Noah-MP is a widely used, community-driven model designed to simulate the coupled energy, water, and carbon cycles at the land surface (Niu et al., 2011; Yang et al., 2011). Developed as an enhancement to the original Noah LSM, Noah-MP incorporates multiple parameterization options for vegetation processes, canopy radiative transfer, multi-layer snow physics, soil thermal and moisture dynamics, and groundwater interactions (Niu et al., 2011; Yang et al., 2011; He et al., 2023). The availability of multiple parameterization options within Noah-MP is particularly valuable for regional applications, because it allows researchers and practitioners to optimize model performance by selecting the most appropriate physical representations for specific geographical and climatological conditions (Hong et al., 2014; Ma et al., 2017; He et al., 2023).

In this study, Noah-MP 4.0.1 was forced with meteorological data from NLDAS-2. NLDAS-2 provides high-quality, observation-based atmospheric forcing at hourly temporal resolution and 0.125° (∼12km) spatial resolution, spanning from 1979 to the present (Xia et al., 2012a). The forcing dataset includes hourly integrated precipitation, 2 m air temperature and specific humidity, incoming shortwave and longwave radiation, 10 m zonal and meridional wind speed, and surface pressure, making it well suited for driving land surface and hydrological models (Xia et al., 2012a, 2014; Nearing et al., 2016). NLDAS-2 has been extensively validated and is widely used in hydrological and climate studies across North America, supporting applications such as streamflow simulation, drought monitoring, and water resource management (Xia et al., 2012a, 2016; Hao et al., 2016; Pascolini-Campbell and Reager, 2024).

Initial conditions for all experiments were established through an experiment-specific spin-up procedure. Each parameterization experiment (default and all four alternatives) was independently initialized from 1 January 1980, and run through 31 December 2013, using NLDAS-2 forcing data with its respective parameterization scheme settings. The model states at the end of 2013 from individual spin-ups were used as initial conditions for the subsequent 2014–2023 analysis, ensuring that the model state entering the analysis period is fully consistent with each experiment's parameterization configuration.

To assess Noah-MP's sensitivity to parameterization schemes, this study evaluates four alternative schemes governing runoff and groundwater generation, surface energy exchange, frozen soil permeability, and subsurface thermal dynamics (Table 1). These schemes were selected based on their perceived influence on snowmelt partitioning, groundwater interactions, and flood generation in prior sensitivity studies (Niu et al., 2011; You et al., 2020; Li et al., 2022; Yang et al., 2023). In the following, we briefly discuss these four schemes and their impact on snow and runoff simulations.

In this study, Kling–Gupta Efficiency (KGE) is employed as the primary performance metric to evaluate Noah-MP hydrological simulation. The KGE metric was developed to address limitations of traditional performance measures such as Nash–Sutcliffe Efficiency and has been widely adopted for calibration and evaluation of hydrological models (Gupta et al., 2009; Liu et al., 2022; Cinkus et al., 2023). KGE is calculated as: 7KGE=1-(r-1)2+(β-1)2+(γ-1)2 where r represents the linear correlation coefficient between observed and simulated values, β is the ratio of simulated to observed means, and γ is the ratio of simulated to observed standard deviations. KGE ranges from negative infinity to 1, with values close to 1 indicating better model performance. Knoben et al. (2019) showed that KGE=-0.414 represents the performance of prediction using the observation mean.

To evaluate the influence of parameterization schemes, a water balance analysis was conducted over the entire Missouri River Basin domain. All 0.125° resolution grid cells within the basin were considered. The analysis was performed using daily values from 2014–2023, with water years defined as 1 October through 30 September of the following year. For each water year, the following components were computed: total precipitation (P), total evapotranspiration (ET), annual maximum SWE (SWE), change in groundwater storage (ΔGW; groundwater storage at the final day minus the first day of the year), change in soil moisture storage (ΔSM; soil moisture at final day minus initial day), surface runoff (Qs), subsurface runoff (Qsb), and total runoff (Qtotal; surface runoff + subsurface runoff). These variables were aggregated spatially across the basin and temporally averaged (sum for fluxes) over a 10 year period to represent long-term hydrologic behavior under each parameterization scheme.

