DCSwinLSTM for spatiotemporal meteorological drought forecasting

Area of Science:

• Meteorology and Climate Science
• Artificial Intelligence and Machine Learning
• Environmental Engineering

Background:

• Accurate meteorological drought prediction is crucial for effective climate-risk management and early warning systems.
• Modeling regional drought evolution across multiple timescales presents significant challenges.
• Existing spatiotemporal prediction models often struggle with the complexities of drought dynamics.

Purpose of the Study:

• To develop and evaluate a novel feature-fusion framework, DCSwinLSTM, for improved multi-timescale meteorological drought prediction.
• To enhance the accuracy and reliability of drought forecasting for climate-risk management and water-resource planning.
• To demonstrate the framework's effectiveness using a global gridded dataset of drought indices.

Main Methods:

• Proposed DCSwinLSTM, a hybrid framework integrating deformable convolution, Swin transformer, and LSTM.
• Employed deformable convolution for boundary-sensitive spatial encoding.
• Utilized Swin transformer for hierarchical multi-scale feature learning and LSTM for temporal dependency modeling.
• Trained and validated the model on a global gridded dataset (1959-2022) of Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) at 3- and 6-month scales.

Main Results:

• DCSwinLSTM consistently outperformed mainstream spatiotemporal prediction baselines across different drought scales.
• Achieved superior performance metrics on SPI-3 and SPI-6 test sets, including low Mean Squared Error (MSE) and high R-squared (R²).
• Demonstrated high accuracy with MSEs of 0.4269 (SPI-3) and 0.2800 (SPI-6), and R² values of 0.5875 (SPI-3) and 0.7265 (SPI-6).

Conclusions:

• The proposed DCSwinLSTM framework offers a reliable solution for multi-timescale meteorological drought forecasting.
• The model's performance supports enhanced risk management and water-resource planning, even in data-scarce regions.
• This advancement contributes to more robust climate adaptation strategies through improved drought prediction capabilities.