Seasonal Forest Disturbance Detection Using Sentinel-1 SAR & Sentinel-2 Optical Timeseries Data and Transformers

Tropical seasonal forests make up 40% of the globally available forest stock and play an essential role in regulating the variability in the global carbon cycle. Therefore, there is a strong need to persistently monitor seasonal forest changes to understand the forest carbon fluctuation and to better conserve biodiversity and enforce laws. In this regard, the advent of the European Space Agency (ESA) Copernicus program avails a dense time-series of both Synthetic Aperture Radar (SAR) and optical images, globally and free of charge, that enables the exploitation of these images for near real-time forest monitoring. Detecting seasonal forest changes in dense time-series, however, is complicated by fluctuation in the detected signal that is induced by forest phenology change. Therefore, forest disturbance detection methods should account for these seasonal fluctuations to make an accurate inference about forest disturbances. In this regard, deep learning approaches designed for sequential data such as Transformers can be used to implicitly learn the natural forest seasonality pattern in the signal to detect forest disturbances. This abstract demonstrates the efficacy of Transformers to detect seasonal forest disturbance in a seasonal dry-forest region in Bolivia.