Misclassification-Aware Hybrid Model for Binary Rainfall Prediction

Abstract

Hybrid framework for rainfall classification that integrates machine learning (ML) and deep neural network (DNN) applied on meteorological data. The primary DNN serves as robust baseline trained with normalized meteorological features from different sites in the target area after simple preprocessing such as dropping unnecessary features, using MinMaxScaler for feature scaling. To further reduce misclassification samples, a correction mechanism was applied using secondary Light Gradient Boosting Machine (LightGBM) model, which was trained only on misclassified samples by the DNN model.

ROC curve (AUC = 0.98), confusion matrix and precision recall curve proved the model differentiation ability. Furthermore, KMeans clustering highlighted the class strong separation for rain / no-rain classes. Learning curves suggested stable training optimization with consistent generation.

This hybrid method achieved overall accuracy of 98%, while both precision and F1-score were 98% with 96% recall.

As the results proves that using DNN and second ML model specifically for the misclassified samples can boost the rainfall prediction model.