A Comparison of LiDAR Wavelengths for the Unsupervised Classification of Trees
MetadataShow full item record
The task of species classification under monochromatic LiDAR is well documented, however to date no direct comparisons of classification power have been made between wavelengths across the same datasets. Additionally it is not clear if combinations of wavelengths can improve relative information gain. In an attempt to address these issues, this study conducted a novel classification utilising three LiDAR wavelengths commonly applied in Airborne Laser Scanning (1550, 1064 and 532nm). An unsupervised classification approach was undertaken using Multivariate Gaussian Mixture Models (MGMM) from which the results were compared using Akaike’s Information Criteria (AIC). Classification without ground data is ill posed due to the lack of absolute accuracy assessment, but this paper will argue that in the absence of ground data, unsupervised classification can offer useful information on a woodland’s class composition and on the comparative explanatory power for different wavelengths. From the AIC of model likelihoods, this study found that the models with 1064nm had the greatest classification power, closely followed by 1550nm and lastly 532nm which performed relatively poorly in comparison to the former. The 1064nm model results also pointed at suitability for use with metrics of variance such as Full-Width-Half-Maximum, a feature not exhibited by models with using 1550 or 532mn wavelengths.