Ensemble Graph Neural Networks for Probabilistic Sea Surface Temperature Forecasting via Input Perturbations

This paper demonstrates that an ensemble of Graph Neural Networks for regional sea surface temperature forecasting, which introduces diversity through spatially coherent input perturbations like Perlin noise rather than model retraining, achieves well-calibrated probabilistic forecasts with improved uncertainty representation at no additional training cost.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Picture: Predicting the Ocean's Mood

Imagine you are trying to predict the weather for your town. You know the ocean is a huge, chaotic place, and its temperature (Sea Surface Temperature, or SST) changes constantly. This matters because it affects fishing, shipping, and even how storms form.

Traditionally, scientists use massive, super-complex physics equations to predict this. It's like trying to calculate the exact path of every single water molecule in the ocean. It works, but it's incredibly slow and expensive, like trying to solve a Rubik's cube by hand while running a marathon.

Recently, scientists started using AI (specifically a type called Graph Neural Networks or GNNs) to learn from past data. These AI models are fast and cheap. However, they have a big flaw: they are deterministic. This means if you ask them the same question twice, they give the exact same answer. They don't know when they are unsure. In the real world, we need to know not just what will happen, but how confident we are that it will happen.

The Problem: The "One-Model" Trap

If you rely on just one AI model, you get one single prediction. But what if that model is slightly off? You have no way of knowing.

To fix this, meteorologists usually use Ensemble Forecasting. Imagine asking 50 different weather experts for their opinion. If they all say "sunny," you can be pretty sure it will be sunny. If 25 say "sunny" and 25 say "rainy," you know there is uncertainty, and you should pack an umbrella just in case.

The problem with AI is that training 50 different AI models is too expensive and slow. It's like hiring 50 different chefs to cook the same soup just to see which one tastes best.

The Solution: The "Whispering" Trick

This paper proposes a clever, cheap trick to get 50 different opinions from just one AI chef.

Instead of training 50 chefs, they take one chef and give them slightly different ingredients before they start cooking.

• The Chef: The Graph Neural Network (the AI model).
• The Ingredients: The current state of the ocean (temperature, wind, etc.).
• The Trick: They add a tiny bit of "noise" (random static) to the ingredients before feeding them to the AI.

Think of it like this:

• Chef A gets the recipe with a tiny pinch of extra salt.
• Chef B gets the recipe with a tiny splash of extra water.
• Chef C gets the recipe with a slightly different temperature.

Because the ingredients are slightly different, the chefs will cook slightly different soups. When you taste all 50 soups and average them out, you get a very accurate prediction. More importantly, if the soups are all very different, you know the recipe is tricky (high uncertainty). If they all taste the same, you know the recipe is solid (low uncertainty).

The Experiment: What Kind of "Noise" Works Best?

The researchers tested different ways to add this "noise" to the ocean data to see which one created the best "ensemble" (group of predictions).

1. Gaussian Noise (The Static): This is like turning on a radio between stations. It's pure, random static. Every pixel of the ocean gets a random number added to it, completely independent of its neighbors.

   • Result: It worked okay, but it felt "artificial." The ocean doesn't work in pure random static; warm water usually stays next to warm water.

2. Perlin Noise (The Smooth Waves): This is a special type of noise that creates smooth, flowing patterns. It's like adding gentle ripples to the water. If one part of the ocean gets a little warmer, the neighbors get a little warmer too.

   • Result: This was the winner. Because the ocean moves in smooth waves and currents, this type of noise mimics reality better. It created a group of predictions that were diverse but still looked like a real ocean.

3. Fractal Perlin Noise (The Fractal Zoom): This is like looking at a coastline. You see the big bays, then the smaller coves, then the tiny rocks. It adds layers of detail.

   • Result: It didn't help much more than the smooth Perlin noise. Sometimes, adding too much tiny detail just confused the AI without adding useful information.

The Findings: What Did They Learn?

• Accuracy: The "noisy" group of predictions was just as accurate as the single, perfect model. They didn't lose accuracy by adding the noise.
• Confidence: The "smooth" noise (Perlin) was much better at telling the user when they should be worried. It gave a realistic range of possibilities. The "random static" noise (Gaussian) was too chaotic and didn't represent the ocean's natural behavior well.
• Cost: They achieved this without training 50 new models. They just took one model and ran it 50 times with slightly tweaked inputs. This is a massive win for speed and cost.

The Takeaway

This paper shows that you don't need a supercomputer to get a "team of experts" for ocean forecasting. You just need one smart AI and a way to gently nudge its inputs with the right kind of randomness.

By using Perlin noise (smooth, wave-like randomness) instead of Gaussian noise (chaotic static), they created a system that can tell us not just what the ocean temperature will be, but how sure we can be about it. This is a huge step toward making AI ocean forecasting reliable enough for real-world use, like helping ships avoid storms or helping fishermen find the best catch.

Technical Summary

Here is a detailed technical summary of the paper "Ensemble Graph Neural Networks for Probabilistic Sea Surface Temperature Forecasting via Input Perturbations."

1. Problem Statement

Accurate regional ocean forecasting is critical for maritime operations, ecosystem monitoring, and climate decision-making. However, traditional numerical ocean models are computationally expensive and often inaccessible for high-resolution regional applications. While Machine Learning (ML) models, particularly Graph Neural Networks (GNNs), offer a computationally efficient alternative, they are predominantly deterministic. This lack of inherent uncertainty quantification limits their utility for operational forecasting, where understanding the probability of different outcomes is essential.

Existing ensemble methods for ML (e.g., training multiple models, generative diffusion models, or hybrid physics-ML systems) often incur prohibitive computational costs or training complexity. There is a need for a lightweight, operationally feasible ensemble strategy that can generate probabilistic forecasts for regional GNNs without requiring the retraining of multiple models or complex generative architectures.

