Integrated Hydroclimate Modeling of Non-Stationary Water Balance, Snow Dynamics, and Streamflow Regimes in the Devils Lake Basin Region

Abstract

The hydrology of the transboundary region encompassing the western Red River Basin headwaters, such as Devils Lake Basin (DLB) in North America, is complex and highly sensitive to climate variability, impacting water resources, agriculture, and flood risk. Understanding hydrological shifts in this region is critical, particularly given recent hydroclimatic changes. This study aimed to simulate and analyze key hydrological processes and their evolution from 1981 to 2020 using an integrated modeling approach. We employed the NASA Land Information System (LIS) framework configured with the Noah-MP land surface model and the HyMAP routing model, driven by a combination of reanalysis and observational datasets. Simulations revealed a significant increase in precipitation inputs and consequential positive net water storage trends post-1990, indicating increased water retention within the system. Snow dynamics showed high interannual variability and decadal shifts in average Snow Water Equivalent (SWE). Simulated streamflow exhibited corresponding multi-decadal trends, including increasing flows within a major DLB headwater basin (Mauvais Coulee Basin) during the period of Devils Lake expansion (mid-1990s to ~2011). Furthermore, analysis of decadal average seasonal hydrographs indicated significant shifts post-2000, characterized by earlier and often higher spring peaks and increased baseflows compared to previous decades. While the model captured these trends, validation against observed streamflow highlighted significant challenges in accurately simulating peak flow magnitudes (Nash–Sutcliffe Efficiency = 0.33 at Mauvais Coulee River near Cando). Overall, the results depict a non-stationary hydrological system responding dynamically to hydroclimatic forcing over the past four decades. While the integrated modeling approach provided valuable insights into these changes and their potential drivers, the findings also underscore the need for targeted model improvements, particularly concerning the representation of peak runoff generation processes, to enhance predictive capabilities for water resource management in this vital region.

1. Introduction

The hydrology of the Northern Great Plains (NGP), particularly in the Devils Lake Basin (DLB), a headwater and terminal lake basin inside western Red River Basin, represents a critical nexus of climate variability, land use practices, and water resource management challenges. This area, encompassing a significant portion of North America’s Prairie Pothole Region (PPR), is characterized by unique glacial landscapes, extensive agriculture, and high sensitivity to climatic shifts, making it a crucial area for hydro-climatic research [1,2,3].

Episodic shifts between wet and dry periods dramatically influence the region’s landscape and water resources [4]. For instance, a prolonged wet spell commencing since 1993 led to what has been termed a “novel climatic continuum” a persistent period of above-average moisture that has fundamentally altered the region’s hydrology [5] causing significant wetland expansion in areas like Nelson County within the PPR [6], converting vast tracts of agricultural land into water bodies, which severely impacted local farm economies and highlighted the delicate balance between agricultural productivity and hydrological hazards [5,7]. Such changes underscore the vulnerability of prairie ecosystems and the communities reliant upon them to climatic extremes. Further research into cold agricultural regions, such as the South Tobacco Creek basin and the La Salle Basin in Manitoba, demonstrates the complexity driven by seasonal snowmelt and rainfall patterns [8,9]. Mahmood et al. [8] identified distinct hydro-climatic patterns influenced by wet/dry seasonal combinations, emphasizing the critical role of antecedent soil moisture and seasonal precipitation timing in governing runoff generation and streamflow response, particularly the importance of snowmelt dynamics in these northern latitude basins, where the unique geomorphological features of the region add layers of complexity [10,11]. Distinct hydroclimatic phases were further identified using process-based model [12] and cold region Budyko framework [13]. The prairie potholes, remnants of the last glaciation, act as crucial seasonal wetlands, trapping meltwater and rainfall, modulating runoff, recharging groundwater, and supporting significant biodiversity [14,15,16,17].

