Airborne multispectral LiDAR point cloud classification with a feature Reasoning-based graph convolution network

• Published in: International journal of applied earth observation and geoinformation Vol. 105; p. 102634
• Main Authors: Zhao, Peiran, Guan, Haiyan, Li, Dilong, Yu, Yongtao, Wang, Hanyun, Gao, Kyle, Marcato Junior, José, Li, Jonathan
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 25.12.2021; Elsevier
• Subjects: Deep learning; Feature reasoning; Graph convolution network; Multispectral LiDAR; Point cloud classification; Deep learning; Multispectral LiDAR; Point cloud classification; Feature reasoning; Graph convolution network
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2021.102634

•A state-of-the-art FR-GCNet was applied for airborne multispectral LiDAR (MS-LiDAR) point cloud classification.•The FPS-KNN sampling strategy was used to quickly and effectively obtain training samples.•A novel feature reasoning module was designed to enhance features and further improving point classification results.•The sufficient ablation study and comparative experiments were implemented. This paper presents a feature reasoning-based graph convolution network (FR-GCNet) to improve the classification accuracy of airborne multispectral LiDAR (MS-LiDAR) point clouds. In the FR-GCNet, we directly assign semantic labels to all points by exploring representative features both globally and locally. Based on the graph convolution network (GCN), a global reasoning unit is embedded to obtain the global contextual feature by revealing spatial relationships of points, while a local reasoning unit is integrated to dynamically learn edge features with attention weights in each local graph. Extensive experiments on the Titan MS-LiDAR data showed that the proposed FR-GCNet achieved a promising classification performance with an overall accuracy of 93.55%, an average F1-score of 78.61%, and a mean Intersection over Union (IoU) of 66.78%. Comparative experimental results demonstrated the superiority of the FR-GCNet against other state-of-the-art approaches.