An Improved DeepLabV3+ Algorithm for Identifying Key Deformation Areas in InSAR Images

Abstract

Conventional surface deformation monitoring predominantly depends on SAR and interferometric imagery while neglecting the comprehensive analysis of InSAR-processed deformation results, consequently limiting the precise detection of critical deformation zones in the derived data. In view of the difficulties in interpretation caused by the RGB multicolor coding of the interferometric synthetic aperture radar (InSAR) deformation rate map and the problems of low efficiency and poor accuracy of the existing automatic recognition methods, this study proposed an improved DeepLabV3+ architecture, integrated with MobileNetV2 backbone network, self-attention module, and multilevel self-attention feature fusion mechanism, to improve the accuracy and efficiency of automatic deformation detection. This study takes the main urban area of Daqing city, Heilongjiang province as the study area, and uses short-baseline InSAR technology to obtain surface deformation data and make dataset. Through ablation experiment and comparative analysis, the improved model has achieved significant improvement in indicators, such as recall (increased by 5.38%), mean intersection over union (mIoU, increased by 3.34%), pixel accuracy (PA, increased by 1.40%), and reasoning speed (shortened by 47 ms) compared with other mainstream semantic segmentation models, which can effectively identify the key surface displacement areas in InSAR images and provide reliable technical support for geological disaster prevention and control.