基于无人机可见光和LiDAR数据的单木树种识别

Single Tree Species Identification Based on Visible Light and LiDAR Data of UAV

以亚热带林业实验中心年珠实验林场为研究区,以无人机可见光和LiDAR数据为数据源进行树种识别。基于CHM和可见光数据进行单木分割,对可见光数据和LiDAR数据进行特征提取,构建多特征集合;基于单木对象选择随机森林和支持向量机2种分类器进行分类识别,并利用混淆矩阵对不同数据源不同特征组合的12种方案进行精度评价,比较不同特征组合和分类器对树种分类精度的影响。结果表明:将基于CHM分割和多尺度分割结合的单木分割效果较好,满足单木树种识别需求。支持向量机的精度高于随机森林分类器,经过随机森林特征筛选之后精度优于未进行特征筛选的结果,总体平均精度提高1.45%,可见光和LiDAR数据结合较仅使用单一数据源平均精度提高了6.01%。特征筛选能减少维度灾难,有效难避免过多特征造成的冗余现象,进一步提高分类器的性能和效率。相对于随机森林分类器,支持向量机在对于多维的样本集以及训练样本有限的情况下,能够表现出更好的性能。多源数据结合能将不同数据源优势有效结合,提高分类精度。

The Nianzhu experimental forest farm of subtropical forestry experimental center was selected as the research area, and the visible spectral and LiDAR data of UAV were used as data sources for tree species identification in this study. Based on CHM and visible spectral data, single-wood was firstly carried out, and then extracted features of visible spectral data and LiDAR data to build more characteristic collection. Based on object single-wood to choose RF and SVM classifier to classify recognition, and using confusion matrix to evaluate the accuracy of 12 schemes with different data sources and different feature combinations. To compare different characteristics of the combination and classifier to the influence of tree species classification accuracy. The results show that the single tree segmentation based on CHM segmentation and multiresolution segmentation is effective,which can meet the needs of single tree species identification;and the results show that the precision of SVM is higher than that of random forest classifier, and the precision after RF feature selection is better than those without feature selection. The overall accuracy is improved by 1.45% on average, and the average precision of visible spectral and LDAR data integration has improved the average accuracy by 6.01% compared with using only a single date source. It can be seen that feature selection can reduce dimension disaster, effectively avoid redundancy caused by too many features, and the performance and efficiency of classifiers are further improved. In contest to the RF classifier, SVM performs better in the case of limited multidimensional sample sets and training samples. The combination of multi-source data can effectively combine the advantages of different data sources and improve the classification accuracy.