Standardized precipitation evapotranspiration index (SPEI) estimated using variant long short-term memory network at four climatic zones of China.

• LSTM-type models outperformed empirical methods and classical machine learning models. • BiLSTM was recommended for SPEI-3, and CNN-LSTM was suitable for SPEI-6 and SPEI-12. • Rn-based models performed best for estimating multiscale SPEI throughout China. • RH-based was superior to T-based ML models, but empirical methods differed with zones. Although the accurate prediction of the Standardized Precipitation Evapotranspiration Index (SPEI) is considered meaningful in reducing drought losses, its wide applications are limited to substantial meteorological data requirements. Considering Long Short-Term Memory network (LSTM) has proved its potential in estimating drought index, the concern is justified regarding the performance of its variants for estimating SPEI using limited meteorological input at the national level. Therefore, this study established the SPEI models using empirical methods, SVM, RNN, LSTM, BiLSTM, and CNN-LSTM, respectively, to cope with different data-missing scenarios. Based on a comprehensive comparison among different methods for multiscale SPEI estimation at four climatic zones of China, the results showed that BiLSTM was the most recommended model for estimating SPEI at 3-month timescale, with R2, NSE, and RMSE ranging 0.916–0.997, 0.907–0.997, and 0.143–0.353, respectively. Whereas CNN-LSTM was more suitable for other timescales, with R2, NSE, and RMSE being 0.904–0.999, 0.858–0.989, and 0.145–0.365 for estimating SPEI at 6-month timescale, respectively, and 0.858–0.998, 0.795–0.991, and 0.081–0.568 for estimating SPEI at 12-month timescale. Generally, the accuracy performance of SPEI methods can be ranked from best to worst as LSTM-type models, SVM, RNN, and empirical methods. The exception was that H-S exceeded RNN3 in TCZ and MPZ by 4.1–30.0 % for R2, 4.1–65.0 % for NSE, and 2.3–19.6 % for RMSE, respectively. Moreover, this study found that the accuracy performance of machine learning models for SPEI estimation got worse with the number of independent variables decreased. Overall, the variants of LSTM exerted excellent performance for multiscale SPEI estimation, which provided the most accurate prediction of meteorological, agroecological, and hydrological droughts throughout China. [ABSTRACT FROM AUTHOR]