Machine learning applications on air temperature prediction in the urban canopy layer: A critical review of 2011–2022

Abstract

Air temperature within the urban canopy layer is one of the most critical variables that impact the environmental sustainability of cities. With advantages in computational speed, machine learning approaches have been increasingly applied for urban air temperature prediction. Yet the current status and limitations of machine learning applications remain largely unclear. This study critically reviews 45 articles that use machine learning for urban air temperature prediction between 2011 and 2022. By separating the studies into two groups, we find that tree-based models fit for spatial prediction and neural network models work for temporal prediction. For spatial prediction, the root-mean-square error (RMSE) varies from 0.10 to 2.60 °C, and there is no clear relationship between RMSE and spatial resolution. Prediction errors of temporal studies tend to increase with the time horizon, with an RMSE of 1.35 ± 0.60 °C (4.02 ± 0.93 °C) for 4-h (5-days) ahead prediction. We provide three key recommendations to enhance the reliability and performance of future machine learning applications: 1) reporting metadata, quality control process, and model evaluation in detail, 2) fusing the temperature data from different sources and diversifying predictor variables, and 3) improving machine learning by stacking and integration with physical-based models.