Abstract
Despite the abundance of arable land in Africa, food insecurity persists as a significant challenge, largely due to increasingly unpredictable rainfall patterns driven by climate change. Accurate daily rainfall prediction is therefore critical for agricultural planning and food security. Deep learning (DL) offers powerful tools for modeling complex, nonlinear, and temporal dynamics in climate data. In this study, we conduct a comprehensive comparison of four single DL models: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Artificial Neural Network (ANN), and Recurrent Neural Network (RNN), alongside three hybrid architectures (RNN + ANN, LSTM + ANN, and LSTM + RNN). These models were selected for their proven ability to capture both spatial and sequential dependencies in meteorological datasets, while the hybrid models were intended to leverage complementary strengths of the single learning models. Daily rainfall and associated meteorological variables (relative humidity, wind speed, and pressure) were obtained from NASA's MERRA-2 reanalysis, chosen for its long-term consistency, global coverage, and high spatial resolution. The dataset spans January 1, 1980, to December 31, 2024, and was preprocessed using standard scaling with an 80/20 train-validation split. Model performance was evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Huber loss, which provide robust measures of predictive accuracy and error sensitivity. Results reveal significant spatial variability in rainfall dynamics across the studied cities. Abuja and Libreville exhibited the highest variability, while Rabat showed consistently low and stable rainfall. Weak correlations among cities highlight the diversity of local rainfall regimes. Performance varied by location, but overall, single DL models, particularly RNN, outperformed hybrid models in most cities. The LSTM-ANN hybrid showed superior results only in Abuja (MSE = 50.0173, RMSE = 7.0723, MAE = 2.5242, Huber loss = 2.2478). Relative humidity emerged as the most influential predictor in most cities, whereas temporal persistence of rainfall dominated in Pretoria and Rabat. These findings underscore that while hybrid DL models can enhance performance in highly complex rainfall regimes, single models, especially RNN, remain more reliable and effective across diverse African climates.