Extended Abstract
Introduction: Dust storms are among the most significant climatic hazards in the arid and semi-arid regions of Iran, accompanied by consequences such as reduced visibility, severe air pollution, threats to public health, decreased agricultural productivity, and damage to urban and rural infrastructure. In recent years, particularly in the western regions of the country, an increasing trend has been observed in the intensity, frequency, and spatial extent of this phenomenon. This alarming trend results from a combination of natural factors, such as recurring droughts, declining soil moisture, and strong winds, as well as human-induced drivers, including unbalanced land use changes, excessive exploitation of water resources, and unsustainable land management practices. Given the widespread impacts of this phenomenon on the environment and the livelihoods of local populations, accurately predicting the number of dusty days within specific time periods is of great importance as a critical tool for damage mitigation and informed operational and managerial decision-making. Achieving this goal requires the use of advanced data-driven methods and artificial intelligence algorithms that can be effective in identifying complex, nonlinear, and non-deterministic patterns.
Materials and methods: In this study, a nonparametric predictive model based on the K-Nearest Neighbors (KNN) algorithm was developed. The Particle Swarm Optimization (PSO) metaheuristic algorithm was employed as a catalyst to optimize the model structure and enhance its prediction accuracy. The input data included the Frequency of Dust Storm Days (FDSD) index from 26 synoptic stations located in 11 provinces across western Iran, covering the long-term period from 1981 to 2020. To construct the predictive model, lagged values of the FDSD index over the four previous time steps were used as input variables to accurately capture the temporal patterns of this phenomenon. Initially, the base KNN model was implemented by adjusting the k parameter. Subsequently, the PSO algorithm was applied to optimize key model parameters, including the number of influential neighbors and the weighting of input variables. The models’ performance was evaluated using four statistical indicators: the correlation coefficient (r), root mean square error (RMSE), mean absolute error (MAE), and Nash–Sutcliffe efficiency (NS) coefficient, to assess the model’s accuracy and stability in predicting the FDSD index.
Results and Discussion: The results show that the KNN-PSO model outperformed the standalone KNN model. The application of the PSO algorithm allowed for automatic and optimal determination of key KNN parameters, such as the optimal number of neighbors (k) and weighting of input variables. For instance, at Abadan station, the correlation coefficient (R) of the KNN-PSO model increased to 0.997, while the RMSE decreased to 0.113. In contrast, the KNN model values were 0.654 and 0.437, respectively, indicating a significant improvement in prediction accuracy using the hybrid model. Comparison between observed and predicted values confirmed the improved model performance with an increased frequency of dust storm days. Among the stations studied, Abadan, which recorded the highest FDSD values, showed the highest agreement between the observed and predicted data. Overall, high-dust stations, such as Abadan, Ahvaz, Masjed Soleyman, Bostan, Sarpol-e Zahab, and Bandar Mahshahr, exhibited strong correlations between actual and predicted values. In scatter plots, these predictions closely followed the 1:1 line (f(x) = x), indicating the high efficiency of the KNN-PSO model. Furthermore, results revealed that using lagged FDSD indices from previous seasons did not enhance model performance, and simpler models utilizing only one-step lag yielded more accurate predictions.
Conclusion: Overall, the results demonstrate that the hybrid KNN-PSO model can significantly enhance the accuracy of predicting the frequency of dust storm days, particularly at stations with high occurrence rates, such as Abadan. By leveraging the capability of the PSO algorithm to automatically and optimally determine the sensitive parameters of the KNN model, this approach improves predictive performance compared to the base model. The findings indicate that integrating metaheuristic optimization algorithms such as PSO with simple data-driven models such as KNN not only increases prediction accuracy and efficiency in climatically challenging regions but also enhances the stability and generalizability under varying climatic and spatial conditions. Therefore, the use of such hybrid approaches can be considered an effective strategy for improved monitoring and management of climate-related hazards, including dust storms, in arid and semiarid regions.