University of Isfahan Geography and Environmental Planning 2008-5362 36 3 2025 09 23 Evaluation and Prediction of Seasonal Precipitation in Different Climatic Regions of Iran Using a Hybrid Deep Learning Model (Case Study: Rasht and Yazd) Evaluation and Prediction of Seasonal Precipitation in Different Climatic Regions of Iran Using a Hybrid Deep Learning Model (Case Study: Rasht and Yazd) 35 64 29932 10.22108/gep.2025.145951.1735 FA Aynaz Vafaei M.Sc. student, Department of Water Science and Engineering, Faculty of Agriculture, University of Tabriz, Tabriz, Iran Mohammad Taghi Sattari Ph.D., Department of Water Engineering, Faculty of Agriculture, University of Tabriz, Tabriz, Iran 0000-0002-5139-2118 Journal Article 2025 07 16 Abstract The primary objective of this study was to evaluate the stability and generalizability of machine learning models under varying climatic conditions and identify the optimal approach for seasonal precipitation prediction. This research is novel in its first-time application of the hybrid CNN-BiLSTM model in the arid climate of Yazd, alongside a comparison with the humid climate of Rasht. The findings provided valuable insights for enhancing early warning systems and improving water resource management across different climates. To achieve this, we assessed the performance of 4 models: Support Vector Machine (SVM), Deep Neural Network (DNN), Bidirectional Long Short-Term Memory (BiLSTM), and the hybrid Convolutional Neural Network BiLSTM (CNN-BiLSTM). This evaluation utilized 28 years of climatic and precipitation data (1995–2022) from Rasht and Yazd stations. Two input scenarios were designed: the first included climatic parameters (minimum temperature, maximum relative humidity, average wind speed, and sunshine hours), while the second combined these parameters with 3 rainfall time lags. The dataset was divided into training and testing subsets using a ratio of 70:30. The results indicated that the second scenario significantly outperformed the first due to the inclusion of rainfall lags. In Yazd, the CNN-BiLSTM model achieved the best performance with a correlation coefficient of 0.89, a Root Mean Square Error (RMSE) of 7.48 mm, and a Nash–Sutcliffe efficiency of 0.76. In Rasht, the same model produced even better results, with a correlation coefficient of 0.94, an RMSE of 61.89 mm, and a Nash–Sutcliffe efficiency of 0.90. These findings underscore the effectiveness of hybrid deep learning architectures in capturing complex spatio-temporal precipitation patterns across diverse climatic conditions. Keywords: Arid and Humid Climate, Precipitation Prediction, Seasonal Modeling, Deep Learning, CNN-BiLSTM. Introduction Rainfall is a fundamental component of the hydrological cycle, playing a crucial role in water supply, agricultural planning, flood control, drought mitigation, and environmental management. Accurate seasonal rainfall forecasting remains a significant challenge in water and climate sciences, directly influencing management decisions, sustainable development, and food security. With the intensification of climate change and increasing irregularities in rainfall patterns, the demand for timely and reliable forecasts of this complex climatic variable has become increasingly critical. The nonlinear, complex, and stochastic nature of rainfall resulting from the interaction of multiple climatic and atmospheric factors makes its prediction a challenging scientific and technical task. Factors like temperature, humidity, wind speed, solar radiation, rainfall intensity and duration, evaporation, and large-scale teleconnection phenomena, such as the Madden–Julian Oscillation (MJO), significantly affect rainfall variability, creating highly intricate interdependencies. Recent advances in machine learning and deep learning have enabled more effective modeling of these nonlinear and complex relationships. Traditional methods, such as linear regression and time series models, are limited by their linear assumptions and their inability to capture dynamic dependencies. In contrast, modern models—such as Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), Recurrent Neural Networks (RNNs), and particularly advanced hybrid architectures like CNN-BiLSTM—have shown promising accuracy in modeling and forecasting climatic variables. However, most existing research has focused on humid and semi-humid regions, with relatively few studies addressing arid and semi-arid zones, particularly in Iran. These regions characterized by low and irregular rainfall require more sophisticated modeling and detailed analysis. Additionally, comparative evaluations of models across contrasting climatic zones can provide valuable insights into model robustness and