Urban change detection of remote sensing images via deep-feature extraction

Urban change detection based on remote sensing images holds significant importance in environmental monitoring and emergency management. However, it poses several challenges including large disparity errors, diverse types of changes, and a substantial difference between the number of changed and unchanged areas. In this study, we propose an efficient model called BiUnet-Dense for extracting deep features by integrating the advantages of Bi-Unet, Dense Block, and long short-term memory (LSTM) networks. Building upon the classical architecture of U-Net, Bi-Unet utilizes bi-temporal images to compare and extract features. The incorporation of modified dense connections reduces network parameters while mitigating gradient disappearance through maximizing feature reuse. Additionally, LSTM facilitates information transmission from earlier to later using cell states to provide more meaningful feature vectors. We implemented our model on two datasets: Onera Satellite Change Detection (OSCD) and Change Detection Data of Guangzhou (CD_Data_GZ). Quantitative and qualitative results demonstrate that our method significantly improves detection effectiveness with enhanced F1-score and Kappa while effectively reducing false-positive detections as well as identifying labeling errors.