A NEW MACHINE LEARNING TECHNIQUE FOR ACCURATELY FORECASTING EXTREME PRECIPITATION FROM LANDFALLING ATMOSPHERIC RIVERS

Atmospheric rivers transport large volumes of water vapor outside the tropics and when landfalling, produce large quantities of rain that replenish aquifers, contribute to beneficial increases in snowpack, yet can cause flooding and damage. Accurate forecasts of precipitation during landfalling atmospheric rivers are critical because they play a large role in water supply and flooding. However, the intensity, location, and duration of atmospheric rivers are poorly forecast in all currently operational numerical weather prediction models beyond forecast Day 7, resulting in a significant decision support services gap. We will employ machine learning techniques in developing a family of products that improve Earth system decision support models in general and the outlook for landfalling atmospheric rivers at 7-21 days in particular. Today’s numerical weather prediction models produce skillful forecasts accurately predicting atmospheric rivers one week in advance, with rapidly diminished skill in Week 2. There exists additional information in the observing systems’ datasets that are complementary to traditional weather forecasts and can significantly extends the accuracy and precision of the predictions—in terms of timing, location, intensity, and duration of landfalling atmospheric rivers—when these data are combined with deterministic or ensemble model predictions through the use of machine learning techniques.