🤖 machine learning

Volumetric Radar Echo Motion Estimation Using Physics-Informed Deep Learning: A Case Study Over Slovakia

This study demonstrates that while a physics-informed deep learning model can estimate altitude-specific motion fields from volumetric radar data over Slovakia, the resulting nowcasting accuracy does not significantly improve over 2D composite-based methods because meaningful vertical variability in precipitation motion is rare in the region.

Original authors: Peter Pavlík, Anna Bou Ezzeddine, Viera Rozinajová

Published 2026-03-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Peter Pavlík, Anna Bou Ezzeddine, Viera Rozinajová

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

The Big Idea: Predicting the Rain's Next Move

Imagine you are trying to predict where a flock of birds will fly in the next hour. You have a camera that takes pictures of them every few minutes.

Most weather forecasters (and computer models) currently look at a flat, 2D map of the birds. They calculate the average speed and direction of the whole flock and guess where they will be next. This works pretty well for simple flocks.

But what if the birds are flying in a giant, 3D cloud? What if the birds at the top are flying north, while the birds at the bottom are flying east? If you only look at the flat map, you miss that complexity.

This paper asks a simple question: Does looking at the birds in 3D (at different heights) actually help us predict their movement better than just looking at the flat map?

The Experiment: The "Slovakia Flock"

The researchers built a super-smart computer brain (a Deep Learning model) to test this. They used radar data from Slovakia, which acts like a giant 3D scanner, taking pictures of rain clouds at 16 different heights, from the ground up to 8 kilometers high.

They trained their AI to do two things:

The 2D Team: Look at a flat, averaged picture of the rain and guess the wind direction.
The 3D Team: Look at the rain at every single height level separately and guess the wind direction for each layer independently.

The Surprising Results: "The 3D Team Got Confused"

You might expect the 3D team to win because they have more information. But the results were a bit of a twist:

1. The Rain in Slovakia is "Sticky"
The researchers found that in Slovakia, rain clouds are usually very "cohesive." Think of a rain cloud like a solid block of Jell-O rather than a loose pile of feathers. The wind blowing the top of the Jell-O is almost exactly the same as the wind blowing the bottom.

The Finding: Even though the AI could see different wind speeds at different heights, it turned out that for 99% of the storms in Slovakia, the wind was moving in the same direction at all heights. The "3D" information was mostly redundant.

2. The "Ghost Splitting" Problem
Here is where it gets funny. Because the 3D AI was trying so hard to find differences where there weren't any, it started making mistakes.

The Analogy: Imagine you are pushing a long, heavy sofa across a room. If you push the front leg slightly faster than the back leg, the sofa starts to twist and split.
The Result: The 3D model estimated that the top of a rain cell was moving slightly slower than the bottom. When the computer combined these layers back into a flat map to show the forecast, the single rain cell looked like it had split into two separate cells.
The Consequence: This "splitting" made the forecast look like it covered more area. In the math, this looked like a "success" because the model "detected" more rain. But in reality, it was just a glitch—a hallucination caused by the model overthinking the 3D data.

The Verdict: Is 3D Worth It?

For Slovakia? No.
The researchers concluded that for the weather patterns in Slovakia, the extra complexity of 3D modeling isn't worth the trouble.

The 2D model (flat map) was just as accurate, if not better.
The 3D model was more expensive to run and created "ghost" rain cells that didn't exist.
The "gain" in accuracy was mostly an illusion caused by the model splitting storms apart.

The Takeaway:
Just because you can see in 3D doesn't mean you need to. If the rain is moving like a solid block of Jell-O, a flat map is perfectly fine.

However...
The authors add a crucial footnote: This might be different in places with wild weather. If you are in a region with massive thunderstorms where the wind at the top is blowing one way and the wind at the bottom is blowing the opposite way (like a tornado or a hurricane), then 3D modeling would be a superpower. But for the steady, "Jell-O-like" rain in Slovakia, the simple 2D approach wins.

Summary in One Sentence

The researchers tried to use a 3D super-vision AI to predict rain in Slovakia, but they discovered that the rain there moves so uniformly that the 3D model just got confused and started inventing fake, split-up storms, proving that sometimes a simple 2D map is all you need.

1. Problem Statement

Precipitation nowcasting (predicting rain/snow intensity over the next few hours) traditionally relies on 2D radar composites (e.g., CMAX or CAPPI) and assumes that precipitation motion is uniform across all altitudes. However, atmospheric motion is inherently three-dimensional; vertical wind shear and varying flow at different heights can cause precipitation systems to move differently at various altitudes.

While some studies suggest that modeling volumetric (3D) motion fields could improve nowcasting accuracy, it remains unclear whether the added complexity of estimating independent motion fields for multiple altitude layers yields practical benefits over standard 2D approaches, particularly in regions with specific climatological characteristics. This paper investigates whether explicitly modeling altitude-dependent advection improves nowcasting performance compared to vertically pooled 2D data.

