Samson et al. (2025) Comparative study of single and hybrid deep learning models for daily rainfall prediction in selected African cities

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-11-28
• Authors: Timothy Kayode Samson, F. O. Aweda
• DOI: 10.1038/s41598-025-26739-x

Research Groups

• Statistics Programme, College of Agriculture, Engineering and Science, Bowen University, Iwo, Nigeria
• Institute of Aerospace Technologies, University of Malta, Msida, Malta
• Physics Programme, College of Agriculture, Engineering and Science, Bowen University, Iwo, Nigeria

Short Summary

This study comprehensively compares single (CNN, LSTM, ANN, RNN) and hybrid (RNN+ANN, LSTM+ANN, LSTM+RNN) deep learning models for daily rainfall prediction in five diverse African cities. It finds that single deep learning models, particularly RNN, generally outperform hybrid models across most locations, although hybrid models can be superior in specific complex rainfall regimes.

Objective

• To conduct a comprehensive comparative analysis of single and hybrid deep learning models for daily rainfall prediction in selected African cities (Nairobi, Abuja, Pretoria, Rabat, and Libreville).
• To identify the most suitable deep learning architecture and influential meteorological predictors for diverse African climates to advance data-driven rainfall forecasting for robust prediction systems.

Study Configuration

• Spatial Scale: Five selected African cities (Nairobi, Abuja, Pretoria, Rabat, Libreville) representing diverse geographic locations and climates across Africa. Data resolution: 0.5° latitude × 0.625° longitude.
• Temporal Scale: Daily rainfall prediction. Data collected from January 1, 1980, to December 31, 2024.

Methodology and Data

• Models used:
  • Single Deep Learning Models: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Artificial Neural Network (ANN), Recurrent Neural Network (RNN).
  • Hybrid Deep Learning Models: RNN + ANN, LSTM + ANN, LSTM + RNN.
• Data sources: NASA’s Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) reanalysis dataset.
  • Input features: Daily rainfall, relative humidity, wind speed, pressure, lagged rainfall values (previous 10 days), day of the month, and month of the year.
  • Performance metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Huber loss.

Main Results

• Significant spatial variability in daily rainfall dynamics was observed across the studied cities, with Abuja and Libreville showing the highest variability and Rabat exhibiting consistently low and stable rainfall.
• Weak correlations among the daily rainfall patterns of the selected cities suggest that local and regional atmospheric conditions primarily drive rainfall, rather than a single continental-scale system.
• Single deep learning models, particularly the Recurrent Neural Network (RNN), generally outperformed hybrid models in most cities (Nairobi, Pretoria, Rabat).
• The LSTM–ANN hybrid model demonstrated superior performance only in Abuja (MSE = 50.0173 mm², RMSE = 7.0723 mm, MAE = 2.5242 mm, Huber loss = 2.2478 mm).
• Artificial Neural Network (ANN) showed the best performance in Libreville (MSE = 184.1970 mm², RMSE = 13.5719 mm, MAE = 8.4073 mm, Huber loss = 7.8178 mm).
• Relative humidity emerged as the most influential predictor of rainfall in Abuja, Libreville, and Nairobi.
• Temporal persistence of rainfall (lagged values from previous days) was the most important feature in Pretoria (previous day's rainfall) and Rabat (previous two days' rainfall).
• Temperature generally had a positive impact on rainfall in most cities, but a strong negative contribution in Rabat, consistent with its Mediterranean climate characterized by dry, hot summers.

Contributions

• Provides a comprehensive, systematic comparative analysis of single and hybrid deep learning models for daily rainfall prediction specifically tailored to diverse African urban environments, addressing a gap in existing literature that often overlooks the spatio-temporal richness of African climate data or focuses on coarser temporal scales.
• Offers context-specific insights into the performance of various deep learning architectures, demonstrating that simpler single models (like RNN) can often be more reliable than complex hybrid models for localized and noisy meteorological datasets in Africa.
• Identifies key meteorological predictors (e.g., relative humidity, temporal rainfall persistence) that are most influential for rainfall forecasting in different African climatic zones, aiding in the development of more accurate and robust early warning systems and agricultural planning tools.

Funding

No specific funding projects, programs, or reference codes were provided in the paper.

Citation

@article{Samson2025Comparative,
  author = {Samson, Timothy Kayode and Aweda, F. O.},
  title = {Comparative study of single and hybrid deep learning models for daily rainfall prediction in selected African cities},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-26739-x},
  url = {https://doi.org/10.1038/s41598-025-26739-x}
}