Enhanced Change Detection Method in Historical Districts: A Lightweight Visual Transformer Integration Model with Context-Aware Local Feature Augmentation

Abstract. Due to rapid urbanization and the continuous increase in building stock, significant challenges arise for historic district preservation. To overcome the persistent challenge of insufficient small-scale unauthorized structure detection in dense historic districts—a critical limitation of existing deep learning-based change detection frameworks—this paper introduces a Siamese network integrated with a lightweight visual transformer, effectively resolving subtle change omission in complex scenarios. The model utilizes context-aware local enhancement to capture high-frequency local information, significantly improving its accuracy in identifying changed regions. Within the change detection network, a CNN feature extractor first performs downsampling on the input image pair to preliminarily extract feature information. Subsequently, a semantic extraction module extracts and enhances semantic information from the feature maps. Finally, a prediction module calculates the differences between the features of the two images and generates the change prediction results. The reasrech comprehensively validated the model on the public LEVIR-CD dataset. Experimental results demonstrate significant improvements in performance metrics compared to other models. The findings indicate that the improved model also performs excellently on this dataset, verifying its effectiveness and robustness, and showcasing its ability to substantially reduce both omissions and false detections. This study offers a solution for high-accuracy remote sensing change detection by improving deep learning-based models.