Abstract ID: 243
Forecasting Subseasonal Extreme Precipitation in the Contiguous United States Using a Convolutional Neural Network
Lead Author: Ty Dickinson
University of Oklahoma School of Meteorology, United States of America
Keywords: extreme precipitation, artificial intelligence
Abstract: Extreme precipitation events across multiple timescales are natural hazards that poses a significant risk to life, with a commensurately large cost through property loss. Unfortunately, extreme precipitation, especially extended-duration precipitation, remains as one of the most challenging hazards to forecast. This study uses a convolutional neural network and novel explainability methods to explore the forecasting capability and important features leading to these extreme precipitation events. We examine events from a publicly available database of 14-day extreme precipitation events across the contiguous United States (CONUS) created by the authors. Thresholds for both total precipitation and the duration of the precipitation are used to identify events with sufficient length to accentuate the synoptic and subseasonal contributions to the extreme event. Using the developed database, our neural network is subsequently trained on binary occurrences of extreme conditions throughout the CONUS. Important precursor variables identified using lag compositing, including geopotential height, horizontal wind components, and specific humidity, were used as predictors for the neural network. The model is trained on ERA5 data between 1 January 1950 and 31 December 1997, while next two decades are split roughly in half for validation and testing. Qualitatively, the model captures the general region of extreme precipitation in the testing dataset among several examples. The neural network capably identified extreme precipitation events related to both extratropical cyclones (e.g., December 2015 in the Pacific Northwest) and tropical cyclones (e.g., Hurricanes Florence and Joaquin); however, the model comparatively struggled on events comprised purely of wintry precipitation. Skill is quantified in an object-oriented framework (e.g., centroid offset, area ratio, etc.) and compared across regions in the CONUS. We also employ explainability methods such as layerwise relevance propagation and to glean physical insights from our developed model. These methods are consistent with previous findings that emphasize the importance of mid-tropospheric circulation patterns responsible for driving extreme precipitation events, including an upstream anomalously deep trough and anomalous strengthened zonal flow. Details on the positioning and orientation of these features depend on the location of the extreme precipitation event in the CONUS. Feedback from product users and its impact on model architecture and product design are also discussed as time allows.
Co-authors:
Jason C. Furtado (University of Oklahoma School of Meteorology)
Michael B. Richman (University of Oklahoma School of Meteorology)