generalizability, which are essential for enhancing forecasting systems. Thus, the main objective of this study was to assess the performance and generalizability of 4 machine learning and deep learning models—SVM, DNN, BiLSTM, and CNN-BiLSTM—in predicting seasonal rainfall in two climatically distinct regions: Rasht (humid) and Yazd (arid). This research pioneered the application of the CNN-BiLSTM model in the arid climate of Yazd and compares its performance with that in the humid climate of Rasht, offering practical recommendations for flood early-warning systems and water resource management. Materials & Methods The dataset utilized in this research comprised seasonal climate and rainfall data from the meteorological stations of Rasht and Yazd over a 28-year period (1995–2022). The climatic variables included minimum temperature, maximum relative humidity, average wind speed, and sunshine duration, along with seasonal rainfall data. Two input scenarios were designed to assess the impact of past rainfall data on model performance: Scenario 1 included only the climatic parameters, while Scenario 2 combined these with 3 rainfall time lags. These scenarios aimed to evaluate how different input configurations influenced the predictive capabilities of the models. The dataset was divided into 70% for training and 30% for testing, following standard practices in machine learning research. The four main models employed in the study were: SVM (Support Vector Machine): A classic machine learning model known for its effectiveness in modeling nonlinear relationships DNN (Deep Neural Network): Capable of learning complex patterns and features from high-dimensional data BiLSTM (Bidirectional Long Short-Term Memory): Effective at capturing long- and short-term dependencies from both forward and backward temporal directions CNN-BiLSTM (Convolutional Neural Network – BiLSTM): A hybrid architecture where the CNN extracts spatial and short-term features, while the BiLSTM models temporal dependencies The CNN component employs convolutional filters to reduce noise and extract local patterns, while the BiLSTM component captures bidirectional temporal relationships. This hybrid design enabled the model to effectively reconstruct complex rainfall patterns across diverse climatic conditions. The performances of the models were evaluated using 4 criteria: Correlation Coefficient (R), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Nash-Sutcliffe Efficiency (NSE). Each of these metrics reflected different aspects of the model's accuracy and reliability. Research Findings The results clearly demonstrated the effectiveness of Scenario 2, which incorporated past rainfall data, in enhancing model performance. The inclusion of lagged rainfall significantly improved prediction accuracy across both regions. In the humid and complex climate of Rasht, the CNN-BiLSTM model achieved impressive results with a correlation coefficient (R) of 0.94, a Root Mean Square Error (RMSE) of 61.89 mm, and a Nash-Sutcliffe Efficiency (NSE) of 0.90. The model effectively captured intricate and variable rainfall patterns, providing accurate seasonal forecasts. In Yazd characterized by sporadic and irregular rainfall, the CNN-BiLSTM model still demonstrated high accuracy, achieving an R of 0.89, an RMSE of 7.48 mm, and an NSE of 0.76. These results highlighted the model's robustness, even under challenging climatic conditions. The BiLSTM model ranked second due to its bidirectional temporal memory followed by the DNN and SVM models, which were less effective in capturing complex temporal dependencies. The differences in performance were particularly evident during extreme rainfall events and periods of seasonal variability. These findings confirmed the importance of incorporating past rainfall data as model input, as well as the superiority of deep and hybrid architectures in addressing temporal and nonlinear relationships. The results aligned with similar studies and underscored the value of memory-based and hierarchical feature extraction models. Discussion of Results & Conclusion Incorporation of rainfall time lags significantly improved the predictive accuracy of all models, especially in irregular climates like that in Yazd. Notably, the CNN-BiLSTM model surpassed all other models in both regions, highlighting the benefits of hybrid deep learning architectures for managing complex spatiotemporal data. This study enhanced the broader understanding of intelligent forecasting systems for climate-related applications. It illustrated the adaptability and robustness of the CNN-BiLSTM model across various climatic conditions, providing reliable support for decision-making in water management, agricultural planning, and