Deep Learning Approach For Post-Flood Soil Deformation Mapping Using Insar Data

Deep learning techniques proved their robustness in many applications of automatic classification. As a first application in the context of post-flood soil deformations using InSAR imagery, we considered a transfer learning approach based on two successive fine-tunings of the pre-trained convolutional neural network AlexNet. The first fine-tuning is processed using a very large publicly available dataset of InSAR interferograms. Subsequently, the second fine-tuning is done using a small dataset of post-flood landslide InSAR interferograms. This latter dataset was built based on three major landslide events that happened in 2017. The main objective is to automatically distinguish between the fringes of landslides and the fringes of noise and atmospheric artifacts on the flattened interferogram. Preliminary classification experiments show encouraging results.