Selective Kernel Res-Attention UNet: Deep Learning for Generating Decorrelation Mask With Applications to TanDEM-X Interferograms

Decorrelation is one of the main limitations for InSAR. Masking decorrelated pixels crucial retrieving information from SAR interferograms. However, traditional masking methods, manually drawing masks time-consuming and may be unfeasible when decorrelation areas are with complicated blurred boundaries. Setting a single coherence threshold also difficult, if not impossible, to mask out all without losing valid phases. Here, we propose deep-learning segmentation network (Mask Net) based on Selective Kernel Res-Attention UNet, generating applications TanDEM-X We conduct several experiments determine training strategy parameters, including sample size, batch loss function down-sampling scheme, optimize performance. Afterwards, compare performance Mask Net other classical networks. Our evaluation metrics show that outperforms best networks by IoU 6.32% F1 Score 3.97%, respectively. It possesses fastest inferring speed, 0.4505s size 1024-by-1024 pixels, which at least ~50% faster than applied three interferograms Klauea crater in Hawaii, metropolitan region Wuhan, Muztagata Glacier China. results comparing method, can clearly regions, rarely causing exhibits better networks, especially those complex boundaries, less computational time.