Regionalization of Statistical Forecasts of Field Scale Resolution Runoff Modeling using National Water Model Outputs through Unsupervised Cluster Analysis

Abstract:

The National Oceanic and Atmospheric Administration’s (NOAA) National Water Model (NWM) is a state-of-the-art high resolution distributed model of hydrologic processes across the continental United States developed to simulate current and projected land surface flows and provide streamflow forecasting guidance to the National Weather Service. Operationalization of the NWM model offers new opportunities to improve tools for predicting agricultural runoff risk. However, the NWM is calibrated with the objective of flow forecasting and the resolution is coarser than agricultural field scale, resulting in the need to develop new methods for postprocessing NWM output for this application. A joint initiative between NOAA and the Cooperative Institute for Great Lakes Research (CIGLR) funded by the Great Lakes Restoration Initiative (GLRI) is developing a statistical model to postprocess NWM outputs to improve runoff risk forecast tools used in agricultural applications. The statistical model is tuned using relatively sparse edge-of-field (EOF) runoff observations, requiring regionalization approaches. Due to the sparsity of EOF sites, watersheds with similar hydrologic responses must be identified so the statistical model can be applied on a regional scale to areas without EOF measurements. This study implements principle component analysis (PCA) and K-means unsupervised clustering techniques to regionalize the development of the statistical model. The PCA and K-means cluster analysis is applied to average annual values of select NWM output variables that are spatially aggregated across HUC-10 watersheds. The novelty of this approach lies in the use of NWM modelled variables which provide quantification of hydrologic processes that don’t allow for observational data measurements. Based on the causality analysis, we identify 25 variables from NWM which can have causal influence on the daily EOF runoff. The clusters identified in the analysis form the regional groupings of watersheds with similar runoff responses and allows development of the statistical model over larger domains. Ultimately this will lead to improved representation of runoff for the purpose of runoff risk prediction.