A Method for Constructing Farmland Semantic Point Cloud Maps Based on LiDAR–Camera Fusion

With the development of intelligent agricultural machinery and unmanned farms, agricultural production is gradually evolving toward greater autonomy, reduced labor dependence, and higher precision. To address the limitation of a single sensor in unstructured farmland environments, where three-dimensional metric information and semantic information cannot be captured effectively at the same time, this study proposes a method for constructing a farmland semantic point cloud map based on LiDAR–camera fusion. DeepLabV3+ was used to extract pixel-level semantic information, while ground-plane alignment, normalized projection, effective field-of-view constraints, and spatiotemporal synchronization were integrated to map two-dimensional labels onto three-dimensional point clouds. The proposed method was validated in three plots with different boundary shapes using RTK-GNSS reference boundaries. The experimental results showed that the DeepLabV3+ model achieved an mPA of 90.71% and an mIoU of 82.86% on the test set. The constructed semantic map yielded an overall mean point-position error of 0.192 m and an RMSE of 0.216 m, with an overall area coverage of 97.20%. The proposed method therefore enables three-dimensional semantic labeling and farmland boundary reconstruction while providing metric-scale semantic information for environmental perception and autonomous navigation of intelligent agricultural machinery.