The interplay between natural hydrological processes and anthropogenic modifications, including land use change and drainage practices, further complicates water management. Hydrological modeling in the PPR is notoriously difficult due to the low-gradient topography and the immense number of depressional wetlands that can either be disconnected from or integrated into the river network depending on their storage state (e.g., [18,19]). Van Hoy et al. [12] investigated hydrological shifts in glaciated landscapes, using physically based models to demonstrate how increased precipitation and land use modifications alter water budget partitioning and streamflow generation—processes vital to understanding the hydrodynamics of headwater basin draining to Devils Lake. Complementing this, Zhang et al. [20] highlighted the sensitivity of groundwater systems to climate change, stressing the need for high-resolution modeling incorporating detailed soil and land use data to accurately forecast groundwater recharge and its interaction with surface water, a critical aspect within the PPR’s interconnected system [1,11].

While previous research has significantly advanced understanding of specific hydrological components or the impacts of broad-scale climate shifts, a gap remains in holistically evaluating the coupled land surface-atmosphere interactions and hydrological routing, particularly concerning the influence of upstream topographic features like Turtle Mountain on the downstream features like Devils Lake system using a fully integrated modeling approach. While previous studies in the NGP hydroclimatic have advanced our understanding using field observations [13], analytical solutions, and process-based hydrological models [8,12,21], they have often relied on data from limited climate stations. A gap remains in using a fully integrated, physically based modeling framework forced by a consistent, high-resolution meteorological dataset to diagnose the region’s hydrological response to recent, well-documented climatic shifts. Therefore, the primary objective of this study is to characterize the evolution of water balance components, snow dynamics, and streamflow regimes from 1981 to 2020 in response to significant hydroclimatic variability. We use an advanced land surface and hydrological modeling system in a diagnostic capacity to understand the first-order hydrological responses and to identify key processes and model limitations that can guide future research in this complex region.

2. Study Area

The Red River Basin is a significant geographical and hydrological area that spans across parts of the United States and Canada, including Minnesota, North Dakota, South Dakota, and Manitoba (Figure 1). This basin is characterized by its diverse ecosystem, which includes extensive agricultural lands, forests, and prairie potholes, contributing to its complex hydrology. The region experiences a wide range of climatic conditions, affecting river flow and water levels in lakes such as Devils Lake. The basin’s hydrology is influenced by snowmelt, rainfall patterns, and human activities, making it a crucial area for studying water management and environmental sustainability. It spans over 287,000 square kilometers, extending from the Great Plains in the United States into southern Manitoba, Canada. It is characterized by its flat topography, especially notable in the Red River Valley, where the river flows northward into Lake Winnipeg. The region’s geography includes diverse landscapes, from prairies and wetlands to forests, contributing to its complex hydrological systems. The basin’s orography is relatively gentle, lacking significant elevation changes, influencing its slow-moving river systems and susceptibility to flooding, particularly during spring snowmelt periods.

The Devils Lake Basin, located in north-central North Dakota, is a closed, terminal basin covering an area of about 9800 km². As a terminal lake, its water levels are highly sensitive to the balance between precipitation/runoff and evaporation, with no natural outlet to the sea. Historically, the lake has experienced dramatic fluctuations, with a substantial and well-documented rise in water levels and surface area since the wet period began in the early 1990s, causing extensive flooding of adjacent lands and communities.

Northwest of the Devils Lake lies the Turtle Mountain region, with its varied topography and vegetation, which plays a role in the atmospheric and hydrological dynamics affecting the Red River and Devils Lake basins. Its elevated terrain can influence weather patterns, contributing to precipitation distribution that feeds into these basins. Hydrologically, runoff from the Turtle Mountains flows into surrounding water bodies, impacting water levels and the ecological balance within the basins. This interaction highlights the interconnectedness of regional landscapes and their cumulative effects on water systems and climate in the area. This study focuses on the Mauvais Coulee basin (Figure 1), a major headwater tributary that drains into the Devils Lake system. Mauvais Coulee is a characteristic prairie stream; its flow regime is intermittent, often experiencing zero-flow conditions during late summer, fall, and winter. The hydrology is dominated by a pronounced spring freshet driven by snowmelt runoff, with occasional, smaller peaks generated by intense summer rainfall events.

3. Methods

3.1. Modeling Framework

To investigate the land-surface/atmosphere interactions within the hydrologic routing of a portion of the Red River Basin, encompassing the upstream of Devils Lake and including the Turtle Mountain area, we configured the NASA Land Information System (LIS) model [22,23]. This decision was driven by the hypothesis that the Turtle Mountain region significantly influences hydrological dynamics and precipitation patterns due to its upstream location. LIS is a high-performance land surface modeling and data assimilation system developed by NASA. It integrates satellite and ground-based observational data with land surface models to provide a comprehensive tool for studying Earth’s terrestrial environment. LIS’s flexible framework allows for the simulation and analysis of soil moisture, temperature, and vegetation states across various scales and geographical regions, making it an essential tool for hydrological and environmental research [22,23].

