Street microclimate prediction based on Transformer model and street view image in high-density urban areas

The frequent occurrence of extreme heat highlights the need to provide pedestrians with street microclimatic information. However, existing microclimate prediction methods can't achieve both high computational efficiency and fine accuracy over large areas. This study addresses this issue by employing a deep learning algorithm and a transformer model integrated with street view images. The model was trained and tested in high-density areas of Hong Kong. The results showed that the model can predict hourly mean radiant temperature (MRT) and wind speed with high spatial resolution accurately and efficiently, with high R2 of 0.99 and 0.82 and low RMSE of 1.53 degrees C and 0.14 m/s, respectively. Our model demonstrated higher prediction accuracy than two existing models. Moreover, our model exhibited a strong ability to generalize in new areas and real-life scenarios, particularly for MRT, with R2 of 0.98 and 0.93 and RMSE of 2.34 degrees C and 4.15 degrees C, respectively. The significance of the model lies not only in predicting hourly microclimate in high spatial resolution efficiently using easily acquirable street view images instead of 3D models, but also in facilitating pedestrians to choose walk paths, assisting in accurate building energy estimations, and designing thermally comfortable streets through 2D street view images.