Class Imbalance in Machine Learning for Weather and Rainfall Prediction: A PRISMA-Aligned Systematic Review

Machine learning is increasingly used for rainfall and weather-event prediction, yet meteorological classification datasets are commonly dominated by non-event observations. This imbalance creates a methodological problem: a classifier may achieve high overall accuracy while failing to identify the rainy, severe, or otherwise consequential class. This PRISMA-aligned systematic review synthesizes evidence on class imbalance in machine-learning-based weather prediction, emphasizing rainfall classification, balancing methods, traditional classifiers, and performance measurement. Eligibility criteria covered English-language studies on data-driven weather or rainfall prediction that reported classifier development, imbalance treatment, or minority-sensitive evaluation. The evidence was screened and organized around four questions: why weather classes become imbalanced, how imbalance affects learning algorithms, what balancing methods can and cannot achieve, and which evaluation measures provide credible evidence of minority-event skill. The literature indicates that classifier superiority is sensitive to class prevalence, validation design, leakage control, decision threshold, and metric choice. Synthetic oversampling can improve recall and F1-score, but it may reduce precision, increase false alarms, distort local structure, or produce optimistic estimates if applied before data splitting. The review proposes a leakage-safe comparative framework in which balancing is restricted to training data and all models are evaluated on one untouched distribution. It concludes that imbalance should be treated as a central property of weather prediction rather than a secondary preprocessing issue.