The core of our LIS model utilized the NOAH-MP 4.0 land surface model (LSM) [24], operating at a 1 Km horizontal resolution and a 30 min timestep to accurately capture the complex interactions at play. Noah-MP is an advanced land surface model that represents various land–atmosphere interactions and hydrological processes. It is designed to improve the simulation of the water cycle and energy fluxes between the land surface and the atmosphere. Noah-MP incorporates multiple parameterizations of key hydrological and biophysical processes, including vegetation dynamics, soil moisture, and snowpack, allowing for a detailed representation of terrestrial environments in climate and weather models. For this diagnostic study, we used a standard, uncalibrated configuration of Noah-MP to assess the performance of a physically based model structure without site-specific tuning.

3.2. Model Forcing, Validation Data, and Static Parameters

The modeling framework was driven by a suite of high-resolution meteorological and satellite-based datasets (

Table 1. Datasets used for model forcing and validation.

Dataset	Source/Name	Spatial Resolution	Temporal Resolution	Period Used	Purpose
Meteorological Forcing	MERRA-2	~50 km	Hourly	1981–2020	Primary atmospheric data (temperature, wind, etc.)
Meteorological Forcing	NLDAS-2	~12 km	Hourly	1981–2020	Downscaling of MERRA-2 and precipitation when IMERGE is unavailable.
Precipitation	GPM IMERG	10 km	Hourly	2000–2020	Precipitation input
Vegetation	MODIS—derived	5 km	Monthly	2000–2020	Dynamic Green Vegetation Fraction (GVF)
Streamflow Validation	USGS	Point Guage	Daily	1981–2020	Model evaluation

). To accurately represent atmospheric fields crucial for configuring the land surface-hydrological model, we integrated the MERRA-2 dataset as our primary meteorological forcing source. This was further refined by overlaying it with the higher resolution NLDAS-2 reanalysis dataset. This process adjusts the coarser MERRA-2 data to match the finer spatial patterns of NLDAS-2. Precipitation data, critical for hydrological modeling, was sourced from the GPM IMERG observational dataset, offering high-accuracy rainfall measurements essential for our simulations.

For the period from 2000 to 2021, the model’s Green Vegetation Fraction (GVF) was derived from MODIS (Moderate Resolution Imaging Spectroradiometer) observations data assimilated into the model framework. For the earlier period of 1981 to 2000, we relied on the default climatology provided by MODIS GVF data. This approach ensured a comprehensive representation of vegetation dynamics over the study period.

3.3. Model Execution and Evaluation

Acknowledging the importance of model initialization in regions affected by permafrost, and soil moisture dynamics we designated a 40-year spin-up time for the model. This duration allowed the system to attain an equilibrium state, a prerequisite for reliable simulation outcomes in such environments [25]

The hydrological routing within the LIS framework was facilitated by the Hydrological Modeling and Analysis Platform version 2 (HyMAP-2) routing model, set to a matching 30 min timestep. HyMAP is a hydrological routing model that simulates the movement of water across the Earth’s surface. Integrating with Noah-MP LSM provides a comprehensive view of the water cycle. It represents riverine and floodplain dynamics, including streamflow, inundation, and river transport processes. HyMAP’s ability to simulate surface water dynamics at various spatial and temporal resolutions supports detailed studies of hydrological impacts on regional and global scales [26].

