| Abstract | Optical and radar remote sensing are both used to map surface water features. Since use cases range considerably in the literature between applications, a direct comparison is warranted to assess how well each perform in a wide range of geographic settings using a range of classification methods. Thus, surface water maps generated from Sentinel-1 Synthetic Aperture Radar (S1SAR) and Sentinel-2 Multispectral Instrument (S2MSI) imagery were compared across four machine learning techniques and eight diverse image areas in Canada. Additionally, the polarizations and multispectral bands were varied to understand their effect. The results were validated using high resolution satellite imagery, and analysis of variance was calculated. S2MSI consistently produced higher accuracy surface water maps compared to S1SAR. Contrary to previous understanding, the cross-polarization did not produce significantly more accurate surface water maps than like-polarization, and the same was true for dual and single polarization. The introduction of an additional band of multispectral imagery improved accuracy significantly. In flooded conditions, dual polarization produced the best results, and for the detection of ice, cross-polarization produced the best results. These findings will increase the quality and efficient generation of surface water maps for water resource management, climate change impact studies, and other disciplines. |
|---|
