| Summary: | Deep learning (DL)-based change detection (CD) methods for high-resolution (HR) remote sensing images can still be improved by effective acquisition of multi-scale feature and accurate detection of the edge of change regions. We propose a new end-to-end CD network, named the Multi-Scale Residual Siamese Network fusing Integrated Residual Attention (IRA-MRSNet), which adopts an encoder-decoder structure, introduces the Multi-Res block to extract multi-scale features and deep semantic information, and uses the Attention Gates module before the skip connection to highlight the change region features. Considering that the residual connection and attention module benefits the edge feature extraction, we proposed an IRA unit, consisting of the Res2net<sup>+</sup> module, the Split and Concat (SPC) module, and the Channel Attention Module (CAM), which can make the CD results better through finer-grained multi-scale feature extraction and adaptive feature refinement of the feature map channel dimension. The experimental results show that the <i>F</i>1 and <i>OA</i> values of our network model outperform other state-of-the-art (SOTA) CD methods on the Seasonal Change Detection Dataset (CDD) and the Sun Yat-Sen University Change Detection Dataset (SYSU-CD), and the number of parameters and the calculated amount are reduced significantly.
|
|---|