SIS - Satellite Imagery Solar Energy Analysis

Project description

This project explores how satellite imagery and AI can be used to estimate solar radiation in the Netherlands. The main goal is to build a prototype that combines Sentinel-2 surface features, CAMS atmospheric data, SARAH-3 radiation measurements and DEM terrain information. The system processes these datasets, trains AI models to estimate Global Horizontal Irradiance (GHI) and visualizes results in an interactive web dashboard. The dashboard shows average radiation values and a ROI-based comparison map, with explainable layers such as clouds, greenness and terrain.

Rather than predicting exact future outputs, the prototype focuses on transparency, learning and exploration. It demonstrates how radiation varies between neighbourhoods and why. The work contributes to the Sustainable Data & AI Applications research group and supports research into the relationship between solar production and potential grid congestion.

Context

Solar installations in the Netherlands are expanding rapidly, while parts of the electricity grid are becoming congested. Understanding how sunlight varies across regions is essential for planning where solar capacity makes sense and where the grid may experience stress.

However, radiation is influenced by multiple factors such as clouds, land type, urban structures and terrain. These effects are spread across space and time and they are not easy to interpret from raw datasets. Many AI approaches can be powerful but they are often too complex or “black boxes” and difficult for non-experts to understand.

This internship project is carried out at the Fontys Sustainable Data & AI Applications research group. It focuses on building a prototype that combines open satellite data with climate datasets and applies AI in a transparent way. The work follows the Fontys AI Project Methodology and FAIR principles, ensuring that processing steps, assumptions and limitations are documented and reproducible.

By delivering an interactive demo, the project allows researchers, students and stakeholders to explore how radiation differs across Eindhoven and what factors contribute to those differences. It is positioned as an exploratory, educational support tool, a foundation for future research rather than a final decision-making system.

Results

The project delivered a functional end-to-end prototype for solar radiation estimation. A complete preprocessing pipeline was implemented for Sentinel-2, CAMS, SARAH-3 and DEM datasets. These data sources were resampled, aligned and prepared for AI modeling.

Several model iterations were developed. Iteration 0 and 1 tested classical machine-learning models and showed that meaningful spatial and temporal radiation patterns can be captured when evaluated against NASA POWER and SARAH-3. Iteration 2 introduced a deep-learning approach based on the clearness-index concept, using two U-Net-style models with a strict temporal validation setup.

The prototype dashboard, publicly deployed, allows users to select a region in Eindhoven and see:

• Daily radiation estimates
• The average GHI value for the area
• An ROI-based map comparing neighbourhoods to the city average
• Explainable layers (clouds, greenness and terrain)

View the prototype dashboard

A simple illustrative grid-risk indicator was also implemented, linking estimated solar power to an assumed grid capacity. This makes the potential relationship between sunlight and congestion more understandable, while clearly stating that it is not using real grid capacity data.

Validation confirmed that the concept is feasible but accuracy is limited by dataset size and coarse grid resolution. The tool therefore emphasizes insight rather than prediction.

Additional outcomes include a structured research paper, accepted conference abstract, improved literature review and reusable documentation aligned with FAIR principles. Together, these results create a foundation that future students and partners can extend: connecting to real grid data, expanding to new regions and refining evaluation.