Stand species identification using GF-2 imagery and airborne LiDAR data based on SVM classifier

Abstract

Stand species recognition is one of the key topics in forest research. The new GaoFen-2 (GF-2) satellite imagery contains detailed spectral and texture features. LiDAR data can provide abundant forest structure information. Using deep learning algorithms to identify forest types from optical and LiDAR data can improve recognition accuracy effectively. In this paper, the GaoFeng forest farm of Guangxi Zhuang Autonomous Region is taken as the experimental area, the mean and variance of GF-2 red-green-blue triple band spectra are taken as the spectral characteristics, and the corner two matrix, entropy, contrast and correlation in the gray level co-occurrence matrix are taken as the texture characteristics, a canopy height model (CHM) derived from LiDAR data is built as the structural characteristics, comparing the classification accuracy of using spectral feature, texture feature and structure information or spectral feature, texture feature based on SVM deep learning algorithm to classify in different resolution and different classification levels. The results show that the accuracy of the specific tree species classification by SVM classifier using 2 m-resolution fused GF-2 image combined with spectral, texture and structure features is high, the total classification accuracy is 95.65%, and Kappa coefficient reaches 0.93. As a contrast, the results of 4m-resolution fused GF-2 image using spectral and texture feature classification are poor. Combining the spectral and texture information of forest and the structural information reflected by LiDAR point cloud data can achieve significantly better classification results than each single feature. Compared with several traditional supervised classifiers, the new method needs less specialization and is more beneficial to the large-scale promotion of forestry classification.