What question did this study set out to answer?

The main goal is to develop an LSTM-based model for predicting SPI and SPEI drought indices.

July 14, 2026Natural Hazards Review0 citations

Deep Learning LSTM-Based Model for Predicting SPI and SPEI Drought Indices

MFMohammad Hadi FattahiIslamic Azad University, Tehran TDTajeddin DerikvandIslamic Azad University of Marand FRFarhang RahmaniIslamic Azad University, Tehran

Key Points

The main goal is to develop an LSTM-based model for predicting SPI and SPEI drought indices.
Utilized statistical data from 1972 to 2020 across two Iranian watersheds.
Compared the proposed LSTM model with control models using various methodologies.
Evaluated model performance with metrics like mean absolute error (MAE), root mean squared error (RMSE), and R-squared (RS).
LSTM model achieved MAE of 0.057, RMSE of 0.079, and RS of 0.0053, indicating high accuracy.
The proposed LSTM model outperformed traditional time-series models in predictive accuracy.
Dynamic analysis showed better preservation of time-series properties by the LSTM model compared to ARIMA.

Abstract

Abstract Prolonged drought conditions significantly exacerbate the depletion of water resources. This increases the importance of accurate prediction and preparedness to deal with the negative impacts of drought. This work proposed a deep learning LSTM-based model for predicting drought indices—the Standard Precipitation Index (SPI) and Standardized Precipitation-Evapotranspiration Index (SPEI). To achieve this, two watersheds in Iran were studied, and statistical data from 1972 to 2020 were employed. Several linear and nonlinear methodologies were used to evaluate the proposed model and control models’ results. The mean absolute error (MAE), root mean squared error (RMSE), and R-squared (RS) findings demonstrated that the suggested model outperformed the control models in terms of accuracy ( MAE = 0.057 , RMSE = 0.079 , and RS = 0.0053 ). Similarly, the findings of the difference percentage of Lyapunov exponent (LE) and approximate entropy (ApEn) suggested that the proposed LSTM model was superior compared with the time-series models. Furthermore, a nonlinear dynamic analysis of the ARIMA and LSTM model results demonstrated that the suggested model conserved the dynamic properties of the base time-series better, resulting in more accurate results.

Ask AI

Helpful

Bookmark

View Full Paper