Super-resolution of near-surface temperature utilizing physical quantities for real-time prediction of urban micrometeorology

The present paper proposes a super-resolution (SR) model based on convolutional neural network and applies it to the near-surface temperature in urban areas. SR incorporates skip connection, channel attention mechanism, separated feature extractors for inputs of temperature, building height, downward shortwave radiation, horizontal velocity. We train with sets low-resolution (LR) high-resolution (HR) images from building-resolving large-eddy simulations (LESs) city, where resolutions LR HR are 20 5 m, respectively generalization capability is confirmed LESs another city. estimated fields more accurate than those bicubic interpolation image that takes only as its input. Except input, height most important reconstruct enables reduce errors near boundaries. considers appropriate boundary each information. analysis weights indicates importance increases radiation becomes larger. contrast between sun shade strengthened increase solar which may affect distribution. short inference time suggests potential proposed facilitate real-time prediction metropolitan areas by combining an LES model.