The performance of the HyMAP routing model in simulating streamflow was evaluated against observational data from the USGS gauge on the Mauvais Coulee River, located in Cando, ND, USA (station 05056100). This gauge provided a high-quality dataset with over 90% data completeness for the study period. Model performance was quantified using the Nash–Sutcliffe Efficiency (NSE) metric, which is calculated as follows:

$$NSE=1-{\displaystyle \frac{\sum _{t=1}^T{\left(Q_{sim,t}-Q_{obs,t}\right)}^2}{\sum _{t=1}^T{\left(Q_{obs,t}-{\overline{Q}}_{obs}\right)}^2}}$$

where $Q_{sim}$ is the simulated streamflow, $Q_{obs}$ is the observed streamflow, and ${\overline{Q}}_{obs}$ is the mean of the observed streamflow. NSE values range from −∞ to 1. An NSE of 1 indicates a perfect match between simulated and observed data. An NSE of 0 indicates the model simulation is as accurate as the mean of the observed data, while values less than 0 indicate that the observed mean is a better predictor than the model.

This configuration enabled us to simulate surface water dynamics with precision, providing valuable insights into the hydrological impacts of land-surface/atmosphere interactions within the specified domain of the Red River Basin.

4. Results

4.1. Water Balance Analysis

Figure 2 illustrates the simulated annual water balance components for the study domain from 1981 to 2020. Input fluxes, shown as positive bars, consist of rainfall (dark blue) and snowfall (light blue), while output fluxes, shown as negative bars, include evapotranspiration (ET; green), sublimation (purple), and streamflow (yellow). The numerical labels above each year’s bar stack indicate the net annual change in system storage (in mm).

Over the 40-year period, the water balance partitioning reveals distinct characteristics. Rainfall was the primary input, averaging 71% of the total annual input, with snowfall contributing the remaining 29%. Notably, the relative contribution of snowfall exhibited high interannual variability (ranging from 13% to 49%), reflecting fluctuating winter precipitation patterns crucial in this region. Among the output fluxes, ET dominated, accounting for an average of 92% of the total output, clearly visible as the largest negative component (green bars) nearly every year. Streamflow represented a much smaller but variable output pathway (7.3% on average), while sublimation was almost negligible (0.7%). These simulated flux partitions are consistent with findings reported for similar northern glaciated landscapes, where ET is also the dominant water loss pathway [12].

The timeseries reveals a significant temporal shift in the water balance. A marked increase in total annual inputs (sum of blue bars) is apparent, particularly after the early 1990s. Average annual precipitation inputs rose from 489 mm/yr during the relatively dry 1981–1990 period to 615 mm/yr over the wetter 1991–2020 period. This increased water availability strongly influenced system storage. As indicated by the numerical labels, substantial net storage gains occurred frequently during wetter years (e.g., +150 mm in 1993, +145 mm in 1999, +150 mm in 2009), suggesting enhanced infiltration and potential groundwater recharge. Conversely, pronounced dry years, such as 1981 (−84 mm), 1988 (−91 mm), 2012 (−101 mm), and 2017 (−61 mm), resulted in net storage deficits. Intriguingly, the magnitude of storage gains during wet years often exceeds the storage losses during dry years throughout the simulation, suggesting a trend towards increased long-term water retention within the basin’s subsurface systems under the observed hydroclimatic conditions.

4.2. Capturing the Snow Dynamics

To investigate cold-region processes, Figure 3 presents the simulated Snow Water Equivalent (SWE) time series for the study domain from 1981 to 2020. The plot displays SWE at multiple temporal resolutions: daily (blue line), monthly average (red line), 5-year moving average (green line), and 10-year moving average (yellow line).

The daily and monthly SWE series clearly depict a pronounced seasonal cycle characteristic of the region’s climate. Snow accumulation typically occurs from late autumn through winter, leading to peak SWE values generally between February and early April, followed by rapid ablation during the spring melt period, usually complete by late April or early May. Significant interannual variability is evident in the magnitude of peak seasonal SWE (height of blue/red peaks), with notably high accumulations simulated during the mid-to-late 1990s (e.g., peaks exceeding 200 mm around 1996/1997).

Longer-term variations in snowpack conditions are highlighted by the 5-year and 10-year moving averages. Both averages show a distinct rise into the late 1990s and early 2000s, before exhibiting a gradual decline towards the end of the simulation period (2020). This pattern suggests underlying decadal variability potentially linked to shifts in winter precipitation amounts or temperatures influencing snow accumulation and persistence.

