Basin-wide snow depth (SD) maps can support operational water supply assessments, but their availability is limited by measurement costs (airborne) or sampling constraints (satellite and drone). We present Swath-random forest (RF), a methodology that trains random forests on SD measured within a narrow swath (<10% of a basin) to extrapolate basin-wide depths. Using 68 LiDAR surveys from eight basins in Colorado and California, we evaluate two predictor cases: (a) physiography plus prior full-basin snow-depth maps and (b) physiography alone. For the first case, Swath-RF with 2-km-wide swaths reproduces basin-wide depth with low extrapolation absolute bias (0.019 m) and RMSE (0.21 m), and represents snow volume across topographic gradients and across dissimilar years. Errors are 2-3 times larger when using physiography alone. Swath-RF enables basin-wide mapping more frequently or across more basins, but at the cost of accuracy; applicability to other regions will depend on snow climate, physiography, and data availability.

In snow-dominated basins, snowmelt timing exerts a dominant control on the timing of streamflow generation. Data-driven models that rely on precipitation and temperature represent snowpack evolution and streamflow but without direct information on the timing and area of a basin that is actively melting. Here, we introduce melt fraction, a satellite-derived indicator of basin-scale snowmelt progression derived from Sentinel-1 runoff onset timing and MODIS snow cover. We evaluate the utility of melt fraction to improve streamflow simulations across 35 snow-dominated CAMELS-US basins over a seven-year period (2017-2023 water years) using a long short-term memory framework. Using leave-one-basin-out cross-validation, a model (Mf-PT) that includes melt fraction and meteorological data (temperature and precipitation) consistently improved predictive skill relative to a baseline model with meteorological predictors alone. Across full water years, median NSE increased from 0.65 to 0.82, and median normalized RMSE decreased from 0.75 to 0.61. Improvements were most pronounced in late summer (July-September), when discharge becomes increasingly storage-controlled. During late summer, Mf-PT improved NSE from 0.57 to 0.77 and NRMSE decreasing from 0.65 to 0.44. The effectiveness of melt fraction did not vary systematically between low-flow and high-flow years, with Mf-PT showing similar performance gains across both conditions, while the baseline model showing reduced skill during low-flow years. Individual components of melt fraction - runoff onset and snow disappearance - each improved performance in some basins, but their gains were less consistent than when combined, highlighting the value of integrating complementary constraints on snow depletion timing to better represent spring and summer flows. A model trained on gauged basins with physically-meaningful snowmelt timing information from satellite observations offers a scalable pathway for improved streamflow simulations in ungauged snow-dominated regions worldwide.

In the Western United States, water supply forecasting has traditionally relied on snow water equivalent measurements at ground-based stations due to their strong correlations with streamflow volume during spring and summer. However, stations are sparse and sample a small area, prompting interest in spatially complete - but costly - basin-wide mapping from airborne surveys or future satellite missions. Here we show that adding strategic measurements at snow hotspots - localized areas with untapped information for predicting streamflow - consistently outperforms spatially complete surveys that provide basin-average snowpack, both in basins with and without existing stations. While both improve forecast skill, hotspot monitoring increases correlations with streamflow volume by 11-14% (median) across 390 basins, compared to 4% from basin-wide surveys. These findings hold across snowpack datasets, skill metrics, and statistical models. The greatest gains in water supply prediction come from leveraging existing stations and expanding snow measurements to the right places, rather than everywhere.

The Ensemble Streamflow Prediction (ESP) framework combines a probabilistic forecast structure with process‐based models for water supply predictions. However, process‐based models require computationally intensive parameter estimation, increasing uncertainties and limiting usability. Motivated by the strong performance of deep learning models, we seek to assess whether the Long Short‐Term Memory (LSTM) model can provide skillful forecasts and replace process‐based models within the ESP framework. Given challenges in implicitly capturing snowpack dynamics within LSTMs for streamflow prediction, we also evaluated the added skill of explicitly incorporating snowpack information to improve hydrologic memory representation. LSTM‐ESPs were evaluated under four different scenarios: one excluding snow and three including snow with varied snowpack representations. The LSTM models were trained using information from 664 GAGES‐II basins during WY1983–2000. During a testing period, WY2001–2010, 80% of basins exhibited Nash‐Sutcliffe Efficiency (NSE) above 0.5 with a median NSE of around 0.70, indicating satisfactory utility in simulating seasonal water supply. LSTM‐ESP forecasts were then tested during WY2011–2020 over 76 western US basins with operational Natural Resources Conservation Services (NRCS) forecasts. A key finding is that in high snow regions, LSTM‐ESP forecasts using simplified ablation assumptions performed worse than those excluding snow, highlighting that snow data do not consistently improve LSTM‐ESP performance. However, LSTM‐ESP forecasts that explicitly incorporated past years' snow accumulation and ablation performed comparably to NRCS forecasts and better than forecasts excluding snow entirely. Overall, integrating deep learning within an ESP framework shows promise and highlights important considerations for including snowpack information in forecasting.

Seasonal snow plays a critical role in hydrological and energy systems, yet its high spatial and temporal variability makes accurate characterization challenging. Historically, satellite remote sensing has had limited success in mapping snow depth and snow water equivalent (SWE), particularly in global mountain areas. This study evaluates the temporal and spatial accuracy of recently developed snow depth retrievals from the Sentinel-1 (S1) C-band spaceborne radar and their utility within a data assimilation (DA) system for characterizing mountain snowpack. The DA framework integrates the physics-based Flexible Snow Model (FSM2) with a Particle Batch Smoother (PBS) to produce daily snow depth maps at a 500 m resolution using S1 snow depth data. The S1 data were evaluated from 2017 to 2021 in and near the East River Basin, Colorado, using daily data at 12 ground-based stations for temporal evaluation and four LiDAR snow depth surveys from the Airborne Snow Observatory (ASO) for spatial evaluation. The analysis revealed significant inconsistencies in temporal and spatial errors of S1 snow depth, with higher spatial errors. Errors increased with time, especially during ablation periods, with an average temporal RMSE of 0.40 m. In contrast, the spatial RMSE exceeded 0.7 m, and S1 had poor spatial agreement with ASO LiDAR (R2 < 0.3). Experiments with DA window sizes showed minimal performance differences for full-season and early-season windows. Joint assimilation of S1 snow depth with MODIS Snow Disappearance Date (SDD) yielded similar temporal errors but degraded performance in space relative to assimilating S1 alone, while SDD assimilation alone performed best spatially. While S1 may perform better in other regions or snow conditions, our findings are consistent with findings from other error analyses across the Western US and suggest S1 has limited potential to improve snow DA in the East River Basin, specifically. Future work should address retrieval biases, refine algorithms, and consider other snow datasets in the DA system to improve snow depth and SWE mapping in diverse snow environments globally.

A declining spring snowpack is expected to have widespread effects on montane and subalpine forests in western North America and across the globe. The way that tree water demands respond to this change will have important impacts on forest health and downstream water subsidies. Here, we present data from a network of sap velocity sensors and xylem water isotope measurements from three common tree species (Picea engelmannii, Abies lasiocarpa and Populus tremuloides) across a hillslope transect in a subalpine watershed in the Upper Colorado River basin. We use these data to compare tree- and stand-level responses to the historically high spring snowpack but low summer rainfall of 2019 against the low spring snowpack but high summer rainfall amounts of 2021 and 2022. From the sap velocity data, we found that only 40 % of the trees showed an increase in cumulative transpiration in response to the large snowpack year (2019), illustrating the absence of a common response to interannual spring snowpack variability. The trees that increased water use during the year with the large spring snowpack were all found in dense canopy stands – irrespective of species – while trees in open-canopy stands were more reliant on summer rains and, thus, more active during the years with modest snow and higher summer rain amounts. Using the sap velocity data along with supporting measurements of soil moisture and snow depth, we propose three mechanisms that lead to stand density modulating the tree-level response to changing seasonality of precipitation: Topographically mediated convergence zones have consistent access to recharge from snowmelt which supports denser stands with high water demands that are more reliant and sensitive to changing snow. Interception of summer rain in dense stands reduces the throughfall of summer rain to surface soils, limiting the sensitivity of the dense stands to changes in summer rain. Shading in dense stands allows the snowpack to persist deeper into the growing season, providing high local reliance on snow during the fore-summer (early-summer) drought period. Combining data generated from natural gradients in stand density, like this experiment, with results from controlled forest-thinning experiments can be used to develop a better understanding of the responses of forested ecosystems to futures with reduced spring snowpack.

