Evaluating the Application of Artificial Intelligence in Landslide Susceptibility Mapping Using XGBoost and SHAP Methods in Khalkhal County, Iran

Description

This study evaluates landslide susceptibility in Khalkhal County, Iran, using XGBoost and SHAP methods. A landslide inventory of 152 points and 12 conditioning factors were analyzed. The model achieved 83.70% accuracy and 89.98% AUC. SHAP analysis identified distance to roads as the most influential factor (1.441), followed by elevation, distance to faults, NDVI, and rainfall. The susceptibility map shows 40% of the area falls into high-very high classes, concentrated near roads, faults, and streams. This research provides an interpretable framework for landslide risk management and land-use planning.

Steps to reproduce

1. Data Collection: - Compile landslide inventory (152 points) through field surveys and Google Earth imagery interpretation. - Extract 12 conditioning factors: elevation, slope, aspect, profile curvature, TWI, NDVI, rainfall, distance to roads, distance to faults, distance to rivers, lithology, and land use using ArcGIS 10.7 and ENVI 5.6. - Resample all raster layers to 50×50 m cell size and reproject to a unified coordinate system. 2. Data Preprocessing: - Split the dataset into training (70%) and testing (30%) sets using random stratified sampling. - Standardize the input features using StandardScaler from scikit-learn. 3. Model Development: - Implement the XGBoost classifier using the xgboost library in Python. - Optimize hyperparameters using Grid Search with 5-fold cross-validation. - Train the model on the training dataset. 4. Model Evaluation: - Evaluate model performance on the test set using accuracy, precision, recall, F1-score, and AUC metrics. - Generate the ROC curve. 5. Model Interpretation: - Apply SHAP (SHapley Additive exPlanations) using the shap library. - Calculate global and local SHAP values to identify and interpret the contribution of each conditioning factor. 6. Susceptibility Mapping: - Apply the trained model to the entire study area to predict landslide susceptibility probabilities. - Classify the susceptibility map into five classes (Very Low, Low, Moderate, High, Very High) using the Quantile classification method. - Produce the final susceptibility map using ArcGIS. 7. Software and Libraries: - Python 3.x with libraries: scikit-learn, xgboost, shap, numpy, pandas, matplotlib. - GIS software: ArcGIS 10.7 or QGIS for spatial data processing and visualization.

Institutions

• Islamic Azad University, Tehran

  Tehran, Tehran

Categories

Licence