The Noah-MP model captures these dynamics using multi-layer snowpack physics that account for processes like accumulation, compaction, melting, and sublimation. In the Devils Lake Basin (DLB), snow accumulation is primarily governed by solid precipitation and snow redistribution driven by wind transport, which is influenced by surface roughness and vegetation. Although the model includes snow sublimation, which can account for approximately 20% of snow loss, it does not currently represent canopy interception—a limitation in forested or shrub-dominated regions. Other processes known to be important in the region, such as wind-driven snow redistribution [27], are implicitly represented through the model’s physics but not explicitly tracked. The simulated snow dynamics are broadly consistent with the findings of Van Hoy et al. [12], who emphasized the role of snow redistribution and sublimation in modulating water availability in glaciated northern landscapes. A more detailed comparison of our SWE simulations with other regional studies is provided in the Discussion.

4.3. Streamflow Simulations

Initial simulations revealed a consistent overestimation of baseflow by HyMAP. To address this, a simple baseflow correction was applied by subtracting the long-term mean simulated baseflow from the daily time series before comparison.

Figure 4 compares the observed daily streamflow (blue line) with the baseflow-corrected simulated streamflow (red line). Quantitatively, the comparison yielded a Nash–Sutcliffe Efficiency (NSE) of 0.33 over the entire period for the annual streamflow. The very low NSE value (where 1 indicates a perfect match and values below 0 indicate the model is worse than using the observed mean) highlights significant limitations in the model’s ability to capture the observed daily variability and magnitude of streamflow.

Streamflow results confirm several aspects of the model performance suggested by the metrics. While the simulation generally captures the timing of seasonal runoff events (red peaks often align temporally with blue peaks), substantial discrepancies in magnitude are apparent.

• Baseflow: Despite the correction, simulated baseflows often remain higher than observed, particularly during prolonged dry periods where observed flows approach zero. This suggests the simple mean correction may be insufficient or that the model struggles with simulating low-flow recession dynamics.
• Peak Flows: The most significant deficiency is the systematic and severe underestimation of peak flows. This is visible for nearly all runoff events but is particularly stark during high-flow years. For instance, during the major flood event of 2011, the observed peak discharge exceeded 30 m³/s, whereas the simulated peak barely reached 5 m³/s. Similar dramatic underestimations occurred during other high flow periods.

This consistent underperformance in capturing peak flows, especially extremes, points to challenges in accurately simulating key runoff generation processes within the coupled LIS/HyMAP framework. Potential contributing factors may include difficulties in representing infiltration-excess runoff during intense rainfall, frozen soil infiltration, inadequate simulation of melt rates and water release during rain-on-snow events, or insufficient representation of upstream hydrological controls such as wetland storage dynamics or reservoir operations that can significantly modulate flow regimes in the Mauvais Coulee Basin. These limitations impact the direct usability of the streamflow simulations for applications highly sensitive to peak flow magnitudes, such as flood risk assessment, and suggest areas requiring further model development or parameter refinement.

While event-based peak flow simulations present challenges (as discussed for Figure 4), the modeling framework can also be used to explore longer-term, multi-decadal hydrological trends. These trends are critical for understanding cumulative water balances, particularly within the closed Devils Lake Basin (DLB). To this end, we examined the simulated streamflow for the Mauvais Coulee watershed, a tributary located near Cando, ND, USA, within the DLB.

Figure 5 presents the simulated streamflow time series for Mauvais Coulee from 1981 to 2020. The simulation reveals distinct multi-decadal patterns: (1) 1981–mid-1990s: An initial period characterized by generally declining average streamflow. (2) Mid-1990s–~2012: A subsequent, prolonged period exhibiting a clear trend of increasing average streamflow. (3) Post-2012: An apparent recent reversal, suggesting a shift towards decreasing mean annual flows and potentially reduced year-to-year variability.

These simulated long-term shifts for this DLB tributary likely reflect the watershed’s integrated response to the broader hydroclimatic changes observed regionally. The transition to increasing flows around the mid-1990s aligns temporally with the documented step-increase in regional precipitation inputs (Figure 2) and periods of higher snow accumulation (Figure 3), suggesting increased water availability drove higher runoff. This simulated increase in tributary discharge from the mid-1990s through the early 2010s corresponds directly with the period of dramatic expansion and rising levels of Devils Lake [7], highlighting the contribution of such watershed responses to the basin’s overall water balance. The mechanisms likely involve complex interactions between climate drivers and landscape factors (e.g., evolving wetland storage, land use), consistent with processes explored by Van Hoy et al. [12] in similar settings. The potential post-2012 decline warrants further attention, as it could signify a stabilization or shift in the regional hydroclimate or landscape response, with implications for the future water balance of Devils Lake.

