نقش مدل‌های طبقه‌بندی داده‌محور در تفکیک رخدادهای گردوغبار وابسته و غیروابسته به چریان جتی در غرب و جنوب‌غرب ایران

Dust outbreaks are among the most critical environmental and meteorological hazards in western and southwestern Iran. Because dust occurrence and intensity are controlled by interacting synoptic-scale dynamics and near-surface conditions, improving dust monitoring and decision support benefits from combining a dynamical perspective with data-driven classification methods. The present study evaluates the performance of machine-learning classifiers in identifying and separating dust events that are dependent on versus independent of upper-level jet-stream conditions, with specific attention to the 250-hPa jet-stream level.

Hourly observations from 29 synoptic meteorological stations in the study region were compiled for the period 2010–2024. Dust occurrences were not identified solely based on reduced visibility. Instead, dust events were first detected using dust-related present-weather codes reported in synoptic (SYNOP/WMO) observations. Horizontal visibility was then used as an auxiliary quantitative indicator to stratify event intensity/type and to define the classification labels (e.g., separating dust-storm conditions associated with marked visibility reduction such as 200–1000 m from cases of suspended dust). To describe the atmospheric environment, dynamical–thermodynamic predictors were extracted from global reanalysis products over the same period. Upper-level circulation and jet-stream conditions were characterized at 250 hPa using geopotential height and horizontal wind (u and v components, or their derived wind speed as a jet indicator). Near-surface predictors included surface pressure, 2-m air temperature, and 10-m wind, representing the meteorological controls most directly linked to dust emission and transport. Reanalysis variables were collocated to station locations and temporally aligned with station reports.

Each dust-related sample was assigned to one of two synoptic regimes: “jet-stream present” and “jet-stream absent,” determined from the concurrent 250-hPa flow. For each regime, feature selection was performed using the Boruta algorithm to retain truly informative predictors while eliminating weak and noisy variables. The selected predictors were then used to train and test several supervised classification models, including Random Forest, XGBoost, Support Vector Machine (SVM), neural networks, and logistic regression. Model performance was assessed using Accuracy, AUC, and F1 score, and comparative diagnostics were summarized via radar plots. (Where applicable in the analysis workflow, SHAP-based interpretation was used to support physical understanding of the dominant predictors.)

The findings reveal a strong regime dependence in both the dominant predictors and the most skillful classifiers. In the jet-present regime, predictors describing upper-level dynamics and the pressure field were consistently emphasized, including 250-hPa geopotential height and 250-hPa horizontal wind, together with surface pressure and near-surface wind. In this regime, ensemble tree-based models, particularly XGBoost and Random Forest, showed the highest skill for the binary (two-class) discrimination between dust-storm conditions (e.g., 200–1000 m visibility) and suspended dust. This performance is consistent with the more organized synoptic patterns typically associated with jet-stream influence, which enhance the learnability of nonlinear links between upper-level circulation, pressure configuration, and near-surface wind forcing.

In contrast, in the jet-absent regime and in the four-class formulation, the relative contribution of upper-level dynamical predictors decreased, and the primary weight shifted toward near-surface variables, particularly surface pressure, 2-m temperature, and 10-m wind. Under these conditions, SVM with an RBF kernel exhibited the most stable performance across the four dust conditions, indicating that a smooth nonlinear decision boundary driven by near-surface thermodynamic–dynamic structure can be advantageous when upper-level jet constraints are weaker.

Overall, the results demonstrate that explicitly separating synoptic regimes and then using regime-appropriate predictors and model families can improve both the accuracy and the physical interpretability of dust-event classification in western and southwestern Iran. This regime-based, data-driven framework provides practical guidance for developing regional dust monitoring and decision-support systems, while also clarifying how upper-level jet-stream variability modulates the relative importance of upper-level versus near-surface meteorological controls.