A Hierarchical Extraction Method of Impervious Surface Based on NDVI Thresholding Integrated With Multispectral and High-Resolution Remote Sensing Imageries

Abstract

Research on extraction of impervious surface has developed for decades, but it is still quite challenging to obtain impervious surface information with high accuracy, especially from multispectral remote sensing imageries. Linear spectral mixture analysis (LSMA) is a major method for estimating impervious surface areas. When LSMA is conducted, an important phenomenon is often neglected: the spectral signature is frequently interfered by the dominant component (impervious surface, vegetation, or soil) within a mixed pixel, namely the effects of mixed-pixel spectral interference. The primary objective of this study is to design a hierarchical extracting method based on normal difference vegetation index (NDVI) thresholding to eliminate these effects and improve mapping accuracy. In this method, NDVI was used and divided into different values within the range of 0.1–0.9 at intervals of 0.05. Every NDVI value was used to divide the Landsat OLI data into two segmented layers that were respectively unmixed using LSMA. The critical step was to seek out an optimal NDVI threshold by accuracy assessment, through which, a segmented layer that produced impervious surface extraction with best mapping accuracy could be identified. The results demonstrated that NDVI = 0.70 was the optimal threshold, indicating that Landsat image with NDVI ≤ 0.70 was not affected by spectral interference. Therefore, LSMA was implemented to extract impervious surface based on the Landsat image with NDVI ≤ 0.70, and the remaining area was classified by Gaofen-2 imagery to produce a complete map using object-based image analysis. This hierarchical method is effective to remove or minimize the effects of mixed-pixel spectral interference with a promising accuracy improvement of 29.3% on average. Furthermore, it provides a novel idea to enhance urban impervious surface mapping by LSMA based on NDVI thresholding.