An improved lightweight irrigation canal segmentation network with direction perception for agricultural UAVs

The inspection for irrigation canals based on unmanned aerial vehicles (UAVs) is an important and challenging task in the field of modern agriculture. Specifically, accurate segmentation of irrigation canals from UAV images faces several challenges such as complex background textures, vegetation occlusions, and varying lighting conditions, which can lead to blurred canal boundaries and discontinuous features. To improve the accuracy and robustness of the image segmentation, an improved lightweight semantic segmentation network (named GEA-UNet) is proposed in this paper. In the proposed model, a direction perception attention module is presented to enhance orientation sensitivity. In addition, an edge detection auxiliary module is designed for refined boundary learning, and a context-aware segmentation module is proposed to capture local and global features of the irrigation canals. Evaluation results on the self-constructed irrigation canal dataset show that the proposed GEA-UNet model achieves an accuracy of 98.9%, mean Intersection over Union of 85.4%, and F1-score of 92.2%, outperforming other mainstream semantic segmentation models. Path extraction experiments using sliding projection and RANSAC regression further showed that the proposed method reduces the average angular error to 1.27$$^{\circ }$$ and the average fitting time to 3.67 ms, significantly enhancing the navigation accuracy and efficiency for agricultural UAVs. This work provides an effective and efficient solution for autonomous UAV-based canal inspection, contributing to intelligent decision-making and precision management in modern irrigation systems.