Uncovering Drought Causative Mechanisms to Develop Long-Lead Drought Forecasting Using Entropy and Scientific Machine Learning Models for the Continental U.S.

Abstract

Drought forecasting can be facilitated by understanding drought causative mechanisms that unveil predictors which can explain much variability of drought. However, our understanding of drought causative mechanisms is less than complete. Therefore, this study uncovered the impacts of Rossby wave packets (RWPs) and atmospheric rivers (ARs) on meteorological drought to contribute to the understanding of implications of internal atmospheric variability that are least understood. Then, Burg, configurational, and relative entropy spectral analyses, respectively, hereafter called BESA, CESA, and RESA, which are equipped with maximal overlap discrete wavelet transform (MODWT), were developed for long-term meteorological drought forecasting for the Continental United States (CONUS). For soil moisture drought forecasting for CONUS, neural networks embedded in ordinary differential equations (NeuralODEs) belonging to Scientific Machine Learning (SciML) were developed. The increase in the occurrence of recurrent RWPs (RRWPs) with low phase speed escalated drought severity, area, and duration. Furthermore, the drying ARs affected drought by increasing the severity and area of extreme droughts. Besides, although the drying ARs contributed more to extreme droughts than did RRWP, the co-occurrence of ARs and RRWPs contributed more than did each individually. Among entropy-based forecasting models, BESA outperformed CESA and RESA. BESA forecasted self-calibrated Palmer Drought Severity Index (scPDSI) up to median lead times of 12 months with RMSE less than 0.8 and a modified NSE greater than 0.6 across CONUS. Although RESA forecasted slightly less accurate scPDSI than BESA, it forecasted the detailed behavior better than did BESA at a longer lead time and showed applicability for long-term drought forecasting. However, CESA lost its accuracy faster than other entropy-based models as lead time increased. NeuralODEs forecasted soil moisture drought index (SMDI) and were benchmarked against long short-term memory (LSTM) neural networks and LSTM stacked with a one-dimensional convolutional layer (C-LSTM). NeuralODEs outperformed the LSTM-based models at all lead times, covering 80% of the watersheds for a 6-month lead time or more. Besides, risk analysis based on the copula methods for drought characteristics forecasted up to a 12-month lead time showed the potential for applicability of BESA and NeuralODEs to drought management.