Rethinking the Default: Fine-Tuning a Land Surface Model for Snow and Flood Prediction in the Missouri River Basin

When the 2019 Midwest Spring Flood tore through more than 50 levees and caused nearly 12 billion dollars in damage across the Missouri River Basin, it was a stark reminder that the models used to forecast snowmelt and runoff are not academic luxuries. They are the tools that water managers, farmers, and emergency planners rely on to anticipate when a mountain snowpack will turn into a flood wave. A new study published in Hydrology and Earth System Sciences now shows that one of the most widely used land surface models in the United States can be made substantially better at that job, but only if users stop treating its default settings as one-size-fits-all.

The model in question is Noah-MP, the Multiparameterization land surface model that serves as the hydrologic core of the US National Water Model. Noah-MP’s defining feature is its flexibility: for key physical processes such as runoff generation, surface heat exchange, frozen soil behavior, and snow and soil temperature dynamics, users can choose among several alternative parameterization schemes. In theory, this lets modelers tailor the physics to local conditions. In practice, most applications simply run the default configuration or apply a single scheme across enormous domains, because rigorous guidance on which scheme works where has been scarce.

A team led by Eunsaem Cho of NASA’s Goddard Space Flight Center and Eunsang Cho of Texas State University set out to fill that gap. They evaluated the default Noah-MP setup against four alternative schemes, each swapping in a different representation of one physical process: runoff and groundwater dynamics, the surface exchange coefficient for heat, frozen soil permeability, and the snow and soil temperature time scheme. The test bed was the Missouri River Basin, a deliberately demanding choice. The basin stretches from alpine tundra in the Rocky Mountains through semi-arid Great Plains steppe to humid subtropical lowlands, and its snow regimes range from deep montane forest snow to shallow, ephemeral prairie snow.

The evaluation covered 50 USGS Hydro-Climate Data Network subbasins, selected to have at least 25 years of continuous streamflow records and minimal dam influence, so that the gauges reflect natural hydrology rather than reservoir operations. Simulations ran from 2014 to 2023, driven by the NLDAS-2 meteorological forcing dataset at roughly 12 kilometer resolution, and were compared against USGS streamflow observations and the University of Arizona’s gridded snow water equivalent dataset, which blends station measurements from the SNOTEL and COOP networks with modeled estimates. Evapotranspiration was checked against the GLEAM dataset.

The headline result is that every alternative scheme improved runoff simulation somewhere, but each did so through a different physical pathway. Measured by the Kling-Gupta Efficiency, a standard hydrological skill score that balances correlation, bias, and variability, the runoff and groundwater experiment gained 0.19 in its best representative subbasin, the surface exchange experiment gained 0.38, the frozen soil experiment gained 0.44, and the snow temperature experiment gained 0.14. Those numbers may look modest, but in operational hydrology a shift of that size can be the difference between catching and missing a flood peak.

The mechanisms behind the gains are instructive. The default groundwater scheme ties recharge to the moisture state of the deepest soil layer alone, which flattens groundwater variability and weakens simulated baseflow. The alternative TOPMODEL scheme with an equilibrium water table instead derives the water table from the total water deficit of the entire soil profile, restoring realistic seasonal groundwater swings and improving peak runoff in years like 2014, 2018, and 2023. The surface exchange fix tackles a different flaw: the default scheme systematically overestimates evapotranspiration in vegetated and cropland areas, and the alternative Chen97 formulation, which treats thermal roughness differently, cools the surface and dials evapotranspiration back to realistic levels, leaving more water in the soil before flood events.

The frozen soil result may be the most consequential for winter hydrology. The default scheme lets frozen soils stay more permeable than they really are, so snowmelt and rain infiltrate when they should run off the surface. The alternative Koren99 scheme applies a nonlinear reduction in soil water diffusivity as ice content rises, which sharply curtails infiltration under frozen conditions. In the Bear Den Creek subbasin, this change alone lifted the skill score by 0.44 and finally allowed the model to reproduce the rapid surface runoff surges of 2019 and 2023 that the default completely missed.

Snow told a subtler story. The snow and soil temperature experiment produced the highest average annual maximum snow water equivalent of any configuration, 48.9 millimeters against the default’s 33.4, and it was by far the best at correcting the systematic underestimation of snowpack in the high-snowfall Rockies: 14 subbasins landed within 10 percent of observed snow water equivalent, compared with just four for the default. Yet better snow did not automatically mean better runoff. In subbasins where the default already overestimated April-to-July flows, the improved snowpack made the overestimate worse, while in subbasins with a pre-existing low bias the snow fix translated directly into better streamflow. The lesson is that snow accuracy and runoff accuracy are linked but not identical targets.

From all this, the team distilled practical recommendations. The runoff-groundwater and surface exchange schemes proved broadly effective across climates and land covers, with the strongest gains in hot-summer humid continental climates and croplands. The frozen soil scheme delivered targeted improvements in cold semi-arid climates, vegetation mosaics, and prairie snow regions. The snow temperature scheme, despite its clear snow benefits, failed the study’s improvement criteria for overall runoff in every classification. And notably, the default configuration remains the best choice in subarctic climates and high-elevation forest snow regimes, suggesting Noah-MP is genuinely well tuned for those environments even as it struggles elsewhere.

The authors are careful to note the limits of their findings: the analysis covers one basin and one decade, no explicit river routing was applied, and scheme interactions can be nonlinear, so combinations may behave in ways no single-scheme test predicts. Still, the message for operational hydrology is clear. With snowmelt floods growing more erratic and billions of dollars riding on seasonal water forecasts, the era of running land surface models on factory settings is ending. The physics options were always there; now there is evidence for when to use them.

Subject of Research: Evaluation of alternative parameterization schemes in the Noah-MP land surface model for snow and runoff simulation in the Missouri River Basin

Article Title: Basin-scale evaluation of the Noah-MP land surface model for runoff and snow generation in the Missouri River Basin: insights and recommendations for parameterization scheme selection

Article References: Cho, E., Cho, E., Vuyovich, C., Li, B., & Jacobs, J. M. (2026). Basin-scale evaluation of the Noah-MP land surface model for runoff and snow generation in the Missouri River Basin: insights and recommendations for parameterization scheme selection. Hydrology and Earth System Sciences, 30(20), 6359-6380. https://doi.org/10.5194/hess-30-6359-2026

Image Credits: AI Generated

DOI: 10.5194/hess-30-6359-2026

Keywords: Noah-MP, land surface model, Missouri River Basin, snowmelt, runoff, hydrology, parameterization, flood forecasting, snow water equivalent, groundwater, frozen soil, National Water Model