Table 2 summarizes the 10 year mean values of each water balance component across the five parameterization schemes evaluated. Mean annual precipitation for the five experiments was 534.0 mm, of which 464–480 mm (87 %–90 %) was lost through evapotranspiration and 64–80 mm (12 %–15 %) was routed as total runoff. Variability among experiments was mainly expressed in the ET and runoff, with ET differences of up to 16 mm balanced by corresponding changes in runoff. For all experiments, the water-budget residual – defined as Residual=P-ET-ΔGW-ΔSM-Qtotal – was below 1 %, indicating tight water-balance closure in the simulations.

Table 2 demonstrates how key hydrological components differed by Noah-MP model parameterization scheme. The RUN experiment, which assumes an equilibrium water table, produced the largest increase in groundwater storage (ΔGW=+2.8mm). The SFC experiment resulted in the lowest evapotranspiration (ET=463.6mm), because its alternative scheme limits latent heat flux. This reduction in ET contributed to the highest total runoff (Qtotal=80.2mm), largely through increased subsurface runoff. In contrast, the INF experiment yielded the highest surface runoff (Qs=27.9mm), likely reflecting reduced infiltration capacity under frozen soil conditions. The increased surface runoff was partially offset by lower subsurface flow, maintaining a total runoff value comparable to other schemes. The STC experiment produced higher surface, subsurface and total runoff than the default configuration, reflecting altered snow dynamics under its alternative snow/soil temperature time scheme. Notably, the STC experiment produced the highest average annual maximum SWE (48.9 mm) among all schemes, substantially exceeding the default value (33.4 mm). This difference reflects the altered snow accumulation and melt timing under the implicit snow/soil temperature time scheme, which delays snowmelt and prolongs snowpack retention. The enhanced snow storage under the STC scheme is consistent with the higher surface, subsurface, and total runoff relative to the default.

Runoff simulation performance was evaluated by comparing monthly Noah-MP total runoff with observed streamflow at the 50 USGS HCDNs (Fig. 2). Overall, runoff simulation performed well in high-elevation area in west of the Rocky Mountains and near the basin outlet in the southeastern portion of the Missouri River Basin. However, sensitivity to parameterization varies by location. For example, the SFC experiment provided relatively higher KGE in lowland subbasins compared to the default experiment (Fig. 2c). The RUN experiment also shows improved performance in lowland areas as well as in high-elevation subbasins in the western part of the domain (Fig. 2b). The INF experiment produced better simulation results in the northern part of the Missouri River Basin, where grassland dominates (Fig. 2d).

To highlight these differences, we generated ΔKGE maps that display the improvement in KGE for each experiment relative to the default (Fig. 3). The location of improvements differed by experiment. The RUN experiment demonstrated widespread performance improvements across most of the study domain (Fig. 3a). However, the effectiveness of this experiment is diminished in the subbasins located near the outlet of the Missouri River Basin, which is located at the bottom right of the domain. The SFC experiment performance improvements were concentrated in the lower subbasin regions (Fig. 3b). The INF experiment did not exhibit a clear spatial pattern of improvement, with relatively better performance observed in the northern part of Missouri River Basin (Fig. 3c). Finally, the STC experiment showed moderate improvements in subbasins near the outlet of the domain, while substantial improvements were limited (Fig. 3d). In terms of the fraction of subbasins showing improvement, the RUN experiment improved 58 % of subbasins (29 of 50), the SFC experiment improved 64 % (32 of 50), the INF experiment improved 64 % (32 of 50), and the STC experiment improved 50 % (25 of 50). “No Improvement” refers to subbasins where ΔKGE is zero or negative, encompassing both no change and performance degradation relative to the default. The absolute KGE values for all five schemes across each HCDN subbasin are provided in Fig. S1 in the Supplement, allowing direct comparison of scheme performance at the individual subbasin level.

To assess how alternative parameterization schemes affect hydrological processes, a comprehensive evaluation of hydrometeorological variables was conducted across multiple representative subbasins. Figure 4 shows precipitation, evapotranspiration, groundwater storage, soil moisture, SWE and runoff time series for four selected subbasins. These subbasins were selected because they showed notable improvements in KGE values and exhibit process-specific response for each experiment. Their geographic locations are shown in Fig. 1f. The selected subbasins represent diverse hydroclimatic settings: Battle Creek Keystone Basin (warm-summer humid continental, evergreen needleleaf forest, montane forest snow), Gasconade River Basin (humid subtropical, grassland, ephemeral snow), Bear Den Creek Basin (warm-summer humid continental, vegetation mosaic, prairie snow), and Dinwoody Creek Basin (subarctic, shrubland, tundra snow). For comparison with USGS streamflow observations, simulated runoff at each 0.125° grid cell was aggregated by computing the area average of all grid cells within the basin boundary. Soil moisture reported throughout this study (e.g., Fig. 4) represents the full 2.0 m soil column, integrated across all four Noah-MP soil layers.

