Canopy3D-Net: Semantic segmentation of fruit tree canopies based on 3D point clouds

Precise segmentation of tree canopies and extraction of their semantic information are crucial for estimating crop water requirements and pesticide application needs. However, accurately segmenting densely clustered fruit tree canopies from 3D point clouds remains challenging due to the irregular geometry and variable point density characteristics typical of agricultural environments. To address this, we propose a semantic segmentation network for fruit tree canopy point clouds, named Canopy3D-Net. The network architecture incorporates three key modules: (1) a height-aware random sampling method designed to ensure efficient sampling while directing the model's attention toward the upper regions of the canopy; (2) a local multi-feature fusion module that enriches local contextual features and enables the network to autonomously learn optimal channel-wise feature representations; and (3) a hybrid loss function integrated to mitigate class imbalance during canopy classification. Evaluated on a self-built citrus orchard dataset (456 samples), our model achieves high performance (mIoU = 0.849, OA = 0.938, IoU_Canopy = 0.832, IoU_Background = 0.968). It also demonstrates strong generalizability on the public Semantic3D dataset (mIoU = 0.752, OA = 94.5 %, IoU_Canopy = 89.1 %). The results indicate that Canopy3D-Net delivers outstanding segmentation performance for large-scale canopy point clouds, thus providing an efficient and accurate solution for point cloud processing in agricultural and forestry applications.