4.4. Decadal Shifts in Storm Hydrograph

Complementing the analysis of multi-decadal trends at specific locations (Figure 5), Figure 6 investigates how the average seasonal pattern of streamflow has evolved across consecutive decades within the study region. The figure presents decadal mean hydrographs, calculated as the average monthly streamflow over the water year (October–September) for four periods: 1981–1990 (blue line), 1991–2000 (orange line), 2001–2010 (green line), and 2011–2020 (red line).

While all decades display the characteristic seasonal flow regime of the region—marked by low baseflow in winter (approximately December–March) and a distinct spring freshet—significant interdecadal differences in both the magnitude and timing of flows are clearly evident:

• Spring Peak Magnitude & Timing: The most dramatic feature is the exceptionally high average spring peak observed during the 2001–2010 decade (green line), indicative of intensified runoff generation during that period. The subsequent 2011–2020 decade (red line) also shows elevated peak flows compared to the earlier decades. Concurrently, there is a distinct shift towards earlier spring peaks in the later decades. The peaks for 2001–2010 and 2011–2020 center in mid-April, whereas the peaks for 1981–1990 (blue line) and 1991–2000 (orange line) occur later, generally in late April or early May.
• Baseflow Levels: Both winter (December–March) and summer/autumn (June–September) baseflows exhibit noticeable increases in the two more recent decades (2001–2010 and 2011–2020; green and red lines) compared to the earlier decades (1981–1990 and 1991–2000; blue and orange lines).
• Hydrologic regime Modality: The cold region NGP hydrologic system transitions from a unimodal streamflow hydrograph (only snowmelt streamflow system in 1981–1990) to a bimodal streamflow hydrograph (streamflow from both snowmelt and rainfall in 2001–2010). This is consistent with Dumanski et al.’s [28] Smith Creek study.

These pronounced shifts in the average seasonal hydrograph strongly suggest an evolving, non-stationary hydrological regime over the 40-year study period. The combination of higher and earlier spring peaks in recent decades likely reflects changes in winter climate conditions, potentially involving increased overall winter precipitation leading to larger snowpacks (consistent with increased inputs in Figure 2 and SWE variability in Figure 3) coupled with shifts in spring temperatures causing earlier and possibly more rapid snowmelt. The potential for more frequent rain-on-snow events could also contribute to these higher, earlier peaks. Furthermore, the observed increase in baseflow throughout the year in the later, wetter decades could signify greater contributions from subsurface storage (reflecting the net storage gains seen in Figure 2), potentially indicating rising water tables or enhanced groundwater discharge, possibly modulated by changes in vegetation and evapotranspiration patterns. These findings align with the expected hydrological responses to climate variability and change in northern prairie environments (e.g., [12]) and highlight the value of examining changes in the full seasonal flow distribution, not just annual totals, or long-term trends.

5. Discussion

5.1. A Diagnosed Shift Toward a Wetter, Non-Stationary Hydrological Regime

This study used an integrated modeling framework in a diagnostic capacity to understand the first-order hydrological responses to significant hydroclimatic shifts in the Devils Lake Basin region over a 40-year period. The results consistently point to a non-stationary system that transitioned to a wetter regime after the early 1990s. The primary driver for this shift is the marked increase in annual precipitation inputs, which rose from an average of 489 mm/yr in the 1981–1990 decade to 615 mm/yr over the 1991–2020 period (Figure 2). This increased water availability is the foundational driver for the other observed changes in the basin’s water balance, snowpack, and streamflow.

A key finding is the fundamental change in the seasonal streamflow hydrograph, which evolved from a classic unimodal, snowmelt-dominated system in the 1980s to a bimodal system in the 2001–2010 decade, where summer rainfall contributes a second distinct peak (Figure 6). This transition signifies a regime shift where summer convective storms, falling on wetter antecedent conditions, have become an increasingly important driver of annual streamflow. This finding is consistent with studies in other prairie watersheds, such as Smith Creek, Saskatchewan, where a similar shift from unimodal to bimodal hydrographs was observed in response to increased rainfall and wetland storage [28]. This alteration of the seasonal pattern is a clear hydrological fingerprint of the “novel wet climatic continuum” described for this region since 1993 [5].