For the subbasins shown in Fig. 4, the time series comparisons demonstrate how the alternative schemes address limitations of the default setting. In Battle Creek Keystone (Fig. 4a), the default shows limited groundwater variability. The RUN experiment produces clear seasonal and interannual groundwater variability which in turn improved simulation of peak runoff in 2014, 2018, and 2023 and increases KGE by 0.19. The 2019 peak was still lower than observed but is more accurately represented than in the default scheme. The RUN experiment also shows that the improved groundwater storage dynamics influenced soil moisture variability, with more pronounced wet and dry cycles compared to the default. In the Gasconade River (Fig. 4b), the SFC experiment reduces evapotranspiration relative to the default and raises KGE by 0.38. It corrects the systematic underestimation of peak runoff from 2014–2023, including the largest flood event in 2017. Comparison with GLEAM evapotranspiration confirms that the default scheme systematically overestimates ET in this subbasin, while the SFC experiment produces more realistic ET estimates. The reduced evapotranspiration in the SFC experiment led to increases in soil moisture, contributing to wet conditions prior to flood events. In Bear Den Creek (Fig. 4c), the default fails to capture the rapid increases in surface runoff, while the INF experiment resolves this limitation. It improves KGE by 0.44 with better simulation of the 2019 and 2023 peaks. In Dinwoody Creek (Fig. 4d), the STC experiment enhances SWE simulation and produces a KGE improvement of 0.14. The alternative SWE aligns more closely with UA SWE and results in more realistic runoff, particularly during the 2015, 2019, and 2023 floods. The delayed snowmelt under the STC experiment produces a later but more sustained infiltration signal, resulting in distinct seasonal soil moisture patterns relative to the default. Collectively, these results show how each experiment resolves distinct weaknesses of the default and improves runoff representation across diverse hydroclimatic conditions.

Figure 5 presents the KGE differences between the default and alternative parameterization schemes, illustrating how model performance varies with basin characteristics such as climate, land cover, and snow type. The RUN and SFC experiments showed broad effectiveness across multiple climate and land cover classes, with particularly notable improvements in Hot-Summer Humid Continental climates (0.18 and 0.28, respectively) and Cropland areas (0.22 and 0.24, respectively). The INF experiment demonstrated targeted improvements in Cold Semi-Arid climates (0.14), Vegetation Mosaic land cover (0.34), and Prairie Snow class (0.23), while the STC experiment provided only modest improvements across most classifications. In contrast, the default scheme performed best in Subarctic climates, Evergreen Needleleaf Forest, and high-elevation snow classes including Montane Forest and Boreal Forest.

While previous results focused on overall runoff evaluation, this section examines the snow component of the Noah-MP model and assesses how alternative schemes impact SWE and snowmelt-driven runoff simulation. The analysis focused on the maximum SWE for each water year and total runoff from April–July (Q), as these metrics are important to snowmelt-driven runoff generation (Lapides et al., 2022; Gottlieb and Mankin, 2024). Relative bias was calculated as the difference between the average of simulated values and the average of observed values (simulated minus observed), divided by the average of observed values.

Figure 6 shows how the relative bias of SWE varies among subbasins under different parameterization schemes. To examine whether these results relate to SWE variability, we calculated the mean annual maximum SWE for each HUC8 subbasins in Missouri River Basin over the 2014–2023 period (Fig. S2). Overall, the default, RUN, SFC, and INF experiments produced similar spatial patterns, with systematic underestimation of SWE across most high-snowfall regions in the Rocky Mountains (Fig. S2). In contrast, the STC experiment effectively mitigated this underestimation issue. The number of subbasins with relative bias within ±10% (indicated by bold boundaries in Fig. 6) was notably higher for the STC experiment, with 14 subbasins, compared with the default, RUN, SFC, and INF experiments, which had four, four, eight, and six subbasins, respectively. This indicates that while most schemes perform comparably across the basin domain, the STC experiment provides an advantage in snow-dominated subbasins where annual SWE is high and snowmelt-driven runoff is the dominant source of streamflow.

