Remote sensing image change detection method based on dual-branch multi-level feature difference interactive learning

Abstract

Remote sensing (RS) image change detection (CD) is a key technology in environmental monitoring and geographic information systems (GIS). It can reveal the dynamic changes of surface features and is of great significance in fields such as urban planning, disaster assessment, and ecological research. However, the pseudo-change problem, that is, the image differences caused by non-actual surface changes, often affects the accuracy of detection, leading to false alarms and omissions, which limits the effectiveness of the CD technology. Traditional dual-branch CD methods often focus on basic feature extraction. This method independently processes the feature extraction of the bi-temporal phases and lacks a comparative interactive learning process for the features of the bi-temporal phases, thereby weakening its ability to identify pseudo-changes in complex environments. To solve the above problems, we propose a RS image CD method based on dual-branch multi-level feature difference interactive learning (DMFDIL). The model is built based on the siamese convolutional neural network (CNN) of deep learning and includes three parts: the dual-branch cooperative coding module (DCM), the dual-branch difference decoding module (DDDM), and the change output module (COM). Among them, the DCM innovatively introduces the tri-attention mechanism. Through this mechanism, the model can effectively interact on multi-level features, enhancing the ability to capture subtle changes in RS images, especially in distinguishing real changes from pseudo-changes. The DDDM, on the other hand, focuses on further optimizing the detection capability of the model by identifying real changes from pseudo-changes and integrating feature information at different scales. Finally, the validation was carried out on three public datasets, and the results were better than other popular methods. The experimental results on the LEVIR-CD dataset show that the proposed DMFDIL model achieved 95.80% in precision (Pre), 94.54% in recall (Rec), 95.16% in F1-score (F1), 91.10% in Intersection over Union (IoU), and 99.07% in overall accuracy (OA), which are significantly better than those of the state-of-the-art (SOTA) approaches. This method provides a new technical approach in the field of RS image CD, especially in improving detection accuracy and dealing with pseudo-change problems, and has important application value and broad application prospects.