Realtime Prediction in Disturbed Landscapes: Identifying Highest Priority Disturbance Characteristics Impacting Streamflow Response in a CONUS-Scale Operational Model

Abstract:

The "next generation" of hydrologic prediction systems - targeting unified, process-based, real-time prediction of the total water cycle - bring with them an increased need for real-time land surface characterization. Climatologically-derived estimates may perform well under stationary conditions, however disturbance can significantly alter hydrologic behavior and may be poorly represented by mean historical conditions. Fortunately, remote sensing and on-the-ground observation networks are collecting snapshots of these land characteristics over an increasing fraction of the globe. Given the computing constraints of operating a large-domain, real-time prediction system, to take advantage of these data streams we need a way to prioritize which landscape characteristics are most important to hydrologic prediction post-disturbance.

To address this need, we setup a model experiment over the contiguous US using the community WRF-Hydro system with the NoahMP land surface model to assess the value of incorporating various aspects of disturbed landscapes into a real-time streamflow prediction model. WRF-Hydro will serve as the initial operational model for the US National Weather Service's new national water prediction effort, so use of WRF-Hydro allows us to leverage both an existing CONUS-scale model implementation and a short research-to-operations path. We first identify USGS GAGES-II basins that experienced more than 25% forest loss between 2000 and 2013. Based on basin disturbance type, geophysical setting, and climate regime, we formulate a conceptual model of which "disturbed" landscape characteristics we expect to dominate streamflow response. We test our conceptual model using WRF-Hydro by modeling a baseline (no disturbance) case, and then bringing in empirically-derived model state shifts representing key disturbance characteristics (e.g., leaf area index, rooting depth, overland roughness, surface detention). For each state update and each basin, we quantify prediction skill improvement across a range of hydrologic timescales. Finally, we compare our model findings to our original conceptual model of dominant controls based on basin attributes. The results allow us to identify opportunities and gaps in real-time products to support post-disturbance streamflow prediction.