Abstract:
Drought is one of the most ecologically and economically devastating natural phenomena affecting the United States. It is estimated that a single drought can cause the U.S. economy over a billion dollars in damage, and that sustained or repetitive drought can lead to widespread degradation of ecosystem health. Accordingly, many drought indices exist that monitor the current status of drought so that farmers, local governments, and other stakeholders can respond. However, methods for forecasting drought conditions weeks or months in advance are less common, limiting our ability to prepare ahead of time. The ability to forecast drought over weekly to monthly lead times may enable more effective drought risk reduction and mitigation, promoting greater resiliency to climate extremes that are expected to intensify with global warming. While effective methods of forecasting drought conditions do exist, they do not directly translate to existing drought indices, making it difficult to compare current drought conditions with model forecasts. Here, we introduce DroughtCast, a novel machine learning framework for forecasting the United States Drought Monitor (USDM). DroughtCast operates on the knowledge that recent anomalies in soil moisture and meteorology drive future changes in drought conditions over the continental USA domain. We use gridded surface meteorology interpolated from regional weather stations and satellite observed soil moisture data from the NASA SMAP (Soil Moisture Active Passive) mission as key inputs into a recurrent neural network machine learning framework to accurately forecast the USDM from 1 and 12 weeks into the future. Precipitation and soil moisture were found to be the most important features in forecasting drought conditions over a broad range of climate regimes relative to other model predictors. A regional case study of the 2017 Northern Plains Flash Drought also showed that the model was able to effectively forecast the timing, progression, and severity of the very extreme drought event up to 12 weeks in advance. The DroughtCast framework was found to provide efficient and accurate regional forecasts of the USDM with relatively long lead times to better inform drought mitigation planning.