Graph-Enhanced Multimodal Valuation: Integrating Climate Physical Risk into Green Mortgage Analytics with LLMs and Knowledge Graphs

Abstract

Traditional property valuation and mortgage underwriting pipelines largely ignore explicit modelling of physical climate risk such as floods, heat stress, and extreme weather, despite growing evidence that these hazards materially affect asset values, default probabilities, and portfolio resilience. This omission creates a structural blind spot for banks and housing finance companies attempting to align with emerging green finance, ESG, and climate disclosure regimes. Green mortgages and sustainable housing finance products still tend to focus narrowly on energy efficiency metrics while underweighting location-specific climate hazards and resilience features. This paper proposes a climate-aware valuation framework that combines multimodal large language models, geospatial APIs, and Neo4j-based knowledge graphs to inject structured climate risk signals into property valuation workflows. The architecture comprises four layers: (i) a multimodal extraction layer using document understanding models such as LayoutLM/vision-enhanced LLMs to extract building attributes, materials, and layout features from loan and property documents; (ii) a geospatial integration layer that enriches each property with hazard and climate indicators from platforms such as ISRO Bhuvan and OpenWeatherMap; (iii) a knowledge graph construction layer that encodes properties, hazards, resilience features, and valuation events as nodes and relationships; and (iv) a GraphRAG-based risk scoring layer that performs reasoning over the graph to derive composite climate risk scores and climate-adjusted valuations. Experimental results on a hybrid (semi-simulated) dataset of residential properties show that the proposed system improves document-level extraction F1 by approximately 5–8 percentage points over text-only baselines, reduces valuation RMSE by 10–15 per cent when climate features are included, and lowers hallucinated risk explanations by leveraging graph-constrained retrieval. The framework is designed to be compatible with BFSI IT constraints. It can support green mortgage origination, portfolio climate stress testing, and ESG reporting by providing explainable, queryable climate risk signals linked to underlying evidence.