Abiyev, R., Mohammed, M., & Abizada, R. (2025). Rainfall prediction using stacked deep learning networks. Modeling Earth Systems and Environment, 11(5), 324.
In a newly published article in Modeling Earth Systems and Environment, Near East University (NEU) researchers Prof. Dr. Rahib Abiyev, Mansur Mohammed, and Rufat Abizada present an advanced artificial intelligence approach designed to improve the accuracy of rainfall forecasting—an area that is critically important for disaster preparedness, agriculture, and water resource management.
Titled “Rainfall prediction using stacked deep learning networks,” the study introduces a Stacked Bidirectional Long Short-Term Memory (SBiLSTM) model for rainfall time-series prediction. Unlike conventional forecasting methods that often struggle with the complex and nonlinear nature of rainfall patterns, this approach processes the time series in both forward and backward directions, allowing the model to capture richer temporal relationships. By stacking multiple BiLSTM layers, the network can learn deeper patterns from historical rainfall data. The authors also incorporate data preprocessing steps—such as handling missing values and preparing training samples using an overlapping sliding window approach—to strengthen model reliability.
Using meteorological data from Nigeria (both nationwide data and the local Kaduna region), the researchers compared their SBiLSTM model with other widely used deep learning architectures, including LSTM, CNN, BiLSTM, and MLP. The results demonstrate strong predictive performance. For Nigeria overall, the proposed model achieved low error values and a high goodness-of-fit (MAE: 4.4158, RMSE: 9.962, R²: 0.987), indicating a very close match between predicted and observed rainfall patterns. For the Kaduna region, the model also performed well (MAE: 51.831, RMSE: 162.934, R²: 0.831), reflecting the greater variability and challenge of local-scale prediction.
By improving rainfall forecast accuracy—especially in regions where climate variability and extreme weather events create high uncertainty—this research highlights how deep learning can support earlier warnings, better planning, and stronger climate resilience in meteorological applications.
About the researchers
Prof. Dr. Rahib H. Abiyev is a researcher at Near East University (NEU), Northern Cyprus, with expertise spanning soft computing, control systems, digital signal processing, and artificial intelligence. His research focuses on developing intelligent, data-driven methods that enhance prediction, classification, and decision-making in complex real-world systems.
He has made significant contributions to deep learning and neural network–based time-series forecasting (including rainfall prediction), as well as advanced fuzzy systems, such as type-2 and type-3 fuzzy and fuzzy neural networks, for dynamic system control and system identification. His work also extends to computer vision, medical image and video analysis, and IoT-enabled intelligent systems, reflecting a strong commitment to translating rigorous computational models into practical engineering and healthcare applications.
For collaboration and inquiries, please contact Prof. Dr. Rahib H. Abiyev at [email protected].
Abstract
Accurate rainfall prediction is a critical yet complex challenge in meteorological forecasting, essential for disaster preparedness, agricultural planning, and water resource management. Rainfall forecasting relies on the relationship between predictors and the rainfall being forecasted. In this study, a Stacked Bidirectional Long Short-Term Memory (SBiLSTM) network is proposed for rainfall time-series prediction. Leveraging bidirectional processing, the SBiLSTM model effectively captures complex temporal dependencies in both forward and backward directions, uncovering intricate relationships within the time series data. By stacking multiple BiLSTM layers, the model learns deeper and more complex patterns and relationships within the data, thereby enhancing the reliability and accuracy of rainfall forecasts. Meteorological data from Nigeria was used to train and evaluate the model. The methodology involves preprocessing the data and employing a cross-validation approach to train the SBiLSTM-based time series model. The performance of the SBiLSTM model was evaluated through simulations and compared against other deep learning architectures, including Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), BiLSTM, and Multilayer Perceptron (MLP) models. The obtained results demonstrate that the SBiLSTM model achieved low mean absolute error (MAE) and root-mean-squared error (RMSE), along with a high R2 value for rainfall prediction of Nigeria, with scores of 4.4158, 9.962 and 0.987, respectively. For the local Kaduna region, the model achieved MAE, RMSE and R2 values of 51.831, 162.934, and 0.831, respectively. The results obtained indicate that the SBiLSTM model achieves superior predictive performance, highlighting its potential to improve the accuracy of rainfall forecasting in meteorological applications.
