A novel information changing rate and conditional mutual information-based input feature selection method for artificial intelligence drought prediction models

• Published in: Climate dynamics Vol. 58; no. 11-12; pp. 3405 - 3425
• Main Authors: Li, Qiongfang, Han, Xingye, Liu, Zhennan, He, Pengfei, Shi, Peng, Chen, Qihui, Du, Furan
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.06.2022; Springer; Springer Nature B.V
• Subjects: Accuracy; Artificial intelligence; China; Climatic changes; Climatology; Computing time; Correlation analysis; Correlation coefficient; Correlation coefficients; Dependent variables; Drought; Droughts; Earth and Environmental Science; Earth Sciences; Evapotranspiration; Evapotranspiration-precipitation relationships; Geophysics/Geodesy; Independent variables; Influence; Methods; Modelling; Oceanography; prediction; Prediction models; Technology application; Weather forecasting; China; Yunnan-Guizhou Plateau; Feature selection; Drought-driving factors; Information changing rate; Conditional mutual information
• ISSN: 0930-7575; 1432-0894
• DOI: 10.1007/s00382-021-06104-0

The efficient and accurate selection of primary drought-driving factors as the independent variables of drought prediction model is critical in improving drought prediction accuracy. In this study, a novel feature selection method based on information changing rate and conditional mutual information (ICR-CMIFS) was proposed and evaluated by the comparison with other feature selection methods from feature selection, simulation, and classification aspects; two artificial intelligence drought prediction models, which treated the factors selected by ICR-CMIFS, correlation analysis (CA) and mutual information maximum (MIM) respectively as independent variables and standardized precipitation evapotranspiration index (SPEI) in 3/6/12-month time scales as dependent variables, were established; the superiority of ICR-CMIFS over CA and MIM methods in the selection of primary climatic drought-driving factors in Yunnan-Guizhou Plateau (YGP) was tested by the performance of the two models. The results revealed: the ICR-CMIFS was superior to the other feature selection methods; both artificial intelligence drought prediction models with the independent variables selected by ICR-CMIFS performed better in terms of correlation coefficient, Nash–Sutcliffe coefficient, root-mean square error and model computing time than by the MIM and CA methods. The outputs can provide an innovative approach in selecting primary drought-driving factors and improving drought prediction accuracy.