Comparing Regression Techniques for Temperature downscaling in Different Climate Classifications

In recent years, downscaling techniques have emerged as practical methods in numerous fields, including climatology trend simulation. Therefore, identifying the optimal regression technique is critical for assessing, simulating, and predicting climate patterns. This paper aims to identify the optimum regression techniques for downscaling among ten commonly used methods in climatology, including SVR, LinearSVR, LASSO, LASSOCV, Elastic Net, Bayesian Ridge, RandomForestRegressor, AdaBoost Regressor, KNeighbors Regressor, and XGBRegressor. Historical data from Can_ESM5 were collected from four synoptic stations based on the Köppen climate classification. To achieve this goal, the data were classified based on the Köppen climate classification system, including A (tropical), B (dry), C (temperate), and D (continental). Additionally, to enhance the performance of downscaling accuracy, eliminate redundant information, and overcome overfitting, Mutual Information (MI) was employed on the Can-ESM5 dataset. The downscaling performance was evaluated using the Coefficient of Determination (DC) and Root Mean Squared Error (RMSE). In conclusion, SVR had superior performance in tropical and dry climates with DC and RMSE values of 0.89, 0.02 °C and 0.93, 0.01 °C, respectively. On the other hand, LassoCV with RandomForestRegressor had the best results in temperate and continental climates with DC and RMSE values of 0.87, 0.04 °C and 0.88, 0.03 °C, respectively. The outstanding performance of the optimum downscaling methods relies on the network structure in consideration of the suitability of those networks with the target variable and climate type.