Project Summary

During this project at ESA Φ-lab, I developed a pipeline for producing high-resolution solar surface irradiance maps based on Sentinel-2 imagery. The work combined cloud height retrieval through the Asterisk algorithm, parallax correction, and cloud index mapping to improve the spatial accuracy of irradiance estimates. To further enhance resolution, I applied deep learning super-resolution models trained on CAMS data, enabling low-resolution irradiance inputs to be upscaled with improved spatial detail. The outcomes show clear improvements over baseline interpolation methods and highlight the potential of combining physics-based approaches with machine learning to refine EO-based solar resource assessments. This collaboration contributes to the wider goal of using high-resolution EO data to support renewable energy applications and climate monitoring.

Development Tools

• Data: Sentinel-2 MSI imagery, CAMS irradiance data, Cloudnet profiles, BSRN station data.
• Algorithms: Asterisk (cloud height via parallax), SenSeIv2 (cloud/shadow masking), Heliosat methods (cloud index), U-Net and SegFormer models (super-resolution).
• Software & libraries: PyTorch, segmentation-models-pytorch, HuggingFace, rasterio, xarray, geopandas, Satpy.
• Infrastructure: DelftBlue HPC (GPU/CPU parallelisation), local development on MacOS.

Development Outputs & Description

The project carried out during my stay at ESA Φ-lab focused on developing a novel approach to improve the spatial resolution of surface solar irradiance (SSI) estimates. Accurate high-resolution irradiance data are essential for applications such as renewable energy forecasting, climate monitoring, and the study of atmospheric processes. Geostationary (GEO) satellites such as Meteosat provide continuous temporal coverage of clouds and atmospheric conditions, but at kilometre-scale resolution. While this temporal density is advantageous, the coarse resolution prevents detailed representation of small cloud structures, which are critical drivers of solar irradiance variability. On the other hand, Low Earth Orbit (LEO) satellites such as Sentinel-2 provide data at 10–60 m resolution, offering much greater spatial detail, though with limited revisit frequency. The project set out to bridge this gap by combining Sentinel-2 observations with GEO-based irradiance products and applying machine learning super-resolution models.

The first stage of the work focused on retrieving irradiance directly from Sentinel-2. To achieve this, several processing steps were developed. Cloud masks were generated using the SenSeIv2 algorithm, a state-of-the-art deep learning method capable of distinguishing between cloudy and clear pixels as well as cloud shadows. Building on this, cloud heights were estimated using the Asterisk parallax method, which exploits the slight viewing-angle differences between Sentinel-2’s spectral bands to reconstruct cloud altitude. These estimates were validated against ground-based radar and lidar measurements from Cloudnet stations located in Cabauw (Netherlands), Palaiseau (France), Hyytiälä (Finland), and Granada (Spain). Once cloud masks and heights were obtained, they were used to construct a high-resolution cloud index following the Heliosat methodology. This was complemented with a shadow index derived from geometric projection of clouds onto the surface. Combining these elements with a clear-sky irradiance model allowed the generation of surface solar irradiance maps at 10–20 m resolution, which were directly comparable to CAMS data derived from Meteosat.

The second stage of the project involved applying machine learning to downscale irradiance maps from GEO resolution to Sentinel-2 resolution. To do so, a dataset was constructed pairing CAMS irradiance maps with co-located Sentinel-2–based irradiance estimates derived from the first stage of the pipeline. These pairs were used to train convolutional neural networks, in particular U-Net and SegFormer architectures, to learn how to reconstruct high-resolution irradiance patterns from coarse inputs. The results demonstrated that deep learning significantly outperformed classical interpolation methods such as bilinear resampling. Notably, the models were able to reproduce small-scale irradiance variability caused by broken cloud fields, which is typically lost in GEO-based products. Evaluation was carried out using standard error metrics (MAE, RMSE, correlation) as well as power spectral density analysis to verify improvements across different spatial frequencies.

The outcomes of the collaboration include the development of a complete end-to-end pipeline, from Sentinel-2 preprocessing to machine learning super-resolution of GEO data. The project delivered new high-resolution irradiance maps, validated cloud height retrievals, and a trained set of models capable of enhancing GEO irradiance products. The work has also produced a curated dataset of Sentinel-2 scenes, CAMS irradiance maps, and ground-based observations from Cloudnet and BSRN stations, which together provide a valuable benchmark for future studies.

In terms of collaboration, the project aligned well with ESA Φ-lab’s mission to explore the integration of artificial intelligence and Earth Observation for climate applications. Working within the Φ-lab environment provided the opportunity to test novel approaches such as parallax-based cloud height estimation at scale, and to combine them with advanced deep learning methods for super-resolution. The project outcomes will feed into the broader discussion on how to bridge physics-based EO methods and machine learning, ultimately contributing to more accurate and actionable information for energy and climate stakeholders.

Looking ahead, the methodology developed here opens several future research directions. One avenue is the extension of the framework to other LEO missions beyond Sentinel-2, such as MODIS, EarthCARE, or TRUTHS, which would expand temporal coverage and allow cross-sensor validation. Another direction is the integration of physics-informed machine learning, where radiative transfer constraints are embedded within neural networks to improve generalisation. Finally, the combination of this framework with real-time GEO data streams would enable the production of continuous, high-resolution irradiance maps across continental scales. These future developments will further strengthen the role of Earth Observation in supporting the energy transition and in providing critical insights into atmospheric processes.