💻 computer science

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

This study demonstrates that deep learning models, particularly when trained on mixed wind-direction data and optimized sensor placements, significantly outperform traditional Kriging interpolation in reconstructing complex rooftop wind fields from sparse sensor measurements, offering a robust framework for urban air mobility and wind control applications.

Original authors: Yihang Zhou, Chao Lin, Hideki Kikumoto, Ryozo Ooka, Sibo Cheng

Published 2026-03-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yihang Zhou, Chao Lin, Hideki Kikumoto, Ryozo Ooka, Sibo Cheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out exactly how the wind is blowing across a flat rooftop. This is crucial for things like landing drones, setting up solar panels, or just making sure people on the roof don't get blown off their feet.

The problem? The wind on a roof is chaotic. It swirls, separates, and changes direction instantly. To map it perfectly, you'd need hundreds of wind sensors covering every inch of the roof. But that's too expensive and impractical. So, engineers have to use just a few sensors (maybe 5 to 30) and try to guess what the wind is doing in the empty spaces between them.

This paper is like a cooking competition to see who can make the best "Wind Map" using the fewest ingredients.

The Contestants

The researchers tested four different "chefs" (methods) to reconstruct the full wind map from sparse data:

The Traditional Chef (Kriging): This is the old-school method. It's like a very careful cartographer who draws lines between known points based on the assumption that the wind changes smoothly and predictably. It's reliable but can get confused if the wind does something weird (like a sudden swirl).
The Copycat (UNet): A deep learning model that acts like a student who has seen thousands of wind maps. It tries to "copy" the patterns it learned to fill in the blanks. It's great at seeing the big picture and the small details.
The Artist (CWGAN): Another AI, but this one is more creative. It doesn't just copy; it imagines what the wind should look like based on the rules of physics and turbulence. It tries to generate a realistic, artistic version of the wind field.
The Visionary (ViTAE): A newer type of AI that looks at the whole picture at once (like a bird's-eye view) rather than just looking at small patches. It's good at connecting distant parts of the roof that might be influencing each other.

The Training: Learning from One Direction vs. Many

The researchers tested these chefs in two different "kitchens":

Single-Direction Training (SDT): The chefs only practiced with wind blowing from one direction (say, straight North). When they were tested on wind from the East, they struggled. It's like a chef who only knows how to cook Italian food and is asked to make sushi.
Mixed-Direction Training (MDT): The chefs practiced with wind coming from North, Northeast, and East. When tested, they were much better at guessing the wind patterns, even if the wind came from a new angle. The big takeaway: To be a good wind predictor, you need to learn from all the different ways the wind can blow, not just one.

The Results: Who Won?

When sensors are very scarce (only 5 sensors): The Traditional Chef (Kriging) actually did a decent job if the wind was predictable. However, the AI Chefs struggled a bit because they didn't have enough clues.
When sensors are moderate to many (15+ sensors): The AI Chefs (especially UNet and CWGAN) completely crushed the Traditional Chef. They could reconstruct the swirling, chaotic wind patterns that Kriging missed.
The "Mixed" Advantage: When the AI chefs were trained on mixed wind directions, they became superstars. They outperformed the traditional method by a huge margin (up to 32% better at keeping the shape of the wind map correct).

The "Where to Put the Sensors" Puzzle

The paper also asked: Does it matter where we put the sensors?

Uniform Placement: Putting sensors in a neat grid (like a checkerboard).
Optimized Placement (QR Method): Using a smart math trick to figure out the best spots to put sensors. It's like a detective realizing that the most important clues are usually found near the corners or where the wind swirls, not in the boring middle.

The finding: Using the smart "Optimized Placement" made the system much more robust. If a sensor gets knocked slightly out of place (by wind or rain), the system using smart placement still works well, whereas the grid system might fail.

The "Time" Question

Finally, they asked: Should we average the wind data before we try to map it, or after?

After (Post-averaging): Reconstruct the wind map for every single second, then average them. (More accurate, but takes more computing power).
Before (Pre-averaging): Average the wind data first, then reconstruct the map. (Faster, but slightly less accurate).
Verdict: Doing it "After" is better, but if you are in a hurry or have limited computer power, doing it "Before" is a perfectly fine backup plan.

The Bottom Line

If you want to know the wind on a rooftop using only a few sensors:

Don't just use old math (Kriging) if you have enough sensors; use Deep Learning (AI).
Train your AI on all wind directions, not just one.
Put your sensors in the smartest spots, not just a neat grid.
If you have a lot of sensors, the AI is your best friend. If you have very few, the old math might still hold its own, but the AI is catching up fast.

This study proves that by using real-world wind tunnel data (not just computer simulations) and smart AI, we can build reliable, real-time wind maps for safer and smarter rooftop operations.

1. Problem Statement

Real-time reconstruction of rooftop wind fields is critical for the safe operation of drones, urban air mobility, and wind control systems. However, rooftop flows are characterized by strong nonlinearity, flow separation, and cross-directional variability, making accurate reconstruction from sparse sensor data extremely challenging.

Existing approaches face three primary limitations:

Data Source Bias: Most studies rely on Computational Fluid Dynamics (CFD) simulations for training, which may contain systematic biases and lack the natural turbulence variability and measurement noise found in real-world deployments.
Methodological Gaps: Traditional interpolation methods (e.g., Kriging) assume spatial stationarity, which fails in complex rooftop environments with recirculation zones. Conversely, deep learning studies often lack systematic comparisons between different architectures and rarely test generalization across different wind directions.
Real-World Constraints: Many studies ignore sensor position uncertainties (offsets/failures) and do not optimize sensor placement for specific datasets, often assuming ideal uniform grids.