Data from the NASA Cyclone Global Navigation Satellite System (CYGNSS) mission have shown promise for the retrieval of soil moisture, and many soil moisture products using CYGNSS data have been developed. In this work, we present a merged product that combines several CYGNSS soil moisture products using a Minimum Variance Estimator (MVE). The MVE identifies an optimal weighted averaging scheme based on the error covariance characteristics of the CYGNSS soil moisture products. The error covariance matrix is computed using two reference datasets: soil moisture data from the Soil Moisture Active Passive (SMAP) radiometer and in situ soil moisture data. The results from each of these provide insights into both the performance of the merged product and the individual input CYGNSS products. Overall, the merged product offers better performance than any individual CYGNSS product while also offering better temporal resolution than SMAP. The results of this work also demonstrate that the use of the MVE is a compelling technique for soil moisture applications.

Machine learning (ML) has emerged as an effective tool for estimating snow depth and snow water equivalent at unsampled times and locations. Airborne lidar surveys are particularly useful for ML applications: the high-resolution, high-precision snow depth data allow for algorithm training and testing to an extent and spatial resolution not previously possible. Here, we train a random forest model to estimate snow depth relative to a nearby Snotel site using static physiographic data and dynamic (i.e., time-dependent) snowpack data as predictor variables and lidar for the target variable. The model output is daily, 50 m resolution snow depth for basins that have both lidar and Snotel data in Colorado. We evaluated multiple approaches for random forest training: using historic lidar data in a basin (temporal transfer), using lidar data from other basins in a region (spatial transfer), and both together. The three approaches yield RMSE values ranging from 0.37 to 0.44 m at 50 m resolution, achieving lower errors compared to ML studies at higher resolutions. Model error decreases when outputs are upscaled, with RMSE values of 0.17 and 0.10 m for the 4 km and basin scales, respectively. The model scenario which includes both temporally and spatially transferred lidar data is the most robust to the number and timing of lidar surveys used in model training. This framework extends the spatial footprint of Snotel and the temporal coverage of lidar by leveraging the strengths of the two data sets, with applications for water resource management and validation of gridded snow products.

Automated snow station networks provide critical hydrologic data. Whether point observations represent snowpack at larger areas is an enduring question. Leveraging the recent proliferation of airborne lidar snow depth data, we revisit the question of snow station representativeness at multiple scales surrounding 111 stations in Colorado and California (USA) from 2021–2023 (n=476 total samples). In about 50 % of cases, station depths were at least 10 cm higher than areal-mean snow depth (from lidar) at 0.5 to 4 km scales. The nearest 50 m lidar pixels had lower bias and were more often representative of the areal-mean snow depth than coincident stations. The closest 3 m lidar pixel often agreed with station snow depth to within 10 cm, suggesting differences between station snow depth and the nearest 50 m lidar pixel result from highly localized conditions and not the measurement method. Representativeness decreased as scale increased up to ∼6 km, mainly explained by the elevation of a site relative to the larger area. Relative values of vegetation and southness did not have significant impacts on site representativeness. The sign of bias at individual snow stations is temporally consistent, suggesting the relationship between station depth and that of the surrounding area may be predictable. Improving understanding of snow station representativeness could allow for more accurate validation of modeled and remotely sensed data.

