Multispectral Airborne LiDAR Point Cloud Classification With Maximum Entropy Hierarchical Pooling

Abstract

The demand for accurate airborne LiDAR point cloud classification has increased with improved resolutions of land cover map products. Although existing deep learning-based methods are capable of classifying airborne LiDAR point clouds, these methods indeed have a limited capability to extract the local features and suffer from global and local information losses with the commonly used pooling approaches. Therefore, we present a deep learning-based optimal homogeneous neighbor selection (HNS) and hierarchical pooling by exploiting maximum entropy, called MEHPool. The module is designed to directly extract sufficient homogeneous neighbor points for each point, followed by a designed graph pooling (GP) layer that encapsulates the selected homogeneous neighbor points into small-size graphs to build hierarchical features. The plug-and-play module consisting of an HNS module, two GP layers, and three graph neural networks (GNNs) can be easily embedded into various networks for point cloud classification and produces the architecture MEHPool-Net in this letter. Our experimental results show that the proposed MEHPool-Net realizes effective performance for multispectral airborne LiDAR point cloud classification, consistently outperforms four other deep learning methods, and confirms the superiority of the GP module compared with five other pooling methods.