Future drought prediction using time-series of drought factors and the US drought monitor data based on deep learning over CONUS

Predicting future drought conditions is crucial for preventing massive agricultural and/or hydrological resource damage caused by drought. This study predicts future (in this case, 3-month forecast lead time) drought conditions in the contiguous United States, especially focusing on five different dry and drought severity classes indicated by the United States Drought Monitor (USDM) during 2000-2020. A deep learning model was trained using the time-series of USDM and four different types of drought-related variables (i.e., hydro-meteorological variables) such as precipitation and temperature from Phase 2 of the North American Land Data Assimilation System. UNet, one of the image-to-image translation techniques, was used as a basic deep learning architecture to consider the spatial characteristics (extents of each drought severity class) of drought across the continent. As drought classes in USDM are ordinal, the loss function of the deep learning model was set to be able to consider ordinal problems utilizing the cross-entropy loss function. The results of the proposed model were compared to the existing seasonal drought outlooks provided by the National Oceanic and Atmospheric Administration Climate Prediction Center. The performance for the validation period (2 years) showed an overall accuracy of about 65%. When compared to the seasonal outlooks, it demonstrated about a 6% improvement in terms of overall accuracy for changing drought conditions. Future research will further discuss the performance of the proposed model with other comparable reference data and the impact of each input variable to predict future drought conditions.