Abstract:

Short-term soil moisture prediction is essential for drought monitoring, land “surface model initialization, and improving the accuracy of NOAA’s environmental forecasting systems. This NERTO project was motivated by the operational need to better represent soil moisture dry-down behavior in NOAA models such as HRRR and Noah-MP, which rely on accurate land-surface states but often lack sufficient observational constraints. The study focused on U.S. Climate Reference Network (USCRN) stations across the South west, a region characterized by highly dynamic soil moisture conditions and strong land “atmosphere coupling, making it an ideal domain for evaluating model skill. Using USCRN multi-depth soil moisture, meteorological data, and soil characteristics, I developed a data-driven Long Short-Term Memory(LSTM) framework to forecast soil moisture recession at depths of 5″100 cm. The model used a 7-day lookback window to generate 1″30-day ahead predictions. Results show that the LSTM captures depth-dependent drying behavior, achieving KGE values of 0.87″0.97 and RMSE values of 0.015″0.029, with the highest skill in deeper soil layers where moisture changes more gradually. The model generalized well across stations and reproduced realistic dry-down trajectories, revealing consistent patterns between soil depth, atmospheric drivers, and forecast skill. These findings demonstrate the value of machine-learning approaches for improving short-term soil moisture prediction and highlight their potential to support NOAA operations by enhancing land-surface model initialization, bias detection, and drought-related decision-making. The results presented are from the NOAA EPP/MSI CSC NERTO graduate internship project conducted under the mentorship of Dr. Dave Turner, NOAA/OAR/Global Systems Laboratory, and Dr. Michael Barlage, NOAA/NWS/Environmental Modeling Center. This NERTO experience aligns with the NOAA Cooperative Science Center in CESSRST-II, supporting the Center’s goal of conducting NOA A mission-aligned research to understand and predict changes in land and water systems. This project, titled Short-Term Soil Moisture Dry-Down Prediction Using LSTM to Enhance Land-Surface Model Initialization and Verification, addressed the research question: How accurately can an event-based, multi-depth LSTM framework forecast soil moisture recession across USCRN sites, and how can these predictions support NOAA land-surface modeling? The work provides value to the scientific community and program stakeholders by demonstrating a scalable, data-driven approach for improving soil moisture representation in NOAA modeling systems, supporting drought monitoring, environmental prediction, and land” atmosphere process understanding.