Meteosat Third Generation imagery improves CNN-based SSI retrieval
Listen
Radio episode about this paper
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.
Gordei Pribětkin, Piia Post, Velle Toll
Institute of Computer Science, University of Tartu · Centre for Climate Research, Institute of Physics, University of Tartu
physics.ao-ph, cs.LG, physics.data-an
Submitted: 2026-07-30
Updated: 2026-11-09
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 77/100
The gist: Accurate Surface Solar Irradiance (SSI) estimation is increasingly important for photovoltaic energy monitoring and forecasting.
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
Summary
Accurate Surface Solar Irradiance (SSI) estimation is increasingly important for photovoltaic energy monitoring and forecasting. p. 0 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. p. 0 The hybrid SEVIRI–FCI model significantly outperformed the SEVIRI-only model under overcast and cloudy conditions, reducing RMSE by 8. p. 02W m−2 and 5.7W m−2, respectively. p. 0 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. p. 0 Compared with physics-based SARAH-3, the hybrid model yielded skill scores of 35 % under overcast conditions, 21 % under cloudy conditions, and 20 % overall. p. 0 These results show that higher-resolution MTG/FCI imagery improves CNN-based SSI retrieval when clouds dominate irradiance variability, but also indicate that higher spatial resolution alone is insufficient to address clear-sky limitations in machine-learning-based SSI retrieval. p. 0
Model Architecture and Data Integration
The proposed model integrates images from both SEVIRI, at 1 km and 3 km nadir resolution, and, importantly, the higher spatial resolution (0.5 km and 1 km nadir) FCI images. p. 2 The general architecture is a multi-resolution, multi-branch CNN where image modalities are separated by the imager and spatial resolution. p. 4 Each branch CNN module follows the ResNet architecture [25], with a similar model being proposed specifically for SSI retrieval in other works. p. 4 The model retrieves SSI estimations from the stabilized clear-sky index predicted by the model by solving Equation 4 for SSI, as described as follows. p. 4
Data Preprocessing and Input Features
The input to the model includes satellite image patches with physical sizes determined by their nadir resolutions, such as 51 km × 51 km for bands with 1 km px−1 and 3 km px−1 resolution. The images are reprojected to a metric grid using EPSG:3301 coordinate reference system, and then cropped and resampled to obtain square patches with odd pixel sizes where the site is located exactly in the centre. p. 3 Auxiliary numerical inputs include latitude, longitude normalized by the geographical bounding box of the region of interest, normalized clear-sky SSI (SSIcs), and trigonometric encodings such as Esin and Ecos. p. 3
Performance Analysis Across Sky Conditions
The performance metrics are calculated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Bias Error (MBE). The test dataset in each fold is split by sky condition using the Variability Index (VI) and daily CSI to define categories like Clear, Partly cloudy, Cloudy, and Overcast. p. 6 The results show that the HYBRID model generally outperforms the SEVIRI-only model in overcast, cloudy and all-sky conditions by approximately 15 %, 6 % and 6 %, respectively. p. 7 However, in clear conditions, our models were, on average, 38 % worse compared to the physics-based SARAH-3 baseline. p. 0
Discussion of Key Findings
The study concludes that higher-resolution MTG/FCI imagery can improve CNN-based SSI retrieval when clouds are the dominant source of irradiance variability. The lack of improvement under clear skies shows that higher spatial resolution alone is not sufficient to solve all limitations of machine-learning-based SSI retrieval. p. 0 Furthermore, the bias share of error is the highest in clear-sky conditions, suggesting that the main issue limiting clear-sky performance is likely the bias of the model. p. 8 The across-seed relative standard deviation of RMSE was approximately 30 % in clear-sky conditions, compared to values of 3 % - 10 % for other conditions. p. 9
Limitations and Future Work
Multiple limitations must be considered for the interpretation of our results, including the small geographical extent in Northern Europe which limits the generality of our results. Additionally, limited FCI data availability necessitated a temporal overlap between test and train splits due to shared general weather patterns between the splits. p. 3 Future work should focus on evaluating CNN-based SSI retrieval models with longer FCI time series and independent test years in a larger geographic domain. p. 5 The authors believe that a promising research avenue would be to extend the proposed architecture for SSI forecasting with a considerable lead time. p. 0
Declaration of Generative AI and AI-assisted Technologies in the Manuscript Preparation Process
During the preparation of this work, the lead author used ChatGPT to provide suggestions on text wording in selected paragraphs. Additionally, the Cursor IDE was used to assist with conceptualizing some aspects of the study methodology, data analysis, and creation of plots and tables. p. 1 The authors critically reviewed and edited the content as needed and take full responsibility for the content of the published article.
