Improving snow simulations through improved representations of vegetation conditions: insights from high resolution simulations over California

Deep in the Sierra Nevada, the fate of California’s water supply is written in snow. Each winter, storms pile snowpack onto the high mountains, and each spring, that snow melts into the rivers that fill the reservoirs sustaining cities, farms, and ecosystems across the state. Yet the computer models that forecast this frozen reservoir have a persistent blind spot: they systematically underestimate how much snow actually accumulates in the mountainous American Southwest. Now, a team of hydrologists at the University of Arizona has traced much of that missing snow to two overlooked culprits — imprecise precipitation data and, surprisingly, the way models portray the forests above the snow. Their findings, published in The Cryosphere, could reshape how the nation’s flagship water model simulates snowpack in a warming West.

The study, led by Aniket Gupta with co-authors Ali Behrangi, Mohammad Farmani, Patrick Broxton, and Guo-Yue Niu, focused on the Sacramento and San Joaquin River Basins, the two great watersheds that drain the Sierra Nevada and the southern Cascades. Mountain snowpack supplies roughly 70 to 80 percent of the annual runoff in the western United States, making accurate snow estimates a matter of water security rather than academic curiosity. Previous work had shown that the National Water Model, which runs the Noah-MP land surface model at 1-kilometer resolution, underestimates snow water equivalent — the amount of water held in the snowpack — by 20 to 30 percent compared with ground observations from the SNOTEL network. Even sophisticated global reanalyses such as ERA, MERRA-2, and GLDAS miss the mark, largely because their coarse grids blur the sharp elevation gradients that govern where snow falls and lingers.

The Arizona team suspected that the problem lay in two intertwined weaknesses: the precipitation forcing fed into the model and the representation of snow–vegetation interactions. Forests are not passive backdrops to snowpack. Canopies intercept falling snow, some of which sublimates straight back into the atmosphere or drips to the ground as throughfall. Canopies also cast shade that alters the energy budget of the snow beneath, reducing the solar radiation reaching the surface while enhancing longwave emission from the warmer vegetation. These biophysical feedbacks are notoriously difficult to capture at basin scale, and few studies had quantified their influence across an entire mountainous watershed.

To disentangle these effects, the researchers ran eight model experiments using Noah-MP version 5.0 over the two basins at 1-kilometer resolution for the years 2015 through 2019. They paired two forcing datasets — the Analysis of Record for Calibration (AORC), the operational forcing of the National Water Model, and the North American Land Data Assimilation System phase 2 (NLDAS-2), statistically downscaled from roughly 12 kilometers to 1 kilometer using the WRF-Hydro Meteorological Forcing Engine — with two vegetation schemes. One scheme prescribed monthly leaf area index (LAI) and green vegetation fraction from a lookup table, as the National Water Model does. The other employed Noah-MP’s dynamic vegetation module, which prognostically simulates LAI and vegetation cover from a coupled photosynthesis and phenology model, allowing leaves to grow and senesce in response to simulated carbon budgets, temperature, and water stress.

The downscaling step proved decisive. The Mountain Mapper algorithm blends PRISM monthly precipitation climatology with real-time hourly rain gauge observations, computing site-specific bias ratios that are interpolated to the model grid and multiplied by the NLDAS-2 hourly fields. Temperature was downscaled using the NCAR lapse rate, while radiation, pressure, and humidity were adjusted for elevation. The result: the downscaled NLDAS-2 data showed systematically higher precipitation and colder temperatures in the mountains than AORC, producing more snowfall. In the Sacramento Basin, AORC underestimated daily-mean precipitation by about 0.1 millimeters per day relative to PRISM, with the deficit concentrated at the highest elevations where snowfall is greatest. NLDAS-2, by contrast, aligned closely with PRISM between 1,000 and 3,000 meters and matched available gauge observations from the GHCN and California Department of Water Resources networks.

