Dual framework for rainfall prediction: a multi-seed machine and deep learning evaluation across Pakistan's climatic...

Area of Science:

• Meteorology
• Climate Science
• Data Science

Background:

• Accurate rainfall prediction is crucial for Pakistan's agriculture, hydrology, and climate adaptation due to monsoon and drought susceptibility.
• Increasing climate variability necessitates reliable, data-driven forecasting systems for nonlinear atmospheric dynamics.

Purpose of the Study:

• To introduce an extensive hybrid machine learning-deep learning (ML-DL) framework for classifying daily rainfall events and predicting multiple forecast horizons (RainDay0-RainDay5).
• To evaluate the performance of various ML and DL models, enhanced with Variational Mode Decomposition (VMD) and optimization algorithms, for rainfall forecasting in Pakistan.

Main Methods:

• Utilized a dataset of 130,230 daily meteorological records from 10 Pakistani cities (1990-2025).
• Employed machine learning models (Extra Trees, Gradient Boosting, Ridge, Naïve Bayes) with TimeSeriesSplit validation and SMOTE.
• Integrated Variational Mode Decomposition (VMD) for signal processing and Particle Swarm Optimization (PSO) with genetic algorithms for hyperparameter tuning.
• Developed and tested hybrid VMD-augmented deep learning models, including VMD-LSTM, VMD-GRU, VMD-PSO-GRU (VPG), and a multi-head VMD-GRU-Attention model.

Main Results:

• Ensemble ML models provided robust baseline classification.
• VMD-augmented deep learning frameworks significantly enhanced robustness and temporal learning consistency.
• VMD-LSTM and VMD-GRU showed strong performance in specific cities (Hyderabad, Jamshoro, Quetta, Thatta).
• The VMD-PSO-GRU (VPG) framework achieved superior classification, while the multi-head VMD-GRU-Attention model excelled in multi-horizon forecasting (RainDay0-RainDay5).
• The system demonstrated high accuracy in short-term (0.872 in Hyderabad for RainDay0) and steady medium- to long-term performance (around 0.80 at RainDay5).

Conclusions:

• The proposed hybrid ML-DL framework offers improved stability, noise robustness, and generalization capabilities for rainfall prediction.
• The VMD-augmented models, particularly VMD-GRU-Attention with PSO-GA optimization, are effective for operational rainfall early-warning systems in Pakistan's diverse climatic conditions.