Comparative robustness and interpretability analysis of MLP and random forest for multi-class weather classification in tropical regions

Penulis : Yoyok Dewantoro,Joko Triloka,Ahmad Zarkoni,Tomy Ivan Sugiharto

The stability of machine learning models remains a critical challenge in weather type classification, particularly when applied to low-variance meteorological datasets characterized by correlated atmospheric parameters. Neural network models, such as the Multi-Layer Perceptron (MLP), are known to capture nonlinear relationships effectively but may be sensitive to training–testing data partitioning. In contrast, ensemble methods like Random Forests (RFs) are designed to reduce variance through aggregation. This study systematically evaluates the stability and generalization capability of MLP and RF classifiers for multi-class weather classification (sunny, cloudy, rainy) using historical meteorological data. Experiments were conducted under three data split scenarios (80:20, 70:30, 60:40) and validated using 5-fold and 10-fold cross-validation. While MLP achieved accuracy above 96% across all scenarios, RF consistently outperformed MLP with accuracy between 98% and 99%. Importantly, cross-validation results reveal that RF demonstrates superior stability, with standard deviation values ranging from 0.00 to 0.01, compared to 0.01 for MLP. These findings confirm that ensemble-based methods provide more robust and consistent performance for meteorological classification tasks characterized by multivariate dependence and limited variance variability.