[Paper Review] Super-Resolution of Near-Surface Temperature Utilizing Physical Quantities for Real-Time Prediction of Urban Micrometeorology

Yuki Yasuda, Ryo Onishi+3 more|arXiv (Cornell University)|Aug 2, 2021

Urban Heat Island Mitigation4 citations

TL;DR

This paper proposes a physics-informed super-resolution (SR) model using a convolutional neural network with skip connections and channel attention to enhance near-surface temperature resolution from low-resolution (20 m) to high-resolution (5 m) in urban areas. Trained on building-resolving large-eddy simulations (LES) in Tokyo and generalized to Osaka, the model outperforms bicubic interpolation and temperature-only SR models, with building height being the most critical input for reducing errors near building boundaries and improving thermal accuracy under varying solar radiation.

ABSTRACT

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

Motivation & Objective

• To develop a real-time, high-resolution prediction method for urban near-surface temperature using super-resolution (SR) techniques.
• To improve the accuracy of low-resolution (LR) urban micrometeorological simulations by incorporating physical quantities beyond temperature.
• To evaluate the generalization capability of the SR model across different urban environments, such as Tokyo and Osaka.
• To investigate the role of physical inputs—especially building height and solar radiation—in enhancing SR performance and physical consistency.
• To enable real-time high-resolution (HR) temperature prediction by combining the SR model with low-resolution LES simulations.

Proposed method

• The SR model employs a U-Net-like architecture with skip connections to preserve spatial details during feature reconstruction.
• It uses separated feature extractors for four input modalities: near-surface temperature, building height, downward shortwave radiation, and horizontal wind velocity.
• A channel attention mechanism dynamically weights the importance of each input feature, with building height showing high attention under strong solar radiation.
• The model is trained on paired low-resolution (20 m) and high-resolution (5 m) temperature fields from building-resolving large-eddy simulations (LES) in Tokyo.
• Generalization is tested on an independent LES dataset from Osaka, confirming robustness across different urban morphologies.
• Inference time is short, enabling potential real-time integration with LR micrometeorological models.

Experimental results

Research questions

• RQ1 Can a deep learning-based super-resolution model trained on one urban area generalize to another urban area with different morphology?
• RQ2 How do physical quantities like building height and solar radiation influence the model’s ability to reconstruct high-resolution temperature fields?
• RQ3 To what extent does incorporating physical inputs improve super-resolution accuracy compared to temperature-only models?
• RQ4 How does the attention mechanism reflect physical relationships, such as the contrast between sunlit and shaded urban surfaces?
• RQ5 Can the model achieve real-time high-resolution prediction when coupled with low-resolution LES simulations?

Key findings

• The proposed SR model significantly outperforms bicubic interpolation and temperature-only SR models in reconstructing high-resolution temperature fields across both Tokyo and Osaka simulations.
• Building height is the most influential input after temperature, reducing errors particularly near building boundaries by enabling the model to recognize thermal gradients at urban structures.
• The model’s attention mechanism assigns higher weights to building height as downward shortwave radiation increases, reflecting the growing thermal contrast between sunlit and shaded areas.
• The model generalizes well to a different city (Osaka) without retraining, indicating strong transferability across urban morphologies.
• The inference time is sufficiently short to support real-time high-resolution prediction when combined with low-resolution LES models.
• The model implicitly learns physical relationships from LES data, even without explicit physics constraints in the loss function, suggesting emergent physical consistency from training data.