A Hybrid GPR-GAM Model for Enhanced Spatio-Temporal Climate Prediction in Kenya

Abstract

Climate change presents growing challenges in regions like Kenya, where diverse terrain and climatic variability complicate accurate environmental forecasting. Traditional climate models often fall short in capturing both the non-linear relationships among climatic variables and the spatial dependencies inherent in such data. To address these limitations, this study introduces a novel hybrid model that integrates Gaussian Process Regression (GPR) and Generalized Additive Models (GAM) to enhance spatio-temporal climate prediction.The model was developed by combining the structured, interpretable components of GAM with the spatially aware, probabilistic strengths of GPR, using climate data collected from the Google Earth Engine covering the period 2015–2024. Model parameters were estimated through generalized cross-validation and optimized using the L-BFGS algorithm. Results indicate that the hybrid model significantly improves predictive accuracy compared to standalone GPR or GAM approaches, achieving an RMSE of 1.27°C and an R² of 0.91. These findings demonstrate the model’s effectiveness in capturing Kenya’s spatial and climatic heterogeneity. The study recommends the hybrid model’s application in climate-sensitive sectors such as agriculture, infrastructure development, and early warning systems, with future work focusing on scalability and real-time deployment.