Observations recorded by the NASA Cyclone Global Navigation Satellite System (CYGNSS) mission have demonstrated significant sensitivity to soil moisture, motivating the development of several soil moisture products. An assessment of these products was conducted by the CYGNSS science team. The results of the assessment showed that the accuracy of each product may vary based on environmental factors such as surface roughness, vegetation, or terrain complexity. The varied responses led to an effort to combine these products into a single blended product, such that the best features of each product can be used to construct one optimum product. To achieve a preliminary result, five different CYGNSS soil moisture products were combined using a minimum variance estimator (MVE). This approach used in situ soil moisture data to compute the covariance matrix of the soil moisture error for the products. From this, a weighted averaging scheme was derived that minimizes the variance of the blended soil moisture values and therefore the root-mean-square error (RMSE). The performance of the blended product produced by this work appears to be comparable to the performance of the Soil Moisture Active Passive (SMAP) radiometer products.

Soil moisture data with both a fine spatial scale and a short global repeat period would benefit many hydrologic and climatic applications. Since the radar transmitter malfunctioned on NASA’s Soil Moisture Active Passive (SMAP) in 2015, SMAP soil moisture has been downscaled using numerous alternative fine-resolution data. In this paper, we describe the creation and validation of a new downscaled 3 km soil moisture dataset, which is the culmination of previous work. We downscaled SMAP enhanced 9 km brightness temperatures by merging them with L-band Cyclone Global Navigation Satellite System (CYGNSS) reflectivity data, using a modified version of the SMAP active–passive brightness temperature algorithm. We then calculated 3 km SMAP/CYGNSS soil moisture using the resulting 3 km SMAP/CYGNSS brightness temperatures and the SMAP single-channel vertically polarized soil moisture algorithm (SCA-V). To remedy the sparse daily coverage of CYGNSS data at a 3 km spatial resolution, we used spatially interpolated CYGNSS data to downscale SMAP soil moisture. 3 km interpolated SMAP/CYGNSS soil moisture matches the SMAP repeat period of ~2–3 days, providing a soil moisture dataset with both a fine spatial scale and a short repeat period. 3 km interpolated SMAP/CYGNSS soil moisture, upscaled to 9 km, has an average correlation of 0.82 and an average unbiased root mean square difference (ubRMSD) of 0.035 cm3/cm3 using all SMAP 9 km core validation sites (CVSs) within ±38° latitude. The observed (not interpolated) SMAP/CYGNSS soil moisture did not perform as well at the SMAP 9 km CVSs, with an average correlation of 0.68 and an average ubRMSD of 0.048 cm3/cm3. A sensitivity analysis shows that CYGNSS reflectivity is likely responsible for most of the uncertainty in downscaled SMAP/CYGNSS soil moisture. The success of 3 km SMAP/CYGNSS soil moisture demonstrates that Global Navigation Satellite System–Reflectometry (GNSS-R) observations are effective for downscaling soil moisture.

Snow is a critical component of global climate and provides water resources to over 1 billion people worldwide. Yet current measurement methods and modeling techniques lack the ability to fully capture snow characteristics such as snow water equivalent (SWE) and density across variable landscapes. In recent years, physics-informed machine learning (ML) methods have demonstrated promise for combining data-driven learning and physical information. However, this capability has not been widely explored within snow hydrology. Here, we develop a "hybrid" model that applies ML informed by outputs from a physical model and assess whether it provides more accurate estimations of SWE and snow density. We trained and evaluated models at 49 SNOw TELemetry locations spanning a range of snow climates in the western US using 9 years of daily data. The research addressed two questions. In the first, the performance of the hybrid model was compared against a plain neural network (long short-term memory, Long-Short Term Memory), a high-quality physical model, and a statistical snow density model. The second question focused on how regionally trained hybrid models compared to a westwide model as well as their transferability between multiple snow regions. The results showed that combining physical information and ML reduced SWE Root Mean Square Error by 35% compared to a physical model and 51% compared to a neural network. Additionally, regional training only provided minimal benefits compared with a westwide model. These findings indicate that a hybrid approach can yield more accurate snowpack characterization than either physical snow models or ML alone. Seasonal mountain snow provides important water resources for communities throughout the western United States and other global regions. Measuring this snow is difficult due to the large area and variable terrain found in most mountain ranges. This study uses a specialized computer model to produce better estimates of important snowpack information such as snow water equivalent (SWE) and snow density. The computer model combines a neural network (a type of model meant to mimic the human brain) with snowpack estimates from a mathematical model based on physics. When used together, this study found that this "hybrid" approach of the computer and mathematical models was better than other traditional ways of estimating SWE and snow density. A hybrid approach combining machine learning and physical model output was developed and trained over a wide range of snow conditions The hybrid model estimated snow density and snow water equivalent with improved accuracy relative to data-driven and physical models Local training of the hybrid model did not improve snow estimates over a model trained at all study sites across the western United States

