USE a variety of meteorological and geographic data to build a LightGBM temperature predicted model and quantify the impact of feature on temperature changes based on SHAP

Temperature change is a complex atmospheric phenomenon. Even the ERA5-Land atmospheric reanalysis temperature dataset with the highest accuracy currently has errors with the actual observed temperature. This study captures the nonlinear relationship between ERA5-Land temperature and various geographic and climate data such as population density, land use type, DEM, air pressure, and relative humidity by constructing a LightGBM model. The experimental results show that compared with ERA5-Land, the constructed model has a Mean Absolute Error is reduced by 5% and Root Mean Squared Error is reduced by 13%, Mean Squared Error is reduced by 24%, Pearson correlation coefficient increased by 2%, R-squared is increased by 3%. A machine learning model with strong versatility and accurate prediction has been obtained, and the correction of daily temperature data in the Yellow River Basin from 2015 to 2019 has been realized. After the model training is completed, SHAP is used to analyze the impact of various feature data on the prediction of temperature by the machine learning model, sort them according to the degree of impact, and analyze the interaction between various feature values to study how different features affect the model prediction of temperature. The results show that in the process of model prediction, except for the ERA5-Land temperature data used for correction, the greatest impact on the model prediction is air pressure and DEM.