Theoretically, improved SWE representation should lead to better runoff simulation. However, we found that SWE improvements do not universally translate into improved April–July runoff across all subbasins. To investigate this, we identified subbasins where both SWE and April–July runoff improved (blue circles) and those where SWE improved but runoff did not (red circles) under STC experiment. As shown in Fig. 7, both groups are spatially distributed across the Missouri River Basin without a clear regional pattern, suggesting that the translation of SWE improvements into runoff changes is mediated by multiple hydrological factors beyond SWE alone. This spatial distribution motivates a closer examination of the direction of runoff bias under the default scheme.

Figure 8 shows the April–July runoff relative bias for the Default and STC experiment across SWE-improved subbasins, grouped by runoff improvement status. In subbasins where both SWE and runoff improved (Fig. 8a), the default scheme predominantly underestimated runoff, and the STC experiment effectively reduced this negative bias. In contrast, the subbasins where SWE improved but runoff did not, the default scheme already overestimated runoff in most cases, and the STC scheme further exacerbated this positive bias. These results suggest that the effectiveness of the STC scheme in improving runoff simulation depends critically on the direction of the pre-existing runoff bias in the default configuration, rather than on the degree of SWE improvement alone.

This study evaluated the performance of four alternative parameterization schemes in comparison with the default Noah-MP configuration to assess their impacts on runoff generation and other hydrological processes. The analysis revealed three limitations in the default parameterization scheme that were better addressed by the alternative schemes.

The first area for improvement was the limited groundwater dynamics in the default groundwater scheme. Groundwater simulated by the default scheme does not always exhibit strong seasonality, with minimal variability throughout the year (Fig. 4a). In the default scheme, groundwater recharge is governed by the hydraulic conductivity and matric potential of the lowest soil layer. This makes groundwater storage dynamics sensitive to the moisture state of the deepest soil layer only. Overestimation of evapotranspiration depletes moisture in the lower layers, suppressing recharge and dampening groundwater storage variability (Li et al., 2021; Cai et al., 2014). Maina et al. (2025) addressed a related issue by integrating the ParFlow model with Noah-MP. In this study, the alternative TOPMODEL with equilibrium water table scheme overcame this limitation by determining the water table depth from the total water deficit of the entire soil profile.

The second area for improvement involved systematic overestimation of ET in vegetated areas, particularly those dominated by cropland. Previous research has demonstrated that the default Noah-MP crop parameterization does not fully represent dynamic cropland processes, which can contribute to overestimation of ET (Liu et al., 2016). Furthermore, Noah-MP's consistent overestimation of gross primary productivity in vegetated areas exacerbates this issue by elevating ET in both cropland and grassland environments (Ma et al., 2017). The alternative surface exchange coefficient scheme addresses this limitation by using a different function for calculating surface exchange coefficients for heat transfer. Unlike the default scheme, the alternative approach excludes zero-displacement height considerations and dynamically calculates thermal roughness length. This results in ET estimates that better align with observed values by enhancing the efficiency of heat and moisture exchange between the land surface and atmosphere.

A third area for improvement relates to winter-season surface runoff generation under frozen-soil conditions. The default Noah-MP configuration does not fully capture the physical processes associated with frozen and partially frozen soils, including ice lens formation, changes in snowpack thermal conductivity, and alterations in water retention capacity. The default frozen soil permeability scheme tends to predict more permeable frozen soils than typically observed in natural systems, particularly in conditions where ice layers inhibit infiltration (Koren et al., 1999; Niu and Yang, 2006). The alternative scheme overcomes this limitation through a nonlinear weighting function based on volumetric ice content, which reduces soil water diffusivity quickly as ice content increases. This reduced diffusivity effectively limits infiltration capacity under frozen soil conditions, causing a greater fraction of precipitation and snowmelt to be partitioned into surface runoff rather than subsurface flow. This is particularly relevant for improving winter-season runoff estimates, which are often underestimated by the default scheme due to excessive infiltration under frozen soil conditions. A more detailed, event-based evaluation of rain-on-snow processes specifically is left for future work.

