Applying machine learning-based building morphology statistics to the source area model for wind speed estimation on heterogeneous urban surface roughness

 Wind speed estimation tools are the foundation for evaluating thermal and air quality conditions in highly heterogeneous urban environments. The source area model is a promising tool for this purpose. In this study, we examined the applicability of the model and potential for extension by integrating LiDAR observations, and a machine learning algorithm to estimate building morphology statistics.

 With focusing on the coastal urban area and the dense urban center in Japan, the source area model was tested in terms of the reproducibility of the key wind profile parameter; the power law coefficient. An empirical coefficient, which determine the trimming area of buildings, used in building morphology statistics calculation, was examined.

 The horizontal resolution of building morphology statistics, including the mean, maximum, and standard deviation of the building height, was tested to determine an optimal format for the source area model. A machine learning system (ML), an attention U-net-based algorithm was applied to generate height-related parameters (i.e., the above-mentioned parameters and the frontal area index). The source areal model estimated the power law coefficient of the wind reasonably well under O(20 m)-resolution of the ML-based parameters.

 Although our approach relied on high-accuracy building footprint and digital elevation data available in Japan for ML, the results offer a promising pathway for extending model utility to a wider region using a generalized building morphology statistics format.