Deep network based approaches to mitigate seasonal effects in SAR images for deforestation monitoring

Abstract

Most deforestation monitoring programs are based on optical images, which are severely affected by clouds, especially in tropical regions. As an alternative, Synthetic Aperture Radar (SAR) data are minimally affected by atmospheric conditions. However, one of the challenges inherent in such approaches is the effect of seasonal rainforest variations over the SAR images. Recently proposed stabilization algorithms mitigate this effect with significant accuracy gains, however, at the cost of higher computational resources, needed to accommodate long temporal SAR image sequences. This work addresses this issue and presents two alternative solutions that attain similar or better accuracy at a much lower computational requirements. One solution is based on a ResUnet, while the second combines an LSTM and a UNet. Both approaches were tested using raw and stabilized pairs as well as raw sequences. Experiments have indicated that both approaches can mitigate the seasonality effect using a much shorter SAR image sequences than modern stabilization methods require. The results with the multitemporal input outperformed those with the preprocessed bitemporal set by 4.3% in Recall, 1.7% in Precision, and 2.7% in F1-Score, also delivering the best deforestation probability maps.