StarCD-Net: A Remote Sensing Change Detection Method Combining StarNet and Differential Operators

In the past decades, remote sensing image change detection techniques have rapidly developed and achieved remarkable success in several fields. Although deep learning models have demonstrated good performance in change detection tasks, they still face challenges when dealing with some complex scenes, such as possible insufficient feature extraction and pixel loss at the target edges. To overcome these problems, we propose a new model, StarCD-Net, that combines StarNet and differential operators, which uses StarNet for feature extraction and aggregates the extracted features by Multi-Layer Feature Aggregation Module (MLFA) to obtain rich information. To address the problem of loss of edge details, we design the Detailed Feature Enhancement (DFE) Module to enhance the feature edges. In addition, the Hierarchical Feature Fusion Module (HFF) fuses the current layer with deeper features to obtain the global semantic information of the change object. StarCD-Net was tested on three challenging remote sensing image datasets, and its F1 and IOU reached 91.48% and 84.3% in LEVIR-CD, respectively, and its F1 was 0.94% and 0.59% higher than the second place in WHU-CD and SYSU-CD, respectively, which is significantly better than other change detection techniques and shows quite an outstanding performance.