Comparing Hyperspectral and Multispectral Imagery for Land Classification of the Lower Don River, Toronto

Abstract

Urban greenspace is important for the health of cities. Up-to-date databases and information are vital to maintain and record growth in cities. Despite detailed mapping of urban land cover through high resolution imagery, medium resolution data should not be ignored. During the last decade, advances in spaceborne hyperspectral sensors have proven to be beneficial over multispectral sensors for land cover monitoring due to their increased spectral resolution. The objective of this research was to compare Earth Observing-1 (EO-1) Hyperion hyperspectral data to Landsat 5 Thematic Mapper (TM) and Satellite Probatoire d’Observation de la Terre (SPOT) 5 multispectral data for land cover classification in a dense urban landscape. For comparative analysis, aerial orthorectified imagery provided by the Toronto and Region Conservation Authority (TRCA) was used as a ground truth method for accuracy assessment. This study utilized conventional and segmented principal components (CPCA and SPCA) for data compression on the Hyperion imagery, and used principal components analysis (PCA) as a visual enhancement technique for multispectral imagery. Image processing including the generation of the normalized difference vegetation index (NDVI), and mean texture was also performed for both Landsat and SPOT sensors. An unsupervised ISODATA classification was generated on all images to produce a land cover classification map for a portion of the Lower Don River in Toronto, Ontario, Canada. Experiments conducted in this research demonstrated that hyperspectral imagery produced a higher overall accuracy (5-6% better) than multispectral data with the same resolution for defining vegetation cover. However, SPOT generated greater accuracy results than Landsat and Hyperion for vegetation classes. iii