Interpretable machine learning for urban heat mitigation: Attribution and weighting of multi-scale drivers

Strategies to mitigate urban heat are often analysed using complex models. Here, we propose a machine learning method for highly interpretable emulators of such models to efficiently inform actionable levers to mitigate urban heat. We first pre-classify parameters according to scale and modifiability into driving- (large-scale & non-modifiable), local- (scale-bridging & intractable) and urban (small-scale & highly modifiable) features, and use exclusively urban features for cooling potential assessment. Using land-use type (LUT) -specific submodels, we estimate urban temperatures. Demonstrating this framework for a heat wave (HW) period in 2019 in Zurich, Switzerland, we find that LUT-splitting statistically significantly increases accuracy. Furthermore, prediction performance for HW events increases substantially when three or more days of past HW data is included in the training set. Finally, our framework identifies emissivity and surface albedo as actionable levers. Air temperature is predicted to decrease by 1.68°C and 1.32°C for emissivity and albedo increases of 0.1 and 0.5, respectively. Once trained, the emulator evaluates new mitigation scenarios in seconds to minutes, versus hours to days per scenario for the underlying physics-resolving simulations.