Landslide susceptibility mapping using machine learning models and analytical hierarchy process along NH-513, Pasighat to Yingkiong, Arunachal Pradesh, India
Year: 2026
Type: article
Abstract: mapping of landslide susceptibility (LSM) along mountain highways in tectonically active regions demands methodologies capable of capturing complex non-linear relationships among geoenvironmental conditioning factors. The present study presents a rigorous comparative investigation of three frameworks—a machine learning-based artificial neural network model (ANN), random‑forest model (RF). In addition, and the conventional AHP multi‑criteria framework (AHP)—for delineating landslide-prone zones along National Highway 513 (NH-513) from Pasighat to Yingkiong (123 km), Arunachal Pradesh, India. Twelve geoenvironmental causative factors were integrated in a GIS environment. A georeferenced landslide inventory of 230 events was partitioned into 70% training and 30% testing subsets. Spatial autocorrelation between training and testing samples was assessed and mitigated through a five-fold spatial block cross-validation (SBCV) scheme applied during hyperparameter tuning. From field observations, with the conventional random split used for final benchmark evaluation alongside a discussion of potential AUC inflation. The ANN model—a multilayer perceptron (MLP) with two fully connected hidden layers (64 and 32 neurons), ReLU activation. Notably, Adam optimiser, and L2 regularisation—achieved a predictive rate AUC of 0.954, substantially outperforming AHP (AUC = 0.551) and demonstrating competitive performance with RF (AUC = 0.980; p = 0.087). A learning-curve analysis confirmed that model performance stabilises near the full training-sample count, providing evidence of sample adequacy. Prediction uncertainty was quantified via bootstrap resampling (n = 100), revealing higher uncertainty in transitional moderate-susceptibility zones. Feature importance analysis consistently identified slope gradient as the dominant conditioning factor and aspect as the least influential.
Source: Discover Geoscience
Authors: Mrinaljyoti Adhyapok, Mrinal Kumar Dutta
Language: English
Subfield: Management, Monitoring, Policy and Law
Field: Environmental Science
Domain: Physical Sciences
Open Access status: diamond
Funder: National Remote Sensing Centre
OpenAlex ID: W7223049727