Physics super-resolution of thermal environment maps in urban area by deep neural network with attention

Abstract

Real-time micrometeorology predictions are important for heat-stroke risk reduction and safe drone logistics in urban streets. However, the extensive calculation cost hinders operational real-time predictions of micrometeorology. In order to solve this, it is proposed to use a super-resolution simulation system that can provide high-resolution (HR) predictions at the low cost for low-resolution (LR) simulations. The system utilizes the super-resolution that converts LR maps into HR ones by means of machine-learning technologies. The super-resolution technology has been advancing rapidly and its neural networks can now learn physics to some extent. It has, however, the room for improvement. We propose the Pixel Attention Super-Resolution Network (PA-SRN), which gives attention weights to features and space at the same time. The training and test data for the deep neural networks were generated from the building-resolving urban micrometeorology simulations with a multi-scale atmosphere-ocean coupled model named the Multi-Scale Simulator for the Geoenvironment (MSSG). The proposed PA-SRN model has improved learning accuracy compared to a conventional algebraic interpolation and other existing neural network models. In addition, the analysis of attention weight distributions has clarified how the proposed neural network learns the physics in the micrometeorology data.