High-resolution global water body datasets underestimate the extent of small rivers

Abstract

ABSTRACT Several 30-m high-resolution global water body datasets have been developed in recent years and raised prospects for studying small (<300 m wide) rivers. River mapping accuracy is a key factor in determining the applicability and usefulness of a certain water body dataset, and accuracy assessment commonly requires numerous water validation points. However, in practice, the number of water validation points is limited, and their spatial distribution is not sufficiently representative (most points are located in lakes, coastal areas, or large rivers). As a result, although current water body datasets perform well to represent global surface water distribution, their ability to map small rivers remains poorly quantified. In this study, we propose that stable and comprehensive river centerlines derived from the Global River Widths from Landsat (GRWL) dataset can be used to generate sufficient and representative validation points, and to estimate the river mapping producer’s accuracies. Six 30-m high-resolution, state-of-the-art global water body datasets were evaluated, including four global water body datasets (G1WBM, GIW, GSW, and GSWD) and two land-cover datasets (GlobeLand30 and FROM-GLC). The results show that the small river mapping producer’s accuracies (0.51–0.94) estimated for these datasets were considerably lower than their surface water producers’ accuracy (0.86–0.99). The estimated river mapping producer’s accuracies increase with river width, suggesting that these datasets perform better in mapping large rivers than small rivers. When river width increases from 90 m to 300 m (and above), the average river mapping producer’s accuracy of the six water body datasets increases from 0.79 ± 0.12 to 0.94 ± 0.03. Our analysis urges caution when using current global water datasets to study rivers, particularly small rivers narrower than 300 m.