Deep Learning Time Series Prediction Strategies for Efficiently Emulating Noah Land Surface Model Soil Moisture Dynamics

Land surface models including the North American Land Data Assimilation System (NLDAS) run of the Noah soil hydrology model are crucial for monitoring water resources and crop health, operational awareness surrounding drought and flood conditions, and for initializing numerical weather prediction algorithms. These models must evaluate heavily parameterized differential equations at small time steps in order to remain numerically stable when discretely integrated, so the computational cost of running models on large domains such as the 27,245,580 point NLDAS-3 grid can be prohibitive, especially for near real-time ensemble forecasts. Sequence to sequence neural networks including LSTMs and 1D CNNs offer a way to represent complex relationships between input forcings and output states as a composition of efficient matrix operations, and can make predictions at coarser time steps by learning the nonlinear form of the already-integrated governing equations. As such, this presentation will explore the accuracy and efficiency of deep learning techniques for emulating Noah-LSM's 4-layer soil water dynamics at multiple temporal resolutions out to a 2-week horizon, using NLDAS-2 forcings over the conterminous United States,. Emphasis will be placed on model architectures, loss function properties, and learning strategies that contributed to the best results.