2. Methodology

The study proposes a learning-from-observation framework using high-resolution Particle Image Velocimetry (PIV) data from wind tunnel experiments.

Data Source

Experiment: Conducted in a boundary layer wind tunnel at the University of Tokyo using a 1:200 scale rectangular building model.
Measurements: Instantaneous horizontal velocity fields ( $u, v$ ) at $z/H = 1.05$ for three wind directions: $0^\circ$ , $22.5^\circ$ , and $45^\circ$ .
Dataset: 8 seconds of time-resolved data per run ( $\Delta t = 0.001$ s), resulting in ~8,000 snapshots per realization.

Reconstruction Models

The study benchmarks Kriging interpolation (traditional baseline) against three deep learning architectures representing distinct modeling philosophies:

UNet: A deterministic encoder-decoder architecture with skip connections, optimized for detailed spatial reconstruction.
Conditional Wasserstein GAN (CWGAN): A generative adversarial network that learns nonlinear turbulent characteristics and perceptual quality.
Vision Transformer Autoencoder (ViTAE): A hybrid model combining Transformer attention mechanisms (for global dependencies) with CNNs (for local features).

Experimental Strategies

Training Strategies:
- Single-Direction Training (SDT): Trained on one direction, tested on others (assessing cross-direction generalization).
- Mixed-Direction Training (MDT): Trained on a mix of all directions, tested on independent realizations (assessing robustness to diverse flow patterns).
Sensor Configurations: Evaluated 5, 10, 15, 20, 25, and 30 sensors.
- Uniform Placement: Sensors placed via Voronoi tessellation.
- Perturbed Placement: Sensors shifted by $\pm 1$ grid cell to simulate installation errors.
- QR-Optimized Placement: Sensor positions selected via QR decomposition of POD (Proper Orthogonal Decomposition) modes to maximize linear independence of measurements.
Evaluation Metrics: Structural Similarity Index (SSIM), Geometric Mean Bias (MG), Normalized Mean Square Error (NMSE), and Factor of 2 (FAC2).

3. Key Contributions

Experimental Validation: Unlike most prior work, this study trains and validates models on real experimental PIV data, ensuring robustness to measurement noise and real-world turbulence.
Systematic Architectural Comparison: Provides a comprehensive benchmark of UNet, ViTAE, and CWGAN against Kriging, analyzing trade-offs between accuracy, computational cost, and geometric bias.
Training Strategy Insights: Demonstrates that Mixed-Direction Training (MDT) is essential for deep learning models to generalize across flow conditions, whereas Kriging degrades under MDT due to violated stationarity assumptions.
Sensor Optimization: Introduces a QR decomposition-based sensor selection method that significantly enhances system robustness against sensor position perturbations compared to uniform grids.
Practical Guidelines: Offers specific recommendations for method selection based on sensor density, training data availability, and deployment constraints (e.g., sensor uncertainty).

4. Key Results

Performance: Deep Learning vs. Kriging

Under SDT (Single Direction): Kriging outperforms deep learning at very low sensor counts (5 sensors) due to the lack of diverse training data for the networks. However, as sensor density increases (>20 sensors), deep learning surpasses Kriging in error reduction (NMSE) and reliability (FAC2).
Under MDT (Mixed Direction): Deep learning models consistently outperform Kriging across all metrics and sensor counts.
- Improvements: Up to 32.7% improvement in SSIM, 24.2% in FAC2, and 27.8% in NMSE compared to Kriging.
- Generalization: MDT training improved SSIM by up to 173.7% compared to SDT for deep learning models, proving the necessity of multi-directional data.

Architecture Comparison

UNet: Best for geometric accuracy (MG closest to 1.0) and balanced performance.
CWGAN: Achieves the highest structural similarity (SSIM) and robustness to perturbations but incurs significantly higher computational costs (45.6x UNet) and exhibits higher geometric bias.
ViTAE: Offers the best efficiency (lowest GFLOPs) with competitive accuracy, suitable for resource-constrained edge devices.

Robustness & Optimization

Sensor Perturbation: Deep learning models generally show better robustness to sensor position offsets ( $\pm 1$ grid) than Kriging, though sensitivity increases with sensor density.
QR Optimization: Using QR-based sensor placement enhances robustness by up to 27.8% under perturbations. However, it involves trade-offs, sometimes slightly reducing SSIM while significantly improving error metrics (NMSE) and bias (MG).

Temporal Averaging

Post-averaging (reconstructing then averaging) generally yields better performance than pre-averaging (averaging inputs then reconstructing). However, pre-averaging is a viable alternative when temporal data is limited or computational cost is a constraint.

5. Significance and Conclusion

This study bridges the gap between theoretical deep learning applications and practical engineering deployment for wind field reconstruction.

Practical Impact: The findings provide a roadmap for deploying wind monitoring systems. For scenarios with limited sensors and single-direction data, traditional Kriging remains viable. However, for real-world, multi-directional applications, deep learning (specifically UNet for balance or CWGAN for high fidelity) is superior, provided the training data includes diverse wind directions.
Deployment Strategy: The study emphasizes that sensor optimization and training strategy must be co-designed. QR-based optimization is recommended when sensor placement uncertainty is high, while uniform grids suffice for fixed, precise installations.
Future Directions: The work highlights the need for physics-informed objectives, online learning, and validation in full-scale atmospheric conditions to further extend these methods beyond controlled wind tunnel environments.

Code Availability: The source code and trained models are publicly available on GitHub, facilitating reproducibility and further research.