Comparative Evaluation of Machine Learning Models for Predicting Monsoon-Induced Slope Failures in Lateritic Terrain

Abstract

The stability of cut slopes in lateritic terrain presents a significant challenge for infrastructure projects, particularly during intense monsoon seasons. This study presents a machine learning framework to predict the Factor of Safety (FoS) of laterite slopes along the Konkan Railway alignment in India. To overcome the limitations of a small field dataset (n=10), a Monte Carlo simulation was employed to generate a physically constrained, augmented dataset (n=210) that preserves the multivariate statistical distributions of the original field samples. A comprehensive preprocessing pipeline including StandardScaler, Principal Component Analysis (PCA), and Recursive Feature Elimination (RFE) was applied. Five machine learning algorithms—Decision Tree, Random Forest, Gradient Boosting, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—were evaluated using repeated 10-fold cross-validation. The results demonstrate that Gradient Boosting combined with RFE feature selection achieved the highest predictive performance (R² = 0.929, MSE = 0.0004).

Repository Structure

ml-slope-stability-laterite/
├── README.md                    # This file
├── LICENSE                      # MIT License
├── requirements.txt             # Python dependencies
├── .gitignore                   # Git ignore rules
├── data/
│   ├── raw/                     # Original field measurements (n=10)
│   │   └── 01b_original_data.csv
│   ├── augmented/               # Monte Carlo augmented dataset (n=210)
│   │   └── 01_augmented_dataset.csv
│   └── processed/               # Analysis outputs
│       ├── 02_descriptive_statistics.csv
│       ├── 05_pca_loadings.csv
│       ├── 07_rfe_feature_rankings.csv
│       ├── 08_model_performance_cv.csv
│       ├── 09_hyperparameter_tuning.csv
│       └── 12_cv_predictions.csv
├── src/
│   └── run_analysis.py          # Complete reproducible pipeline
├── notebooks/
│   └── exploration.ipynb        # Interactive exploration notebook
├── figures/                     # Publication-quality figures (300 DPI)
│   ├── fig0_mc_distribution_validation.png
│   ├── fig1_pca_scree_plot.png
│   ├── fig2_feature_importance.png
│   ├── fig3_decision_tree_pruned.png
│   ├── fig4_ccp_alpha_selection.png
│   ├── fig5_model_comparison.png
│   ├── fig6_correlation_heatmap.png
│   └── fig7_actual_vs_predicted.png
├── paper/
│   └── paper/Comparative Evaluation of Machine Learning Models for Pre-dicting Monsoon-Induced Slope Failures in Lateritic Terrain A Data Augmentation Approach.docx
└── docs/
    └── METHODOLOGY.md           # Detailed methodology documentation

Key Results

Model	Features	CV R²	CV MSE	Train R²
Gradient Boosting	RFE Top-5	0.929 ± 0.033	0.00036	0.997
KNN (k=5)	RFE Top-5	0.906 ± 0.040	0.00048	1.000
Random Forest	RFE Top-5	0.845 ± 0.067	0.00077	0.953
Decision Tree	RFE Top-5	0.668 ± 0.264	0.00151	0.915
SVM (RBF)	RFE Top-5	0.347 ± 0.322	0.00297	0.509

Top-5 Features (RFE): Cohesion, Friction Angle, Plastic Limit, Fines Percentage, Plasticity Index

Quick Start

Prerequisites

python >= 3.11
pip install -r requirements.txt

Run the Full Analysis

# Clone the repository
git clone https://github.com/sushantwaidande/ml-slope-stability-laterite.git
cd ml-slope-stability-laterite

# Install dependencies
pip install -r requirements.txt

# Run the complete pipeline
python src/run_analysis.py

This will:

1. Load the original field data (n=10)
2. Generate Monte Carlo augmented samples (n=210)
3. Apply preprocessing (IQR outlier capping, StandardScaler)
4. Perform PCA and RFE feature selection
5. Train and evaluate 5 ML models with 10-fold CV (5 repeats)
6. Perform hyperparameter tuning via GridSearchCV
7. Generate all figures and output CSVs

Reproducibility

All random seeds are fixed (random_state=42) for exact reproducibility. Running the script will regenerate all results, figures, and data files from scratch.

Methodology

Data Collection

Laterite soil samples were collected from two sites along the Konkan Railway alignment:

• Site 1 (Ratnagiri): Basalt parent rock, n=6 samples
• Site 2 (Karwar): Granite parent rock, n=4 samples

Laboratory tests determined 12 parameters: gravel, sand, silt, clay percentages; liquid limit; plastic limit; plasticity index; specific gravity; moisture content; cohesion; friction angle; and fines percentage.

Factor of Safety Calculation

The target variable (FoS) was calculated using the infinite slope model under saturated monsoon conditions:

FoS = [c' / (γ × H × sin α × cos α)] + [(1 - ru) × tan φ' / tan α]

Where:

• c' = effective cohesion (kPa)
• γ = unit weight (18 kN/m³)
• H = cut height (12 m)
• α = slope angle (45°)
• ru = pore pressure ratio (0.45, representing monsoon saturation)

Monte Carlo Data Augmentation

To overcome the small sample limitation, 200 synthetic samples were generated from multivariate normal distributions defined by the mean vectors and covariance matrices of the site-specific data. Physical constraints were enforced:

• Grain size fractions normalized to sum to 100%
• Liquid limit constrained to exceed plastic limit
• All parameters clipped to realistic geotechnical ranges

Citation

If you use this code or data in your research, please cite:

@article{author2026slope,
  title={Comparative Evaluation of Machine Learning Models for Predicting Monsoon-Induced Slope Failures in Lateritic Terrain: A Data Augmentation Approach},
  author={[Author Name]},
  journal={Geosciences},
  volume={16},
  year={2026},
  publisher={MDPI}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

This research was conducted as part of the Doctor of Engineering program at The George Washington University, School of Engineering and Applied Science.