Improving Atmospheric River Forecasts With Machine Learning

This study tests the utility of convolutional neural networks as a postprocessing framework for improving the National Center for Environmental Prediction's Global Forecast System's integrated vapor transport forecast field in the Eastern Pacific and western United States. Integrated vapor transport is the characteristic field of atmospheric rivers, which provide over 65% of yearly precipitation at some western U.S. locations. The method reduces full‐field root‐mean‐square error (RMSE) at forecast leads from 3 hr to seven days (9–17% reduction), while increasing correlation between observations and predictions (0.5–12% increase). This represents an approximately one‐ to two‐day lead time improvement in RMSE. Decomposing RMSE shows that random error and conditional biases are predominantly reduced. Systematic error is reduced up to five‐day forecast lead, but accounts for a smaller portion of RMSE. This work demonstrates convolutional neural networks potential to improve forecast skill out to seven days for precipitation events affecting the western United States. Machine learning methods are data‐driven algorithms that improve by examining massive amounts of existing data. We explore the utility of a computer‐vision machine learning technique to reduce error in numerical weather forecasts of the characteristic field for atmospheric rivers (ARs). ARs are long narrow corridors of anomalous vapor transport capable of providing both beneficial and hazardous precipitation. Therefore, accurately forecasting AR events is extremely important from a water supply and flood protection standpoint. We show significant forecast improvements by applying machine learning postprocessing for lead times ranging from 3 hr to seven days, making the predictions more valuable to stakeholders affected by AR events. The GFS forecast field of integrated vapor transport is used for a convolutional neural network‐based forecast postprocessing methodThe machine learning algorithm reduces the full‐field root‐mean‐square error and improves the correlation with ground truthAn error deconstruction shows that the dominant improvements come from the reduction of random error and conditional biases