Get started
Work

From Surface to Canopy to Action: A Systematic Review of Foundation Models, Neural Operators, and Agentic Decision Support for Urban Heat Island and Health-Risk Management

Year: 2026
Type: other
Abstract: Extreme heat is a growing health risk in cities, but most data-driven urban heat research relies on satellite land surface temperature (LST), which differs from the near-surface air temperature that people are exposed to. At the same time, three families of AI methods are reshaping environmental prediction and decision support: geospatial and Earth-observation foundation models; physics-informed machine learning and neural operators; and agents based on large language models (LLMs) that carry out geospatial analysis. These literatures have developed largely separately, and no review has assessed whether together they can support decision-making on urban heat and health. This systematic review assesses whether current evidence can support an uncertainty-aware, planner-facing decision-support system for urban heat and health that is grounded in canopy-layer air temperature. It covers four evidence streams: (1) the relationship between surface and canopy-layer urban heat islands, and methods for measuring and mapping urban air temperature; (2) AI prediction of urban thermal fields, including machine learning, foundation models, physics-informed models and neural operators; (3) AI for heat-health risk, early warning and heat-mitigation planning; and (4) LLM-based agentic systems for geospatial and hazard decision support. It then evaluates the streams jointly, examining how often four properties required by such a system occur together in the same study: an air-temperature target, predictive uncertainty, spatial or external validation, and explicit representation of urban geometry. Tropical South and Southeast Asian megacities are treated as a regional case, to assess how well the evidence transfers to these settings. Peer-reviewed journal articles published in English between 2016 and 2026 were identified in Scopus and ScienceDirect and screened with AI assistance, with all decisions made or confirmed by the author. The review is reported according to PRISMA 2020 and uses a structured narrative synthesis informed by SWiM, without meta-analysis. The expected outcomes are a map of where the evidence is mature and where it is missing across the four streams; a gap analysis of how far the required properties occur together, including in South and Southeast Asian cities; and a research agenda, including a candidate integration architecture for future development and testing. The protocol is registered retrospectively. Searches, screening and data extraction were complete at registration, and this is declared in this registration.
DOI: 10.17605/osf.io/t98h7
Open Access status: green
OpenAlex ID: W7223148473
Cites:
Cited by:
Related to: