Drought modelling by standard precipitation index (SPI) in a semi-arid climate using deep learning method: long short-term memory

Drought modelling is an important issue because it is required for curbing or mitigating its effects, alerting the people to the its consequences, and water resources planning. This study investigates the capability of a deep learning method, long short-term memory (LSTM), in forecasting drought calculated from monthly rainfall data obtained from four stations of Iran. The outcomes of LSTM compared with extra-trees (ET), vector autoregressive approach (VAR) and multivariate adaptive regression spline (MARS) methods in forecasting four drought indices, SPI-3, SPI-6, SPI-9 and SPI-12, taking into account numerical criteria, root-mean-square errors (RMSE), Nash-Sutcliffe efficiency and correlation coefficient together with the visual methods, time variation graphs, scatter plots and Taylor diagrams. The overall results showed that the LSTM method performed superior to the ET, VAR and MARS in forecasting drought based on SPI-3, SPI-6, SPI-9 and SPI-12. The RMSE of ET, VAR and MARS was improved by about 17.1%, 12.8% and 9.6% for SPI-3, by 10.5%, 6.2% and 5% for SPI-6, by 7.3%, 4.1% and 6.2% for SPI-9 and by 22.2%, 27% and 10.6% for SPI-12 using LSTM. The MARS method was ranked as the second best, while the ET provided the worst results in forecasting drought based on SPI.

Automated summary

This study investigated the effectiveness of the LSTM method in predicting drought based on monthly rainfall data from four stations in Iran. The results showed that LSTM outperformed other methods, including ET, VAR, and MARS, in forecasting drought indices.