Supervised and unsupervised machine learning approaches for tree classification using multiwavelength airborne polarimetric LiDAR

As an active remote sensing tool, Light Detection And Ranging (LiDAR) with laser source offers many advantages over passive and radar remote sensing, enabling a wide range of applications across fields such as forestry, agriculture, urban planning, and environment. Current studies have mainly employed commercial non-polarimetric LiDAR for forest surveying and monitoring using point cloud data. A Multiwavelength Airborne Polarimetric LiDAR (MAPL) has been developed. The MAPL system has dual-wavelengths and dual-polarization, and offers full waveform recording capability. With its unique characteristics, it has been used in vegetation remote sensing for better target detection and identification. In this work, field data were collected from five different tree species, including deciduous trees (ash and maple) and coniferous trees (Austrian pine, ponderosa pine, and blue spruce). Subsequently, supervised (Decision-Tree) and unsupervised (clustering) machine learning (ML) methods were developed for tree species classification, based on the peak intensities and full width at half maxima (FWHMs) of the MAPL waveforms. The Decision-Tree approach shows a re-substitution error of 0.14 % and a k-fold loss error of 0.57 % for 2,106 tree samples; and the clustering methods provide accuracies at about 80 %. Furthermore, the results indicate that both peak intensities and FWHMs of the MAPL waveforms are potent features for the ML approaches. In addition, the supervised method has higher accuracy, while clustering is less labor intense and may be applied to large scale remote sensing. The adopted methods enable expeditious and accurate data processing and can be expanded to other classification applications.