Develop Machine-Learning Vegetation Phenology Model for UFS Regional Model to Advance S2S Forecasts of Extreme Hydrometeorological Events over Southwest US

The southwest US has experienced persistent, severe droughts and episodic, catastrophic floods in the past decade. Foreseeing the occurrence of these events on lead times beyond the medium range has been challenging for existing NWS forecast models. A potential contributor to the lack of forecast skills is the deficiency in model representations of land surface processes, including snow, soil moisture and vegetation. The project funded by NOAA Weather Program Office Innovation seeks to develop a machine-learning-based dynamic vegetation model, and couple the model with NOAA Unified Forecast System (UFS) Regional Model to produce experimental forecasts at lead times of week 4-8. The vegetation model is based on long-short term memory (LSTM), and it is trained using archived North American Land Data Assimilation System-2 (NLDAS-2) meteorological forcing in conjunction with MODIS normalized difference vegetation index (NDVI) product. The presentation offers an overview of vegetation phenology model development, the adaption of the UFS regional model workflow, and the experimental strategy for evaluating the impacts of the enhancement.