2. Methodology

Dataset and Domain

• Region: The study focuses on the North Atlantic region surrounding the Canary Islands, an area characterized by strong upwelling dynamics and high spatiotemporal variability.
• Data Sources:
  • Oceanographic: Copernicus Marine Service (CMEMS) Level 4 Sea Surface Temperature (SST) data (0.05° resolution, 1982–2023).
  • Atmospheric Forcing: ERA5 wind components ($u_{10}, v_{10}$) from the Copernicus Climate Data Store.
  • Static Data: NOAA ETOPO bathymetry.
• Preprocessing: Data was aggregated to daily means and interpolated to a consistent grid.

Model Architecture: Adapted SeaCast

The authors adapted SeaCast, a hierarchical Graph Neural Network (GNN), to this specific region. The architecture follows an Encoder-Processor-Decoder paradigm on a hierarchical mesh graph:

• Encoder: Maps latitude-longitude grid data to a hierarchical mesh graph.
• Processor: Updates latent representations via message passing across multiple mesh levels (81, 27, and 9 resolution levels) using interaction networks.
• Decoder: Projects the processed latent states back to the original grid.
• Training: The model was trained deterministically to minimize Mean Squared Error (MSE) over a 17-year period (2003–2019) using an autoregressive approach (predicting one step ahead).

Ensemble Strategy: Inference-Time Input Perturbation

Instead of retraining multiple models (which is computationally expensive), the authors implemented a homogeneous ensemble approach where diversity is introduced only during inference:

1. Single Trained Model: One GNN model is used for all forecasts.
2. Input Perturbation: The initial ocean state ($X_0$) is perturbed with noise before being fed into the model.
3. Aggregation: Multiple perturbed forecasts are generated, and the final probabilistic forecast is the mean of these ensemble members.

Noise Generation Strategies

The study systematically evaluated three types of noise to induce diversity:

1. Gaussian Noise: Spatially uncorrelated random noise with varying standard deviations ($\sigma$).
2. Perlin Noise: Spatially coherent noise with smooth transitions, generated at different spatial resolutions (e.g., $2\times3\times3$vs.$2\times12\times12$).
3. Fractal Perlin Noise: Multi-octave Perlin noise incorporating parameters for persistence, lacunarity, and scale to introduce multi-scale complexity.

Evaluation Metrics

Forecasts were evaluated over a 15-day horizon using:

• Deterministic Metrics: Root Mean Squared Error (RMSE) and Bias.
• Probabilistic Metrics:
  • Continuous Ranked Probability Score (CRPS): Measures the accuracy of the entire probability distribution.
  • Spread-Skill Ratio: Compares the ensemble dispersion (spread) against the error of the ensemble mean (skill). A ratio near 1 indicates perfect calibration.

3. Key Contributions

1. Efficient Ensemble Framework: Proposed a computationally lightweight ensemble method for regional GNNs that generates probabilistic forecasts via inference-time input perturbations, avoiding the cost of multi-model training.
2. Systematic Noise Analysis: Provided a detailed comparison between unstructured (Gaussian) and spatially coherent (Perlin/Fractal Perlin) noise. The study demonstrates that spatial coherence is more critical than noise complexity for achieving well-calibrated uncertainty.
3. Operational Feasibility: Showed that carefully designed input perturbations can yield calibrated probabilistic forecasts for high-resolution regional ocean models without additional training costs.

4. Results

Deterministic Skill (RMSE)

• Ensemble forecasts generally showed slightly higher RMSE than the single deterministic model at short lead times due to the explicit perturbation of observations.
• However, as the forecast horizon extended (up to 15 days), the RMSE of the ensemble mean converged with the deterministic model, indicating that the ensemble approach does not degrade long-term accuracy.

Probabilistic Performance (CRPS and Calibration)

• Spatial Coherence Matters: Ensembles generated with spatially coherent perturbations (Perlin noise) significantly outperformed those with purely random Gaussian noise in terms of CRPS and calibration, especially at longer lead times.
• Resolution Impact: Lower-resolution Perlin noise (larger spatial patterns) achieved better calibration and lower CRPS than high-resolution or fractal variants. This suggests that large-scale geophysical correlations are better captured by smoother, lower-frequency perturbations.
• Fractal Noise Limitations: Fractal Perlin noise, which adds fine-scale details via octaves, did not improve performance and sometimes degraded it, likely because the added high-frequency details were physically inconsistent or too noisy for the model to utilize effectively.
• Calibration: The best configurations (e.g., low-resolution Perlin noise) achieved a Spread-Skill ratio close to 1.0 by day 15, indicating that the ensemble dispersion accurately reflected the actual forecast error. In contrast, Gaussian noise configurations often resulted in under-dispersion (ratio < 1) or poor calibration.

5. Significance and Conclusion

This work demonstrates that ensemble learning for GNNs does not require expensive multi-model training. By leveraging the physical consistency of spatially coherent noise (Perlin noise), researchers can generate well-calibrated probabilistic forecasts for regional ocean systems using a single trained model.

Key Takeaways:

• Noise Structure > Noise Complexity: The spatial structure of the perturbation is more important than the mathematical complexity of the noise generation. Smooth, large-scale perturbations better represent geophysical uncertainty than random or highly detailed fractal noise.
• Operational Viability: This approach bridges the gap between fast, deterministic ML models and the need for uncertainty quantification in operational oceanography, making probabilistic forecasting accessible for resource-constrained regional applications.
• Future Directions: The authors suggest future work could involve training with autoregressive steps, perturbing forcing variables (wind), or exploring parameter perturbations to further enhance diversity and accuracy.