Comparative study on landslide susceptibility mapping based on unbalanced sample ratio.

Area of Science:

• Geosciences
• Geographic Information Systems
• Machine Learning

Background:

• Landslide Susceptibility Mapping (LSM) is crucial for hazard assessment.
• The impact of unbalanced sample datasets on LSM model performance is not fully understood.
• Optimizing sample ratios is key to improving the accuracy of LSM.

Purpose of the Study:

• To investigate the effect of varying sample ratios on LSM model performance.
• To identify the optimal sample ratio for different machine learning models in LSM.
• To enhance the reliability and scientific basis of landslide susceptibility assessments.

Main Methods:

• Utilized the Zigui-Badong section of the Three Gorges Reservoir area for research.
• Employed 12 landslide susceptibility mapping factors and five training sample sets with ratios from 1:1 to 1:16.
• Applied C5.0, Support Vector Machine (SVM), Logistic Regression (LR), and Convolutional Neural Network (CNN) models for analysis.

Main Results:

• CNN, SVM, and LR models performed better with a 1:2 sample ratio compared to balanced sets.
• The 1:2 sample ratio demonstrated superior performance for these models in landslide susceptibility mapping.
• The C5.0 model consistently exhibited overfitting, requiring further investigation.

Conclusions:

• Unbalanced sample sets are vital for training effective LSM models.
• A sample ratio of 1:2 is identified as optimal for CNN, SVM, and LR models in this study.
• Findings contribute to more scientific and reliable landslide susceptibility zoning.