It is important to note, however, that the identified areas for improvement in the default scheme, and the gains offered by alternative schemes, reflect the specific hydroclimatic conditions of the Missouri River Basin. These findings may not be directly transferable to regions with substantially different climate regimes, land cover characteristics, or snow dynamics, as the performance of land surface parameterization schemes is known to be region-dependent (Gan et al., 2019; You et al., 2020; Li et al., 2022; Hu et al., 2023; Chang et al., 2020; Hamitouche et al., 2025). A more detailed discussion of these limitations is provided in Sect. 6.4.

Two key criteria were used to recommend alternative parameterization schemes (Table 3). First, at least 50 % of the subbasins within each environmental classification had to show improvement relative to the default to ensure sufficient representation. Second, the experiment needed to demonstrate an average KGE improvement of at least 0.1 within that classification. The distribution of the 50 HCDN subbasins across these classifications is summarized in Fig. S3. It should be noted that classifications represented by relatively few subbasins warrant more cautious interpretation.

The recommendations summarized in Table 3 show that the RUN and SFC experiments are broadly effective across climate, land cover, and snow classifications. The INF experiment shows more limited but targeted improvements, mainly in Warm-Summer Humid Continental, Vegetation Mosaic and Prairie Snow classifications. In contrast, the STC experiment does not meet the criteria in any classification, reflecting its limited overall runoff improvement discussed in Sect. 5.4. The default is retained for Subarctic climates as well as Evergreen Needleleaf Forest, Shrublands, Montane Forest, Boreal Forest, and Tundra environments. Among these, the forest-related classifications (Evergreen Needleleaf, Montane, and Boreal Forest) may benefit from alternative canopy-related parameterizations such as interception or dynamic vegetation, which none of the four experiments targeted.

The effects of the recommended parameterization schemes in this study are consistent with previous research. For example, Hamitouche et al. (2025) showed that the TOPMODEL scheme with an equilibrium water table (RUN experiment) can improve model performance by reducing runoff underestimation and yielding a higher KGE than the default setting on a global scale. Chang et al. (2020) reported that the Chen97 surface exchange coefficient scheme (SFC experiment) enhances latent and sensible heat flux simulations, resulting in more accurate evapotranspiration estimates than the Monin–Obukhov scheme over a subtropical forest in southern China. Improvements related to frozen soil processes were also highlighted by Hu et al. (2023), who showed that Koren99's lower soil permeability (INF experiment) can better represent frozen soil hydrology by capturing stronger soil moisture variability and producing more realistic runoff responses in permafrost regions of the Tibetan Plateau.

An insight from this study is that improvements in overall hydrological performance do not always coincide with better representation of snow processes in Noah-MP. The STC experiment exemplifies this complexity, demonstrating limited improvements in overall runoff simulation (Fig. 3d) while achieving the most substantial enhancements in SWE representation (Fig. 6e). As demonstrated in Figs. 7 and 8, this disconnect is primarily governed by the direction of the pre-existing runoff bias in the default scheme, whereby SWE improvements translated into better runoff in most of the subbasins where the default configuration underestimated runoff. This finding suggests that the STC scheme's impact on runoff is conditional rather than universal, as its influence on snow accumulation and melt processes does not propagate consistently through the broader suite of hydrological processes that govern total runoff. Similarly, Sthapit et al. (2022) reported that although snowpack dynamics strongly influence runoff and soil moisture, mismatches and temporal lags in the simulated response indicate that improvements in snow representation do not always translate into better streamflow performance.

These findings highlight the importance of considering application needs when selecting a parameterization scheme. In snow-dominated basins where water supply forecasting, flood risk assessment, or ecological management depends on snowpack dynamics, the selection of parameterization schemes should prioritize snow process accuracy even when overall runoff metrics show modest improvements. In basins where snow plays a minor role, however, total runoff metrics may remain the most relevant criterion.

