Interpretable machine learning for urban heat mitigation: Attribution and weighting of multi-scale drivers
The study proposes a machine learning emulator that is highly interpretable for urban heat mitigation models. By classifying parameters into driving, local, and urban features, the method uses only urban features to assess cooling potential. Applied to a 2019 heat wave in Zurich, the approach improves prediction accuracy, especially when land‑use type splitting and past heat‑wave data are included, and identifies emissivity and surface albedo as actionable levers that could reduce air temperature by up to 1.68 °C and 1.32 °C respectively.
The study proposes a machine learning emulator that is highly interpretable for urban heat mitigation models. By classifying parameters into driving, local, and urban features, the method uses only urban features to assess cooling potential. Applied to a 2019 heat wave in Zurich, the approach improves prediction accuracy, especially when land‑use type splitting and past heat‑wave data are included, and identifies emissivity and surface albedo as actionable levers that could reduce air temperature by up to 1.68 °C and 1.32 °C respectively.
This research demonstrates how interpretable AI can accelerate scenario evaluation for Swiss urban climate policy, reducing computation time from hours to minutes.
Original source record. Open the original record to verify the underlying announcement.