COMPARATIVE EVALUATION OF MACHINE LEARNING MODELS FOR RAINFALL-INDUCED SLOPE STABILITY PREDICTION IN BIHAR EMBANKMENTS USING FACTOR OF SAFETY

Abstract

Bihar sits at the heart of the Indo-Gangetic Plain and bears the full force of Hi malayan river flooding every year. The Bihar Water Resources Department has built over 3,500 earthen embankments along the Kosi, Gandak, Bagmati, Kamla-Balan, Burhi Gandak and Ganga river systems to protect more than 40 million people, yet roughly 28% of these structures suffer some form of failure — breaching, slumping or seepage-induced piping — each monsoon season. The economic toll reaches up to Rs. 2,400 crores annually, and between 2 and 5 million people are displaced from their homes as a direct result. The core problem is straightforward but difficult to solve at scale: site-by-site geotechnical assessment costs Rs. 50,000–2,00,000 per embank ment and takes 1–2 weeks, making comprehensive pre-monsoon screening of the full network financially and logistically impossible. This thesis builds a machine learning framework aimed squarely at that gap. A synthetic dataset of 10,000 slope stability scenarios was generated using the Infinite Slope Model, drawing on 14 years (2011–2025) of daily gridded rainfall data from IMD’s RF25 product (0.25◦×0.25◦ resolution). Eight input features were used: effec tive cohesion (c′), friction angle (ϕ′), unit weight (γ), slope angle (β), slope height (z), saturation ratio (m derived from IMD rainfall), mean daily rainfall and maximum daily rainfall. Each Factor of Safety was classified into three stability classes per IS:7894 1975: Stable (FoS ≥1.50), Marginal (1.25 ≤ FoS<1.50) and Unstable (FoS<1.25). Five classifiers were implemented and compared under a single unified frame work: Logistic Regression, Random Forest, SVM with RBF kernel, XGBoost and LightGBM.Thedatasetwassplit70/15/15(train/validation/test) with stratification pre served across all partitions. Class imbalance — the Marginal class accounts for only 13.2% of scenarios — was handled by applying SMOTE exclusively to the training set. All models were tuned via GridSearchCV or RandomizedSearchCV and validated using Repeated Stratified K-Fold Cross-Validation (5 folds × 5 repeats = 25 splits). iv The SVM with RBF kernel proved the strongest classifier, reaching 97.1% accu racy, a macro F1-score of 0.953 and an AUC-ROC of 0.997 — ahead of all baselines by 2.3–11.7 percentage points in accuracy. Cross-validation confirmed the result was not a lucky split: 99.40% ± 0.002 across 25 folds. The dataset labels were independently checked against Bishop’s Simplified Method (86% class agreement, r = 0.998) and the Fellenius Method (94%, r = 0.996). The best model was also validated against the Durgawati Earthen Dam (Kaimur, Bihar), where it produced a mean FoS = 1.864 — consistent with published Bishop’s values of 1.771–2.017. SHAP analysis confirmed the model had learned genuine geotechnical relation ships rather than statistical noise: slope angle ( |SHAP| = 0.438) and slope height (0.184) dominate predictions, with cohesion (0.137) and friction angle (0.112) captur ing soil strength. Rainfall features contribute 8.5% in total, confirming that integrating real IMD data adds measurable predictive value beyond geotechnical parameters alone.