2. Methodology

A. Dataset and Preprocessing

Data Source: Volumetric radar reflectivity data from the Slovak national weather radar network (4 C-band Doppler stations) collected between June 2018 and August 2022.
Resolution: 1 km horizontal resolution, 16 vertical levels (500m to 8000m), 5-minute temporal resolution.
Preprocessing:
- Denoising: Applied polarimetric filtering (using $\rho_{HV}$ ) and morphological operations to remove non-meteorological echoes and noise.
- Conversion: Converted reflectivity (dBZ) to rain rate (mm/h) using the Marshall-Palmer relationship.
- Vertical Reduction: Reduced 16 vertical levels to 8 levels via max-pooling to balance computational cost and information retention.

B. Model Architecture: 3D Motion Field U-Net

The authors propose a Physics-Informed Deep Learning framework based on a 3D U-Net architecture.

Input Strategy: Instead of treating time as a channel, the model treats time as an explicit third dimension. The volumetric radar data is decomposed into horizontal slices, and the model processes sequences of these slices.
Architecture: A standard U-Net encoder-decoder structure using 3D convolutions (kernel $3\times3\times3$ ) to capture spatiotemporal features.
Output: The model predicts a time-invariant horizontal motion field ( $u_x, u_y$ ) for each altitude slice independently.
Inference: The model processes each altitude slice independently (no vertical coupling during inference), stacking the resulting 2D motion fields to form a pseudo-3D field.

C. Training Objective (Physics-Informed)

To ensure the estimated motion fields are physically plausible and stable, the loss function combines data-driven and physics-informed terms:

Sequence-Consistent Loss ( $L_{MF}$ ): Minimizes the extrapolation error across a sequence of past and future frames, ensuring the motion field explains the evolution of the entire sequence, not just a single step.
Multi-Scale Pooling: Evaluates loss at multiple spatial scales to prevent overfitting to small-scale noise.
Physics-Regularization ( $L_{PI}$ ): Penalizes the divergence of the motion field ( $\nabla \cdot u$ ) to enforce the incompressibility assumption (mass conservation), guiding the network toward physically consistent advection patterns.
Total Loss: $L = (1-\beta)L_{MF} + \beta L_{PI}$ .

D. Baseline Comparison

A controlled experiment was conducted using an architecturally identical model trained on vertically pooled 2D CMAX data. This ensures that any performance differences are solely due to the use of volumetric information rather than architectural changes.

3. Key Contributions

Novel Framework: Development of a physics-informed 3D U-Net that estimates independent horizontal motion fields for multiple altitude layers directly from volumetric radar data.
Comprehensive Evaluation: A rigorous comparison between 3D volumetric motion estimation and 2D composite-based estimation using a large, real-world dataset from Slovakia.
Climatological Insight: A detailed analysis of the vertical coherence of precipitation motion in a mid-latitude continental climate, challenging the assumption that 3D motion modeling is universally necessary.
Artifact Identification: Identification of "non-physical cell-splitting" artifacts caused by independent altitude-wise motion estimation when collapsed back to 2D for evaluation.

4. Results

A. Quantitative Performance

Continuous Metrics (MAE, MSE, Bias): The volumetric (3D) model did not outperform the 2D baseline. In fact, it showed slightly higher errors at longer lead times (>30 mins) and developed a systematic positive bias (overestimation).
Categorical Metrics (Precision, Recall, ETS): The 3D model showed improved Recall and Equitable Threat Score (ETS) at longer lead times, suggesting better detection of precipitation events.
The Trade-off: The improvement in Recall was found to be largely driven by non-physical artifacts. Because motion fields were estimated independently for each layer, slight discrepancies in speed/direction between layers caused precipitation cells to "split" or stretch when the volume was collapsed into a 2D composite. This increased the spatial area of predicted rain (higher Recall) but did not reflect true storm evolution.

B. Motion Field Correlation Analysis

High Vertical Coherence: The study computed Pearson correlations between motion fields at different altitudes. Results showed extremely high correlation (often >0.9 for adjacent layers, >0.3 even for distant layers) across the dataset.
Seasonal Trends: Contrary to the hypothesis that summer convective storms would show high vertical variability, the summer months exhibited the highest correlations.
Rare Variability: Cases with significant vertical variability in horizontal motion were found to be rare in the Slovak dataset. Even in selected "outlier" events with low correlation, the motion differences were often attributable to data noise or radar occlusion rather than physical shear.

5. Significance and Conclusion

Main Conclusion: For the Slovak radar dataset, the additional complexity of 3D motion field estimation is not justified. The vast majority of precipitation events exhibit vertically coherent horizontal motion, meaning 2D composites capture the necessary dynamics. The marginal gains in categorical metrics are artifacts of the model's design rather than genuine predictive skill.
Implications for AI in Meteorology: This study highlights the importance of critically assessing the value of 3D data representations. In regions with strong vertical wind shear or complex terrain (where 3D motion varies significantly), volumetric models may be beneficial. However, in mid-latitude regimes dominated by stratiform or weakly convective systems, 2D approaches remain more efficient and robust.
Future Work: The authors suggest that future models should incorporate vertical coupling constraints or physical regularization to prevent the "cell-splitting" artifacts observed when independent layers are aggregated. They also note that the framework remains applicable to climates with stronger vertical variability.

In summary, while the proposed physics-informed deep learning model successfully learned to estimate altitude-wise motion fields, the study concludes that vertical coherence in horizontal precipitation motion is the norm in the studied region, rendering the computationally expensive 3D approach unnecessary for standard nowcasting tasks in this specific climatological context.