Modeling the Onset of Drought Periods Using Explainable Machine Learning Models Enhanced by Bayesian Optimization

This study develops an optimized machine learning-based computational framework for assessing drought conditions in water-scarce regions. The pattern of drought periods is highly non-linear, especially in arid climates, hindering water management planning. The proposed framework integrates a precipitation index (PI) with Gaussian process regression, regression tree, and ensemble tree models to predict drought occurrences using an optimal autoregressive modeling approach specifically designed for areas with limited precipitation. A historical precipitation data set from Kuwait International Airport (1958–2020) was used. Kuwait is characterized by insufficient rainfall and vulnerability to water shortages. To validate the autoregressive model, the historical PI time series data sets were analyzed for stationarity using the Mann–Kendall test. The partial autocorrelation function test revealed a strong relationship with lagged PI back to 4 and 3 months for 12- and 24-month drought monitoring periods, respectively. Partial autocorrelation was compared to other feature selection techniques such as RReliefF and minimal redundancy maximal relevance. A Bayesian optimization algorithm tuned the model parameters to boost results reliability. The Gaussian process regression model outperformed other machine learning (ML) models, achieving a strong correlation between observed and predicted drought events, with coefficients of determination (R2) ranging from 0.87 to 0.92, and mean absolute error ranging from 0.107 to 0.052 for the respective 12- and 24-month monitoring periods. The SHapley Additive exPlanations (SHAP) technique was used to interpret the relevance of model forcing and enhance the transparency of the results. The SHAP analysis results emphasize the sensitivity of the PI value to previous months’ drought status, in particular, the 1-month lagged index exhibited the most significant contribution to the ML models’ performance at the 12- and 24-month timescales. This computational framework aims to provide water managers in water-scarce regions with effective and reliable tools for drought monitoring to assist in developing resilient predictive water management strategies.