Building Damage Identification in Various Natural Disasters From Interferometric SAR Coherence and Machine Learning

This study presents a methodology for building damage identification in natural disasters using interferometric SAR coherence and machine learning technique. We used ALOS/PALSAR-2 images observed at before and after natural disasters in Japan such as earthquakes, landslides, debris flows and flooding. We confirmed that the interferometric coherences were significantly decreased in collapsed buildings compared to non-damaged buildings due to debris of the damaged buildings. Supervised learning based on convolutional neural network (CNN) was performed to the coherence images to identify the collapsed buildings. The results show that the severely damaged areas in the disasters were accurately identified in the building damage mapping although some of intact buildings were falsely classified especially in mountainous areas because the coherences were significantly lower due to radar shadowing effect in SAR and misalignment of two images in steep slopes.