Reconstruction of landslide events in urban setting using LLM-based Agentic AI with multimodal data
4 affiliations
- 1Department of Civil and Environmental Engineering, The Hong Kong Polytechnic University, 11 Yuk Choi Rd, Hung Hom, Kowloon, Hong Kong Special Administrative Region
- 2Department of Civil Engineering, National Yang Ming Chiao Tung University, Taiwan
- 3Research Organization of Information and Systems, The Institute of Statistical Mathematics, 10-3 Midori-cho, Tachikawa, Tokyo, 190-8562, Japan
- 4The Graduate University for Advanced Studies (SOKENDAI), 10-3 Midori-cho, Tachikawa, Tokyo, 190-8562, Japan
Abstract
This study explores the application of agentic artificial intelligence (AI) systems, powered by large language models (LLMs), to automate key components of post-landslide investigations. Two core tasks are addressed: extracting technical information from witness reports using a customized Retrieval-Augmented Generation (RAG) pipeline, and estimating landslide geometry from site images via fine-tuned YOLOv8 segmentation model combined with chain-of-thought (CoT) prompting in a multimodal LLM. Validation with historical landslide events shows that the information extractor produces accurate, source-attributed summaries that align closely with professional engineering reports, while the geometry estimation workflow infers landslide dimensions and debris volumes consistent with official records. By leveraging pre-trained foundation models and tool-based reasoning, agentic AI systems offer a solution to the challenge of applying data-hungry deep learning frameworks in the data-scarce domain of geotechnics. The results highlight the potential of agentic LLM-based systems to support efficient, interpretable, and scalable geohazard investigations, advancing automation in geotechnical engineering.
Related
- Multimodal Interaction and Residual Fusion for Cross-Well Acoustic-Impedance Reconstruction
- Landslide Susceptibility Modelling and Infrastructure Exposure in the Zillertal, Austrian Alps
- Machine-learning modelling of regional landslide susceptibility in eThekwini, South Africa
- Landslide Susceptibility Modeling and Mapping: A Comparison of Frequency Ratio and Decision Tree Models
- Brief communication: Landsifier 2 – Python-based Software for Automated Landslide Type Classification
BibTeX
@article{pang2026reconstruction,
title = {{Reconstruction of landslide events in urban setting using LLM-based Agentic AI with multimodal data}},
author = {Hansong Pang and Man Kong Lo and Yat Fai Leung and Stephen Wu},
journal = {SOILS AND FOUNDATIONS},
year = {2026},
month = may,
doi = {10.1016/j.sandf.2026.101789},
url = {https://doi.org/10.1016/j.sandf.2026.101789},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Reconstruction of landslide events in urban setting using LLM-based Agentic AI with multimodal data
AU - Hansong Pang
AU - Man Kong Lo
AU - Yat Fai Leung
AU - Stephen Wu
PY - 2026
DA - 2026/05/22
JO - SOILS AND FOUNDATIONS
DO - 10.1016/j.sandf.2026.101789
UR - https://doi.org/10.1016/j.sandf.2026.101789
AB - This study explores the application of agentic artificial intelligence (AI) systems, powered by large language models (LLMs), to automate key components of post-landslide investigations. Two core tasks are addressed: extracting technical information from witness reports using a customized Retrieval-Augmented Generation (RAG) pipeline, and estimating landslide geometry from site images via fine-tuned YOLOv8 segmentation model combined with chain-of-thought (CoT) prompting in a multimodal LLM. Validation with historical landslide events shows that the information extractor produces accurate, source-attributed summaries that align closely with professional engineering reports, while the geometry estimation workflow infers landslide dimensions and debris volumes consistent with official records. By leveraging pre-trained foundation models and tool-based reasoning, agentic AI systems offer a solution to the challenge of applying data-hungry deep learning frameworks in the data-scarce domain of geotechnics. The results highlight the potential of agentic LLM-based systems to support efficient, interpretable, and scalable geohazard investigations, advancing automation in geotechnical engineering.
SN - 0038-0806
SN - 1881-1418
SN - 2524-1788
ER -