FOR DETECTION OF FOREST DEGRADATION BY SELECTIVE LOGGING

Forest degradation by selective logging is considered one of the main causes of biodiversity loss and CO2 emissions in tropical regions. However, persistent cloud cover limits the detection of selective logging using optical satellite systems in the Brazilian Amazon. We develop a novel approach to detect selective logging using one-year time-series (TS) from Sentinel-1 RADAR data (C-band), based on state-of-art cloud computing using Google Earth Engine. The method consists of two temporal TS reductions. The first reduces the TS for the median monthly record while the second one computes annual statistics like mean, standard deviation, and amplitude. The result is a composite band used for classifying the annual TS through the application of a machine-learning algorithm (CART). Classification showed 69% overall accuracy within five classes; however, the misclassification of the degradation class was 54%. The classification accuracy has increased to 79% with the removal of the regrowth class, with 74% of the degradation correctly classified.