DRMNet: Difference image Reconstruction enhanced Multi-resolution Network for optical change detection

Change detection in satellite images is an emerging research area as it has a wide range of applications natural resource monitoring, geo-hazard detections, urban planning, etc. Identifying physical changes on the ground and avoiding spurious due to other reasons like co-registration issues, change illumination conditions, sun angle, presence cloud fog challenging task. This work proposes multitask learning based model where two parallel pipeline architectures predict map image difference. The proposed takes their difference input provides them Backbone Network (BN). output BN fed into Multi-Scale Attention Module (MSAM) for effective identification multi-temporal very high-resolution aerial images. In another path, down-sampled passed Deconvolution with Sub-pixel Convolution (DSCM) generate Two loss functions are utilized paths train overall end-to-end supervised setting. A comprehensive set experiments have been carried out, results reveal that DRMNet achieved F1 Score improvement 1.66% CDD, 1.61% SYSU, 0.14% LEVIR-CD datasets. It score 86.11% BCDD dataset new test image. Our implementation available at https://github.com/chouhan-avinash/DRMNet.