Abstract Extreme precipitation events are projected to increase in frequency across much of the land‐surface as the global climate warms, but such projections have typically relied on coarse‐resolution (100–250 km) general circulation models (GCMs). The ensemble of HighResMIP GCMs presents an opportunity to evaluate how a more finely resolved atmosphere and land‐surface might enhance the fidelity of the simulated contribution of large‐magnitude storms to total precipitation, particularly across topographically complex terrain. Here, the simulation of large‐storm dominance, that is, the number of wettest days to reach half of the total annual precipitation, is quantified across the western United States (WUS) using four GCMs within the HighResMIP ensemble and their coarse resolution counterparts. Historical GCM simulations (1950–2014) are evaluated against a baseline generated from station‐observed daily precipitation (4,803 GHCN‐D stations) and from three gridded, observationally based precipitation data sets that are coarsened to match the resolution of the GCMs. All coarse‐resolution simulations produce less large‐storm dominance than in observations across the WUS. For two of the four GCMs, bias in the median large‐storm dominance is reduced in the HighResMIP simulation, decreasing by as much as 62% in the intermountain west region. However, the other GCMs show little change or even an increase (+28%) in bias of median large‐storm dominance across multiple sub‐regions. The spread in differences with resolution amongst GCMs suggests that, in addition to resolution, model structure and parameterization of precipitation generating processes also contribute to bias in simulated large‐storm dominance.

Despite the myriad of remote sensing techniques currently used to map surface water, a gap remains in our ability to rapidly map inundation from flooding at the required temporal resolution to understand how floods evolve. However, the recent launch of constellations of Global Navigation Satellite System-Reflectometry (GNSS-R) satellites can provide data at a more frequent temporal repeat than a single satellite. These L-band instruments could provide moderate spatial resolution inundation maps at a better temporal resolution than other satellite radars. This paper describes a retrieval algorithm for flood inundation mapping using GNSS-R data from the Cyclone GNSS (CYGNSS) constellation. The algorithm employs a simple dielectric model to retrieve fractional inundation from an observation of reflectivity, requiring parameterizations of soil surface and water roughness as well as ancillary soil moisture data. Here, we give a brief overview of the model, describe our parameterization scheme, and present inundation maps using CYGNSS data. We describe four case studies (from the Amazon, Mozambique, Mali, and Australia) and compare the CYGNSS inundation maps to other surface water data (SWAMPS, PALSAR-2, Dartmouth Flood Observatory, MODIS, and the Global Surface Water Explorer). We identify sources of uncertainty in the CYGNSS inundation maps and discuss possible reasons for discrepancies between the inundation retrievals. We introduce the data portal, which houses the CYGNSS inundation maps, for use by the science community.

