Improving Atmospheric River Forecasts with Machine Learning

August 27, 2019

CW3E graduate student Will Chapman, along with co-authors Aneesh Subramanian, Luca Delle Monache, Shang Ping Xi, and Marty Ralph, published a paper in Geophysical Research Letters entitled “Improving Atmospheric River Forecasts with Machine Learning”.

Machine learning methods are data-driven algorithms that improve by examining massive amounts of existing data. This study explores the utility of a computer-vision machine learning technique to reduce error in numerical weather forecasts of integrated vapor transport (IVT), 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. The study presents a forecast post-processing method (dubbed “ARcnn”), which relies on machine learning to correct the IVT field output from the National Center for Environmental Prediction’s Global Forecast System (GFS) for the Eastern Pacific and Western North America. Results show significant forecast improvements after applying machine learning postprocessing for lead times ranging from 3 hours to 7 days, making the predictions more valuable to stakeholders affected by AR events. Figure 1 shows an example of a 4-day forecasted AR event that is corrected by the machine learning method.

Figure 1: Forecasts and analysis valid for IVT fields on 29 November 2017. (a) MERRA-2 analysis field with the IVT = 600 kg m^-1s^-1 contour (solid) and dominant storm axis (dotted) as determined by IVT > 350 kg m^-1s^-1 raw image moment. (b) GFS 96-hour forecast with the MERRA-2 600 IVT contour and dominant storm axis. (c) ARcnn-IVT 96-hour forecast with the MERRA-2 600 IVT contour and dominant storm axis. (d) Difference between ARcnn-IVT and GFS. (e) Difference between GFS and MERRA-2 IVT field. (f) Difference between GFS and MERRA-2 IVT field.

The method offers improved prediction for integrated vapor transport events affecting the Western United States and could lead to better preparation for forecasted precipitation. Figure 2 (Figure 3 in the paper) shows the improvement of four error metrics over the raw GFS forecast. The method reduces full field root mean squared error (RMSE) at forecast leads from 3 hours to 7 days (9-17% reduction), while increasing correlation between observations and predictions (0.5-12% increase). This represents a ~1-2-day lead time improvement in RMSE at 7 days.

Figure 2: Region of Interest average temporal evolution of (a) Bias, (b) CRMSE, (c) RMSE, and (d) PC of raw GFS, ARcnn, persistence (Pers), and climatology (Climo) forecasts. Resampled bootstrap variance intervals are shown for each forecast.

Chapman, W. E., Subramanian, A. C., Delle Monache, L., Xie, S. P., & Ralph, F. M. (2019). Improving Atmospheric River Forecasts with Machine Learning. Geophysical Research Letters, 46. https://doi.org/10.1029/2019GL083662