Interpreting Oil-Spill Mapping DNN: A Case Study with Hybrid-pol SAR data

The study in this paper presents a semantic segmentation deep neural network model for oil-spill mapping. A comprehensive analysis on the performance of this model via Grad-CAM model interpretation strategy is proposed here. Hybrid polarimetric features generated from the ALOS PALSAR full-pol data of the Deepwater Horizon oil-spill event are used for model development and assessment. Corresponding oil-spill masks are created using OTSU multi-thresholding method and used as targets. A semantic segmentation model based on UNet-inspired Encoder-Decoder architecture is developed and trained. The trained model performed well on testing samples with a mean IoU of 0.91. Grad-CAM activation maps of the Encoder and Bridge layers indicate effective class-discriminative mapping performance.