.png){kind=icon}
.png){kind=icon}
Return
Feature Selection and Data Decomposition-Driven Drought Forecasting: Designing Hybrid Deep Learning Models
.png){kind=icon}
PRE
AI
A
M
M
M
A
DOI：10.1007/s11269-026-04866-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Drought is one of the most destructive natural hazards, threatening agriculture, water resources, food security, and regional economies. This study develops a hybrid deep learning framework for multi-timescale drought forecasting in three hydro-climatically contrasting regions of Queensland, Australia: Darling Downs, Rockhampton, and Townsville. Eighteen hydro-meteorological variables were used to compute the Standardized Precipitation Evapotranspiration Index (SPEI) at 1-, 3-, 6-, and 12-month time scales. To improve data quality, Multivariate Variational Mode Decomposition (MVMD) and De-Mixing Multivariate Variational Mode Decomposition (DMVMD) were combined with LASSO feature selection before training eight deep learning models: BiGRU, BiCNN, BiLSTM, CNN-LSTM, CNN-BiGRU, TCN, PINN, and LNN. The proposed framework consistently improved forecasting accuracy across all regions and time scales, with BiGRU demonstrating the most consistent performance among the evaluated models. For example, under the proposed DMVMD framework, the BiGRU model achieved NSE = 0.9757, KGE = 0.7923, WI = 0.9936, MAE = 0.1192, RMSE = 0.1500, and R² = 0.9757 for SPEI-12 in the Darling Downs, while similarly high accuracy was obtained in Rockhampton and Townsville. The results demonstrate that integrating multivariate signal decomposition with feature selection substantially enhances drought forecasting and provides a reliable framework for drought early-warning and climate-resilient water resources management.
Keywords:
Drought forecasting
MVMD
DMVMD
Deep learning
SPEI
LASSO
Temporal convolutional network
Physics-informed neural network
Journal
IF:
4.7
Papers:
8.1K
Citations:
1.6W
Organization
Cited Papers