Feature selection methods in extracting impervious surface from Landsat TM image

Abstract

Impervious surface(s) (ISs) as an important ecological index has been widely used in the study of urban ecological environment evaluation, urban hydrology and urban vegetation mapping. In this paper, methods in feature selection, including principal component analysis (PCA) and optimum index factor (OIF), as well as indices approaches such as the normalized difference impervious surface index (NDISI) and normalized difference built-up index (NDBI) were evaluated. A new approach, experimental band combination (EBC), was proposed to extract ISs in Taiyuan city of Shanxi Province, China, using Landsat Thematic Mapper (TM) imagery. The same classification procedure was applied on original TM image, PCA image, OIF image and EBC image respectively to extract IS; the results were also compared with those extracted from the aforementioned indices approaches. Accuracy assessments were conducted on each of them using 253 randomly selected sampling points. By comparison and analysis, it was found that the EBC method obtained the highest producer’s accuracy (87.63%), user’s accuracy (90.43%) and overall accuracy (87.75%), with the Kappa statistic of 0.85 in IS extraction.