CRediT Authorship Contribution Statement
Gordei Pribõtkin: Conceptualization, Methodology, Software, Data curation, Formal analysis, Investigation, Resources, Validation, Visualization. Piia Post: Conceptualization, Methodology, Supervision. p. 0 p. 0 Velle Toll: Conceptualization, Methodology, Supervision. p. 0 p. 0
Declaration of Competing Interest
We declare that G.P. was supported by STACC OÜ, which sourced institutional research funding and provided computational resources. p. 0 The funder had no role in the study design, data analysis, interpretation of results, or preparation of the manuscript. p. 2
Funding
STACC OÜ supported G.P. with regular base funding for R&D from the Ministry of Education and Research (Estonia). p. 0
Acknowledgements
The authors gratefully acknowledge the Estonian Environment Agency for providing quality-controlled ground-based irradiance measurements used in this study. The authors also acknowledge EUMETSAT for access to Meteosat Second Generation SEVIRI and Meteosat Third Generation FCI data, and the CM SAF/EUMETSAT SARAH-3 surface solar radiation dataset used for baseline comparison. p. 1
References
[1] Gaëtan Masson, Melodie de l’Epine, Izumi Kaizuka, and Junko Okazaki. Trends in photovoltaic applications 2025. p. 5 Technical report, IEA Photovoltaic Power Systems Programme, 2025. p. 0 [4] C Rigollier, M Lefèvre, and L Wald. The method heliosat-2 for deriving shortwave solar radiation from satellite images. p. 1 Solar Energy, 77(2):159–169, 2004. p. 0 [5] U. Pfeifroth, J. Drücke, S. Kothe, J. Trentmann, M. Schröder, and R Hollmann. Sarah-3 – satellite-based climate data records of surface solar radiation. p. 1 Earth System Science Data, 16(11):5243–5265, 2024. p. 2 [8] Manajit Sengupta, Yu Xie, Anthony Lopez, Aron Habte, Galen Maclaurin, and James Shelby. The national solar radiation data base (nsrdb). p. 1 Renewable and Sustainable Energy Reviews, 89:51–60, June 2018. p. 0 [13] K. R. Schuurman and Angela Meyer. p. 1 Surface solar radiation: Ai satellite retrieval can outperform heliosat and generalizes to other climate zones. p. 1 International Journal of Remote Sensing, 46(8):3331–3362, 2025. p. 1 [17] Shanlin Chen, Chengxi Li, Yuying Xie, and Mengying Li. p. 1 Global and direct solar irradiance estimation using deep learning and selected spectral satellite images. p. 1 Applied Energy, 352:121979, 2023. p. 1 [18] EUMETSAT. p. 1 Rapid scan high rate seviri level 1. p. 25 image data - msg (msg15-rss), 2009. p. 0 Collection ID: EO:EUM:DAT:MSG:MSG15-RSS. p. 0 [19] EUMETSAT. p. 1 Fci level 1c high resolution image data - mtg - 0 degree, 2024. p. 2 Product ID: EO:EUM:DAT:0665. p. 0 [20] EUMETSAT. p. 1 Surface radiation data set - heliosat (sarah) - edition 3. p. 1 Collection ID: EO:EUM:DAT:0863. [21] William F. Holmgren, Clifford W Hansen, and Mark A Mikofski. pvlib python: a python package for modeling solar energy systems. p. 2 Journal of Open Source Software, 3(29):884, 2018. p. 3 [23] EPSG Geodetic Parameter Dataset. p. 3 EPSG:3301 – Estonian Coordinate System of 1997. p. 3 [24] Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and Andrew Ng. Multimodal deep learning. p. 0 pages 689–696, 01 2011. [25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. p. 2 Deep residual learning for image recognition, 2015. p. 1 [26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. p. 4 In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 448–456, Lille, France, 07–09 Jul 2015. p. 0 PMLR. [27] Diganta Misra. Mish: A self regularized non-monotonic neural activation function. p. 4 CoRR, abs/1908.08681, 2019. p. 1 [28] Joshua S Stein, Clifford W Hansen, and Matthew J Reno. The variability index: A new and novel metric for quantifying irradiance and pv output variability. p. 6 2012.
