نوع مقاله : مقاله پژوهشی
نویسندگان
1 استادیار، گروه مهندسی احیاء مناطق خشک و کوهستانی، دانشکده منابع طبیعی، دانشگاه تهران، کرج، ایران.
2 دانشجوی دکتری، گروه مهندسی احیاء مناطق خشک و کوهستانی، دانشکده مناطق طبیعی، دانشگاه تهران، کرج، ایران.
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Dust storms pose significant environmental and economic challenges, particularly in arid regions like Sistan-Baluchestan Province, Iran. This study aims to compare the performance of individual models (GRNN and SVM) with a triple hybrid model (GRNN-SVM-LSTM) for forecasting the frequency of dust storm days (FDSD). Using hourly dust data from eight SYNOP codes of the World Meteorological Organization across five synoptic stations, spanning a 40-year period (1980–2020), the models were evaluated based on key performance metrics: Correlation Coefficient (R), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Nash-Sutcliffe Efficiency (NS). The triple hybrid model outperformed all other approaches, achieving the highest predictive accuracy in seasonal combinations 1 and 2. The SVM model ranked second, while the GRNN model performed relatively better in combinations 1 and 2 compared to combination 4. Overall, the GRNN-SVM-LSTM model demonstrated superior predictive performance for FDSD, with RMSE = 0.523–0.501, R = 0.999–0.989, MAE = 0.441–0.421, and NS = 0.907–0.893. These findings highlight the potential of the proposed model for improving dust storm forecasting and developing early warning systems.
کلیدواژه‌ها [English]
EXTENDED ABSTRACT
Dust storms are among the most significant natural hazards affecting arid and semi-arid regions, particularly in Iran, where they cause considerable environmental, health, and economic damage. These phenomena are driven by a combination of climatic factors and human activities, including land-use changes and deforestation. This study aims to evaluate the performance of a hybrid machine learning model that integrates General Regression Neural Network (GRNN), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) for predicting the frequency of dust storm days (FDSD). The analysis focuses on five meteorological stations in Sistan and Baluchestan Province, utilizing a 40-year dataset (1980–2020) comprising hourly visibility and weather codes defined by the World Meteorological Organization (WMO). The increasing frequency and intensity of dust storms in arid and semi-arid regions, particularly in Iran, necessitates accurate forecasting tools to mitigate their adverse environmental, economic, and health impacts. Traditional methods often fail to capture the complex nonlinear relationships between climatic variables and dust storm occurrences. By integrating advanced machine learning techniques, the GRNN-SVM-LSTM model addresses this limitation, offering a robust framework for improved planning, resource management, and policy development to reduce the negative impacts of these natural hazards. This study enhances the understanding of dust storm dynamics and provides actionable insights for policymakers and environmental managers to formulate effective mitigation strategies. The primary objective is to develop and assess the hybrid GRNN-SVM-LSTM model for predicting FDSD in Sistan and Baluchestan Province. Furthermore, the study compares the hybrid model's performance with individual GRNN and SVM models to improve forecasting accuracy and reliability based on climatic and meteorological variables.
The study investigates the GRNN-SVM-LSTM hybrid model and compares its performance with individual GRNN and SVM models for predicting the FDSD index across five meteorological stations: Zabol, Zahedan, Khash, Iranshahr, and Saravan. The analysis utilizes a 40-year dataset (1980–2020) that includes hourly horizontal visibility data and WMO weather codes. Meteorological observations were recorded every three hours, resulting in eight synoptic reports per day. Three distinct models were employed to predict FDSD: two standalone models, GRNN and SVM, and a hybrid GRNN-SVM-LSTM model. The models' performances were assessed using goodness-of-fit metrics, with the prediction horizon varying from one to four past seasons. Initially, the performance of the individual GRNN and SVM models was analyzed. The next step involved evaluating recurrent neural networks (RNNs) for processing time series and sequential data. Long Short-Term Memory (LSTM) networks, a specialized RNN architecture designed to learn and retain patterns in long-term time series data, were then integrated into the hybrid model. The LSTM architecture includes memory units capable of preserving information over time, making it particularly effective for complex sequential data. Given the multifaceted nature of dust storms, adopting a hybrid model is essential for accurately capturing both linear and nonlinear variables influencing these phenomena. Therefore, this study employs the GRNN-SVM-LSTM hybrid model to provide a novel and comprehensive forecasting approach.
The General Regression Neural Network (GRNN) model, implemented in R, demonstrated the best performance when using FDSD data from two previous seasons. The Root Mean Squared Error (RMSE) improved significantly across all five stations when compared to using data from four prior seasons. Similarly, the Support Vector Machine (SVM) model, also implemented in R, achieved optimal performance with FDSD data from one or two previous seasons. The GRNN-SVM-LSTM hybrid model outperformed both standalone models in predicting the FDSD index across all five stations. The hybrid model exhibited substantial improvements in the correlation coefficient and Nash-Sutcliffe efficiency, highlighting its superior forecasting capabilities. In conclusion, the hybrid GRNN-SVM-LSTM model delivered the most accurate predictions for FDSD in the Sistan and Baluchestan region, surpassing the standalone GRNN and SVM models. This superior performance underscores the potential of integrating advanced machine learning techniques for effective dust storm forecasting.
All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.
Data available on request from the authors.
The authors would like to thank the reviewers and editor for their critical comments that helped to improve the paper. The authors gratefully acknowledge the support and facilities provided by the Department of Reclamation of arid and mountainous regions, Faculty of Natural Resources, University of Tehran, Iran.
The authors avoided data fabrication, falsification, plagiarism, and misconduct.
The author declares no conflict of interest.
)