DroughtCast: A Machine Learning Forecast of the U.S. Drought Monitor
DroughtCast is a machine learning model forecast of the U.S. Drought Monitor (USDM) that provides accurate predictions of USDM drought categories over the contiguous United States with up to a 12-week effective lead time and 9-km spatial resolution. The model uses antecedent SMAP soil moisture, gridMet meteorology, and USDM conditions within a Recurrent Neural Network (RNN) framework to forecast the USDM. The resulting model forecasts are robust, with a reported error of less than 1 USDM category, even for the longest (12-week) lead time. Of all model inputs, precipitation, SMAP soil moisture, and surface air temperature are the most important in producing accurate model forecasts. DroughtCast was designed to strengthen national resiliency to drought by providing accurate and timely USDM forecasts to assist natural resource managers, farmers, and government agencies in making informed decisions about drought risk. The RNN architecture leverages operational satellite data and allows for continuous rapid updates well-suited for operational applications.
Example forecast from DroughtCast. The top row shows the actual USDM conditions, and the bottom row shows the corresponding model forecasts at 2-, 4-, 8-, and 12-week lead times.
Brust, C., Kimball, J.S., Maneta, M.P., Jencso, K., & Reichle, R.H., DroughtCast: A Machine Learning Forecast of the U.S. Drought Monitor. 2021. Frontiers in Big Data, doi:10.3389/fdata.2021.773478
Example USDM Forecasts:
- Colin Brust
- John Kimball
- Arthur Endsley