Improving L-Band SAR Forest Monitoring by Big Data Deep Learning Based on ALOS-2 5 Years Pan-Tropical Observations

Abstract

Forest observation is one prime objective for current and future L-band SAR missions. Despite the tremendous achievements made in the last 3 decades, the reliability of derived products, e.g., forest cover and biomass maps, is still largely hampered by ill-defined backscatter fluctuations. To better understand the rain-induced noise-like variations in tropical forest backscatter timeseries, we demonstrate big data deep learning applications exploiting the unprecedented long-term data sets provided by ALOS-2/PALSAR-2 from 2016–2021. Robust fully automatic deforestation detection is achieved in an operational forest monitoring system by using a CNN-based false alarm suppression method. Using 400 TB of ScanSAR images together with 50 billion auxiliary rainfall data an automatic rainfall prediction and correction procedure for L-band SAR is proposed based on LSTM artificial neural network. First results show the approach s potential to reliably reduce rain related disturbances in L-band forest timeseries in an unsupervised fashion.