11:30 AM - 11:45 AM
[AAS05-10] Physics Super-Resolution of Near-Surface Temperature for Urban Micrometeorology Using Convolutional Neural Networks
Keywords:super-resolution, building-resolving simulation, micrometeorological model, urban area, convolutional neural network, attention mechanism
In future cities, various IoT devices such as drones will constantly access meteorological data and social network information in cloud networks. Each system using IoT devices will provide a variety of services in response to complex changes in weather and society without people being aware of them. Such social services will require real-time predictions for urban micrometeorology.
Our research group has developed a micrometeorological model that can resolve buildings and tree canopies at several meter resolution in urban areas (e.g., [1]). However, the computational cost of such simulations is high, and the real-time prediction is difficult even with a supercomputer. We recently proposed a “super-resolution simulation method” [2] using deep learning. Super-resolution is a technique that infers high-resolution (HR) images from low-resolution (LR) ones. Once a neural network for the super-resolution is trained, it can infer HR fields from the LR micrometeorological simulations at a low computational cost.
The previous study [2] investigated the super-resolution of the near-surface temperature only by using the LR temperature. This study [3] super-resolved the temperature by using not only the LR temperature but also other physical quantities such as the HR building height distribution. This method is called here the physics super-resolution. We propose a new convolutional neural network with an attention mechanism to perform the physics super-resolution. The training and test of the neural networks were performed with the results of the micrometeorological simulations in Tokyo and Osaka. The Osaka data were used only for the test to estimate the generalizability of the neural networks. The HR temperature reproduced by the new model is more accurate than those by the bicubic interpolation and the image super-resolution taking only the LR temperature as input (Figure). The analysis of the attention weight reveals that the importance of building height increases as the downward shortwave radiation is strengthened. The contrast between sun and shade becomes stronger with the increase in solar radiation, which may affect the temperature distribution. The result implies that the proposed neural network learns physics from the training data.
[1] Matsuda, K., Onishi, R. and Takahashi, K. “Tree-crown-resolving large-eddy simulation coupled with three-dimensional radiative transfer model,” J. Wind Eng. Ind. Aerodyn., (2018)
[2] Onishi R., Sugiyama D., and Matsuda K. “Super-resolution simulation for real-time prediction of urban micrometeorology,” SOLA, (2019)
[3] Yasuda, Y., Onishi, R., Hirokawa, Y., Kolomenskiy, D., and Sugiyama, D. “Super-resolution of near-surface temperature utilizing physical quantities for real-time prediction of urban micrometeorology,” Build. Environ., (2022)
Our research group has developed a micrometeorological model that can resolve buildings and tree canopies at several meter resolution in urban areas (e.g., [1]). However, the computational cost of such simulations is high, and the real-time prediction is difficult even with a supercomputer. We recently proposed a “super-resolution simulation method” [2] using deep learning. Super-resolution is a technique that infers high-resolution (HR) images from low-resolution (LR) ones. Once a neural network for the super-resolution is trained, it can infer HR fields from the LR micrometeorological simulations at a low computational cost.
The previous study [2] investigated the super-resolution of the near-surface temperature only by using the LR temperature. This study [3] super-resolved the temperature by using not only the LR temperature but also other physical quantities such as the HR building height distribution. This method is called here the physics super-resolution. We propose a new convolutional neural network with an attention mechanism to perform the physics super-resolution. The training and test of the neural networks were performed with the results of the micrometeorological simulations in Tokyo and Osaka. The Osaka data were used only for the test to estimate the generalizability of the neural networks. The HR temperature reproduced by the new model is more accurate than those by the bicubic interpolation and the image super-resolution taking only the LR temperature as input (Figure). The analysis of the attention weight reveals that the importance of building height increases as the downward shortwave radiation is strengthened. The contrast between sun and shade becomes stronger with the increase in solar radiation, which may affect the temperature distribution. The result implies that the proposed neural network learns physics from the training data.
[1] Matsuda, K., Onishi, R. and Takahashi, K. “Tree-crown-resolving large-eddy simulation coupled with three-dimensional radiative transfer model,” J. Wind Eng. Ind. Aerodyn., (2018)
[2] Onishi R., Sugiyama D., and Matsuda K. “Super-resolution simulation for real-time prediction of urban micrometeorology,” SOLA, (2019)
[3] Yasuda, Y., Onishi, R., Hirokawa, Y., Kolomenskiy, D., and Sugiyama, D. “Super-resolution of near-surface temperature utilizing physical quantities for real-time prediction of urban micrometeorology,” Build. Environ., (2022)