Long lead time drought forecasting using lagged climate variables and a stacked long short-term memory model

Area of Science:

• Environmental Science
• Climate Science
• Data Science

Background:

• Long lead time drought forecasting is crucial for effective risk management and early warning systems.
• Climate change and complex drought assessment pose challenges for traditional forecasting methods.
• Deep learning offers potential solutions for improving long-term drought prediction accuracy.

Purpose of the Study:

• To develop and evaluate a stacked long short-term memory (LSTM) deep learning model for forecasting the Standard Precipitation Evaporation Index (SPEI).
• To assess the model's capability in predicting drought characteristics at lead times up to 12 months.
• To apply the model to the New South Wales region, utilizing hydrometeorological and climatic data.

Main Methods:

• A stacked LSTM deep learning architecture was employed for drought forecasting.
• The Standard Precipitation Evaporation Index (SPEI) was computed using the Climatic Research Unit (CRU) multivariate interpolated grid.
• The model was trained on data from 1901-2000 and tested on data from 2001-2018, with forecasts generated for 1-12 month lead times.

Main Results:

• The stacked LSTM model demonstrated effective drought forecasting capabilities at both short-term and long-term lead times.
• Analysis of drought intensity, onset, spatial extent, and duration showed improved understanding through the model's predictions.
• Statistical metrics including R-squared and root-mean-square error validated the model's performance in predicting drought intensity.

Conclusions:

• The stacked LSTM model is a viable tool for enhancing drought forecasting accuracy, particularly for long lead times.
• Findings provide valuable insights for government agencies in developing drought management and adaptation strategies.
• Further research can explore the model's efficacy at finer temporal scales (days to weeks).