The multi-decadal trends in the Mauvais Coulee tributary, which show a sustained increase in water yield from the mid-1990s to the early 2010s (Figure 5), reflect this new water surplus at the watershed scale. This period of simulated higher discharge corresponds directly with the documented dramatic expansion of Devils Lake [7], highlighting the critical role that tributary responses played in the basin’s overall water balance.

5.2. Model Performance, Limitations, and Implications for Results

While the modeling framework captured key long-term trends and seasonal patterns, it is crucial to acknowledge its limitations, particularly in simulating streamflow. The Nash–Sutcliffe Efficiency (NSE) of 0.33 for annual streamflow at the Mauvais Coulee gauge (Figure 4) indicates moderate model performance in reproducing the magnitude of total annual water yield. However, the NSE for daily streamflow was 0.12, indicating poor model performance in reproducing the magnitude of daily peak flows. This deficiency is a common challenge for large-scale, physically based models that are not specifically calibrated for event-scale discharge in complex landscapes like the PPR [18]. Potential reasons include the model’s difficulty in representing processes like infiltration into frozen soils, rapid runoff from rain-on-snow events, and the intricate storage-and-release dynamics of the region’s numerous wetlands, which are not explicitly represented in the HyMAP routing model.

However, our primary aim is not to optimize model performance for historical flow calibration, but to demonstrate the capability of a fully distributed, weakly coupled model (NoahMP + HyMAP) to simulate surface hydrology in a snow-dominated region under nonstationary climate conditions. Unlike traditional semi-distributed models such as SWAT—which are typically calibrated at the outlet and often achieve higher NSE values, sometimes at the expense of parameter realism—fully distributed, process-based models run over long periods and without calibration generally yield lower NSE values (e.g., [29]). For example, Shabani et al. [30] achieved an NSE of 0.37 using a calibrated SWAT model in the same basin, while Van Hoy et al. [12] reported an NSE of up to 0.5 using the Cold Region Hydrological Model with calibration.

This performance discrepancy is not arbitrary but is rooted in specific, identifiable limitations of the uncalibrated framework when applied to this unique landscape. The primary reason for the systematic underestimation of daily peak flows is a fundamental structural mismatch between a conventional riverine routing model like HyMAP and the dominant “fill-spill” hydrological processes of the PPR’s wetland-dominated landscape [18]. HyMAP is not designed to simulate the immense, dynamic storage and variable connectivity of these depressions, and therefore its inability to reproduce sharp runoff peaks generated by wetland complexes is a physically expected outcome. Furthermore, the model’s performance is likely influenced by the meteorological forcing data, as gridded products can struggle to capture the high intensity and small spatial footprint of the convective summer storms that often generate the highest peak flows. Therefore, this study intentionally used an uncalibrated model to diagnose these inherent structural and data-related challenges. The resulting performance, while low on a daily scale, provides a crucial baseline and reveals that even a perfectly calibrated model would likely struggle without a purpose-built structure for prairie hydrology [19].

Given this limitation, the results of this study should be interpreted as a diagnostic analysis of trends and sensitivities rather than a precise, predictive quantification of daily streamflow. The model’s inability to capture high peak flows means that the simulated runoff component of the water balance (Figure 2) is likely underestimated. This introduces uncertainty into other water balance terms; for instance, the trend of increasing subsurface water storage is likely robust in its direction, but its absolute magnitude may be overestimated, as some water that should have become runoff was instead retained in the model’s soil column.

Further uncertainties stem from the model’s structural components and input data. The use of static vegetation climatology for the pre-2000 period, followed by dynamic MODIS-based GVF, could influence the consistency of simulated evapotranspiration trends. Additionally, the lack of canopy snow interception is a known model limitation that could affect the snowpack mass balance in forested areas like Turtle Mountain.

