With the increasing demand for detecting small water bodies in remote sensing imagery, particularly in Synthetic Aperture Radar (SAR) images, traditional water body extraction models often struggle to capture fine details and accurately delineate boundaries. To address these challenges, this paper introduces the integration of Frequency-Selective Deformable Convolution (FSDC) into the TransUNet architecture, optimizing water body extraction in SAR imagery. FSDC enhances feature representation in both the frequency and spatial domains. It does so through two key components: (1) The Frequency Selection Module, which employs Fourier transform to selectively enhance or suppress features across different frequency bands, thereby emphasizing the unique structure and boundaries of water bodies. (2) The Deformable Convolution Unit, which dynamically adjusts the receptive field via content-based sampling, allowing it to adapt to local variations at multiple scales and improve fine detail capture. After incorporating FSDC into the decoder of TransUNet, experimental results on the NY and C2S-MS datasets show a significant improvement in extraction accuracy, especially in detecting small water bodies. These findings underscore the effectiveness of the FSDC mechanism for small water body extraction from SAR imagery, offering a novel solution for precise water body analysis in remote sensing.

Change detection (CD), a crucial technique for observing ground-level changes over time, is a challenging research area in the remote sensing field. Deep learning methods for CD have made significant progress in remote sensing intelligent interpretation. However, with very high-resolution (VHR) satellite imagery, technical challenges such as insufficient mining of shallow-level features, complex transmission of deep-level features, and difficulties in identifying change information features have led to severe fragmentation and low completeness issues of CD targets. To reduce costs and enhance efficiency in monitoring tasks such as changes in national resources, it is crucial to promote the practical implementation of automatic change detection technology. Therefore, we propose a deep learning approach utilizing heterogeneity enhancement and homogeneity restraint for CD. In addition to comprehensively extracting multilevel features from multitemporal images, we introduce a cosine similarity-based module and a module for progressive fusion enhancement of multilevel features to enhance deep feature extraction and the change information utilization within feature associations. This ensures that the change target completeness and the independence between change targets can be further improved. Comparative experiments with six CD models on two benchmark datasets demonstrate that the proposed approach outperforms conventional CD models in various metrics, including recall (0.6868, 0.6756), precision (0.7050, 0.7570), F1 score (0.6958, 0.7140), and MIoU (0.7013, 0.7000), on the SECOND and the HRSCD datasets, respectively. According to the core principles of change detection, the proposed deep learning network effectively enhances the completeness of target vectors and the separation of individual targets in change detection with VHR remote sensing images, which has significant research and practical value.