Optimal spectral index and threshold applied to Sentinel-2 data for extracting impervious surface: Verification across latitudes, growing seasons, approaches, and comparison to global datasets

Abstract

Many spectral indices have recently been developed for accurate extraction of impervious surfaces. Moreover, there are several 10-m global datasets available containing urban/impervious land cover class claiming to be of high accuracy. Up to date, there was no detailed analysis on the influence of easy-to-calculate spectral index and threshold on the final accuracy at large scale applied to Sentinel-2 scenes. Furthermore, the impact of growing season and the land-use type is unclear and the available global datasets must be validated in terms of their applicability for the accurate extraction of impervious surface for urban ecological applications. We show that the highest accuracy can be obtained by applying mNDVI and UCI thresholds (0.41 and −0.49 respectively) for summer median composites of Sentinel-2A/B acquisitions (highest R2>0.82 and lowest RMSE<10%) if validated against true imperviousness on the areal basis. In cases, where the number of cloud-free scenes is insufficient, an established growing season shall be used. Small artificial patches possess the highest uncertainty at this resolution, but not exceeding 20%. Spectral unmixing applied to pixels extracted using the thresholds do not significantly improve the overall estimates. Only ESA Worldcover 10-m demonstrated the comparable R2 and RMSE metrics among global datasets. Moreover, compared global datasets showed significant differences (up to tens of %) between the impervious surface estimates for selected ten cities, that highlights further evaluations of these data. Our results can successfully be implemented for mapping annual and even seasonal dynamics of imperviousness within the urban environment.

Keywords

• Imperviousness
• Urban environment
• Latitudinal transect
• Spectral unmixing
• Spectral transformation
• Global datasets