Enhancing Deterministic Freezing-Level Predictions in the Northern Sierra Nevada through Deep Neural Networks

February 14, 2026

A new paper entitled “Enhancing Deterministic Freezing-Level Predictions in the Northern Sierra Nevada through Deep Neural Networks” was recently published in the Journal of Hydrometeorology. The study was led by Vesta Afzali Gorooh, with co-authors Agniv Sengupta, Shawn Roj, Rachel Weihs, Brian Kawzenuk, and Luca Delle Monache [all at the Center for Western Weather and Water Extremes (CW3E) at Scripps Institution of Oceanography, University of California, San Diego]. 

The work supports CW3E’s priorities as defined in the CW3E Strategic Plan in Advanced Precipitation and Streamflow Prediction and Forecast-Informed Reservoir Operations (FIRO) by using machine-learning postprocessing to directly improve the forecast skill of freezing level (FZL) height over the Yuba Feather watershed in the northern Sierra Nevada. Accurate FZL forecasts are essential in complex topography, for anticipating runoff generation, and for making real-time storage and flood-management decisions in reservoir systems (Sumargo et al., 2020).

Using the CW3E high-resolution West-WRF reforecast and FZL observations from the California-Nevada River Forecast Center (CNRFC) as training datasets, the authors develop a deep-learning postprocessing framework based on the U-Net architecture. Two model variants are introduced: U-Net-log, which uses a log-cosh error loss tailored to heavy-tailed FZL errors, and U-Net-GMM, which employs a Gaussian mixture model loss to better represent multimodal FZL distributions.

Figure 1. Performance metrics of 6-hourly WRF FZL, linear regression, and U-Net postprocessing (U-Net-RMSE, U-Net-GMM, and U-Net-log) forecasts during December 2016–March 2017 over the Yuba–Feather watershed: (a) Pearson’s correlation coefficient and (b) RMSE vs CNRFC observations as a function of lead time (hours). Figure 4 from Afzali Gorooh et al. (2026).

Across two independent winter seasons (2015–16 and 2016–17), U-Net-based postprocessing reduces the centered RMSE by ~20% and increases correlation by about 10% relative to the raw West-WRF forecasts over the Yuba–Feather watershed (Figure 1). Continuous Ranked Probability Score (CRPS) diagnostics for U-Net-GMM show consistent improvements across lead times up to five days, indicating the benefit of the probabilistic aspects of the forecast. The models also enhance FZL skill during strong AR events and across multiple elevation bands from low- to high-elevation terrain, where baseline model errors and biases are largest, underscoring their relevance for FIRO-related operations and flood risk management.

Citation:

Afzali Gorooh, V., Sengupta, A., Roj, S., Weihs, R., Kawzenuk, B., Delle Monache, L., & Ralph, F. M. (2026). Enhancing Deterministic Freezing-Level Predictions in the Northern Sierra Nevada through Deep Neural Networks. Journal of Hydrometeorology, 27(2), 233-255. https://doi.org/10.1175/JHM-D-25-0052.1

Additional References:

Sumargo, E., Cannon, F., Ralph, F. M., & Henn, B. (2020). Freezing level forecast error can consume reservoir flood control storage: Potentials for Lake Oroville and New Bullards Bar reservoirs in California. Water Resources Research, 56(8), e2020WR027072. https://doi.org/10.1029/2020WR027072