Hyper-Local Temperature Prediction Using Detailed Urban Climate Informatics

The accurate modeling of urban microclimate is a challenging task given the high surface heterogeneity of urban land cover and the vertical structure of street morphology. Recent years have witnessed significant efforts in numerical modeling and data collection of the urban environment. Nonetheless, it is difficult for the physical-based models to fully utilize the high-resolution data under the constraints of computing resources. The advancement in machine learning (ML) techniques offers the computational strength to handle the massive volume of data. In this study, we proposed a modeling framework that uses ML approach to estimate point-scale street-level air temperature from the urban-resolving meso-scale climate model and a suite of hyper-resolution urban geospatial data sets, including three-dimensional urban morphology, parcel-level land use inventory, and weather observations from a sensor network. We implemented this approach in the City of Chicago as a case study to demonstrate the capability of the framework. The proposed approach vastly improves the resolution of temperature predictions in cities, which will help the city with walkability, drivability, and heat-related behavioral studies. Moreover, we tested the model's reliability on out-of-sample locations to investigate the modeling uncertainties and the application potentials to the other areas. This study aims to gain insights into next-gen urban climate modeling and guide the observation efforts in cities to build the strength for the holistic understanding of urban microclimate dynamics.