An Interpretable Machine Learning Method to Inform Urban Adaptation to Heat Waves

As the deadliest climate event over the past 30 years in the US, heat waves are projected to be more intense and dangerous under global warming. Cities, composed of unevenly distributed populations and diverse landscape, suffer from significant adverse impacts of heat waves with varying land surface temperatures across different areas. In addition to the climate, effective adaptation to heat waves also requires an understanding and evidence regarding how urban natural, physical, and social processes influence urban thermal variations. Prior studies that mostly rely on traditional regression methods fall short in interpreting the complex role of these factors in influencing fine-scale thermal variations within and even beyond the physical boundaries of these factors. To address this research gap, we propose an interpretable machine learning method based on graph neural network to disentangle the significant factors that influence local thermal conditions. Our model can explain 35.3% extra variations compared to the baseline ridge regression model. The results of the model can inform urban adaptation strategies, including nature-based solutions, such as evaluating the cooling efficiency of spatially variant greening interventions at fine-scale levels.