基于双分支混合网络与级联线性回归的无人机航拍蔗田作物行与杂草定位方法

Dual-branch Hybrid Network and Cascaded Linear Regression Method for Crop Row and Weed Localization in UAV Aerial Images of Sugarcane Fields

摘要:

针对无人机航拍蔗田超大尺度正射图像中作物行与杂草联合定位易受杂草形态破碎、行间干扰及长距离角度误差累积影响的问题,提出一种双分支混合网络（CANESNet）语义分割与级联（RANSAC）线性回归的作物行中心线和杂草定位方法。构建Segformer与MobileNetV3双分支编码网络,引入 DCNv4 和 UNet 式多级解码结构,以增强作物行连续结构建模和杂草边界细节表达;在此基础上,基于分割掩膜提取作物行特征点,结合旋转投影分组和2次 RANSAC 拟合,实现跨区间行线合并与累计误差抑制。试验结果表明,CANESNet 在蔗田作物行与杂草分割任务中平均像素准确率、平均交并比、准确率和F1分数分别达91.21%、85.50%、96.68%和94.49%,推理速度为52.79 f/s;杂草类别交并比、像素准确率和精确率分别为61.52%、81.19% 和 71.75%。级联RANSAC行线提取成功率为92.65%,实际位置偏差平均值为0.013 m,角度偏差平均值为0.074°。结果表明,该方法能够在无需人工预设作物行数的条件下稳定提取蔗田作物行中心线并定位杂草区域,可为蔗田除草机器人导航线生成和精准除草作业提供环境感知支撑。

Abstract:

Aiming to address the challenges of joint crop-row and weed localization in ultra-large-scale UAV orthophotos of sugarcane fields, where fragmented weed morphology, inter-row interference, and accumulated angular errors over long distances often lead to unstable segmentation and row extraction offsets, a localization method integrating CANESNet semantic segmentation with cascaded RANSAC-based linear regression was proposed. CANESNet adopted a dual-branch encoder combining Segformer and MobileNetV3, with DCNv4 and a UNet-style multi-level decoder introduced to enhance continuous crop-row representation and fine weed-boundary recovery. Based on the segmentation masks, crop-row feature points were extracted and grouped using rotational projection, followed by two-stage RANSAC fitting to merge cross-regional row segments and suppress cumulative fitting errors. Experimental results showed that CANESNet achieved mPA, mIoU, Acc, and F1 values of 91.21%, 85.50%, 96.68%, and 94.49%, respectively, with an inference speed of 52.79 f/s. For the weed category, the IoU, PA, and precision reached 61.52%, 81.19%, and 71.75%, respectively. The cascaded RANSAC method achieved a crop-row extraction success rate of 92.65%, with an average positional deviation of 0.013 m and an average angular deviation of 0.074°. These results indicated that the proposed method can stably extract sugarcane crop-row centerlines and localize weed areas without manually presetting the number of crop rows, providing reliable environmental perception support for navigation-line generation and precision weeding operations in sugarcane fields.