A machine learning-based analysis of the effect of aerosol optical depth and climatic variables on vegetation condition using remote sensing data: A case study of croplands in Shushtar, Khuzestan

Understanding how atmospheric pollution and climate variability shape vegetation health is essential for sustainable crop management in stress-prone regions. In this study, monthly satellite and reanalysis records of NDVI, aerosol optical depth (AOD), air and land-surface temperature, precipitation, wind speed, soil moisture and surface pressure were compiled for the croplands of Shushtar, Khuzestan. Two predictive frameworks were developed: a multiple-linear-regression model to capture first-order (linear) effects and an XGBoost model to resolve higher-order (non-linear) interactions. Variable influence was examined with two complementary tools: one-way ANOVA, used to test the statistical significance of each factor, and tree-based SHAP values, used to rank and interpret their contributions within the XGBoost ensemble. ANOVA confirmed that AOD, soil moisture, temperature and precipitation exert significant control on NDVI, whereas wind speed showed no meaningful effect. Both SHAP analysis and the linear coefficients indicated a consistent, moderate-to-strong negative impact of AOD on greenness, reflecting the light-screening and stress effects of dust. Precipitation and soil moisture carried the largest positive influences, while high temperatures reduced NDVI, especially during the hot, dusty summer months. The XGBoost model outperformed the linear alternative (test R² of ۰.۷۵ versus ۰.۶۸), demonstrating its ability to capture non-linearities and complex variable interactions. SHAP diagnostics further clarified that seasonal and monthly timing explains the bulk of predictable variance, with AOD and moisture variables providing the strongest physical controls once calendar effects are accounted for. By using ANOVA, linear regression, XGBoost, and SHAP together, this study builds a clear toolkit for separating linear and non-linear effects on vegetation. The results show that both air pollution and shifts in weather- and water-related conditions play key roles in determining crop health across Shushtar's semi-arid croplands.

کلیدواژه ها:

NDVI ، AOD ، Satellite dataset ، XGBoost ، SHAP analysis