DEEPURBANMODELLER (DUM): A PROCESS-INFORMED NEURAL ARCHITECTURE FOR HIGH-PRECISION URBAN SURFACE TEMPERATURE PREDICTION

Yuan Zang, G. Huang, Wenjie Liu, Liming Chen, Dan Wu, Chi Wang, Jonathan Li2023 · DOI

TL;DR

DeepUrbanModeller (DUM) is a novel "physics-inspired" neural architecture that predicts urban land surface temperatures at an ultra-high 30-meter resolution. By fusing atmospheric physics with high-accuracy 3D point cloud data, the model achieves an estimation error of less than 0.2 Kelvin, significantly outperforming traditional statistical downscaling methods.

What Problem Does This Paper Address?

As climate change intensifies, cities face escalating risks from extreme heat. To design resilient urban spaces, planners need high-resolution Land Surface Temperature (LST) data. However, we face a "resolution gap":

  1. Satellite data is often too coarse or infrequent.
  2. Dynamic physical models (which simulate heat transfer) are computationally "expensive" and struggle with the sheer complexity of 3D urban geometry.
  3. Standard Machine Learning often ignores the underlying laws of physics, leading to predictions that might be statistically "correct" but physically impossible.

The authors tackle this by creating a framework that is both data-driven and "process-informed," meaning it respects the physical laws of how heat moves through a city.

Methods

The authors introduce a Physics-Inspired Neural Architecture for Modeling (PINAM) called DeepUrbanModeller (DUM). The architecture is split into two specialized branches:

  • Global Physical Feature Interpretation (GPFI): This branch uses a Multi-Layer Perceptron (MLP) to process broad atmospheric variables (like solar radiation, wind, and humidity). Its job is to ensure the model's baseline predictions align with the laws of atmospheric physics.
  • Local Urban Surface Insight (LUSI): This is the "high-res" branch. It uses a deep residual network to process a novel metric called the Urban Detail Orientation Index (UDOI).

The UDOI is the secret sauce. Instead of feeding raw, heavy 3D point clouds directly into the network (which is computationally taxing), the authors derive the UDOI to summarize 3D geometry—capturing building heights, surface roughness, and verticality. These two branches are eventually fused into a latent vector for the final temperature regression.

Datasets

The study focused on Zhang Zhou Harbor, China, covering roughly 150 square kilometers. The model was trained and evaluated using:

  • LST & NDVI: Sourced from NASA’s Landsat satellite (30m resolution).
  • Atmospheric Forcing: Data from the MERRA-2 reanalysis system (covering wind, pressure, and radiation).
  • 3D Structure: A massive 100GB dataset of 3D point clouds collected via UAVs and mobile laser scanning, manually labeled into eight categories (buildings, roads, vegetation, etc.).

Key Insights and Ideas

  • 3D Geometry is Non-Negotiable: Traditional models rely on 2D indices like NDVI (vegetation). This paper argues that 3D "roughness" and building layouts are just as critical for understanding how heat is trapped or reflected in "urban canyons."
  • The "Physics-Aware" Advantage: By splitting the model into global physics (GPFI) and local geometry (LUSI), the network doesn't just look for patterns in pixels; it "understands" that local temperature is a product of both regional weather and local 3D structure.
  • Efficiency through Indexing: The creation of the UDOI allows the model to benefit from high-fidelity 3D data without the massive overhead of processing raw point clouds in real-time, making it more scalable than previous 3D-aware attempts.

Key Results

  • Superior Accuracy: DUM achieved an average Root Mean Squared Error (RMSE) of 0.116K across different seasons. For context, traditional linear regression had an error of ~1.3K, and even Random Forest models hovered around 0.42K.
  • High Resolution: The model successfully downscaled temperature predictions from a 1000m scale to a 30m grid.
  • Ablation Success: The authors proved that adding the LUSI (3D data) branch was vital; models without it had significantly higher errors (~1.1K), proving that 3D city structure is a primary driver of temperature variation.

Limitations and Future Work

  • Geographic Scope: The current model was validated on a single city (Zhang Zhou Harbor). The authors acknowledge that performance needs to be tested across more diverse urban environments globally.
  • Data Bottleneck: Labeling high-precision 3D point clouds is labor-intensive. Future work may need to look into automated labeling or unsupervised feature extraction to scale the model.
  • Variable Expansion: The researchers plan to expand the model to predict other physical quantities, such as surface wind speed and turbulence, which are also critical for urban microclimate modeling.

Summary updated

Abstract

High-resoulution downscaling of surface climate metrics like urban surface temperature, is a crucial and ongoing research challenge in urban climatology and environmental studies. In this study we propose a groundbreaking Physics-Inspired Neural Architecture for Modeling (PINAM) called DeepUrbanModeller(DUM), designed specifically for urban microclimate temperature estimation. DeepUrbanModeller(DUM) harnesses process-based modelling and satellite remote sensing, and draws upon high-accuracy 3D point clouds to deliver precise estimations of urban Land Surface Temperature (LST) at ultra-high resolutions. By incorporating high-accuracy land surface geometric data sourced from 3D point clouds and guided by the principles of atmospheric physics linked to surface temperature, DeepUrbanModeller(DUM) creates a data-driven framework, informed by physical laws, to accurately model high-resolution temperature distributions a task challenging for numerical simulations or conventional machine learning. The DeepUrbanModeller(DUM) design integrates two key components: Global Physical Feature Interpretation (GPFI) and Local Urban Surface Insight (LUSI). The GPFI captures broader urban physical parameters, ensuring the estimates comply with relevant physical laws. The LUSI enhances estimation performance at high-resolution levels by utilizing a newly proposed Urban Detail Orientation Index (UDOI) derived from 3D point clouds. Experimental results demonstrate the DeepUrbanModeller(DUM)’s superior capability in estimating urban LST on a detailed 30-by-30 meter grid, achieving an estimation error of less than 0.2 Kelvin compared to satellite measurements, a performance surpassing traditional methodologies.
BibTeX
@inproceedings{zang2023deepurbanmodeller,
  author = {Yuan Zang and G. Huang and Wenjie Liu and Liming Chen and Dan Wu and Chi Wang and Jonathan Li},
  title = {{DEEPURBANMODELLER (DUM): A PROCESS-INFORMED NEURAL ARCHITECTURE FOR HIGH-PRECISION URBAN SURFACE TEMPERATURE PREDICTION}},
  booktitle = {˜The},
  year = {2023},
  doi = {10.5194/isprs-archives-xlviii-1-w2-2023-71-2023},
}
Figure