Land cover mapping of urban environments using multispectral LiDAR data under data imbalance

Abstract

Fine scale land cover mapping of urban environments is of pronounced importance in urban planning and management. Over the past decade, many researchers have employed airborne LiDAR data for land cover mapping in urban environments. However, LiDAR point cloud data usually provide radiometric information-known as intensity-recorded at an individual spectral wavelength which may not be as informative as geometric information in classification of targets. To examine the suitability of intensity information captured at different wavelengths, the present study aimed at employing multispectral airborne LiDAR data for land cover mapping of an urban area. To this end, a dataset consisting of spectral and geometric information, both extracted from multispectral LiDAR data, was considered and a segment-based classification scheme was developed. Overall, the classification performance using multispectral airborne LiDAR data was in accordance with previous studies, mainly due to the additional spectral information. To address data imbalance, two different data sampling techniques were combined with the classifier, and either the user's or producer's accuracies for most of the classes experienced moderate to large increases through adopting under-sampling aggregation coupled with random forests classifier compared to the original training data. The best overall accuracy of 94.83% was achieved with seven classes.