Drought prediction and understanding the drivers of drought development using a machine learning approach

Over the past decade, droughts have drawn increasing attention due to their substantial agricultural and economic consequences, particularly in the U.S. Great Plains area (e.g., the 2012 Central US event and the 2017 Northern Plains event). Although certain large-scale atmospheric and oceanic patterns are necessary for drought development, land- atmosphere interactions can play an important role in the intensification of drought conditions, especially for flash drought. This study aims to predict drought conditions over the U.S. Great Plains at 1-3-week lead times using a convolutional neural network (CNN) model. To forecast drought categories derived from the US Drought Monitor (USDM), the models are trained using multi-source atmospheric and land-surface variables, including 500 hPa geopotential height, precipitation, wind speed, surface radiation, humidity, and temperature from ERA5, soil moisture from Global Land Evaporation Amsterdam Model (GLEAM) and North American Land Data Assimilation System (NLDAS), and Normalized Difference Vegetation Index (NDVI) from satellite products. Model performance is evaluated to unravel the atmospheric and land-surface processes that drive droughts at different lead times and identify their relative contributions to drought development and intensification.