DeepUrbanDownscale: A physics informed deep learning framework for high-resolution urban surface temperature estimation via 3D point clouds
Linwei Chen, Bowen Fang, Lei Zhao, Yu Zang, Weiquan Liu, Yiping Chen, Cheng Wang, Jonathan Li
Abstract
Accurate high-resolution downscaling of surface climate variables (such as surface temperature) over urban areas has long been a critical yet unresolved research problem in the field of urban climate and environmental sciences. In this paper, we propose a novel physics informed neural network (PINN) based framework: DeepUrbanDownscale (DUD) for high-resolution urban surface temperature estimation. Anchored in process-based modeling and satellite remote sensing, the DUD network leverages the high-precision 3D point clouds to achieve accurate urban land surface temperature (LST) estimation at an ultra-high spatial resolution. This network, ingesting the high-precision land surface geometry information derived from 3D point clouds and guided by the atmospheric physics related to surface temperature, constructs a physics informed data-driven framework to fit high-resolution temperature distribution, which is otherwise difficult to be obtained by physical (numerical) simulations or traditional machine learning. Specifically, the proposed DUD network contains two branches: The Global Feature Perception (GPFP) branch and Local Urban Surface Perception (LUSP) branch. The former considers the broader-scale urban physical parameters, constraining the estimation results in accordance with the relevant physical laws. The latter, by employing a proposed local spatial coefficient index (LSCI), which is based on 3D point clouds, the estimation performance is further improved at a very high resolution. Results from designed experiments demonstrate that the proposed DUD network predicts the urban LST on a 30-by-30 m grid with the estimated error less than 0.2 Kelvin compared to the satellite measurement, which is well below the errors of other traditional methods.