Multi-scale feature learning for 3D semantic mapping of agricultural fields using UAV point clouds

Abstract

• A field mapping framework to support agricultural robot path planning. • LoGA-Net: Fusion of multi-scale attention with geometric prior for farm point clouds. • A local–global feature enhancement strategy to refine differential features. • Aggregation Classifier Module mitigate class imbalance. • 78.6% mIoU in classifying eight field features. Accurate spatial distribution information of field features is critical for enabling autonomous agricultural machinery navigation. However, current perception systems exhibit limited segmentation performance in complex farm environments due to illumination variations and mutual occlusion among various regions. This paper proposes a low-cost UAV photogrammetry framework for centimeter-level 3D semantic maps of agricultural fields to support autonomous agricultural machinery path planning. The methodology combines UAV-captured images with RTK positioning to reconstruct high-precision 3D point clouds, followed by a novel Local-Global Feature Aggregation Network (LoGA-Net) integrating multi-scale attention mechanisms and geometric constraints. The framework achieves 78.6% mIoU in classifying eight critical agricultural categories: paddy field, dry field, building, vegetation, farm track, paved ground, infrastructure and other static obstacles. Experimental validation demonstrates a 5.9% accuracy improvement over RandLA-Net on the Semantic3D benchmark. This advancement significantly enhances perception accuracy in complex agricultural environments, particularly for field boundary delineation and occluded feature recognition, which directly facilitates robust path planning for unmanned agricultural machinery. The framework provides a scalable technical and data-driven foundation for achieving fully autonomous farm operations, ensuring both operational efficiency and environmental sustainability.