Climate modeling for South Asia: statistical and deep learning for rainfall and temperature prediction.

Area of Science:

• Hydrometeorology
• Climate Science
• Data Science

Background:

• Accurate hydrometeorological forecasting is crucial for South Asia's climate-vulnerable economies.
• Evaluating diverse forecasting models on centennial-scale data is essential for robust predictions.

Purpose of the Study:

• To benchmark four forecasting approaches (SARIMA, TDNN, LSTM, XGBoost) for monthly rainfall and temperature across South Asia.
• To develop a hybrid model selection guide based on cross-country, cross-variable performance analysis.

Main Methods:

• Utilized a centennial dataset (1901-2023) from seven South Asian countries.
• Implemented a reproducible pipeline with unified preprocessing, early stopping, and cross-validated evaluation (RMSE, MAPE, R², CV-RMSE).
• Compared classical (SARIMA) and deep learning (TDNN, LSTM, XGBoost) models.

Main Results:

• Model performance varied significantly by variable and region.
• Time-Delay Neural Networks (TDNN) and XGBoost excelled in rainfall forecasting due to nonlinear dynamics.
• Long Short-Term Memory (LSTM) networks were optimal for minimum temperature, capturing long memory.
• Seasonal Autoregressive Integrated Moving Average (SARIMA) models suited stable seasonal series.

Conclusions:

• Developed a hybrid selection guide mapping variables and locations to the most effective forecasting model.
• Generated 2024-2025 forecasts with uncertainty bands to aid critical decision-making in agriculture, water management, and disaster preparedness.
• The framework's scale and diagnostics promote replication and evidence-based policymaking.