SCSQ-Net: A Shared Kernel Point Convolution Semantic Query Network for Weakly Supervised Classification of Multispectral LiDAR Point Clouds

Multispectral LiDAR (MS-LiDAR) point cloud classification holds great potential, but current methods rely heavily on fully supervised learning, requiring costly manual labeling. To address this, we propose a shared convolutional semantic query network (SCSQ-Net), a novel weakly supervised model designed for airborne MS-LiDAR data. It utilizes a shared kernel point convolution (SKPC) backbone to integrate neighborhood information during both encoding and semantic inference. An adaptive dropout activation (ADA) module is introduced to further enhance activation and feature learning across the entire point cloud scene. Our loss strategy combines point affinity loss, weighted cross-entropy loss, and Lovasz-Softmax loss for comprehensive supervision. Evaluated on two large datasets, the SCSQ-Net demonstrated performance comparable to fully supervised models, achieving an average $F1$ -score of 86.83%, an intersection over union (mIoU) of 80.71%, and overall accuracy (OA) of 96.29% using only 0.1% labeled data.