Urban vegetation cover extraction from hyperspectral imagery and geographic information system spatial analysis techniques: case of Athens, Greece

Abstract

The present study aimed at evaluating the performance of two different pixel-based classifiers [spectral angle mapper (SAM) and support vector machines (SVMs)] in discriminating different land-cover classes in a typical urban setting, focusing particularly on urban vegetation cover by utilizing hyperspectral (EO-1 Hyperion) data. As a case study, the city of Athens, Greece, was used. Validation of urban vegetation predictions was based on the error matrix statistics. Additionally, the final urban vegetation cover maps were compared at a municipality level against reference urban vegetation cover estimates derived from the digitization of very high-resolution imagery. To ensure consistency and comparability of the results, the same training and validation points dataset were used to compare the different classifiers. The results showed that SVMs outperformed SAM in terms of both classification and urban vegetation cover mapping with an overall accuracy of 86.53% and Kappa coefficient 0.823, whereas for SAM classification, the accuracy statistics obtained were 75.13% and 0.673, respectively. Our results confirmed the ability of both techniques, when combined with Hyperion imagery, to extract urban vegetation cover for the case of a densely populated city with complex urban features, such as Athens. Our findings offer significant information at the local scale as regards to the presence of open green spaces in the urban environment of Athens. Such information is vital for successful infrastructure development, urban landscape planning, and improvement of urban environment. More widely, this study also contributes significantly toward an objective assessment of Hyperion in detecting and mapping urban vegetation cover.