CVMNet: A CNN–VMamba Hybrid Network for Farmland Segmentation in VHR SAR Images

Farmland segmentation is crucial for precision agriculture and land resource management, and very high resolution synthetic aperture radar (VHR SAR) offers valuable data support. Yet, effectively leveraging this data remains challenging. Convolutional neural networks (CNNs) struggle to capture long-range dependencies, while transformers, though effective in long-range modeling, suffer from high computational costs. Additionally, noise in VHR SAR images often results in ambiguous edge semantics. To address these issues, a farmland segmentation network for VHR SAR, named the CNN-VMamba network (CVMNet), is proposed in this letter. Specifically, a gradient-curvature joint algorithm (GCJA) is applied to VHR SAR images to generate edge strength map, providing enhanced contour information. In parallel, a CNN-based local feature extraction (LEF) block, sensitive to tubular structures, is employed to capture fine-grained spatial details. For efficient long-range modeling with linear computability, a 2-D selective scan (SS2D) mechanism from VMamba is integrated. Furthermore, an SS2D-based shallow-deep feature fusion module (SDFFM) is proposed to facilitate cross-layer feature fusion. Experiments using TerraSAR-X images showed that the proposed method outperforms compared methods, yielding an F1-score of 96.95% and an intersection over union (IoU) of 94.08%. The code is available at https://github.com/onesar/CVMNet.