5.3. Future Research Directions

The findings and limitations of this study highlight several critical paths for future research. The highest priority for developing a predictive, decision-making tool for this region is the calibration and validation of the integrated model, focusing on parameters that govern runoff generation and routing. Future work should also move towards more sophisticated routing schemes that explicitly account for the dynamic storage and connectivity of the thousands of prairie pothole wetlands and larger lakes in the basin, which have been shown to be critical controls on streamflow [18]. Incorporating dynamic land use change, particularly agricultural drainage, would further enhance the model’s realism. Finally, targeted modeling experiments could be designed to better isolate and quantify the specific atmospheric influence of topographic features like Turtle Mountain on regional precipitation and hydrology.

6. Conclusions

This study utilized an integrated modeling framework, coupling the Noah-MP land surface model within NASA LIS with the HyMAP routing model, to simulate and analyze the complex hydrological dynamics of the transboundary region encompassing the upper Red River Basin headwaters, the Devils Lake Basin, and the Turtle Mountain area from 1981 to 2020. The primary objective was to characterize the evolution of water balance components, snow dynamics, and streamflow regimes in response to significant hydroclimatic variability over recent decades.

The simulations successfully captured key aspects of the region’s hydrology and revealed significant temporal shifts. Key findings include:

• A distinct increase in total annual precipitation inputs occurred post-1990, leading to a wetter overall regime compared to the 1980s. This resulted in frequent periods of substantial net-positive water storage within the basin, particularly from the mid-1990s through the 2000s, suggesting increased long-term water retention in subsurface systems.
• Snow dynamics exhibited high interannual variability, confirming the critical role of winter processes. Decadal trends showed average Snow Water Equivalent (SWE) increasing into the late 1990s/early 2000s before declining, highlighting non-stationarity in seasonal snowpack conditions.
• Simulated streamflow in the Devils Lake Basin tributary (Mauvais Coulee) reflected the broader hydroclimatic shifts, showing a marked increase from the mid-1990s to approximately 2015, coinciding temporally with the period of major Devils Lake expansion.
• Analysis of decadal average hydrographs demonstrated a fundamental shift in the seasonal flow pattern, including notably higher spring peak flows (especially in 2001–2012), a consistent shift towards earlier spring peaks (mid-April vs. late April/early May) in the post-2000 decades, and the emergence of a bimodal (snowmelt and rainfall-driven hydrograph).

Collectively, these findings depict a highly dynamic and non-stationary hydrological system responding sensitively to changes in precipitation and temperature regimes. The observed shifts towards increased water storage, higher baseflows, and earlier spring runoff have significant implications for regional water management, agricultural practices, flood risk (particularly concerning Devils Lake), and the functioning of prairie pothole ecosystems.

However, the study also highlighted important limitations of the current modeling framework. While validation against observed streamflow showed moderate skill at annual scale (NSE = 0.33 for Mauvais Coulee River), it also revealed significant deficiencies in simulating daily peak flows, especially during extreme events (NSE = 0.12), indicating challenges in accurately representing rapid runoff generation processes like rain-on-snow interactions or infiltration dynamics on a daily scale. This limitation reinforces that the study should be viewed as a diagnostic analysis of long-trends and sensitivities, and conclusions regarding the magnitude of water balance components carry uncertainty. Furthermore, the simple baseflow correction applied was insufficient to fully rectify mismatches during low-flow periods. Additionally, limitations such as the lack of canopy snow interception in the land surface model configuration may affect simulation accuracy in specific land cover types.

These findings and limitations highlight several critical paths for future research. The highest priority for developing a predictive, decision-making tool for this region is the calibration and validation of the integrated model, focusing on parameters that govern runoff generation and routing. Future work should also move towards more sophisticated routing schemes that explicitly account for the dynamic storage and connectivity of the thousands of prairie pothole wetlands and larger lakes in the basin, which have been shown to be critical controls on streamflow (Shook et al., 2024) [18]. Incorporating dynamic land use change, particularly agricultural drainage, would further enhance the model’s realism. Finally, targeted modeling experiments could be designed to better isolate and quantify the specific atmospheric influence of topographic features like Turtle Mountain on regional precipitation and hydrology. By characterizing the shifts in the region’s hydrology and transparently identifying key model limitations, this diagnostic study contributes to deeper understanding of the region’s evolving hydrology and provides a clear foundation for targeted future research aimed at improving predictive capabilities for sustainable water resource management.