Class Imbalance-Aware SVM Optimization to Improve Daily Rainfall Event Prediction

Abstract

Daily rainfall events can disrupt port operations. This study evaluates class imbalance-aware optimization of an SVM-RBF rainfall-event classifier using Madura Strait sea surface temperature anomalies, the Oceanic Niño Index, and cyclic seasonal features for 2015-2024. The 2015-2022 development set (n = 2,913) comprised 1,917 non-rain (65.8%) and 996 rain (34.2%) observations, while the independent 2023-2024 test set (n = 731) comprised 518 non-rain (70.9%) and 213 rain (29.1%) observations. Data from 2015-2021 were used for initial fitting and 2022 for probability-threshold selection before refitting on the full development period. Five configurations were compared: baseline SVM-RBF, class weighting, threshold tuning, class weighting with threshold tuning, and random oversampling with threshold tuning. Under the predefined F1-first ranking criterion, the class-weighted model achieved F1-score 0.640, recall 0.676, balanced accuracy 0.748, and MCC 0.482, compared with 0.581, 0.531, 0.704, and 0.435 for the baseline. Threshold-tuned models achieved higher recall but more false alarms. Thus, class weighting is the preferred default for this dataset under the stated selection criterion.