CERBERUS: A Three-Headed Decoder for Vertical Cloud Profiles

The paper introduces CERBERUS, a probabilistic three-headed encoder-decoder framework that generates vertical radar reflectivity profiles from satellite and surface data to bridge observational scale mismatches and improve the representation of complex cloud structures in weather and climate models.

The Gist

The Big Problem: The "Flat Photo" vs. The "3D Cake"

Imagine you are trying to understand a complex, multi-layered cake. But all you have is a single, flat photograph of the top of the cake. You can see the frosting and the sprinkles, but you have no idea if there is a layer of chocolate, a layer of fruit, or a hollow space inside.

This is exactly the problem scientists face with weather and climate models.

• Satellites are like that flat camera. They take beautiful pictures of the top of the clouds (the "frosting") from space.
• Radar is like a knife that cuts through the cake to see the layers inside. But radar stations are rare and only exist in a few spots on Earth.

Because we mostly have "flat photos" but need to know the "3D structure" to predict rain, storms, and climate change, there is a huge gap in our knowledge.

The Solution: Meet CERBERUS

The authors created a new AI tool called CERBERUS. The name is a nod to the three-headed dog from Greek mythology that guards the gates of the underworld. In this case, CERBERUS guards the "gates" of the atmosphere, trying to guess what's happening inside the clouds.

How it works:

1. The Inputs: CERBERUS looks at two things:
   • Satellite Photos: It sees the temperature and brightness of the cloud tops from space.
   • Ground Weather: It checks the temperature, wind, and humidity right where the radar is standing.
2. The Magic: It uses these clues to guess the vertical profile of the cloud. It tries to answer: "How high is the cloud? Is it thick or thin? Is it raining inside?"

The Secret Sauce: The "Three-Headed" Brain

Most AI models are like a student taking a test who just writes down one single answer (e.g., "The cloud is 5km high"). If the student is wrong, the model is just wrong.

CERBERUS is different. It has three heads (three prediction outputs) that work together to give a much smarter answer. Instead of guessing one number, it guesses a range of possibilities:

1. Head 1 (The "Is it there?" Head): It calculates the probability that a cloud actually exists at a specific height. (Is this empty air, or is there a cloud?)
2. Head 2 & 3 (The "How big?" Heads): If there is a cloud, these heads guess the shape of the cloud's density. They don't just say "It's 50% dense"; they say, "It's likely 50%, but it could be anywhere between 40% and 60%."

The Analogy:
Imagine you are trying to guess the weather for tomorrow.

• A standard AI says: "It will rain at 2:00 PM."
• CERBERUS says: "There is a 90% chance of rain starting between 1:45 PM and 2:15 PM, but it might be a light drizzle or a heavy downpour."

This "uncertainty" is actually a superpower. It tells scientists, "I'm pretty sure about this part, but this part is tricky, so be careful."

What Did They Find?

The team trained CERBERUS using data from a radar station in Oklahoma (the "Southern Great Plains") and tested it on data it had never seen before.

• It's Good at Simple Clouds: For flat, uniform clouds (like a blanket), CERBERUS is very accurate.
• It Struggles with Complex Clouds: When clouds are messy, layered, or have deep thunderstorms, the AI gets a bit fuzzy.
• The "Fuzziness" is Honest: Here is the best part. When CERBERUS is confused by a complex storm, it doesn't just guess wildly. It admits, "I don't know exactly what's happening here," by showing a wide range of possibilities. This helps scientists know where their models might be failing.

Why Does This Matter?

Think of weather forecasting like trying to drive a car in the fog.

• Old way: You guess where the road is based on a blurry photo. You might crash.
• New way (CERBERUS): You get a map that says, "The road is definitely here, but over there, it might be a ditch or a bridge. Drive carefully."

By turning 2D satellite pictures into 3D cloud guesses, CERBERUS helps scientists build better climate models. It bridges the gap between what we can see from space and what is actually happening inside the sky, helping us predict extreme weather and understand climate change with greater confidence.

Summary

CERBERUS is an AI that looks at satellite photos and ground weather to guess the 3D shape of clouds. Instead of giving a single, rigid answer, it gives a "best guess" along with a "confidence level," admitting when the clouds are too complex to be sure. This helps scientists understand the invisible layers of our atmosphere better than ever before.

Technical Summary

1. Problem Statement

Atmospheric cloud processes are a primary source of uncertainty in weather and climate models. A critical challenge is the scale mismatch between available observations:

• Satellite Data: Provides extensive spatiotemporal coverage but is predominantly 2D (top-of-atmosphere perspectives), lacking vertical resolution.
• Ground-based Data: Provides high-resolution 3D vertical profiles (e.g., via radar) but is limited to sparse, fixed locations.

Existing machine learning approaches (GANs, U-Nets, Masked Autoencoders) often attempt to reconstruct 3D clouds from 2D satellite data but suffer from two main limitations:

1. Determinism: They produce single-point estimates, failing to quantify the inherent ambiguity in mapping a 2D image to a 3D structure.
2. Data Gaps: Many models only target daytime clouds or fail to utilize near-surface meteorological context effectively.

The goal is to develop a probabilistic framework that infers vertical radar reflectivity profiles from geostationary satellite imagery and surface meteorology, explicitly modeling the uncertainty of this inverse problem.