disaster preparedness. To further enrich this research, it is recommended to integrate satellite and radar datasets, expand temporal data coverage, and assess model performance across wider and more diverse geographical regions. These initiatives will strengthen the accuracy and generalizability of seasonal rainfall forecasts. Abstract The primary objective of this study was to evaluate the stability and generalizability of machine learning models under varying climatic conditions and identify the optimal approach for seasonal precipitation prediction. This research is novel in its first-time application of the hybrid CNN-BiLSTM model in the arid climate of Yazd, alongside a comparison with the humid climate of Rasht. The findings provided valuable insights for enhancing early warning systems and improving water resource management across different climates. To achieve this, we assessed the performance of 4 models: Support Vector Machine (SVM), Deep Neural Network (DNN), Bidirectional Long Short-Term Memory (BiLSTM), and the hybrid Convolutional Neural Network BiLSTM (CNN-BiLSTM). This evaluation utilized 28 years of climatic and precipitation data (1995–2022) from Rasht and Yazd stations. Two input scenarios were designed: the first included climatic parameters (minimum temperature, maximum relative humidity, average wind speed, and sunshine hours), while the second combined these parameters with 3 rainfall time lags. The dataset was divided into training and testing subsets using a ratio of 70:30. The results indicated that the second scenario significantly outperformed the first due to the inclusion of rainfall lags. In Yazd, the CNN-BiLSTM model achieved the best performance with a correlation coefficient of 0.89, a Root Mean Square Error (RMSE) of 7.48 mm, and a Nash–Sutcliffe efficiency of 0.76. In Rasht, the same model produced even better results, with a correlation coefficient of 0.94, an RMSE of 61.89 mm, and a Nash–Sutcliffe efficiency of 0.90. These findings underscore the effectiveness of hybrid deep learning architectures in capturing complex spatio-temporal precipitation patterns across diverse climatic conditions. Keywords: Arid and Humid Climate, Precipitation Prediction, Seasonal Modeling, Deep Learning, CNN-BiLSTM. Introduction Rainfall is a fundamental component of the hydrological cycle, playing a crucial role in water supply, agricultural planning, flood control, drought mitigation, and environmental management. Accurate seasonal rainfall forecasting remains a significant challenge in water and climate sciences, directly influencing management decisions, sustainable development, and food security. With the intensification of climate change and increasing irregularities in rainfall patterns, the demand for timely and reliable forecasts of this complex climatic variable has become increasingly critical. The nonlinear, complex, and stochastic nature of rainfall resulting from the interaction of multiple climatic and atmospheric factors makes its prediction a challenging scientific and technical task. Factors like temperature, humidity, wind speed, solar radiation, rainfall intensity and duration, evaporation, and large-scale teleconnection phenomena, such as the Madden–Julian Oscillation (MJO), significantly affect rainfall variability, creating highly intricate interdependencies. Recent advances in machine learning and deep learning have enabled more effective modeling of these nonlinear and complex relationships. Traditional methods, such as linear regression and time series models, are limited by their linear assumptions and their inability to capture dynamic dependencies. In contrast, modern models—such as Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), Recurrent Neural Networks (RNNs), and particularly advanced hybrid architectures like CNN-BiLSTM—have shown promising accuracy in modeling and forecasting climatic variables. However, most existing research has focused on humid and semi-humid regions, with relatively few studies addressing arid and semi-arid zones, particularly in Iran. These regions characterized by low and irregular rainfall require more sophisticated modeling and detailed analysis. Additionally, comparative evaluations of models across contrasting climatic zones can provide valuable insights into model robustness and generalizability, which are essential for enhancing forecasting systems. Thus, the main objective of this study was to assess the performance and generalizability of 4 machine learning and deep learning models—SVM, DNN, BiLSTM, and CNN-BiLSTM—in predicting seasonal rainfall in two climatically distinct regions: Rasht (humid) and Yazd (arid). This research pioneered the application of the CNN-BiLSTM