Comparative Study of Single and Hybrid Deep Learning Models for Daily Rainfall Prediction: Evidence from African Climatic Data

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Despite the availability of arable land in Africa, food insecurity still remains a major challenge. This can be attributed to the increasingly unpredictable rainfall patterns driven by climate change and hence accurate daily rainfall prediction is therefore critical for informed agricultural planning to guarantee food security. Deep Learning has offer a reliable approach in capturing complex dynamics in data and hence this study conducts a comprehensive comparison of four single Deep Learning (DL) models: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Artificial Neural Network (ANN), and Recurrent Neural Network (RNN) alongside three hybrid models (RNN + ANN, LSTM + ANN, and LSTM + RNN) to capture complex rainfall dynamics across selected African cities. Daily rainfall data spanning January 1, 1980 to December 31, 2024 were sourced from the NASA MERRA-2 reanalysis. The features in training these algorithms were the daily relative humidity, wind speed, pressure, the day, the month as well as 10 days lag daily rainfall. Data was preprocessed with 80/20 as training and validation split with data transformation with MinMax Scaler. The Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Huber loss were used in evaluating the models. Results show variability in rainfall patterns in these cities with Abuja and Libreville exhibiting the highest variability in daily rainfall. This is in contrast with Rabat which experiences the most stable and consistently low rainfall. Low correlation in rainfall among the cities highlight diverse local rainfall regimes. Different DL models are found to be most suitable for each of the African cities but in overall the single DL models particularly, the RNN performed better than hybrid model on most of the selected cities. The hybrid of LSTM-ANN excelled only in Abuja (MSE = 50.0173 RMSE = 7.0723, MAE = 2.5242, Huber loss = 2.2478) suggesting dominance of the single model approach in most of the African cities. In most of these cities, relative humidity was found as the most significant driver of rainfall. However, in Pretoria and Rabat, the most important predictors were identified to be the rainfall amounts from the previous day and the previous two days respectively, emphasizing the stronger influence of temporal rainfall persistence in these locations. These findings underscores that while hybrid models can occasionally offer better predictive accuracy especially when dealing with complex rainfall pattern, single deep learning models especially the RNN generally remain more reliable and effective across varied African climates.

Article activity feed