Investigating the performance of random oversampling and genetic algorithm integration in meteorological drought forecasting with machine learning

Drought is one of the most significant natural hazards worldwide, affecting water resources, agriculture, and socio-economic stability. Accurate drought prediction is essential for sustainable water management. However, traditional drought monitoring approaches are limited in dealing with data imbalances and capturing complex temporal patterns. Therefore, this study aims to evaluate the effectiveness of machine learning methods for meteorological drought estimation and to integrate Random Oversampling (ROS) and Genetic Algorithm (GA) methods to improve estimation accuracy. To achieve this objective, monthly rainfall data from the Isparta, Eğirdir, Senirkent, Uluborlu, and Yalvaç stations, positioned in Türkiye’s Lakes Region, were utilized to compute the Standardized Precipitation Index (SPI) over 3-, 6-, 9-, and 12- month intervals. Machine learning (ML) models were developed for Isparta drought estimation using SPI values, and the best performance was observed with Extra Tree Regression (ETR) models. Data imbalances were eliminated by ROS, and hyperparameter tuning of ETR models was performed using GA. The performance of ROS-GA-ETR hybrid models developed for 3-, 6-, 9-, and 12-month periods increased by 48%, 25%, 20%, and 21%, respectively. The study found that the integration of ROS significantly enhanced data balance, leading to more robust model training, while the use of GA for hyperparameter tuning consistently improved model accuracy. Among all tested models, ETR exhibited superior performance (R2 value 0.97 for 12-month period), highlighting its suitability for drought estimation tasks in the study area. The flexibility and generalizability of the methods used in the study make them easily applicable in different geographical regions and climatic conditions. Additionally, these approaches can significantly contribute to water resources management, agricultural planning, and drought risk management in Isparta and globally by providing more accurate and timely drought estimations. Improved estimation accuracy can support decision-makers in developing proactive strategies, optimizing water allocation, reducing agricultural losses, and enhancing the resilience of communities against drought-related impacts. In the future, it is recommended to apply these methods with larger data sets and in different regions.