2. Methodology: The CERBERUS Framework

CERBERUS (Cloud Estimation with vertically-Resolved Beta-distributed Retrievals of Uncertainty and Structure) is a probabilistic encoder-decoder architecture designed to predict a Zero-Inflated Beta (ZIB) distribution for radar reflectivity at 128 vertical altitudes.

Data Sources & Preprocessing

• Target (Ground Truth): Ka-band zenith-pointing radar (KAZR) data from the ARM Southern Great Plains (SGP) site (Jan 2020–Mar 2025). Data is interpolated to 128 altitudes (160m–15km) and smoothed.
• Inputs:
  • Satellite: 30-minute GOES-16 brightness temperatures (13.3µm, 11.9µm, 11.2µm, 8.4µm, 6.8µm, 3.9µm) and visible reflectance (0.65µm), remapped to an $8 \times 8$ grid.
  • Meteorology: Near-surface variables from MERRA-2 reanalysis (10m temp, winds, humidity, surface pressure).
  • Context: Time-of-day and day-of-year (sine-encoded).
• Preprocessing: Non-cloudy values are zero-filled (threshold -60 dBZ), creating a zero-inflated target distribution.

Model Architecture

The model employs a three-headed encoder-decoder structure (illustrated in Figure 1):

1. Encoder:
   • Processes 2D satellite fields via Conv2D blocks (downsampling to a latent vector).
   • Embeds near-surface scalar variables.
   • Combines them using a FiLM-like (Feature-wise Linear Modulation) approach to condition the latent representation.
2. Decoder:
   • Projects and reshapes the latent vector.
   • Utilizes Transposed Conv1D (Conv1D.T) blocks to upsample the latent vector back to the vertical profile dimension.
3. Three Prediction Heads:
   Instead of predicting a single reflectivity value, the model outputs three parameters for each of the 128 altitude levels:
   • $\pi$ (Pi): Probability of non-cloudiness (zero-inflation component).
   • $\alpha$ (Alpha) & $\beta$ (Beta): Parameters defining the Beta distribution for reflectivity values when clouds are present.

Training Strategy

• Loss Function: The model is trained using Negative Log-Likelihood (NLL) of the ZIB distribution, rather than standard Mean Squared Error (MSE).
  $$L(y, (\alpha, \beta, \pi)) = \begin{cases} -\log(\pi) & y = 0 \\ -\log((1-\pi) \cdot B[\alpha, \beta](y)) & y > 0 \end{cases}$$
  This allows the model to learn the probability of "no cloud" separately from the distribution of reflectivity when clouds exist.
• Optimization: Adam optimizer, batch size 100, optimized hyperparameters (dropout, activation functions) via Optuna.

3. Key Contributions

1. Probabilistic 3D Inference: Introduces a framework that outputs a full probability distribution (ZIB) rather than a deterministic point estimate, explicitly quantifying uncertainty in vertical cloud structures.
2. Three-Headed Architecture: Demonstrates that separating the prediction of cloud presence ($\pi$) from the distribution parameters ($\alpha, \beta$) improves performance over single-head deterministic or non-zero-inflated baselines.
3. Multi-Modal Conditioning: Successfully integrates geostationary satellite imagery with near-surface meteorological reanalysis to improve vertical profile reconstruction.
4. Physical Ambiguity Modeling: The model's uncertainty estimates (spread) correlate with physical complexity; it shows higher uncertainty for multilayer and dynamically complex clouds where the 2D-to-3D mapping is non-unique.

4. Results

• Performance Metrics:
  • Classification: ROC-AUC of 0.957 for detecting cloudy altitudes.
  • Regression: $R^2 > 0.6$ for reflectivity prediction.
  • Error Reduction: The ZIB model outperforms deterministic and non-ZI beta baselines, particularly below 8 km altitude (approx. 1 dBZ error reduction).
• Cloud Regime Analysis:
  • Best Performance: Low and thin clouds (Stratocumulus, Cumulus) with the lowest RMSE.
  • Challenging Cases: Deep convective clouds and Nimbostratus show higher RMSE, consistent with the difficulty of these regimes in existing literature (e.g., SatMAE).
  • Uncertainty Quantification: The model correctly predicts wider distributions (higher uncertainty) for complex multilayer clouds and narrow distributions for clear or simple convective cores.
• Feature Importance:
  • Infrared brightness temperatures are the most critical inputs.
  • Near-surface temperature is the most important scalar variable, indicating thermodynamic boundary layer influence.
  • Visible reflectance is the least important (as expected, due to day-only availability).

5. Significance and Future Work

• Bridging Scales: CERBERUS provides a pathway to generate synthetic, model-relevant 3D cloud observations from widely available 2D satellite data, aiding in the evaluation and calibration of global climate models.
• Uncertainty Awareness: By providing uncertainty estimates, the model helps researchers understand where 2D satellite data is insufficient to constrain 3D cloud physics, preventing overconfidence in model outputs.
• Future Directions:
  • Extending the framework to predict microphysical quantities (cloud water content, droplet size).
  • Incorporating global high-resolution model data and Doppler radar.
  • Exploring more expressive architectures like Mixture Density Networks or Transformers for global-scale applications.

In summary, CERBERUS represents a significant step forward in remote sensing by moving beyond deterministic reconstruction to probabilistic inference, effectively capturing the physical ambiguity inherent in deriving 3D cloud structures from 2D satellite observations.