model in the arid climate of Yazd and compares its performance with that in the humid climate of Rasht, offering practical recommendations for flood early-warning systems and water resource management. Materials & Methods The dataset utilized in this research comprised seasonal climate and rainfall data from the meteorological stations of Rasht and Yazd over a 28-year period (1995–2022). The climatic variables included minimum temperature, maximum relative humidity, average wind speed, and sunshine duration, along with seasonal rainfall data. Two input scenarios were designed to assess the impact of past rainfall data on model performance: Scenario 1 included only the climatic parameters, while Scenario 2 combined these with 3 rainfall time lags. These scenarios aimed to evaluate how different input configurations influenced the predictive capabilities of the models. The dataset was divided into 70% for training and 30% for testing, following standard practices in machine learning research. The four main models employed in the study were: SVM (Support Vector Machine): A classic machine learning model known for its effectiveness in modeling nonlinear relationships DNN (Deep Neural Network): Capable of learning complex patterns and features from high-dimensional data BiLSTM (Bidirectional Long Short-Term Memory): Effective at capturing long- and short-term dependencies from both forward and backward temporal directions CNN-BiLSTM (Convolutional Neural Network – BiLSTM): A hybrid architecture where the CNN extracts spatial and short-term features, while the BiLSTM models temporal dependencies The CNN component employs convolutional filters to reduce noise and extract local patterns, while the BiLSTM component captures bidirectional temporal relationships. This hybrid design enabled the model to effectively reconstruct complex rainfall patterns across diverse climatic conditions. The performances of the models were evaluated using 4 criteria: Correlation Coefficient (R), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Nash-Sutcliffe Efficiency (NSE). Each of these metrics reflected different aspects of the model's accuracy and reliability. Research Findings The results clearly demonstrated the effectiveness of Scenario 2, which incorporated past rainfall data, in enhancing model performance. The inclusion of lagged rainfall significantly improved prediction accuracy across both regions. In the humid and complex climate of Rasht, the CNN-BiLSTM model achieved impressive results with a correlation coefficient (R) of 0.94, a Root Mean Square Error (RMSE) of 61.89 mm, and a Nash-Sutcliffe Efficiency (NSE) of 0.90. The model effectively captured intricate and variable rainfall patterns, providing accurate seasonal forecasts. In Yazd characterized by sporadic and irregular rainfall, the CNN-BiLSTM model still demonstrated high accuracy, achieving an R of 0.89, an RMSE of 7.48 mm, and an NSE of 0.76. These results highlighted the model's robustness, even under challenging climatic conditions. The BiLSTM model ranked second due to its bidirectional temporal memory followed by the DNN and SVM models, which were less effective in capturing complex temporal dependencies. The differences in performance were particularly evident during extreme rainfall events and periods of seasonal variability. These findings confirmed the importance of incorporating past rainfall data as model input, as well as the superiority of deep and hybrid architectures in addressing temporal and nonlinear relationships. The results aligned with similar studies and underscored the value of memory-based and hierarchical feature extraction models. Discussion of Results & Conclusion Incorporation of rainfall time lags significantly improved the predictive accuracy of all models, especially in irregular climates like that in Yazd. Notably, the CNN-BiLSTM model surpassed all other models in both regions, highlighting the benefits of hybrid deep learning architectures for managing complex spatiotemporal data. This study enhanced the broader understanding of intelligent forecasting systems for climate-related applications. It illustrated the adaptability and robustness of the CNN-BiLSTM model across various climatic conditions, providing reliable support for decision-making in water management, agricultural planning, and disaster preparedness. To further enrich this research, it is recommended to integrate satellite and radar datasets, expand temporal data coverage, and assess model performance across wider and more diverse geographical regions. These initiatives will strengthen the accuracy and generalizability of seasonal rainfall forecasts.

https://gep.ui.ac.ir/article_29932_c43fa040c2ea0243dcb1ca980e7016b7.pdf