CNN-Based Tree Species Classification Using Airborne Lidar Data and High-Resolution Satellite Image

Abstract

Spatial information of tree species composition of forest and urban vegetation is very important for forest protection and urban management. Tree species classification using remote sensing data is mainly conducted using such classification methods as SVM (Support Vector Machine) and Random Forest. Images used include multispectral/hyperspectral images, LiDAR (Light Detection And Ranging) data, or the combination of them. As a fast-growing and powerful tool having obtained state-of-the-art results in many remote sensing applications, Deep learning (DL) has a great potential in outperforming these existing classification methods and obtaining more accurate classification maps. In this work, three CNN models were employed for tree species classification at individual tree level, using the combination of high-resolution multispectral image (i.e., WorldView-2) and LiDAR data. Compared the traditional object-based classification using random forest (RF) and support vector machine (SVM), the 18-layer ResNet and the 40-layer DenseNet provided significant higher accuracies. The experimental results indicate the advantages of the two CNN models used for tree species classification.