Abstract:
Accurate modeling of climate variability is critical for understanding the impacts of climate change and
supporting data-driven adaptation strategies. Traditional parametric models, while widely used, often struggle
to capture the complex non-linear relationships and spatial dependencies that characterize climate systems, especially in regions with diverse geography such as Kenya. This study aimed to apply two non-parametric
statistical approaches—Generalized Additive Models (GAM) and Gaussian Process Regression (GPR)—to
model spatial and non-linear trends in climate data over Kenya. Daily climate variables, including
temperature and precipitation, were obtained from the ERA5-Land dataset using Google Earth Engine,
spanning the period from 2015 to 2024. GAM was used to model the smooth effects of covariates such
as time, elevation, and precipitation, while GPR was implemented using a Mat´ern covariance kernel to
capture residual spatial autocorrelation. The models were evaluated using RMSE, MAE, and 2
, and
parameter estimation was conducted via penalized likelihood and L-BFGS optimization techniques. The
results demonstrated that GAM effectively captured structured non-linear effects and provided interpretable
smooth functions, while GPR performed better in modeling spatial variability and uncertainty. Both
models outperformed traditional linear approaches, with GPR offering superior accuracy in areas with
high spatial heterogeneity. The findings affirm that GAM and GPR are powerful and complementary
tools for climate modeling in complex environmental contexts. In conclusion, this study confirms the
suitability of non-parametric approaches for climate modeling in data-rich, spatially heterogeneous settings.
Further research is recommended to explore integrated hybrid GAM–GPR models, extend the methodology to
multivariate climate indicators, and evaluate its performance in other regions or under future climate scenarios.
