基于多光谱全波形LiDAR数据与域自适应学习的单树级林业信息提取

With the rapid development of sensors, more and more accurate information can be derived by remote sensing. However, it is still a difficult issue and almost impossible to retrieval single tree level information by traditional remote sensing techniques, which is of great importance to forest inventory and forest management. As a combination of spectral and full waveform LiDAR sensors, multi-spectral full waveform LiDAR can acquire the spectral and full waveform data simultaneously, the combination of which has been proved to be the best data for single tree level information extraction. There exist many problems to remote sensing in forest caused by the field forest inventory, which is cost expensive and time consuming to collect enough samples.. This research aims to retrieval high accurate forest information from multi-spectral full waveform LiDAR based on domain adaption theory in transfer machine learning, which transfers the samples and models from the source domain to target domain and reduces the requirement of labeled samples from field forest inventory. To do this, the single tree level features are firstly extracted from the multi-spectral full waveform LiDAR data by fusing spectral、full waveform and geometric information and segmenting corresponding point cloud with methods based on the N-Cut graph theory. Secondly, two key issues are studied: one is the tree species classification based on the model domain adaptation of random forest classifier and the other is the single tree parameter estimation based on the domain adaptation of random forest regression model. As decision tree is the basis of random forest, solving the model adaption of decision tree can make the model adaption of random forest. In this research, we study two strategies for the decision tree adaptation: “node structure split/reduction” and “node structure transfer by changing decision rules”. Finally, accuracy assessment method is proposed by double-direction learning based cross validation for the forest information retrieval from the multi-spectral full waveform LiDAR based on the random forest model adaption.. With the proposed and studied methods, less samples are essential and more samples with the corresponding learning models can be transferred from the source domain to the target domain, which is helpful to the development of new theory of transfer learning in remote sensing.

针对传统遥感技术难于获取单树结构信息的不足，以及林业遥感外业调查存在样本数量多、成本高、周期长等问题，本项目以多光谱全波形LiDAR为研究对象，以域自适应学习方法为指导，通过研究基于融合“光谱-波形-几何”信息的单树特征提取方法、重点通过研究“分支节点判定规则迁移”和“分支节点分裂/剪枝”两个决策树迁移策略来突破基于随机森林模型域自适应学习的树种分类识别与参数回归估计这两个关键核心问题、提出并发展基于双向学习交叉验证的遥感迁移学习精度评价框架，研究形成基于多光谱全波形LiDAR数据与域自适应学习方法支持下单树级林业信息的高精度提取理论与方法，实现将源域（林业信息通过大量外业调查样本已提取出的区域）标记样本及其分类回归模型迁移到目标域（林业参数待估计区域）中，进而减少林业遥感信息提取对外业调查的大样本量需求。项目成果可充分利用源域样本和模型信息，降低遥感目标参数估计对标记样本量的要求。

快速获取精细林业信息对于实现双碳目标具有重要意义。对于单树级林业信息的提取，机载激光雷达技术具有快速高效、覆盖面广的优势。利用机载激光雷达进行单树级林业信息提取，需要大量的外业调查数据做支撑。在减少外业调查样本数量的情况下，如何利用机载激光雷达数据和域自适应学习方法来高精度提取单树级林业信息，围绕这一问题和需求，本项目主要开展了以下方面的研究：（1）研究提出了基于卷积神经网络和数据融合的林区地形提取方法，该方法首先利用残差Unet神经网络和DSM来预测林区地形，再利用滤波所得激光雷达地面点来修正所预测的地形，进而得到高保真和高精度的林区地形；（2）研究提出了基于随机森林模型和半监督域自适应学习的单木树种分类方法，该方法以随机森林为基础学习模型，首先采用“决策树节点的分裂/剪枝”和“决策树分支节点判定规则的迁移”这两个策略来实现随机森林模型的域自适应，然后再基于半监督学习算法Co-Forest利用未标记单木样本来提高单木树种分类精度；（3）研究了一种基于随机森林回归模型迁移的单树胸径估计方法，该方法通过修改随机森林模型中的回归子树结构，达到整体回归模型迁移的效果，可以有效减少目标域林区对单木样本数量的要求。整体上，本项目实现了大部分的研究目标，研究成果可为智慧林业提供技术支撑。需要说明的是，多光谱全波形LiDAR是一种非常前沿的技术，目前其机载系统仍在研制中，还无机载多光谱全波形LiDAR数据，“利用多光谱全波形LiDAR数据来提高单木林业信息提取精度”这一目标还需进一步的研究工作来实现。