Document Type : Research Paper
Authors
1 Assistant professor, Department of reclamation of arid and mountainous regions Engineering, Faculty of Natural Resources, University of Tehran, Karaj, Iran.
2 Ph.D. candidate, Department of Reclamation of arid and mountainous regions Engineering, Faculty of Natural Resources, University of Tehran, Iran.
Abstract
Keywords
Main Subjects
EXTENDED ABSTRACT
Dust storms, recognized as pervasive and destructive meteorological phenomena in arid and semi-arid regions, intensify desertification and drought, deplete water resources, increase soil salinity, cause traffic accidents in critical zones, and exacerbate respiratory and other health-related issues. These effects collectively impose substantial economic, social, and environmental costs. Such phenomena typically lead to severe visibility reduction, with horizontal visibility often dropping below one kilometer. Monitoring the spatial and temporal variations of dust storms is crucial for risk forecasting, mitigation, and the effective management of this challenge. Mapping land sensitivity to dust generation in Ilam Province using data mining models indicated that approximately 23% and 18% of the area fall under highly sensitive and very highly sensitive categories, respectively. Despite the increasing importance of dust storms and their extensive social, economic, and environmental repercussions, a review of the literature reveals limited research in this field. Addressing these gaps, this study explores the performance of machine learning models and the Element-Free Galerkin algorithm for modeling dust storms. The integration of numerical models with machine learning algorithms to predict the Frequency of Dust Storm Days presents an innovative and transformative approach. Employing the advanced gradient boosting algorithm for dust storm modeling proves to be a highly efficient strategy, offering several advantages. XGBoost demonstrates exceptional computational power and the ability to handle complex data, enabling precise and reliable predictions of storm intensity and timing. Moreover, its capacity to identify nonlinear relationships among variables provides deeper insights into the factors influencing storm occurrences. Its rapid processing and computational efficiency make it particularly suitable for real-time applications, such as developing early warning systems. These benefits collectively contribute to enhanced disaster management, reduced environmental and economic losses, and the formulation of more effective policies for combating dust storms. This novel approach not only significantly improves prediction accuracy but also provides deeper insights, paving the way for new advancements in managing and mitigating the impacts of dust storms. Additionally, employing the Element-Free Galerkin algorithm for dust storm modeling allows for the precise simulation of complex atmospheric flows with reduced computational demands and costs compared to machine learning approaches. The elimination of grid meshing requirements provides greater flexibility in analyzing regions with intricate geometries. This method’s strong capability to minimize numerical errors and model nonlinear interactions further enhances prediction accuracy. Its computational stability and ease of implementation make it an excellent choice for advanced analyses under dynamic atmospheric conditions. These features facilitate the development of more effective strategies for managing the adverse effects of dust storms.
This research focuses on a comparative analysis of dust storm frequency modeling using the XGBoost machine learning model and the mesh-free Galerkin method. The study aims to predict the frequency of dust storm days at eight meteorological stations in Ilam Province (Mehran, Dehloran, Doviraj, Ilam, Ivan- Gharb, Arkavaz-e Malek, Cham-e Gaz, and Kahreh-Heliyan) over a 40-year statistical period (1981–2020). In this study, hourly data on horizontal visibility of less than 1000 meters were employed to detect dust storms for all stations.
This is a machine learning algorithm that belongs to the ensemble learning methods. By combining boosting and regularization techniques, this algorithm constructs a robust and accurate predictive model. In boosting, several weak models (typically decision trees) are trained sequentially, with each new model correcting the errors of the previous one. The algorithm identifies the best split by evaluating all possible division points and uses a sketch algorithm to speed up computations. Specifically, at each training step, gradient descent is applied to update the model parameters and reduce the loss function. The second-order Taylor expansion is used to approximate the loss function and calculate the updated parameters. Furthermore, XGBoost utilizes various optimization strategies, such as column subsampling, row subsampling, and feature importance evaluation, to improve the model’s performance and stability. These strategies allow XGBoost to handle high-dimensional and sparse data more effectively, increase the interpretability of the model, and enhance its overall performance and robustness.
The element-free Galerkin method, similar to the finite element method, uses formulations based on the principle of virtual work or weighted residual approaches. These formulations allow for the derivation of the weak form of differential equations, offering greater stability than the strong form. To correctly apply essential boundary conditions, this method typically uses Lagrange multipliers or penalty methods.
Dust storms exert considerable impacts on air quality, public health, and natural ecosystems, necessitating meticulous analysis and robust predictive models. Advanced machine learning approaches, such as XGBoost, harness large-scale datasets and excel in detecting complex patterns, offering powerful tools for forecasting dust storm behavior. In parallel, numerical methods like the Element-Free Galerkin approach provide accurate and stable outcomes, leveraging advanced mathematical formulations, reduced computational costs, and simplified relationships. The integration of these modern and classical methodologies offers a comprehensive framework to enhance analysis and decision-making in dust storm management. In this study, the performance of the XGBoost machine learning model and the Element-Free Galerkin algorithm in modeling dust storms was rigorously evaluated using statistical indices, including R, RMSE, MAE, and NS. The analysis was conducted across eight meteorological stations in Ilam Province (Mehran, Dehloran, Doyrej, Ilam, Ivan Gharb, Arkwaz Malek, Chamgaz, and Kahreh Helian). Modeling accuracy improves with an increasing frequency of dust storm days, with Mehran station demonstrating the highest concordance between observed and predicted values. Based on the results of t-test analysis, it can be concluded that the mean time series of observed and predicted values are equal for both the XGBoost and EFG models. XGBoost, known for its ability to process large and diverse datasets, enhances prediction accuracy while minimizing computational time. Meanwhile, EFG offers a cost-effective and computationally efficient alternative. Combining these methods not only refines predictive accuracy but also enables detailed scenario analysis and practical solutions for risk mitigation, early warnings, and sustainable development initiatives.
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.
)