Advancing Regional Landslide Hazard Mapping from Heuristic Indexing to Intelligent Ensembles in Langkat Regency, Indonesia Togi Tampubolon (a*), Juniar Hutahean (a), Muhammad Fazar Zuhri (a), Jeddah Yanti (b)(c)
(a) Department of Physics, Faculty of Mathematics and Natural Science, State University of Medan, Jl. Willem Iskandar, Indonesia
*togitampubolon[at]unimed.ac.id
(b) Doctoral Program in Environmental Science, Gadjah Mada University, Jl. Teknika Utara, Indonesia
(c) Department of Geography, Faculty of Mathematics and Natural Science, State University of Makassar, Jl. Daeng Tata, Indonesia
Abstract
Landslide hazard in Langkat Regency is traditionally assessed using a GIS-based heuristic assessment system in accordance with BNPB guidelines. However, these methods rely on subjective expert weights and assume nearly linear relationships among conditioning factors, which makes them inappropriate for the region^s complex terrain. This study addresses that limitation by developing an intelligent ensemble framework integrating Analytical Hierarchy Process-GIS (AHP-GIS) with a hybrid deep learning architecture, a Convolutional Neural Network (CNN) combined with a Random Forest (RF) classifier. Hazard maps were produced at high resolution using a verified landslide inventory (2020-2025) and 11 geo-environmental conditioning factors, including elevation, slope, aspect, plan and profile curvature, NDVI, rainfall, lithology, distance to fault, land use, and TWI, processed in GEE and Google Colab. The CNN+RF ensemble model outperformed individual models (Accuracy: 97.8%, AUC-ROC: 0.657, and F1: 0.0006). Based on the SHAP analysis, slope (0.0685), lithology (0.0542), and plan curvature (0.0483) were significant hazard predictors. The Pearson correlation analysis showed a strong negative correlation (r = −-0.83) between rainfall and distance to fault, suggesting that high-rainfall areas spatially coincide with fault-proximate areas. The ensemble hazard mapping showed that the Moderate-to-High danger zones covered an area of approximately 12.52 km2, mostly in steep and fault-proximate terrains. The sub-districts of Sei Bingai, Bahorok, and Besitang were consistently identified as high-risk areas. This framework provides an accurate and interpretable decision-support tool that can be validated in other landslide-prone locations in Indonesia.
Keywords: Landslide Susceptibility, Hazard Mapping and Explainability, AHP-GIS, Deep Learning, Langkat
