Few-Shot Semantic Segmentation for Remote Sensing ALS Point Clouds via Multidimensional Geometric Feature Embedding
Haijian Liu, Bin Liu, Caiping Yan, Yongwei Miao · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
Semantic segmentation of remote sensing point cloud data is a key task in the fields of 3-D computer vision and ecological remote sensing. To address challenges such as the sparse and uneven distribution of airborne laser scanning (ALS) point clouds, high annotation costs, and limited segmentation accuracy, this article proposes a few-shot semantic segmentation method that integrates meta-learning with multiprototype learning. The approach focuses on the effective representation and extraction of spatial, shape, and multidimensional geometric features. Especially, four types of geometric features—linearity, planarity, sphericity, and surface variation—are computed and extracted from ALS point clouds. A multiprototype generation module subsequently captures intercategory semantic relationships, while a label propagation algorithm operates on a constructed k-nearest neighbor graph to estimate semantic similarity among query points. Furthermore, an alignment regularization loss is introduced to jointly optimize support–query feature distributions, thereby alleviating issues like background mixing and feature sparsity. Experiments on our self-constructed campus ALS dataset (HZNU-3D) demonstrate that under the two-way one-shot setting, the proposed method achieves an average mean intersection over union (mIoU) of 77.05%, outperforming existing few-shot segmentation methods by approximately 2.14%. Moreover, we apply the proposed method to the whole campus scenes (one-way five-shot), achieving accurate segmentation of trees and buildings, which shows its potential for ecological and campus applications. Similar performance gains are observed on the ISPRS benchmark, where the proposed method consistently outperforms existing few-shot segmentation approaches across multiple task settings, achieving up to 43.38% mIoU in the challenging two-way one-shot scenario, which demonstrates its robustness and competitive generalization across ALS datasets with different point densities.