Abstract
Drought prediction plays a vital role in mitigating the adverse effects of climate variability and ensuring sustainable water resource management. In order to classify drought in the northern Iranian provinces of Golestan, Mazandaran and Guilan, this study assessed the effectiveness of machine learning models such as random forest (RF), AdaBoost, decision tree (DT) and transformer. Historical climate data were preprocessed to include lagged features and statistical aggregates for tree‐based models, while normalised data were directly used for the transformer model to capture temporal dependencies. F1‐score, recall, accuracy and precision were used to evaluate the model's performance. The results revealed that RF consistently outperformed other models across all regions, demonstrating superior accuracy, precision and recall. AdaBoost followed closely, while DT provided moderate performance. The transformer model showed limited effectiveness, particularly in Guilan and Mazandaran. Optimal hyperparameters were determined for each model, ensuring robust evaluation and providing a benchmark for future studies. The findings underscore the effectiveness of RF in drought prediction and highlight the regional variability in model performance. This research emphasises the importance of model selection and tuning in achieving reliable predictions and offers insights into the application of machine learning techniques for drought monitoring. Future research should explore advanced models, such as deep learning or hybrid approaches, and consider additional climatic variables to enhance predictive accuracy further.