Large-scale drought forecasting in the U.S. Sothern Plains through a hybrid cluster-based wavelet-machine learning approach

Authors: LEE, SANGHYUN - Oak Ridge Institute For Science And Education (ORISE); DANANDEH MEHR, ALI - Antalya Bilim University; Moriasi, Daniel; MIRCHI, ALI - Oklahoma State University

Submitted to: Water Resources Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/15/2025
Publication Date: 11/8/2025
Citation: Lee, S., Danandeh Mehr, A., Moriasi, D.N., Mirchi, A. 2025. Large-scale drought forecasting in the U.S. Sothern Plains through a hybrid cluster-based wavelet-machine learning approach. Water Resources Research. https://doi.org/10.1029/2024WR039744.
DOI: https://doi.org/10.1029/2024WR039744

Interpretive Summary: This study presents a new method to improve drought forecasting for the U.S. Southern Plains, encompassing Kansas, Oklahoma, and Texas, aiming to provide early warnings and support water planning efforts. Drought forecasts are challenging due to the complexity of processing large climate datasets, especially when computing power is limited. We resolved this issue by identifying 21 regions, each consisting of areas with similar drought conditions, and trained machine learning models only for these regions instead of training all areas. The region-specific trained models are then used to predict drought conditions for the entire homogeneous areas for a given region. Our approach captures essential climate patterns by combining two techniques: a method which helps the model to recognize important weather patterns, and a second method for selecting the best climate data inputs based on local climate conditions. The results showed this model accurately predicts drought conditions across the Southern Plains region one to three months in advance when compared with observed drought maps; with accuracies ranging from 71 to 92%. Our findings suggest that the presented modeling framework can serve as a valuable tool for creating drought forecast maps to help farmers and other water resource managers prepare for droughts in a timely manner.

Technical Abstract: High-resolution gridded datasets provide valuable opportunities to enhance drought forecasting, but applying complex machine learning algorithms across large spatial domains is computationally challenging. This study presents a novel hybrid approach for forecasting the gridded Standardized Precipitation-Evapotranspiration Index (SPEI) across the U.S. Southern Plains (SP), with lead times of 1 and 3 months. We developed a clustering-based method using 21 centroid grid cells, each representing a unique cluster of similar grid cells based on various climatic characteristics, to train and evaluate multilayer perceptrons (MLPs), long short-term memory (LSTM), and genetic programming (GP). Based on the superior performance of the trained MLPs in terms of Nash-Sutcliffe efficienty and root-mean-square error, they were extended to corresponding grid cells for each cluster, enabling spatially adaptive drought prediction at a high resolution. The use of discrete wavelet transform (DWT) further enhanced model accuracy by capturing key temporal patterns in the SPEI series. Notably, our results showed that physical and climatic attributes strongly influenced input selections. While a 12-month lag period worked well in regions with weaker seasonality, areas having strong seasonality benefited from selection of effective lags by using mutual information. For three-month-ahead forecasts, including decomposed potential evapotranspiration in addition to precipitation as inputs improved accuracy in drier regions but decreased accuracy in humid areas. The forecast maps based on the hybrid DWT-MLP models effectively captured the spatial variability of drought, with high correlations to observed values, demonstrating their effectiveness for regional drought early warning systems to inform water resources management adaptations.