CC-GRMAS: A Multi-Agent Graph Neural System for Spatiotemporal Landslide Risk Assessment in High Mountain Asia

Mihir Panchal, Ying-Jung Chen, Surya Parkash · 2025

Read CC-GRMAS: A Multi-Agent Graph Neural System for Spatiotemporal Landslide Risk Assessment in High Mountain Asia by Mihir Panchal, Ying-Jung Chen, Surya Parkash (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-20875.txt.

Introduction

Mihir Panchal, and two co-authors with him, built CC-GRMAS, a multi agent graph neural system for landslide risk checks in High Mountain Asia. Landslides are a growing climate driven hazard there, yet satellite and ground data for timely warning remain fragmented across tools.

Three linked agents handle prediction, planning, and execution, sharing satellite views and environmental signals for real time awareness and response. Local factors feed the risk scores so warnings reflect actual slope and weather conditions rather than broad averages alone. The design targets faster detection and coordinated action in remote terrain where response windows are short and ground access is hard.

Catalog record

CREATORMihir Panchal, Ying-Jung Chen, Surya Parkash
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-20875

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

CC-GRMAS: A Multi-Agent Graph Neural System for Spatiotemporal Landslide Risk Assessment in High Mountain Asia Mihir Panchal, Ying-Jung Chen, Surya Parkash Abstract Landslides are a growing climate induced hazard with severe environmental and human consequences, particularly in high mountain Asia. Despite increasing access to satellite and temporal datasets, timely detection and disaster response remain underdeveloped and fragmented. This work introduces CC-GRMAS, a framework leveraging a series of satellite observations and environmental signals to enhance the accuracy of landslide forecasting. The system is structured around three interlinked agents Prediction, Planning, and Execution, which collaboratively enable real time situational awareness, response planning, and intervention. By incorporating local environmental factors and operationalizing multi agent coordination, this approach offers a scalable and proactive solution for climate resilient disaster preparedness across vulnerable mountainous terrains. 1 Introduction Landslides, driven by climate change, pose significant risks to life and economic stability, particularly in High Mountain Asia (HMA), where complex topography, active seismicity, and shifting precipitation patterns amplify vulnerability [ 1 , 2 ] . More than 1.5 billion people depend on the glaciers and rivers of the HMA, making the impacts of these hazards particularly severe [ 3 , 4 , 5 ] . Key climate change drivers, such as extreme rainfall, glacial retreat, freeze thaw cycles, and permafrost degradation, exacerbate the frequency and intensity of landslides [ 6 , 7 ] . Furthermore, increasing population pressure, expanding infrastructure, and systemic challenges such as limited early warning systems and fragmented disaster response mechanisms increase the region’s exposure to these catastrophic events [ 8 , 9 , 10 ] . Current approaches to landslide risk management in HMA reveal critical technical and operational gaps [ 11 , 12 , 13 ] . Technically, spatial modeling remains weak, with poor integration of multi source data, limited incorporation of contextual knowledge, and reliance on reactive systems [ 14 , 15 , 16 ] . Operationally, challenges include inadequate coordination among stakeholders, low scalability of existing frameworks, insufficient real time monitoring, and weak integration of local and regional actors. These gaps hinder effective prediction, planning, and response, leaving HMA communities vulnerable to increasing climate induced landslide risks and underscoring the need for innovative, scalable solutions [ 17 , 18 ] . To address these challenges, we propose the Climate Change Graph Risk Management and Analysis System (CC-GRMAS), a novel framework leveraging advanced AI techniques for proactive landslide risk management. CC-GRMAS employs a multi agent architecture integrating Graph Neural Networks (GNNs) with attention mechanisms for enhanced spatial modeling, Retrieval Augmented Generation (RAG) enhanced graph databases for knowledge aware decision making, and automated systems for proactive hotspot detection and real time intervention. The framework integrates data sets such as NASA’s Global Landslide Catalog (1,558 HMA events) and supports Sustainable Development Goals (SDGs) 13, 11, and 15 through its modular scalable design. By combining data preprocessing, GNN based prediction, GraphRAG pipelines, and multi agent coordination, CC-GRMAS addresses systemic gaps, offering a robust solution for mitigating landslide risks in HMA. 2 Dataset and Preprocessing The dataset used in this study was derived from the NASA Global Landslide Catalog (GLC), curated by the NASA Goddard Space Flight