A hybrid deep learning approach to predict meteorological drought based on several remote sensing products
摘要
Drought is a natural disaster that is on the rise globally due to the severity of climate change. The accurate prediction of meteorological drought is critically important for implementing essential policies to overcome its disastrous consequences. This study investigates the application of a novel attention-based machine learning model, consisting of ARIMA preprocessing, CNN-BiLSTM, and XGBoost (AttCLX) model for predicting meteorological drought. In this context, remotely sensed rainfall products of CHIRPS, ERA5, and NASA POWER were acquired for predicting non-parametric SPI with time scales of 1, 6, and 12 month(s). Then, Gamma correlation was employed to find the relationship between remotely sensed data with three observational SPIs across different lag times. The inputs of machine learning models were derived based on the correlations, considering three scenarios: the most correlated one, the two most correlated ones, and all three lag times. Finally, three models, ARIMA-based AttCLX, LSTM, and XGBoost, were employed to predict non-parametric SPI-1, -6, and -12. Results showed that the proposed ARIMA-AttCLX hybrid model exhibited the best performance compared to the other models. Moreover, the AttCLX hybrid model demonstrated its best effectiveness in predicting SPI-12, outperforming the prediction of SPI-1 and SPI-6. This superior performance was notable using all three lag times, with MSE = 0.0696