Machine learning transfers heat risk screening to planned new towns

Assessing heat risk in a city that has not yet been built is a stubborn problem for urban planners. Thermal environments, population activity patterns, and the detailed fabric of streets and buildings simply do not exist at the planning stage, so conventional heat risk assessments can only describe danger after development is complete. A new study from Hanyang University’s ERICA Campus in South Korea proposes a way around this: take the heat risk relationships measured in a mature new town and translate them into variables that planners can calculate before a single foundation is poured.

The researchers used Bundang-Pangyo, a developed new town in South Korea, as their reference area. There, they characterized composite heat risk through three components: heat hazard, exposure, and built environment vulnerability. Spatial analysis of this composite measure revealed substantial variation within the city itself, showing that the relationships between urban form and heat risk were not uniform across the reference area.

The core methodological move was to identify planning-stage variables that could serve as a common language between the mature city and unbuilt sites. Land use, road network, and building development characteristics available in planning documents were used as intermediary variables for cross-regional machine learning prediction. The models trained on Bundang-Pangyo’s observed heat risk relationships could then be applied to areas where only plans exist.

The results showed a clear hierarchy of predictive value. Conventional land use and road characteristics alone captured meaningful variation in heat risk, but the inclusion of additional building development and three-dimensional planning information substantially improved performance. In other words, knowing how tall and how densely a district will be built adds real predictive power beyond what flat land use maps can offer.

To test the framework, the team applied it to two planned new towns: Goyang Changneung and Namyangju Wangsuk. The transferred models produced clearly differentiated planning-stage risk profiles for the two sites, demonstrating that the approach can distinguish spatial variations associated with different planned urban configurations rather than flattening every new town into the same risk picture.

SHAP analysis, a technique that quantifies how much each input variable contributes to a model’s predictions, consistently highlighted three variables as most strongly associated with the transferred predictions: planned cooling balance, planned normalized building coverage ratio, and planned impervious surface ratio. Together, these point to a joint importance of ecological cooling capacity and development intensity in shaping planning-stage heat risk.

The study’s limitations follow from its design. The framework was validated against one mature reference area and two planned sites, all within South Korea, and the predictions concern comparative risk screening rather than measured future temperatures. The authors also note that the transferred outputs are associations with planned configurations, not forecasts of realized thermal conditions.

Even with those caveats, the methodological contribution is significant. The study extends urban heat risk assessment from post-development diagnosis toward planning-stage comparative screening, giving planners a quantitative tool to evaluate heat implications of alternative development configurations before construction.

Subject of Research: Transferring heat risk prediction from a mature new town to planned new towns using planning-stage environmental variables and machine learning

Article Title: A cross-regional machine learning transfer prediction framework for heat risk based on environmental characteristics of mature new towns

Article References: Lee, D., Zeng, J., & Kim, H. (2026). A Cross-Regional Machine Learning Transfer Prediction Framework for Heat Risk Based on Environmental Characteristics of Mature New Towns. Frontiers in Environmental Science, 14, 2012398. https://doi.org/10.3389/fenvs.2026.2012398

Image Credits: AI Generated

DOI: 10.3389/fenvs.2026.2012398

Keywords: heat risk, machine learning, new towns, urban planning, SHAP analysis, South Korea, built environment, land use, urban heat, planning-stage screening, cross-regional, machine