Meteosat Third Generation imagery improves CNN-based SSI retrieval

summary

The gist

Accurate Surface Solar Irradiance (SSI) estimation is increasingly important for photovoltaic energy monitoring and forecasting.

In short

A hybrid convolutional neural network model combining data from SEVIRI and higher-resolution FCI satellite imagery was developed to estimate Surface Solar Irradiance (SSI) over Northern Europe. The hybrid model significantly improved accuracy under overcast and cloudy conditions, reducing error by 8.2W m−2 and 5.7W m−2, showing that higher-resolution imagery helps when clouds dominate variability.

Key concepts

Multi-imager Architecture
The model uses data from two different satellites: SEVIRI (lower resolution) and MTG/FCI (higher resolution). This combination allows the system to leverage complementary information from both sources, which is crucial for improving accuracy when one satellite's data is insufficient, especially under cloudy skies.
Multi-resolution CNN
The architecture uses separate Convolutional Neural Network branches for each imager and spatial resolution. This design allows the model to process image patches with different resolutions (e.g., 1 km vs. 0.5 km) independently, enabling it to capture spatial details relevant to the specific data source.
Clear-sky Limitation
The study found that higher spatial resolution alone is not enough to improve accuracy when skies are clear. In clear conditions, the models performed significantly worse than a physics-based baseline. This suggests that the primary challenge in clear skies is likely model bias rather than just image detail.

Terminology used across episodes

This episode discusses

The paper

Meteosat Third Generation imagery improves CNN-based SSI retrieval · Read on arXiv

Gordei Pribětkin, Piia Post, Velle Toll

Institute of Computer Science, University of Tartu · Centre for Climate Research, Institute of Physics, University of Tartu

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Meteosat Third Generation imagery improves CNN-based SSI retrieval".

Mira: Here's the gist of the paper. Accurate Surface Solar Irradiance (SSI) estimation is increasingly important for photovoltaic energy monitoring and forecasting.

Kai: In this work, a multi-imager and multi-resolution convolutional neural network architecture was introduced for 10-minute SSI retrieval over Northern Europe using MSG/SEVIRI and MTG/FCI satellite imagery together with solar-geometry and clear-sky irradiance features.

Mira: The hybrid SEVIRI–FCI model significantly outperformed the SEVIRI-only model under overcast and cloudy conditions, reducing RMSE by 8.2W m−2 and 5.7W m−2, respectively.

Kai: However, under partly cloudy or clear skies, no statistically significant difference in RMSE was observed between the SEVIRI-FCI hybrid and the SEVIRI-only models.

Title and authors: Mira: So, looking at the paper "Meteosat Third Generation imagery improves CNN-based SSI retrieval," the main focus is how combining MSG/SEVIRI and MTG/FCI imagery helps estimate surface solar irradiance using a multi-imager, multi-resolution convolutional neural network architecture.

Kai: Right, so they are showing that the core idea is mixing those different satellite views and using features like solar geometry and clear-sky irradiance to make the deep learning model more reliable at figuring out what's happening on the sun's surface.

Lev: It’s interesting because they built this whole architecture to handle those different input types and resolutions separately in its branches, which lets it use the higher resolution FCI imagery even when the SEVIRI data isn't perfectly available for a specific area.

Kai: That modular design is what allows it to be flexible enough to combine these disparate data sources effectively. They tested this hybrid system over Northern Europe using those combined images and features.

Mira: They found that under overcast and cloudy conditions, this hybrid setup significantly cuts down the error when compared to using only SEVIRI data, reducing the root mean square error by about eight point two Watts per square meter and five point seven Watts per square meter.

Lev: That reduction in error is something that matters for practical application because it shows that incorporating complementary data sources is necessary to get any meaningful result when you're aiming for practical hardware application.

Kai: But here's the catch—the results show they didn't see a statistically significant improvement in accuracy when the sky was clear or partly cloudy compared to just using SEVIRI on its own.

Mira: That suggests that while mixing the data helps under bad weather, it doesn't solve all problems when things are clear.

Kai: It means when you look at the performance scores against a physics-based model like SARAH-three, the ML approach does look better under cloudy conditions with a skill score of twenty one percent, but it falls off significantly when the sun is clear.

The paper's summary: Kai: Now let's get into the actual summary of "Meteosat Third Generation imagery improves CNN-based SSI retrieval." They are looking at what they found about the overall performance of this multi-imager, multi-resolution convolutional neural network architecture.

