基于双分支U-Net的遥感影像稻田分割方法

Dual-branch U-Net-based Method for Paddy Field Segmentation in Remote Sensing Imagery

摘要:

遥感影像中小面积、分布零散的水稻田块难以精准分割。基于线性光谱组合的输入无法挖掘波段间非线性耦合关系，基于堆叠波段影像的输入易引入冗余信息。针对此问题，本文提出一种基于U-Net的改进网络，该网络以RGB与NRG假彩色影像作为双输入，采用双编码器结构提取多模态特征信息，并结合局部金字塔注意力模块与自适应多尺度注意力特征融合模块，显著提升网络对小尺度水稻田块的感知与分割能力。对构建的水稻影像数据集进行实验，表明DFAU-Net在分割精度、鲁棒性和效率上表现优异。其Dice系数、平均交并比和准确率分别达到77.54%、86.34%和91.48%，较多种主流方法具有明显优势。进一步的消融实验验证了LPA模块、AMSADFF模块和双分支结构的有效性。该方法不仅能提高水稻田块的分割精度，也为复杂背景下的小目标分割提供了有效的解决方案。此外，本研究展示了高分辨率遥感影像在农业监测中的潜力，为精准农业、作物监测及产量估算提供了新的技术路径。综合而言，DFAU-Net为解决小规模水稻田块分割难题提供了有效的技术支持，具有广泛的实际应用价值。

Abstract:

Accurate delineation of small and spatially fragmented paddy fields from remote sensing imagery remains a challenging task. Traditional inputs based on linear spectral combinations are limited in their ability to capture nonlinear couplings across spectral bands, while direct stacking of multi-band imagery often introduces redundant information and increases computational complexity. To address these limitations, an enhanced U-Net-based segmentation framework was proposed. The network employed a dual-input strategy, integrating both RGB and NRG false-color images derived from the Gaofen-2 satellite, and incorporated a dual-encoder architecture to extract complementary multimodal feature representations. To further strengthen the discrimination of fine-scale objects, a local pyramid attention (LPA) module that enabled hierarchical aggregation of local contextual cues was designed, thereby improving the model’s sensitivity to small paddy patches with irregular boundaries. In addition, an adaptive multi-scale attention dynamic feature fusion (AMSADFF) module was introduced to dynamically integrate features across multiple scales, mitigating redundancy while preserving the most informative spatial and spectral patterns. By synergizing these mechanisms, the proposed framework—termed DFAU-Net—achieved a robust balance between local detail preservation and global contextual understanding. Experimental evaluations were conducted on a dedicated paddy field dataset constructed from high-resolution Gaofen-2 imagery. Results demonstrated that DFAU-Net consistently outperformed several state-of-the-art segmentation models. Specifically, it achieved a dice coefficient (Dice) of 77.54%, a mean intersection over union (mIoU) of 86.34%, and an overall accuracy (Acc) of 91.48%. These results highlighted the superiority of the method in capturing fragmented and smallscale field patterns, where conventional models tended to fail. Overall, DFAU-Net provided a promising solution for accurate agricultural parcel mapping, with potential applications in crop monitoring, yield estimation, and precision agriculture.