Landslide susceptibility mapping using machine learning models and analytical hierarchy process along NH-513, Pasighat to Yingkiong, Arunachal Pradesh, India

Landslide susceptibility mapping (LSM) along mountain highways in tectonically active Himalayan terrain requires methods that can represent non-linear interactions among terrain, geology, hydrology and anthropogenic controls. This study compares a multilayer perceptron artificial neural network (MLP-ANN), a Random Forest (RF) classifier and an expert-weighted analytical hierarchy process (AHP) for the 123 km NH-513 corridor between Pasighat and Yingkiong, Arunachal Pradesh, India. Twelve conditioning factors were prepared in a GIS environment and linked with a georeferenced inventory of 230 landslide initiation points and 230 non-landslide samples. Multicollinearity was screened using variance inflation factor analysis (all VIF < 5). Model performance was evaluated using a 70:30 train-test split, five-fold spatial block cross-validation (SBCV), ROC-AUC, accuracy, precision, recall, F1-score and Matthews correlation coefficient. The ANN achieved predictive AUC = 0.954, accuracy = 0.936 and MCC = 0.873, while RF achieved predictive AUC = 0.980, accuracy = 0.954 and MCC = 0.908. Conservative SBCV estimates were 0.921 ± 0.028 for ANN and 0.941 ± 0.022 for RF, indicating mild but transparent spatial autocorrelation effects. AHP showed limited discrimination (predictive AUC = 0.551). Bootstrap uncertainty analysis identified transitional and corridor-margin zones requiring field verification. Slope gradient was the dominant factor in both ML models, followed by LULC, geology and drainage density. The results provide a reproducible, uncertainty-aware framework for highway-slope screening in the Eastern Himalaya.