Mira: The main point is that combining different satellite views and features like solar geometry gives the machine learning model a more reliable picture of what's happening on the sun's surface than if it just used one type of data alone.

Lev: They are specifically looking at how they use cross-validation to robustly show that higher resolution FCI imager data can improve SSI retrieval performance.

Kai: They tested this combination over Northern Europe using ten-minute SSI retrieval, comparing different image resolutions and models against ground-based pyranometer measurements from eight Estonian meteorological stations.

Mira: They found that the hybrid model generally outperforms the SEVIRI-only model in overcast, cloudy and all-sky conditions by about fifteen percent, six percent, and six percent respectively.

Lev: That comparison shows that while their machine learning approach is performing better than SARAH-three in cloudy conditions by a bit, it still lags behind the physics model when things are clear.

Kai: However, in other sky conditions they did not observe a statistically significant difference in performance compared to the baseline, and the results show that they were on average thirty-eight percent worse compared to the physics-based SARAH-three baseline in clear conditions.

Mira: That comparison against SARAH-three is important because it shows that while their machine learning approach is performing better than SARAH-three in cloudy conditions by a bit, it still lags behind the physics model when things are clear.

Lev: So the core finding is that higher spatial resolution MTG imagery improves retrieval specifically when clouds are the main source of irradiance variability but it doesn't fix all issues across every single sky condition.

The paper's improvements: Kai: Moving on to what the paper suggests they can do next, they are looking at adding recurrent or attention elements over the embeddings produced by those branches of their model.

Mira: That’s a big idea, Kai, trying to give the model a better memory for how things change from one moment to the next across different satellite inputs.

Lev: That’s what they are hoping for—to move beyond just a static snapshot of the image and get something more predictive. They’re essentially trying to build a forecaster instead of just an estimator.

Kai: It’s about improving the model's ability to understand the temporal evolution of those cloudy or clear sky conditions, which is where they think it falls short right now.

Mira: If you look at this from a deployment standpoint, adding those layers means more computations per time step, but if it leads to better long-term forecasting skill, that’s a trade-off we need to consider.

Lev: Forecasting with lead time means you have to handle much longer data sequences, which puts stress on the entire pipeline, so they’ll need robust error-correction techniques built in.

Kai: The authors are also pushing for extending the proposed architecture to forecast surface solar irradiance with a considerable lead time ahead of the actual measurement.

Mira: That shifts the whole goal from just getting an estimate for right now to predicting what's coming, which is much more useful for energy grid management and forecasting.

Kai: So the paper points toward building a more advanced forecasting tool, not just a better retrieval tool for current conditions.

Mira: And that moves us closer to using these satellite products for much more proactive energy planning instead of reactive monitoring.

Conclusion: Kai: So we've been looking at how they improved surface solar irradiance retrieval using that hybrid model from the paper "Meteosat Third Generation imagery improves CNN-based SSI retrieval."

Mira: The big takeaway is that mixing those different satellite views, especially when you add features like the clear-sky information and solar geometry, gives the machine learning model a more reliable picture of what's happening on the sun’s surface.

Lev: From an engineering standpoint, it shows that incorporating those complementary data sources is necessary to get any meaningful result when you are aiming for practical hardware application.

Kai: They showed that this hybrid setup really cuts down the error when things are overcast or cloudy, reducing the root mean square error by about eight point two Watts per square meter in those specific cases.

Mira: That comparison against a cloudy condition is important because it means while the higher resolution data gives a big boost when clouds are dominating the variability, they didn't see any statistically significant difference in error reduction when it was clear or partly cloudy skies.

Kai: They also compared this hybrid model against a physics-based model called SARAH-three and they saw skill scores of thirty-five percent under overcast conditions and twenty one percent under cloudy conditions.

Lev: The paper also points out that in clear conditions, their CNN-based models were on average thirty-eight percent worse compared to the SARAH-three baseline and they noted that the main factor limiting that performance is likely model bias.

Kai: That bias part is interesting because it suggests that when there’s no cloud variability to confuse things, the model itself might be introducing systematic errors rather than just missing data.

Mira: It really means higher spatial resolution alone isn't enough to solve all the limitations in these machine learning methods for surface solar irradiance estimation, so they suggest incorporating those auxiliary numerical inputs and perhaps adding recurrent layers to handle time aspect more robustly.

Lev: If they actually implement those advanced architectures, it could lead to much more reliable forecasting capability for real-world energy systems which is what I care about most on the hardware side.

More episodes

← Home