Accurate farmland segmentation is essential for precision agriculture, land-use planning, and environmental monitoring, as it facilitates efficient resource management and supports sustainable agricultural development. However, existing methods still struggle with multi-scale variation, complex topological structures, and fine-grained boundaries, which significantly degrade segmentation performance in remote sensing imagery. To address these challenges, we propose TCNet, a novel deep learning framework with a progressive dual-enhancement strategy. The framework integrates a CNN–Transformer hybrid architecture with Multi-Scale Self-Attention Modules (MSAMs) to effectively capture local details and global contextual information across multiple scales. In addition, a Group CBAM Enhancer (GCE) is introduced to progressively refine feature representations, while complementary feature fusion operations, including multiplication, concatenation, and element-wise addition, are employed to enhance multi-scale feature integration. Extensive experiments on the benchmark datasets FGFD and FIT demonstrate that TCNet outperforms state-of-the-art models in key metrics such as Intersection over Union (IoU), F1-score, and Recall. These results demonstrate the robustness and generalization capability of the proposed method, making it a promising solution for accurate farmland mapping in diverse and challenging environments.
Wei et al. (Sun,) studied this question.