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

Abstract

In semi-arid regions, daily relative humidity exhibits strong nonlinear and regime-dependent variability, which limits the robustness of conventional predictive approaches. This study investigates the influence of regularization on predictive stability and generalization within two supervised learning frameworks: Support Vector Regression (SVR) and Radial Basis Function Neural Networks (RBF-NN), using long-term meteorological observations from Fez, Morocco (1985-2022). Seven atmospheric predictors representing thermodynamic, radiative, and moisture-related processes were used as inputs. For SVR, hyperparameters were optimized using Bayesian optimization with cross-validation. While RBF-NN parameters and regularization were determined using grid-search and hold-out validation. Both models were evaluated using an identical data partitioning scheme (70% training, 15% validation, 15% testing) to ensure a consistent comparison. The results indicate that regularization effects differ across the two frameworks. In SVR, the Gaussian RBF kernel achieves the highest predictive performance (R = 0.9890; MSE = 0.0016), reflecting strong representation of relative humidity dynamics under global regularization through parameter C. In contrast, the RBF-NN model exhibits higher sensitivity to regularization parameter selection due to its localized learning structure, with the optimal configuration (λ = 0.01; architecture [7-16-1]) with (R = 0.9603; MSE = 0.0141), while deviation from this range leads to overfitting or underfitting. These results demonstrate that SVR provides more stable predictive performance than RBF-NN under identical climatic conditions. The study contributes to the understanding of how model structure and regularization influence relative humidity prediction under semi-arid climatic conditions.