多光谱LiDAR系统出现为森林资源调查和监测开启了一个新的数据时代。目前，基于多光谱LiDAR数据的应用存在缺乏有效多波段数据融合方法、基于空间几何信息与光谱信息结合的单木特征提取智能化程度低、树木三维点云的整体特征描述能力不足、分类方法上特征抽象与可区分能力差等问题。本项目结合计算机视觉、数值分析、计算几何等领域理论和方法，重点研究（1）多光谱LiDAR多波段点云数据集的融合模型，（2）光谱-空间几何信息结合的森林地形提取以及归一化割的三维单木点云数据分割方法，（3）根据单木点云数据的局部特征抽象，建立单棵树木的整体特征的空间词语袋，通过深层波尔兹曼机模型构建的多层分类器从而实现树种分类。本研究将着重于多光谱LiDAR数据的智能化处理，为多光谱LiDAR技术在森林资源调查提供技术示范与实现方法。

Multispectral LiDAR, which includes several active imaging channels of different wavelengths for day or night mapping of complex environments, opens the doors for multispectral acting imaging of forest resource investigation and monitoring. Currently, the use of multispectral LiDAR data is facing the following issues: lack of an effective means of integrating all channels into a comprehensive and single point cloud, low automation of tree feature extraction algorithms by using spectral and geometric information, and lack of feature descriptors for three-dimensional (3D) tree points and robust tree species classifiers. To solve these issues, this project, with the knowledge of computer vision, numerical analysis, computational geometry and remote sensing-related fields, will research on methods of (1) data fusion of multiple channels, (2) terrain filtering and single tree segmentation by using the integration of spectral and geometric information, and (3) bag-of-spatial-word feature abstraction for representing local features extracted from 3D tree points, followed by a Deep Boltzmann Machine Mode based classifier to successfully distinguish tree species. This project emphasizes on the automation of multispectral LiDAR data and aims at providing a theoretical foundation and prototype in the forest resource monitoring and management.