Regression-based regionalisation of hydrological parameters using catchment’s spectral signature

Water resources management is difficult due to the uncertainties of the parameters controlling the hydrological response and, this uncertainty is even greater in ungauged basins where parameters are generally defined by regionalisation approaches. Among the available methods, one of the most used is the regression-based approach, which relates the most appropriate parameter values to catchment properties, such as physical properties, topographic, land use, soil and geological data. This approach assumes that the hydrological response depends on the catchment attributes and the hydrological response in catchments with similar characteristics is meant to be similar, and traditionally, these properties are derived from cartographic data sources. Since the spectral response of the territory depends on these attributes, this study uses remote sensing techniques to characterise the spectral response and apply it to the regionalisation of hydrological parameters using a machine learning approach with Random Forest.

The study area is a Mediterranean environment in Spain and corresponds to eighteen gauged watersheds in the region of Extremadura, in which we find two bioclimatic variants: a wetter and a drier one. In this study the algorithm is tested in two scenarios for regionalisation, the new approach using the spectral signature of the catchments and the results are compared with the traditional approach using the physical properties from data provided by the European Soil Data Centre. The spectral response of the catchments is studied using images from the Sentinel-1 (S1) and Sentinel-2 (S2) missions of the Copernicus Program of the European Commission. S1 is a synthetic aperture radar (SAR) sensor (C-band ) and S2 is a multispectral sensor working in the visible, near-infrared and shortwave infrared bands. In addition, several spectral indices and texture metrics derived from the grey-level cooccurrence matrix are also used for a better characterization of the watersheds.

The results perform well in both scenarios showing almost the same goodness of fit and the efficiency depends on the climatic environment. In this sense, the prediction in the wetter catchments exhibits better performance than the driest variant.  Specifically in the latter, the spectral regionalisation outperformed the physical scenario. The new spectral approach shows promising results, especially considering the advantage of having continuous coverage of Sentinel data worldwide, which offers new possibilities in areas where no mapping information is available.