Terrain scene category is useful not only for some geographical or environmental researches, but also for choosing suitable algorithms or proper parameters of the algorithms for several point cloud processing tasks to achieve better performance. However, there are few studies in point cloud processing focusing on terrain scene classification at present. In this paper, a novel deep learning framework for 3D terrain scene recognition using 2D representation of sparse point cloud is proposed. The framework has two key components. (1) Initially, several suitable discriminative low-level local features are extracted from airborne laser scanning point cloud, and 3D terrain scene is encoded into multi-view and multimodal 2D representation. (2) A two-level fusion network embedded with feature- and decision-level fusion strategy is designed to fully exploit the 2D representation of 3D terrain scene, which can be trained end-to-end. Experiment results show that our method achieves an overall accuracy of 96.70% and a kappa coefficient of 0.96 in recognizing nine categories of terrain scene point clouds. Extensive design choices of the underlying framework are tested, and other typical methods from literature for related research are compared.

地形场景类别不仅可用于某些地理或环境研究，而且还可用于为多个点云处理任务选择合适的算法或算法的合适参数，以实现更好的性能。然而，目前很少有针对地形场景分类的点云处理研究。本文提出了一种新的深度学习框架，该框架使用稀疏点云的2D表示进行3D地形场景识别。该框架有两个关键组成部分。（1）首先，从机载激光扫描点云中提取几个合适的判别性低层局部特征，然后将3D地形场景编码为多视图和多模态2D表示。（2）设计了嵌入特征和决策级融合策略的两级融合网络，以充分利用3D地形场景的2D表示，可以端到端地对其进行训练。实验结果表明，该方法在识别九种地形场景点云时，总体准确率达到96.70％，kappa系数达到0.96。测试了基础框架的广泛设计选择，并比较了文献中用于相关研究的其他典型方法。