Abstract:

【Objective】 Forest vegetation is a principal part of forest resources. Accurate identification of forest vegetation types is important for research and utilization of forest resources. The combination of different characteristics of remote sensing data has great advantages for determining forest vegetation types and forest parameter estimation, which could be used to classify tree species more effectively. In the face of massive feature data, the use of classifiers that are insensitive to dimensionality could also result in a decrease in classification accuracy. At the same time, nonparametric classifiers [random forest (RF) and support vector machine (SVM)] will improve classification accuracy by adding non-spectral data to the classification process. In this study, the main impact of feature selection on classification results was explored, the importance of different features for tree species classification was studied, and the effectiveness of multi-source data for individual tree species classification was investigated. 【Method】 Two plots (100 m × 100 m) in the Zhonglin District of Maoershan Forest Farm of Northeast Forestry University, Heilongjiang Province, China, were used as the study area, multi-spectral remote sensing charge coupled device (CCD) images and airborne light detection and ranging (LiDAR) data were taken as data resources, and forest resource survey data from 2016 were taken as the basis of forest types classification system. First, LiDAR data were preprocessed and a canopy height model (CHM) was generated using the separated point cloud data. Then, CHM was optimized using the Khosravipour algorithm and individual tree crowns were segmented by region-based hierarchical cross-section analysis; subsequently, an accuracy assessment was performed. Second, 37 features, such as height, intensity and canopy size, were extracted based on airborne LiDAR data and a total of 21 texture and spectral features were extracted based on CCD images. Feature selection was performed using the RF method. Next, two kinds of nonparametric classifiers, including RF and SVM, were used for classification by combining with the segmented image object and selected features and 12 classification schemes. Finally, 40% of the data from each tree species were randomly selected to test the overall accuracy (OA), user’s accuracy, and producer accuracy based on stratified sampling. Classification results were compared and evaluated. 【Result】 The detection accuracy of individual tree crown segmentation was over 80%, which conforms to forestry production requirements. In total, 12 features were retained using only airborne LiDAR data, 11 features were retained using only CCD images, and 11 features were retained by combining with the two datasets after RF feature selection. Then, the importance ranking of mean decrease accuracy was performed. After RF feature selection, the tree species classification results were better than those without RF feature selection. OA can be increased by an average of 3.47%. The average accuracy of combining CCD images and airborne LiDAR was increased by 6.07% compared with the average accuracy of using only CCD images. 【Conclusion】 RF feature selection could optimize features, reduce feature redundancy, and improve tree species classification accuracy. Multi-source data could also improve tree species classification accuracy. When combined with multi-source data, spectral features were the most important feature, and intensity features extracted from LiDAR data were more stable than height features. In the future, the combination of more band images and LiDAR data for different study areas will be considered and further studies will be conducted by adding more crown structure and spectral features.

Key words: light detection and ranging(LiDAR), individual tree crown delineation, random forest, feature selection, support vector machine(SVM)