When the researchers compared simulated snow water equivalent against independent estimates from SNODAS, the University of Arizona snowpack dataset, and the operational National Water Model, the pattern was unmistakable. Simulations driven by downscaled NLDAS-2 produced up to 100 millimeters more peak SWE than AORC-driven runs, bringing modeled snow amounts into far better agreement with ground truth. AORC-driven simulations, mirroring the operational model, fell short of SNODAS under every vegetation configuration. Validation at 23 CDWR snow stations across both basins confirmed the advantage: at 11 stations in the Sacramento Basin and 12 in the San Joaquin, the model consistently performed better with NLDAS-2 forcing. At higher-elevation stations such as CAP, GNL, TUM, and HHM, the combination of downscaled NLDAS-2 and dynamic vegetation produced optimal snow amounts, whereas AORC-driven runs underestimated peak March SWE by roughly 500 millimeters.

But the vegetation experiments revealed something more subtle — and more surprising. The prescribed LAI from the lookup table turned out to be substantially higher than both the dynamically predicted LAI and satellite-derived LAI from MODIS, especially in winter and spring above 1,000 meters. That excess foliage cast too much shade over the snowpack. Counterintuitively, the strongly shaded runs accumulated excessive snow on the ground: by suppressing the episodic mid-winter melt events that punctuate accumulation in these subtropical mountains, the heavy canopy allowed snow to pile up unrealistically. The dynamic vegetation module, producing lower LAI closer to MODIS observations, let more shortwave radiation reach the snow surface, enhancing those episodic melts during the accumulation season and thinning the overall snowpack. Validation against in-situ snow data showed that the dynamic vegetation runs, when driven by downscaled NLDAS-2, performed best of all.

The energy-budget analysis explained why. Noah-MP computes the energy available for snowmelt as the sum of net shortwave and net longwave radiation minus latent and sensible heat fluxes. Across all experiments, net solar radiation dominated the balance. During the accumulation period from December through March, the higher prescribed LAI reduced the shortwave radiation reaching the snow surface through canopy shading, which depends on vegetation density and solar angle. During the melt season from April through July, however, the dynamic vegetation runs showed higher green vegetation fraction than the prescribed runs, so shading became more sensitive to cover fraction than to LAI as the sun climbed higher. The net effect: the shading influence of the canopy on snowmelt and sublimation proved stronger than the interception influence on ground snow beneath the canopy, meaning sparser canopies yield thinner, shorter-lived snowpacks overall.

The findings carry direct implications for operational hydrology. The National Water Model, which uses AORC forcing and prescribed vegetation, inherits both weaknesses identified here. The researchers note that the forthcoming NLDAS-3 dataset, dynamically downscaled to 1-kilometer resolution, could substantially improve snow simulations across the western United States and merits careful evaluation within the National Water Model framework. They also caution that even with dynamic vegetation, NLDAS-2-driven simulations still produced higher SWE than SNODAS, likely because the modeled LAI remains larger than MODIS-derived values — suggesting that further calibration of vegetation parameters, such as type-dependent specific leaf area, could push simulations even closer to observations.

Beyond the operational realm, the study underscores a broader lesson about subgrid heterogeneity in Earth system modeling. Grid-averaged vegetation parameterizations, the authors argue, can obscure critical nonlinear responses in snow accumulation and melt, altering diurnal energy partitioning, runoff generation, and seasonal snow persistence. Because LAI controls snow duration through shading, longwave enhancement, and interception, it also modulates surface albedo and land–atmosphere energy exchanges, potentially feeding back into regional climate. As climate change shifts precipitation patterns and forest structure across the West, capturing these fine-scale vegetation processes may prove essential not just for forecasting next spring’s water supply, but for building physically consistent coupled models of land, water, and atmosphere in snow-dominated basins worldwide.

Subject of Research: Improving snow simulations through improved representations of vegetation conditions: insights from high resolution simulations over California

Article Title: Improving snow simulations through improved representations of vegetation conditions: insights from high resolution simulations over California

Article References: Gupta, A., Behrangi, A., Farmani, M., Broxton, P., & Niu, G.-Y. (2026). Improving snow simulations through improved representations of vegetation conditions: insights from high resolution simulations over California. The Cryosphere, 20(10), 5973-5992. https://doi.org/10.5194/tc-20-5973-2026

Image Credits: AI Generated

DOI: 10.5194/tc-20-5973-2026

Keywords: Improving, snow, simulations, improved, representations, vegetation, conditions, insights, high, resolution, over, California