Assessing Sentinel-1 InSAR Short-Time-Series for Systematic Rainforest Mapping with Deep Learning

Abstract:
Monitoring land cover changes in forest ecosystems with Earth Observation (EO) data has become of paramount importance in the context of preserving biodiversity and mitigating climate change worldwide. Nonetheless, most of the current warning systems rely on optical data and update forest maps no more than once a year – an observational gap that severely affects the monitoring of rainforests, which suffer from persistent cloud cover and rapid deforestation. In this paper, we investigate the capabilities of Sentinel-1 interferometric synthetic aperture radar (InSAR) time series for short-term mapping of endangered Amazon rainforest regions. In particular, we exploit the potential of deep learning and convolutional neural networks (CNNs) to better correlate the presence of forest areas with the evolution in time of both SAR backscatter and interferometric coherence. To this end, we consider a U-Net-like architecture and we train it from scratch to comply with the unique requirements of our problem, and learn higher level features, while preserving pixel-wise spatial localization. A performance benchmark analysis on the impact of different input features is provided, with respect to state-of-the-art methods. Preliminary results of the rainforest mapping show that an overall agreement above 90% can be achieved when compared to the Finer Resolution Observation and Monitoring of Global Land Cover (FROM-GLC) thematic map, outperforming the current state-of-the-art random forest classifiers for the same task and strongly reducing the required computational load.