A Comparative Study of OLS and MLP Models for Spatial Downscaling of Air Temperature

Abstract. High-resolution and accurate air temperature data are critical for climate studies and environmental planning. This study evaluates and compares two spatial downscaling approaches for estimating monthly air temperature in Iran: Ordinary Least Squares (OLS) for its interpretability, and Multilayer Perceptron (MLP) for its predictive power. Both models were trained on data from synoptic stations using predictors such as geographical location, elevation, precipitation, and Normalized Difference Vegetation Index (NDVI). The trained models were then used to generate temperature grids at 5-km spatial resolution for four representative months. The results demonstrate that the MLP model consistently outperforms the OLS model, achieving a higher coefficient of determination (R² = 0.946 vs. 0.899 in May) and a significantly lower Root Mean Squared Error (RMSE). Furthermore, visual analysis of the generated maps reveals the MLP’s superior ability to capture complex local phenomena, such as the moderating effect of the Caspian Sea and sharp temperature gradients in mountainous regions. Overall, results show that while OLS serves as a simple global linear baseline, non-linear models such as MLP are essential for achieving higher accuracy and capturing geographically realistic temperature patterns.