Even the youngest child knows that fresh water is crucial for life, and it’s easy to see and appreciate our reservoirs, lakes, and rivers for the numerous services they provide, not only for drinking water but also for transportation and the health of our ecosystems. But inland surface water is both a friend and a foe. Too much of it can be devastating for communities—floods are one of the costliest natural disasters, and they often disproportionally impact the most vulnerable members of those communities. Too little of it, though, can be just as destructive. Years-long droughts empty reservoirs, increase wildfire risk, and can lead to conflict over remaining water resources. Quantifying the amount and extent of inland surface water is thus important for knowing where we lie in this delicate balance between abundance and scarcity.A variety of approaches to map flood and inundation dynamics already exist, be they stream gage data, hydrologic models, or remote sensing observations from satellites. All of them have advantages and challenges, and none alone provide a complete picture of the extent of surface water at any one particular moment. This presentation will describe a new approach to mapping flooding and inundation dynamics, which can provide complementary information to that which already exists via other sensors, models, or networks. This approach uses spaceborne Global Navigation Satellite System-Reflectometry (GNSS-R) observations to infer surface water extent. Currently, the vast majority of spaceborne GNSS-R data come from the Cyclone GNSS (CYGNSS) constellation, a NASA mission comprised of eight small satellites orbiting the tropics. Here, we will present flood inundation maps derived from CYGNSS data for the full period of record (2017 – present), which are gridded to three km and have a temporal revisit rate of three days. We will discuss the retrieval algorithm, its validation, limitations of our approach, and plans to disseminate the data to the public. Finally, we will comment on the potential of GNSS-R data beyond CYGNSS to provide hydrologic information to the broader research community and other end users.

This dataset contains meteorology and snow observation data collected at sites in the southwestern Colorado Rocky Mountains during water years 2019-2021. Data collection had an emphasis on paired open-forest sites and included three forested elevations. In total, we present 270 snow pit observations, 4,019 snow depth measurements, and three years of meteorological forcing from two weather stations (one in a meadow, the other in an adjacent forest). The dataset is described in a forthcoming publication of the same name: A meteorology and snow dataset from adjacent forested and meadow sites at Crested Butte, CO, USA (Bonner et al., 2022).

Understanding how the presence of a forest canopy influences the underlying snowpack is critical to making accurate model predictions of bulk snow density and snow water equivalent (SWE). To investigate the relative importance of forest processes on snow density and SWE, we applied the SUMMA model at three sites representing diverse snow climates in Colorado (USA), Oregon (USA), and Alberta (Canada) for 5 years. First, control simulations were run for open and forest sites. Comparisons to observations showed the uncalibrated model with NLDAS-2 forcing performed reasonably. Then, experiments were completed to isolate how forest processes affected modelled snowpack density and SWE, including: (1) mass reduction due to interception loss, (2) changes in the phase and amount of water delivered from the canopy to the underlying snow, (3) varying new snow density from reduced wind speed, and (4) modification of incoming longwave and shortwave radiation. Delivery effects (2) increased forest snowpack density relative to open areas, often more than 30%. Mass effects (1) and wind effects (3) decreased forest snowpack density, but generally by less than 6%. The radiation experiment (4) yielded negligible to positive effects (i.e., 0%-10%) on snowpack density. Delivery effects on density were greatest at the warmest times in the season and at the warmest site (Oregon): higher temperatures increased interception and melted intercepted snow, which then dripped to the underlying snowpack. In contrast, mass effects and radiation effects were shown to have the greatest impact on forest-to-open SWE differences, yielding differences greater than 30%. The study highlights the importance of delivery effects in models and the need for new types of observations to characterize how canopies influence the flux of water to the snow surface.

NASA’s Soil Moisture Active Passive (SMAP) mission only retrieved ~2.5 months of 3 km near surface soil moisture (NSSM) before its radar transmitter malfunctioned. NSSM remains an important area of study, and multiple applications would benefit from 3 km NSSM data. With the goal of creating a 3 km NSSM product, we developed an algorithm to downscale SMAP brightness temperatures (TBs) using Cyclone Global Navigation Satellite System (CYGNSS) reflectivity data. The purpose of downscaling SMAP TB is to represent the spatial heterogeneity of TB at a finer scale than possible via passive microwave data alone. Our SMAP/CYGNSS TB downscaling algorithm uses β as a scaling factor that adjusts TB based on variations in CYGNSS reflectivity. β is the spatially varying slope of the negative linear relationship between SMAP emissivity (TB divided by surface temperature) and CYGNSS reflectivity. In this paper, we describe the SMAP/CYGNSS TB downscaling algorithm and its uncertainties and we analyze the factors that affect the spatial patterns of SMAP/CYGNSS β. 3 km SMAP/CYGNSS TBs are more spatially heterogeneous than 9 km SMAP enhanced TBs. The median root mean square difference (RMSD) between 3 km SMAP/CYGNSS TBs and 9 km SMAP TBs is 3.03 K. Additionally, 3 km SMAP/CYGNSS TBs capture expected NSSM patterns on the landscape. Lower (more negative) β values yield greater spatial heterogeneity in SMAP/CYGNSS TBs and are generally found in areas with low topographic roughness (<350 m), moderate NSSM variance (~0.01–0.0325), low-to-moderate mean annual precipitation (~0.25–1.5 m), and moderate mean Normalized Difference Vegetation Indices (~0.2–0.6). β values are lowest in croplands and grasslands and highest in forested and barren lands.

