Drought Identification in the Yangtze River Basin Using CMIP6 Multi-Model Data Fusion: A Comparison of Traditional and Machine Learning Methods

This study compares the advantages and limitations of traditional CMIP6 data fusion methods and machine learning fusion methods when applied to drought identification in the Yangtze River Basin. We consider three traditional fusion methods and five machine learning fusion methods, and calculate drought indices over 3-, 6-, and 12-month periods based on precipitation data from meteorological stations in the Yangtze River Basin (1960–2014) and 15 CMIP6 model datasets. The drought identification index is used to evaluate the performance of the fusion methods. Results indicate that traditional statistical methods have significant limitations in the upper reaches of the basin, where the terrain is highly undulating, but perform better in the middle and lower reaches, which are relatively flat. Among the machine learning methods, neural networks tend to amplify the observational noise, whereas kernel-tuning methods better accommodate nonlinear relationships across different SPI time scales. The prediction performance of all methods decreases from the 12- to 3-month drought indices, but the extent of the decline varies. The Random Forest and Radial Basis Function methods give the smallest reduction in performance, while the Backpropagation and Backpropagation-Adaboost methods produce the largest drop in performance.