Improve Precipitation Zoning Accuracy by Applying Ensemble Learning Models Based on Interpolation and Data Mining Integration

Precipitation is a crucial and variable meteorological phenomenon, widely used in various models, including rainfall-runoff models, and the accuracy of its spatial estimation significantly impacts the results of these models. In the present research, an Ensemble learning model (ELM) was developed based on a step-by-step tree regression to improve the accuracy of precipitation zoning in Golestan, one of the northern provinces of Iran, which has irregular spatial changes, and also to improve the prediction results with integrating six methods of interpolation and two data mining techniques. The results showed that among the interpolation methods, the empirical Bayesian kriging method with RMSE = 132 mm was the most accurate. Using the covariate variable of altitude due to the non-significance of the gradient of precipitation did not improve the results of the empirical 3-D cokriging and Bayesian kriging methods. Instead, combining the results of different interpolation methods with ELM reduced the RMSE value to 110 mm. Spatial estimation using the spatial coordinates of the stations with the M5 model tree by dividing the region into 4 precipitation zones resulted in a prediction with RMSE = 128.7 and R = 0.8. Combining its results with the ELM model of interpolation methods in the final ensemble learning model caused the spatial prediction error to decrease again with the RMSE value to 85.3 mm, the error percentage to 21, and the accuracy of precipitation zoning to increase.