In many parts of the world including the western United States, the allocation of water is governed by complex water laws that dictate who receives water, how much they receive, and when. Because these rules are generally based on the seniority of water rights, they are not necessarily focused on maximizing economic value across the entire economy. The maximization of value from water use economy-wide is a complex optimization problem that must explicitly consider each user’s water demand, willingness to pay (WTP) function, and the feedbacks among users in a coupled natural-human system model. In this study, we distill these complexities into a simple MATLAB® model developed to represent a two-user economy with water-dependent sectors representative of agriculture and industry. We feed the model with realistic values of relative water use, relative willingness to pay, and return flows to explore the relationships among these factors in water-limited systems. We find that the total economic value generated from water-dependent users depends primarily on the total water available in the system. However, for a given volume of water available, economic value is not necessarily maximized when all the water is appropriated to the user with the highest WTP. Rather, total economic value depends on the amount of water available, the relative WTP between the two users, and on the return flows generated from each sector’s water use. While our simple two-user model is a significant abstraction of the complexities inherent in natural systems, our study provides important insights into the coupled natural-human system dynamics of water allocation and use in water-limited environments.

Snowpack provides the majority of predictive information for water supply forecasts (WSFs) in snow-dominated basins across the western United States. Drought conditions typically accompany decreased snowpack and lowered runoff efficiency, negatively impacting WSFs. Here, we investigate the relationship between snow water equivalent (SWE) and April-July streamflow volume (AMJJ-V) during drought in small headwater catchments, using observations from 31 USGS streamflow gauges and 54 SNOTEL stations. A linear regression approach is used to evaluate forecast skill under different historical climatologies used for model fitting, as well as with different forecast dates. Experiments are constructed in which extreme hydrological drought years are withheld from model training, that is, years with AMJJ-V below the 15th percentile. Subsets of the remaining years are used for model fitting to understand how the climatology of different training subsets impacts forecasts of extreme drought years. We generally report overprediction in drought years. However, training the forecast model on drier years, that is, below-median years (P-15, P-57.5], minimizes residuals by an average of 10% in drought year forecasts, relative to a baseline case, with the highest median skill obtained in mid- to late April for colder regions. We report similar findings using a modified National Resources Conservation Service (NRCS) procedure in nine large Upper Colorado River basin (UCRB) basins, highlighting the importance of the snowpack-streamflow relationship in streamflow predictability. We propose an "adaptive sampling" approach of dynamically selecting training years based on antecedent SWE conditions, showing error reductions of up to 20% in historical drought years relative to the period of record. These alternate training protocols provide opportunities for addressing the challenges of future drought risk to water supply planning. Significance StatementSeasonal water supply forecasts based on the relationship between peak snowpack and water supply exhibit unique errors in drought years due to low snow and streamflow variability, presenting a major challenge for water supply prediction. Here, we assess the reliability of snow-based streamflow predictability in drought years using a fixed forecast date or fixed model training period. We critically evaluate different training protocols that evaluate predictive performance and identify sources of error during historical drought years. We also propose and test an "adaptive sampling" application that dynamically selects training years based on antecedent SWE conditions providing to overcome persistent errors and provide new insights and strategies for snow-guided forecasts.