Catchment scale runoff time-series generation and validation using statistical models for the Continental United States

We developed statistical models to generate runoff time-series at National Hydrography Dataset Plus Version 2 (NHDPlusV2) catchment scale for the Continental United States (CONUS). The models use Normalized Difference Vegetation Index (NDVI) based Curve Number (CN) to generate initial runoff time-series which then is corrected using statistical models to improve accuracy. We used the North American Land Data Assimilation System 2 (NLDAS-2) catchment scale runoff time-series as the reference data for model training and validation. We used 17 years of 16-day, 250-m resolution NDVI data as a proxy for hydrologic conditions during a representative year to calculate 23 NDVI based-CN (NDVI-CN) values for each of 2.65 million NHDPlusV2 catchments for the Contiguous U.S. To maximize predictive accuracy while avoiding optimistically biased model validation results, we developed a spatio-temporal cross-validation framework for estimating, selecting, and validating the statistical correction models. We found that in many of the physiographic sections comprising CONUS, even simple linear regression models were highly effective at correcting NDVI-CN runoff to achieve Nash-Sutcliffe Efficiency values above 0.5. However, all models showed poor performance in physiographic sections that experience significant snow accumulation.