机载激光雷达数据的三维深度学习树种分类

针对传统基于激光雷达(Light Detection And Ranging, LiDAR)数据的树种分类方法难以直接且全面地利用点云的三维结构信息的问题，提出一种基于三维深度学习的机载LiDAR数据的树种分类方法。该方法直接从三维数据中抽象出高维特征，而无须将点云转化为体素或二维图像。以塞罕坝国家森林公园内白桦和落叶松两个树种的无人机LiDAR数据为研究对象，对其进行数据滤波，去除噪声和地面点；基于点云距离和改进的分水岭分割的方法提取单木并制作数据集；最终建立由权重共享的多层感知器、最大池、全连接层和softmax分类器组成的深层神经网络，其能自动提取树木的高维特征并实现树种分类。实验结果显示分类总体准确率为86.7%,kappa系数为0.73,最优特征维度为1 024,最优点密度为2 048。与将单木点云投影到二维视图的方法相比，该算法提供了更高的分类精度，且能有效减少计算成本、提高工作效率。

Tree species classification of airborne LiDAR data based on 3D deep learning

Aimed at the problem that the traditional tree species classification method based on LiDAR (light detection and ranging) data is difficult to directly and comprehensively use the 3D structure information of the point cloud, a tree species classification method of airborne LiDAR data based on 3D deep learning was proposed. This method directly abstracts high-dimensional features from 3D data without converting point clouds into voxels or two-dimensional images. Taking the airborne LiDAR data of white birch and larch in Saihanba National Forest Park as the research object, data filtering was performed to remove noise and ground points; the point cloud distance and improved watershed segmentation method were used to extract the individual wood and make a data set. Finally, a deep neural network composed of a weight-sharing multi-layer perceptron, a max pooling, a fully connected layer and a softmax classifier was established, which can extract the high-dimensional features of trees automatically and realize tree species classification. The experimental results show that the overall classification accuracy rate is 86.7%, the kappa coefficient is 0.73, the optimal feature dimension is 1 024, and the most advantageous point density is 2 048. Compared with the method projecting individual tree point cloud to a two-dimensional view, this algorithm provides higher classification accuracy, and can reduce the calculation cost effectively and improve work efficiency.