نوع مقاله : مقاله کامل علمی پژوهشی
عنوان مقاله English
نویسندگان English
Background and objectives: Recurrent droughts and the unprecedented decline in water levels of Lake Urmia in recent decades have transformed this unique aquatic ecosystem into one of the primary centers of environmental crises at both the national and regional levels. This crisis has transcended a mere hydrological challenge, evolving into a prominent environmental issue for the region, which has led to extensive negative feedbacks. The widespread desiccation of the lake bed has not only altered the natural landscape of the region but has also intensified the phenomenon of salt-laden dust generation and transport. These processes have imposed detrimental consequences on public health, the agricultural sector, the local economy, and the ecological sustainability of the area. The dispersion of salt-laden particulate matter, resulting from dust storms, threatens the respiratory health of communities living around the lake and has reduced the yield of agricultural products across vast areas of Azerbaijan. Consequently, the accurate and timely prediction of atmospheric pollutants, such as Particulate Matter (PM₁₀), is of strategic importance. These particles, due to their ability to penetrate deep into the human respiratory system, are considered the primary risk factor for pulmonary diseases. The objective of this research was to develop, train, and evaluate the performance of a data-driven predictive model based on the Extreme Gradient Boosting (XGBoost) algorithm to estimate monthly PM₁₀ concentrations across the Lake Urmia basin, thereby providing a precise tool to facilitate optimal management and decision-making for environmental authorities.
Materials and methods: To achieve the main objective of this research, a comprehensive and multidimensional database was designed and compiled. This database included a wide range of meteorological variables (such as temperature, precipitation, and wind speed), spatial data (including land-use maps and topography), and remote sensing parameters, such as the Aerosol Optical Depth (AOD), which play a fundamental role in identifying the physical and chemical behavior of dust phenomena in the atmosphere. Data pre-processing steps included the removal of outliers, normalization of scales, and the management of missing values using temporal interpolation methods. The XGBoost model, as a machine learning algorithm based on decision trees—which operates with high power in regression and classification problems—was selected to analyze this data. This model was developed based on monthly data during an extensive training period (2000–2020). To ensure the generalizability and stability of the model, the evaluation process was conducted using out-of-sample testing data related to the years 2021–2023, which guaranteed the stability of the model’s performance in the face of new climatic conditions. In addition to standard accuracy metrics (including the Coefficient of Determination (R²), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE)), advanced temporal and spatial analyses were performed on the model’s residuals to identify systemic weaknesses and structural errors. Furthermore, the relative importance of input variables was assessed using the Permutation Importance method to determine which factor (temperature, wind speed, AOD, etc.) plays a more decisive role in PM₁₀ fluctuations.
Results: The results of the performance evaluation showed the high accuracy of the XGBoost algorithm in estimating monthly PM₁₀ concentrations. In the modeling phase, the R² reached 0.84, and in the out-of-sample validation phase, it reached 0.71, indicating the high capability of the model to generalize to new conditions and future climatic variables. The RMSE and MAE values were also reported to be relatively low in both phases (0.402 and 0.315 in the modeling phase, respectively), which reinforces the computational accuracy of the predictions. The temporal stability analysis of the residuals showed that the model maintains good functional stability throughout the months of the year; however, in some spring months, due to convective instabilities and severe temperature differences, the atmosphere exhibits more non-linear behavior, and a slight tendency toward the underestimation of actual PM₁₀ concentrations was observed. The investigation of the spatial distribution of residuals indicated that the highest prediction error was specifically concentrated in the northern and northwestern parts of the basin. These areas overlap with vast salt pans and dried lake beds, indicating that local emission sources in these regions play a more prominent role, which may not be fully reflected in large-scale variables. Based on the variable importance analysis, the AOD index was introduced as the strongest predictor, followed by air temperature, wind speed, and precipitation, which had the highest share in determining PM₁₀ concentrations. This ranking confirms that the direct interaction between atmospheric optical indices and thermal air parameters is the main driver of pollutant changes in the region.
Conclusion: In summary, this study showed that the XGBoost algorithm, as an advanced data-driven approach, has high potential for the accurate modeling and analysis of the temporal and spatial dynamics of Particulate Matter (PM₁₀) concentrations in areas affected by lake desiccation, including the Lake Urmia basin. While classic statistical models face significant limitations when encountering the complex and non-linear relationships of climatic and remote sensing data, the XGBoost model, due to its tree structure and gradient boosting process, was able to identify and extract the patterns governing dust dispersion with high accuracy. The findings of this research provide an efficient and reliable tool for environmental decision-makers and policymakers so that they can better manage the environmental risks caused by salt-laden dust by relying on more accurate predictions. Also, the results of this study allow for the creation of a platform for continuous air quality monitoring, which can serve as a reference for evaluating management interventions in identified centers. Finally, given the desirable performance of the model in this research, the use of this approach can be an important step toward improving environmental monitoring and formulating climate change adaptation strategies in the Lake Urmia basin.
کلیدواژهها English