Center in collaboration with the Global Precipitation Measurement (GPM) mission. This catalog compiles landslide event records from a wide range of sources, including news articles, scientific literature, government reports, and citizen science contributions, focusing on landslides from 2007 to 2020 [ 19 ] . Data were obtained through the NASA Landslide Portal ( 1 1 1 https://gpm.nasa.gov/landslides/ , 2 2 2 https://landslides.nasa.gov/viewer ) (a) Landslides in High Mountain Asia (b) 1558 ground truth data points To ensure data consistency and usability for subsequent modeling, a series of preprocessing steps were performed. The dataset consisted of 1558 landslides events across in the HMA region. Event metadata like landslide profile, source of information and gazeteer point of the landslide were aligned with a structured knowledge graph schema to support the multiagent system. The processed dataset was then represented in a graph database, enabling seamless integration of spatial, temporal, and descriptive attributes for downstream analytical tasks. Table 1: Graph Database Node Type Distribution Node Type Count Percentage Description Event 1,558 61.1% Core landslide event records Source 440 17.2% Information sources and references GazetteerPoint 331 13.0% Geographic reference points LandslideProfile 223 8.7% Landslide characterization profiles Total 2,552 100.0% Complete node inventory The graph database organizes records into several distinct node types. Table 1 summarizes the distribution of nodes in the dataset. Further details of the dataset attributes are provided in Appendix A.1 . 3 Proposed Method This study develops CC-GRMAS, a multi-agent system using graph neural networks and retrieval-augmented generation for improved landslide forecasting and disaster response in HMA. CC-GRMAS models landslide risk through graph-based spatial relationships, enhancing prediction accuracy and enabling coordinated disaster response. The system employs an agentic architecture where each agent specializes in complementary aspects of landslide risk management. The Prediction Agent utilizes Graph Neural Networks with attention mechanisms to model spatial relationships between landslide events across geographic regions. The Planning Agent leverages Large Language Models integrated with Retrieval Augmented Generation to provide context aware risk analysis and climate impact assessments. The Execution Agent coordinates operational responses through automated hotspot detection and response generation workflows. Figure 2: Architecture Diagram We applied feature engineering, including spatial coordinate normalization, temporal encoding, and impact severity quantification, to improve training diversity and regional variation modeling. We then constructed dynamic proximity graphs using distance-based connectivity patterns to enhance spatial relationship modeling for landslide risk assessment. CC-GRMAS will be evaluated against NASA’s Global Landslide Catalog to validate performance across diverse geographic and temporal scenarios. [ 20 , 21 ] . The workflow of the CC-GRMAS multi agent system is shown in Figure 2 . This approach demonstrates that while individual machine learning models offer valuable capabilities [ 22 ] , multi agent coordination with graph based knowledge integration is essential for optimizing performance in complex applications, such as climate induced landslide risk management [ 23 , 24 , 25 ] . Details of the agentic architectures and implementation are provided in the Appendix A.2 . The implementation code is available on GitHub 3 3 3 https://github.com/MihirRajeshPanchal/CC-GRMAS . 4 Conclusion & Pathways to Climate Impact This work introduces CC-GRMAS, a multi-agent system for landslide forecasting and disaster response in HMA. It combines graph neural networks for spatial risk prediction, large language models using retrieval-augmented generation for contextual analysis, and automated response coordination, leveraging localized environmental data and dynamic graph representations to enhance the accuracy and timeliness of landslide risk assessments. CC-GRMAS provides real-time situational awareness, targeted planning, and rapid interventions for vulnerable communities, with a modular design adaptable to various geographies and hazards. It supports SDG 13 (Climate Action) by mitigating disaster impacts, SDG 11 (Sustainable Cities and Communities) by enhancing resilience, and SDG 15 (Life on Land) by protecting ecosystems. The framework equips policymakers and stakeholders with actionable insights to reduce risk and safeguard livelihoods in a changing climate. Appendix A Appendix A.1 Detailed Dataset Description The NASA GLC dataset undergoes continuous updates as new events are reported and verified, adhering to a standardized schema to ensure global comparability and high spatial precision of the data. Each landslide record encompasses a set of attributes across the

Read the complete text