Study on landslide susceptibility mapping with different factor screening methods and random forest models

Area of Science:

• Geosciences
• Environmental Science
• Data Science

Background:

• Landslide susceptibility mapping (LSM) is crucial for predicting landslide occurrences.
• The selection of input factors significantly impacts the accuracy of predictive models.
• Effective factor screening is a critical initial step in building robust LSM models.

Purpose of the Study:

• To evaluate the impact of different factor screening methods on landslide susceptibility prediction accuracy.
• To compare the performance of various factor screening techniques in the context of LSM.
• To identify the most influential factors for landslide prediction in Jingdong County.

Main Methods:

• Construction of a landslide database using 136 landslide events and 11 selected factors from Jingdong County.
• Application of four factor screening methods: information gain ratio (IGR), GeoDetector, Pearson correlation coefficient, and multicollinearity test (MT).
• Development of a random forest (RF) model for LSM, utilizing datasets processed by each screening method, followed by accuracy validation using confusion matrices and ROC curves.

Main Results:

• Factor screening demonstrably enhances the accuracy of LSM models compared to using all original factors.
• The IGR-RF model achieved the highest Area Under the Curve (AUC) value of 0.9334, outperforming the non-screened model (AUC=0.9194).
• The IGR-RF model exhibited superior prediction performance, accurately classifying the largest proportion of landslides into the very high susceptibility zone (51.22%).

Conclusions:

• Factor screening is a beneficial preprocessing step for improving LSM model performance.
• The information gain ratio (IGR) method is highly effective for selecting relevant factors in landslide susceptibility modeling.
• NDVI, elevation, and aspect were identified as the most significant factors influencing landslides in the study area.