Rooftop Wind Field Reconstruction Using Sparse Sensors: From Deterministic to Generative Learning Methods

Yihang Zhou, Chao Lin, Hideki Kikumoto, Ryozo Ooka, Sibo Cheng

TL;DR

Researchers have developed a deep learning framework to reconstruct high-resolution rooftop wind fields using data from just a handful of sparse sensors. By training models on real-world wind-tunnel data rather than idealized simulations, the team demonstrated that architectures like UNet and Generative Adversarial Networks (GANs) significantly outperform traditional interpolation, especially when trained to handle multiple wind directions simultaneously.

What Problem Does This Paper Address?

As cities look toward a future filled with delivery drones and Urban Air Mobility (UAM) vehicles, the rooftop has become a critical piece of infrastructure. However, rooftops are aerodynamically "messy." When wind hits a building, it creates complex phenomena like separation bubbles, recirculating zones, and conical vortices.

For a drone trying to land safely, knowing the wind speed at its specific coordinates is a matter of life and death (or at least, expensive repairs). We can’t cover every square inch of a roof with sensors, so we have to "reconstruct" the full wind field from a few sparse data points. Traditional methods often rely on Computational Fluid Dynamics (CFD), which is too slow for real-time use, or simple interpolation, which struggles with the sharp, non-linear changes in wind speed found near building edges.

Methods

The authors benchmarked a traditional geostatistical method against three distinct deep learning "philosophies":

  1. Kriging (The Baseline): A traditional interpolation method that assumes spatial correlation—essentially, if it’s windy here, it’s probably windy a few inches away.
  2. UNet (Deterministic): An encoder-decoder neural network that treats reconstruction as an image-to-image translation task. It’s excellent at preserving fine-scale spatial details.
  3. ViTAE (Hybrid): A Vision Transformer Autoencoder that combines "global" attention (looking at the whole roof) with "local" convolutional processing.
  4. CWGAN (Generative): A Conditional Wasserstein Generative Adversarial Network. This model uses a "generator" to create wind fields and a "discriminator" to guess if they are real or fake, pushing the model toward high perceptual realism.

Crucially, they also implemented a QR-based sensor optimization strategy. Instead of placing sensors in a uniform grid, they used Proper Orthogonal Decomposition (POD) to identify the most "informative" locations on the roof—the spots that capture the most variance in the wind flow.

Datasets

Unlike many studies that rely on "perfect" synthetic data from computer simulations, this paper uses experimental wind-tunnel data.

  • PIV Measurements: They used Particle Image Velocimetry (PIV) to capture high-resolution (15x15 grid) snapshots of wind flow over a 1:200 scale model of a rectangular building.
  • Wind Directions: Data was collected for three approach angles: 0°, 22.5°, and 45°.
  • Real-World Constraints: The researchers intentionally added "noise" by perturbing sensor positions by ±1 grid cell to test how robust the models are to imperfect installation in the real world.

Key Insights and Ideas

  • The "Sparse Sensor Paradox": Interestingly, when sensors are extremely sparse (e.g., only 5 sensors) and the wind direction is known, traditional Kriging actually outperforms deep learning. The neural networks simply don't have enough context to "learn" the physics from five dots. However, as soon as sensor density increases or wind directions become unpredictable, deep learning takes a massive lead.
  • MDT is Mandatory: The study compared Single-Direction Training (SDT) with Mixed-Direction Training (MDT). They found that training on multiple wind directions isn't just a "nice-to-have"—it's essential. MDT models were far more robust and capable of generalizing to the complex, shifting winds of a real urban environment.
  • Perception vs. Precision: The CWGAN produced the most "realistic-looking" wind fields (highest SSIM), but the UNet often had better "geometric accuracy" (lower bias). This suggests that the choice of model depends on the application: do you need the field to look physically plausible, or do you need the exact number to be right?

Key Results

  • Performance Gains: Compared to Kriging, deep learning models improved the Structural Similarity Index (SSIM) by up to 32.7% and reduced the Normalized Mean Square Error (NMSE) by 27.8%.
  • The Power of Diversity: Switching from single-direction to mixed-direction training led to a staggering 173.7% improvement in SSIM for some models.
  • Speed: Deep learning inference was roughly 13.7 times faster than Kriging, making it much more suitable for the split-second decision-making required by autonomous drones.
  • Robustness: Using the QR-based sensor optimization increased the system's robustness to placement errors by up to 27.8%.

Limitations and Future Work

While the results are promising, the authors note several hurdles before this hits the "real world":

  • Geometry Specificity: The current POD-QR optimization is tailored to a specific building shape. If you move to a L-shaped or circular building, you likely need to recompute the basis.
  • Height and Scale: The study focused on a single height ($z/H = 1.05$). Real drones operate across a 3D volume, requiring multi-height or full 3D reconstruction.
  • Outdoor Validation: Wind tunnels are controlled environments. Future work needs to test these models against "messy" natural atmospheric conditions, including gustiness and thermal effects.

Summary updated

Abstract

Real-time rooftop wind-speed distribution is important for the safe operation of drones and urban air mobility systems, wind control systems, and rooftop utilization. However, rooftop flows show strong nonlinearity, separation, and cross-direction variability, which make flow field reconstruction from sparse sensors difficult. This study develops a learning-from-observation framework using wind-tunnel experimental data obtained by Particle Image Velocimetry (PIV) and compares Kriging interpolation with three deep learning models: UNet, Vision Transformer Autoencoder (ViTAE), and Conditional Wasserstein GAN (CWGAN). We evaluate two training strategies, single wind-direction training (SDT) and mixed wind-direction training (MDT), across sensor densities from 5 to 30, test robustness under sensor position perturbations of plus or minus 1 grid, and optimize sensor placement via Proper Orthogonal Decomposition with QR decomposition. Results show that deep learning methods can reconstruct rooftop wind fields from sparse sensor data effectively. Compared with Kriging interpolation, the deep learning models improved SSIM by up to 32.7%, FAC2 by 24.2%, and NMSE by 27.8%. Mixed wind-direction training further improved performance, with gains of up to 173.7% in SSIM, 16.7% in FAC2, and 98.3% in MG compared with single-direction training. The results also show that sensor configuration, optimization, and training strategy should be considered jointly for reliable deployment. QR-based optimization improved robustness by up to 27.8% under sensor perturbations, although with metric-dependent trade-offs. Training on experimental rather than simulated data also provides practical guidance for method selection and sensor placement in different scenarios.
BibTeX
@inproceedings{zhou2026rooftop,
  author = {Yihang Zhou and Chao Lin and Hideki Kikumoto and Ryozo Ooka and Sibo Cheng},
  title = {{Rooftop Wind Field Reconstruction Using Sparse Sensors: From Deterministic to Generative Learning Methods}},
  booktitle = {Building},
  year = {2026},
  doi = {10.1016/j.buildenv.2026.114480},
}
Figure