Urban vegetation mapping by airborne hyperspetral imagery; feasibility and limitations

Fast urbanization requires complex management of green spaces inside districts and all around the cities. In this context, the use of high-resolution imagery could give a fast overview of species distribution in the considered study zone, and could even permit species recognition by taking advantage of high spectral resolution (i.e. superspectral/hyperspectral imagery). In this study, we aim to explore the feasibility of eight vegetation species recognition inside Kaunas city (Lithuania). The goal is to determine the potential of metric/centimetric spatial resolution imagery with less than hundred bands and a limited spectral interval (e.g. Vis-NIR), to be able to recognize urban vegetation species. The ground truth samples were also limited for some of the considered species. The method included pre-treatments based on vegetation masking and feature selection using Minimum Noise Fraction (MNF). Support Vector Machine (based classifier) showed encouraging performance over Spectral Angle Mapper (SAM), the accuracies were not notably high in term of statistical analysis (i.e. up to 46% of overall accuracy) but the visual inspection showed coherent distribution of the detected species.