.png){kind=icon}
.png){kind=icon}
Return
Prediction of standardized precipitation index based on LSTM and variants of LSTM
.png){kind=icon}
PRE
AI
N
D
V
DOI：10.1016/j.pce.2026.104422.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Drought prediction is vital for early warnings, agricultural planning and management of water resources. The Standardised Precipitation Index (SPI) provides a quantitative measure of the severity of the drought and its extent, enabling preparedness and mitigation actions for short- and long-term drought events. Being solely dependent on precipitation records having spatio-temporal inconsistencies, conventional methods of SPI estimation suffer several computational challenges. In this study, four deep learning models, namely, Long Short Term Memory (LSTM), Stacked LSTM (S-LSTM), Bi-directional LSTM (BiLSTM), and Stacked BiLSTM (S-BiLSTM), were used to predict SPI values at various time scales (SPI-1, SPI-3, SPI-6, SPI-9, SPI-12, SPI-18, SPI-24, SPI-48, and SPI-yearly) for assessing the meteorological drought of Coimbatore station in India, based on 120 years of rainfall data. The short-term predictions (SPI-1) showed poor consistency (R-2 <= 0.11, RMSE similar to 1.0) with LSTM, though BiLSTM and S-BiLSTM experienced slight improvements. For moderate aggregations (SPI-3 to SPI-9), the S-LSTM resulted in improved performance (R-2 approximate to 0.62 to 0.89). For long-term scales (SPI-12 to SPI-24), S-BiLSTM outperformed with the highest consistency (R-2 > 0.94), indicating improved learning of long-range dependencies. The drought predictions at the annual scale were found to be weak due to the masking effect of critical short-term variabilities, resulting in a near-normal abundance. Bidirectional and stacked variants of LSTM models demonstrated better computational reliability with accurate, stable, and generalizable predictions, particularly for time scales SPI-6 to SPI-24. Deep learning improves drought forecasting using multi-source predictors, but marginal data variability must be addressed to ensure transparency and interpretability.
Keywords:
Drought prediction
LSTM
Stacked LSTM
Bidirectional LSTM
Bidirectional stacked LSTM
SPI
Journal
IF:
4.1
Papers:
3.3K
Citations:
6.9K
Organization
Cited Papers