CNN-Based Tree Classification Using Multiwavelength Airborne Polarimetric LiDAR Data

Abstract

LiDAR as a commonly used active remote sensing method has been frequently applied in fields like forestry and agriculture. Many existing studies have utilized commercial non-polarimetric LiDAR for vegetation surveying and monitoring. A multiwavelength airborne polarimetric LiDAR system (MAPL) was developed for vegetation remote sensing. The MAPL has dual-wavelength (1063-nm and 532-nm), dual-polarization (co- and cross-polarization), and full waveform recording, and hence leverages enhanced capabilities for target identification and classification. In this work, the MAPL data from five different tree types, blue spruce, ponderosa pine, Austrian pine, ash and maple, were collected. A convolutional neural network (CNN) approach was adopted to classify the trees. The numerical features, i.e., the peak reflectance intensities and the full width at half maxima (FWHMs), were extracted from the MAPL waveforms as input to the CNN model. Four different scenarios were studied, i.e., SCENARIO 1, dual-wavelength and dual-polarization; SCENARIO 2, single wavelength at 1064-nm with dual-polarization; SCENARIO 3, single wavelength at 532-nm with dual-polarization; and SCENARIO 4, dual-wavelength with co-polarization only. The study reveals that as the number of the CNN hidden layers increases, the tree-classification accuracy also improves following a logistic growth model. Furthermore, when the number of hidden layers is greater than 5, SCENARIO 1 has the highest stability (minimum deviation) and the fastest convergence. The results indicate that both the peak intensity and FWHMs of the MAPL waveforms are potent features for deep learning to classify trees, and CNN is an effective tree classification method in this case. The methodology can be extended to other agricultural and forestry remote sensing applications.