Regularization effects on machine learning models for relative humidity prediction under semi-arid climate conditions

Area of Science:

• * Climatology and Meteorology
• * Machine Learning Applications in Environmental Science

Background:

• * Daily relative humidity in semi-arid regions shows complex, nonlinear variability, challenging traditional prediction methods.
• * Supervised learning models like Support Vector Regression (SVR) and Radial Basis Function Neural Networks (RBF-NN) are explored for improved humidity prediction.

Purpose of the Study:

• * To investigate the impact of regularization on the predictive stability and generalization of SVR and RBF-NN models.
• * To compare the performance of SVR and RBF-NN in predicting relative humidity using meteorological data from Fez, Morocco.

Main Methods:

• * Utilized long-term meteorological observations (1985-2022) with seven atmospheric predictors.
• * Optimized SVR hyperparameters using Bayesian optimization and RBF-NN parameters via grid-search.
• * Employed a consistent data partitioning scheme (70% training, 15% validation, 15% testing) for model evaluation.

Main Results:

• * SVR with a Gaussian RBF kernel achieved superior performance (R=0.9890, MSE=0.0016), demonstrating robust humidity dynamics representation.
• * RBF-NN showed higher sensitivity to regularization, with optimal performance at λ=0.01 (R=0.9603, MSE=0.0141), prone to overfitting/underfitting.
• * SVR exhibited more stable predictive performance than RBF-NN under identical semi-arid climatic conditions.

Conclusions:

• * Regularization effects differ significantly between SVR and RBF-NN, influencing their predictive stability.
• * SVR offers more reliable relative humidity predictions in semi-arid environments compared to RBF-NN.
• * Findings enhance understanding of model structure and regularization for climate prediction.