High Resolution Impervious Surface Estimation

Abstract

Recent studies have attempted to extract impervious surfaces from high-resolution satellite imagery such as Ikonos and QuickBird. These images, however, often lack necessary spectral information due to technological limitations. This study integrates spectral information (temperature and moisture) derived from Landsat-7 ETM+ imagery with Ikonos imagery to derive high-resolution impervious surface information. Furthermore, three popular methods, including linear regression modeling, artificial neural network, and regression tree have been developed and compared using a paired t-test statistic. Analysis of results reveal that Tasseled Cap components particularly greenness and wetness of Ikonos imagery are most important in estimating sub-pixel imperviousness. Also, to some extent the brightness temperature derived from Landsat-7 ETM+ image helps in better estimation of impervious surfaces. Moreover, a comparative analysis indicates that the non-linear approaches yielded statistically better results. Particularly, the regression tree model generated best results with highest Pearson’s r (0.939) and lowest mean absolute error (8.307).