关键词: 高光谱, 激光雷达, 叶绿素, 特征选择, 随机森林, 支持向量机

Abstract:

Objective: Forest stands are the basic units for forest resources monitoring and management. Using remote sensing data to accurately classify forest stand types has always been a hot topic in forestry research. The paper aims to explore the influence of classifiers and data sources (hyperspectral data and LiDAR data) on forest type identification, and to test whether chlorophyll plays an positive role in the identification. Method: This study area is at Laoshan working unit of Mao'ershan forest farm of Northeast Forestry University, and the airborne hyperspectral image and LiDAR point cloud were used as experimental data sources. Based on multi-scale image segmentation, spectral, texture and chlorophyll index features were extracted from hyperspectral images, and height and intensity features were extracted from LiDAR point clouds. Through the feature selection method of RF, features with higher importance were selected. Based on the classifier of RF and SVM, the image segmentation data were used as experimental samples to identify five forest types: broad-leaved mixed forest, Pinus sylvestris var. mongolica forest, Larix gmelinii forest, Pinus koraiensis forest and Quercus mongolica forest, and the classification effects of different data sources and different classifiers were compared. Result: A total of 34 feature variables were extracted from hyperspectral data, and 72 feature variables were extracted from LiDAR data. After feature selection, 11 feature variables were selected from hyperspectral data and LiDAR data, respectively, and normalized differential vegetation index(NDVI) extracted from hyperspectral data was the most important variable among them. Among the six classification schemes, the highest classification accuracy was to used hyperspectral and LiDAR data and RF classifier (overall accuracy was 88.02%), and the lowest classification accuracy was to use LiDAR data and SVM classifier (overall accuracy was 76.19%). The average classification accuracy of multi-source data was 86.22%, which was higher than the average classification accuracy of single data source of 79.98%. The average classification accuracy of RF classifier was 82.92%, which was higher than that of SVM (81.19%). At the same time, the results indicated that the chlorophyll index had a positive impact on stand type identification, and the classification accuracy had improved by about 3.32% after participating in classification. Among the five stand types, the broad-leaved mixed forest had the best classification effect with an average classification accuracy of 92.62%, while the Korean pine forest had the worst classification effect with an average classification accuracy of 49.67%. Conclusion: Compared with single data source, multiple data sources could better improve the accuracy of stand type identification. When using single data source, hyperspectral data had better classification effect than LiDAR data, and spectral characteristics were important factors that influenced stand type identification. In the process of stand type identification, comparing different machine learning models, RF had better classification effect than SVM. Additionally, as a biochemical parameter, chlorophyll had a positive effect on stand identification.

Key words: hyperspectral, LiDAR, chlorophyll, feature selection, random forest(RF), support vector machine(SVM)