A Two-Stage Multi-Scale Network for HighResolution Remote Sensing Images Change Detection

Deep learning-based change detection (CD) has emerged as a prominent research topics in the field of remote sensing. However, existing deep learning-based CD methods still face challenges due to insufficient multi-scale feature extraction and imprecise boundary delineation. To address the aforementioned problems, this paper proposes a novel two-stage multi-scale network (TSMSNet) for CD. In the encoding stage, a probability-gated multi-scale convolutional modulation module (PGMCM) is proposed, which integrates convolutional modulation, strip convolution, and a gating mechanism. This module can comprehensively extract multi-scale features from bitemporal remote sensing images while significantly enhancing boundary information. To further refine multiscale feature extraction a layer-decreasing pyramid pooling structure (LDPPS) is proposed within the skip connection architecture During the decoding stage, a global-local relation-aware feature fusion module (GLRAF) is introduced to enhance the network's capability in reconstructing image features. Finally, to optimize the learning process, a fuzzy-weighted binary cross-entropy loss function is designed to incorporate pixel classification uncertainty based on the degree of fuzziness. This novel loss function enables the network to focus more effectively on difficult-to-classify samples, thereby improving overall detection performance. The performance of the proposed network TSMSNet is validated on four publicly available CD datasets: WHU, GZ, GVLM and HGG. TSMSNet achieves IoU/F1 scores of 83.07%/90.75%, 73.39%/84.65%, 79.65%/88.67%, and 75.58%/86.09% on these four datasets, respectively. Compared with six state-of-the-art methods, TSMSNet shows improvements of at least 5.03%/3.08%, 3.18%/2.15%, 1.69%/1.06%, and 1.75%/1.15% in terms of IoU and F1 score. These experimental results demonstrate the effectivenesss of TSMSNet.