A novel framework for the selection of spectral input to pixel-based river discharge estimation model using Sentinel-2 imagery

While access to discharge data is key to hydrologic studies, it is a serious obstacle in ungauged basins. Currently, Sentinel-2 imagery at high spatiotemporal resolution offers a unique opportunity to infer the relation between pixel-based discharge rate and surface reflectance. One promising approach in this respect has been to find the complex relationship between river discharge and the spectral ratio between two benchmark pixels, namely the wet and dry pixels, whose dynamics resembles river discharge variation. However, this has been challenging due to the adverse impact of soil moisture and mixed land cover on the spectral behavior of the dry pixel. The selection of the wet pixel must also guarantee sufficient sensitivity of its spectral response to water depth fluctuations. To tackle the above issues, in this study, we developed a novel framework that automatizes the selection of the wet and dry pixels by using Sentinel-2 imagery. We also introduced the Normalized Difference Discharge Index (NDDI), as the best band combination, to predict river discharge. We used linear regression with leave-one-out cross-validation as the prediction model, which leverages limited satellite data due to the cloud cover. By implementing the developed framework at multiple gauged points across the continental United States, the best location of the dry pixel was consistently found in urban pixels whose longwave reflectance fall within a certain range. By analyzing the pixel-wised correlation coefficient between surface reflectance at NIR band and river discharge across the studied river widths, we found that the best wet pixels are located along river banks with shallow water depth. These pixels were characterized by the average reflectance higher than the 98th percentile in the green band. Finally, by testing over 4000 band combinations as input to the river discharge prediction model, we found that the normalized difference between B11 and B4 for the wet pixel, as well as the B11 ratioing between the dry and wet pixels yielded the most accurate predictions with R2 = 0.88 and R2 = 0.73, respectively.