Performance Evaluation of Machine Learning Methods for Real-Time Rainfall Classification

Abstract

Reliable real-time rainfall intensity classification is essential for supporting early warning systems and disaster mitigation, particularly in regions vulnerable to hydrometeorological hazards. This study evaluates three machine learning algorithms SVM, Neural Network, and AdaBoost for multiclass rainfall intensity classification using real-time data collected from Internet of Things (IoT)-based sensors. Rainfall intensity is categorized into four classes: no rain, light rain, moderate rain, and heavy rain, based on threshold values defined by BMKG standards. The dataset is imbalanced and dominated by the no rain class, therefore, model performance is evaluated using imbalance aware metrics, including per-class precision and recall, macro F1-score, balanced accuracy, and overall accuracy. Experimental results show that SVM and Neural Network achieve very high overall accuracy of up to 99.46%, however, this performance is mainly influenced by accurate classification of the majority class, leading to low recall for minority rainfall classes. In contrast, AdaBoost provides a more balanced baseline performance, achieving an accuracy of 92.4% and a macro F1-score of 0.714 on the original dataset. To enhance minority class detection, the SMOTE is applied to the training data using an 80:20 train test split. After data balancing, AdaBoost demonstrates improved recall and macro F1-score for light and moderate rain classes, although overall accuracy decreases to 77.1%. These results are acceptable for early warning applications, where sensitivity to rainfall onset is prioritized over majority class dominance. Consequently, balanced AdaBoost, evaluated using time-based data partitioning and imbalance aware metrics, is considered an effective approach for real-time IoT-based rainfall classification.