An interpretable (explainable) model based on machine learning and SHAP interpretation technique for mapping wind...

Area of Science:

• Environmental Science
• Geosciences
• Data Science

Background:

• Wind erosion is a significant threat in global drylands, particularly in the Middle East and Iran.
• Wind erosion hazard maps are crucial for identifying high-risk areas and implementing mitigation strategies.

Purpose of the Study:

• To develop an interpretable machine learning (ML) model for mapping wind erosion hazards.
• To identify key features influencing wind erosion and interpret the predictive model's output using Shapley additive exPlanation (SHAP).

Main Methods:

• Employed four ML models: Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGB), and Quadratic Discriminant Analysis (QDA).
• Utilized Multivariate Adaptive Regression Spline (MARS) for feature selection and Variance Inflation Factor (VIF) to assess multicollinearity.
• Applied SHAP for model interpretability and feature importance analysis.

Main Results:

• Eight features, including elevation, soil bulk density, precipitation, aspect, slope, soil sand content, vegetation cover (NDVI), and lithology, were identified as most effective.
• RF model highlighted elevation and soil bulk density as the most important features.
• All ML models demonstrated high accuracy (AUROC > 90%, PR > 90%), with SVM performing slightly better. SVM results indicated moderate, high, and very high wind erosion hazard classes covering 20.9%, 23%, and 16.6% of Hormozgan Province, respectively.
• SHAP analysis confirmed soil sand content and elevation as primary contributors to the model's predictions.

Conclusions:

• This study pioneers the application of interpretable ML models for wind erosion hazard mapping in Southern Iran.
• Elevation and soil properties are critical factors in wind erosion.
• Incorporating model interpretability is essential for a deeper understanding of predictive model outputs in environmental studies.