While this study advances understanding of parameterization selection in the Noah-MP model, several limitations should be acknowledged regarding the generalizability of our findings. The current analysis is confined to the Missouri River Basin for the period from 2014–2023, which may limit the broader applicability of the recommended parameterization scheme. The literature consistently shows that the performance of land surface parameterization schemes varies considerably across different climatic and geographic settings (Gan et al., 2019; Li et al., 2022). For example, Hu et al. (2023) showed that the Koren99 frozen soil scheme (INF experiment) is particularly effective in permafrost regions of the Tibetan Plateau. Chang et al. (2020) found that the Chen97 surface exchange coefficient scheme (SFC experiment) improved ET simulations over subtropical forests in South China. Hamitouche et al. (2025) further documented substantial global variability in the performance of the TOPMODEL equilibrium water table scheme (RUN experiment). Taken together, these findings underscore that scheme recommendations derived from one region should be applied with caution elsewhere. Further validation across diverse geographical settings is therefore necessary to establish the robustness of the alternative parameterization approaches identified in this study.

This study focused on a subset of runoff and snow related parameters within the Noah-MP model. Future work should extend the analysis to include additional snow-related processes, such as snow thermal conductivity and snow cover fraction parameterizations. Snow thermal conductivity regulates the transfer of energy through the snowpack and thereby influences the timing of snowmelt, while snow cover fraction controls the representation of sub-grid snow heterogeneity and affects surface energy balance (Yang et al., 2023; Cao et al., 2024; Abolafia-Rosenzweig et al., 2025). Incorporating these processes would enable a more comprehensive assessment of Noah-MP sensitivity to snow dynamics.

An additional limitation of this study is that no river routing scheme was applied when aggregating grid-cell runoff to each basin. Instead, runoff was computed as an area average within each basin boundary. This approach follows established practice in continental-scale land surface model evaluation, where gridded runoff has been area-averaged and compared against USGS streamflow over 961 small basins ranging from 23–10 000 km2 without explicit channel routing (Xia et al., 2012b). Routing effects on simulated streamflow are also known to diminish at monthly and annual timescales relative to sub-daily and daily resolutions, and to be more pronounced for larger contributing areas (Cortés-Salazar et al., 2023). Because this study evaluates runoff at a monthly resolution across relatively small, HUC-10-scale basins, the effect of omitting explicit channel routing is expected to be limited. Future studies could incorporate explicit routing to better capture the timing of peak flows, particularly for larger or more topographically complex basins.

The Noah-MP model offers multiple physics parameterization schemes across various land surface processes, where interactions among different schemes can substantially affect model performance. Such interactions can be complex and non-linear: changes in ET, soil moisture, or snowmelt driven by one scheme may amplify or dampen the response of another, and compensating errors between schemes can be disrupted when any one of them is changed (Cuntz et al., 2016). Prior multi-physics ensemble studies have reported that certain scheme combinations produce emergent behaviors not predictable from individual scheme evaluations alone (Hong et al., 2014; Gan et al., 2019). Future investigations should implement methodologies that evaluate different combinations of physics schemes, thereby delivering a more comprehensive understanding of model uncertainty and overall performance.

As an alternative strategy to reduce the complexity arising from interactions among multiple physical processes, future studies could adopt a process-oriented evaluation framework that systematically compares multiple parameterization schemes within a single physical process. This approach would help isolate process-specific uncertainties while identifying parameterizations that provide the greatest performance improvements. For example, the current frozen-soil parameterization in Noah-MP does not fully account for preferential flow through soil macropores, which may contribute to biases in infiltration and runoff generation during frozen conditions. Agnihotri et al. (2023) evaluated combinations of supercooled liquid water and soil hydraulic property schemes in Noah-MP over the Mississippi River Basin and demonstrated that representing higher frozen-soil permeability to better account for preferential flow significantly improved spring streamflow simulations. In addition, Abolafia-Rosenzweig et al. (2026) tested 256 combinations of Noah-MP physics options across 790 SNOTEL stations in the western United States and identified an updated configuration that reduced both snow accumulation and ablation rate errors relative to the default. Such process-focused evaluations could be extended to other land surface processes to further improve the physical realism and predictive capability of Noah-MP.

In addition to these parameterization schemes, other improvements in model physics may help enhance model performance. This study revealed several key areas where process-based enhancements could address current model limitations, including the underestimation of evapotranspiration and SWE, along with the misrepresentation of rapid surface runoff generation and groundwater dynamics. Incorporating more sophisticated algorithms for snow processes, improving groundwater-surface water interactions, and refining the representation of rapid surface flow processes could substantially enhance the model's ability to simulate realistic hydrological responses.