Performance Metrics by Fold
Table 8 presents the per-fold RMSE for each held-out test site, by sky condition, for the SEVIRI-only and HYBRID models and SARAH-3. The HYBRID model generally outperforms the SEVIRI-only model in overcast, cloudy and all-sky conditions with 8. p. 72W m−2, 5.7W m−2, and 4.9W m−2 RMSE improvement, respectively. p. 7 This shows that our model successfully integrates complementary data from the FCI imager. p. 7
Discussion of Results
We found that the model using both FCI and SEVIRI imagers consistently outperforms a model using only SEVIRI images in overcast, cloudy and all-sky conditions by approximately 15 %, 6 % and 6 %, respectively. Our models generally outperform SARAH-3 in more cloudy conditions, with the HYBRID model yielding a statistically significant improvement over the baseline overall with a 20 % skill score and skill scores of 35 % and 21 % in overcast and cloudy conditions. p. 7 However, in other sky conditions, we did not observe a statistically significant difference in performance. p. 7 In clear conditions, our models were, on average, 38 % worse compared to the physics-based SARAH-3 baseline. p. 0 The main factor in poorer clear-sky performance is likely model bias. p. 8
The small geographical extent in Northern Europe limits the generality of our results and does not allow us to study the generalization ability of our models. The authors suggest that for example, adding recurrent or attention elements over the embeddings of images produced by the branches of the SSI retrieval model could be a promising research avenue. p. 1 The paper concludes that higher-resolution MTG/FCI imagery can improve CNN-based SSI retrieval when clouds are the dominant source of irradiance variability. p. 0
Gordei Pribõtkin: Conceptualization, Methodology, Software, Data curation, Formal analysis, Investigation. Piia Post: Conceptualization, Methodology, Supervision. Velle Toll: Conceptualization, Methodology, Supervision.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve SSI retrieval accuracy by implementing a multi-imager and multi-resolution convolutional neural network architecture that integrates data from both MSG/SEVIRI and MTG/FCI satellites alongside solar-geometry and clear-sky irradiance features. This hybrid 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.
-
Enhance model generalization by employing cross-validation across multiple training seeds to robustly judge the uncertainty of obtained results, as noted in the paper:
We use cross-validation to robustly show that higher-resolution FCI imager data can improve SSI retrieval performance.
-
Develop a conditional prediction capability by training models with separate branches for each imager and spatial resolution, allowing the model to handle missing modalities by appending binary missingness flags. This architecture enables the model to
allow for missing modalities
which is critical when combining data from different satellites. -
Optimize clear-sky performance by incorporating auxiliary numerical inputs such as normalized clear-sky SSI (SSInorm) and phase information (sine and cosine of day/year phases). The paper notes that
the predictions of our CNN-based models are more biased, and the training outcomes are much more unstable, when compared to cloudy sky conditions
under clear skies. -
Establish a robust baseline comparison by evaluating new ML models against both physical models (like SARAH-3) and gridded meteorological products. The results show that
Our ML-based models outperform the SARAH-3 baseline under cloudy conditions,
while acknowledging thatgridded meteorological products can be a much stronger baseline in clear conditions compared to cloudy conditions for SSI prediction.
Sources
Related papers
- Anthropogenic Heat in Urban Climate Systems: Forcing, Sensitivity, and Feedback
- NORi: An ML-Augmented Ocean Boundary Layer Parameterization
- A Checklist to assess the energy and carbon impacts of ML/AI applications in Earth System Modeling
- Skillful Thunderstorm Forecasts Learned Directly from Observations
- Wind vector retrieval from miniaturized wave-enabled sea-surface drifters
- Stochastic coupling of climate variables and ice volume over the Late Pleistocene glacial cycles