Traditional pesticide application methods often result in excessive deposition and drift in non-target areas, leading to severe environmental pollution. This study proposes a precision variable-rate spraying system for greenhouse asparagus based on real-time canopy semantic segmentation and volume estimation. To address the challenges of occlusion and sparsity in asparagus point clouds, an improved RangeNet++ architecture is introduced, incorporating Lightweight Channel Attention (LCA) and Efficient Spatial Attention (ESA) mechanisms to optimize feature extraction. An adaptive feature fusion module is further designed to enhance boundary delineation. Based on the semantic segmentation, an octree-based 3D grid model is constructed to calculate canopy volume in real-time, driving a CAN-bus controlled variable-rate spraying robot. Field experiments demonstrated that the proposed method significantly outperforms the standard RangeNet++, particularly in asparagus segmentation, where the Intersection over Union (IoU) increased by 13.2% and overall accuracy improved by 2.1%. For online canopy volume estimation, the system achieved accuracies of 90.07%, 89.73%, and 88.34% at robot travel speeds of 0.2, 0.4, and 0.6 m/s, respectively, demonstrating robust performance under dynamic operational conditions. Pearson correlation analysis confirmed a strong positive correlation between the applied spray volume and canopy density. Furthermore, filter paper deposition tests revealed that variable-rate spraying significantly improved uniformity: the Coefficient of Variation (CV) of deposition in the outer, middle, and inner canopy layers was reduced from 94.2%, 53.9%, and 70.0% (conventional spraying) to 38.5%, 49.8%, and 46.5% (variable spraying), respectively. The proposed system effectively mitigates localized excessive deposition while maintaining adequate coverage, providing a viable engineering solution for precision plant protection in facility agriculture.
Tan et al. (Mon,) studied this question.