Enhancing Earth Observation applications with Synthetic Data

GMV successfully concluded Synthetic Data for Earth Observation (SD4EO), a research initiative in collaboration with the ARTEC group of the University of Valencia, funded by the European Space Agency’s FutureEO programme. The project aimed to demonstrate the benefit of integrating physically- and AI-based synthetic data into Earth Observation (EO) applications. SD4EO focused on two types of simulation:

• Physically Based Rendering. The ARTEC group used Unity graphics engine to generate realistic synthetic images by accurately simulating light behaviour and sensor characteristics.

   

• AI-driven simulation. GMV implemented advanced conditional diffusion models, starting with random statistical distribution and iteratively shaped by constraints, allowing features to evolve and resemble the target image or signal.

  The synthetic data was integrated with real EO data into AI-driven analytic pipelines to explore their potential in enhancing performance in target categorisation applications:

1. Categorization of crop fields. Integrating synthetic data generally maintained or slightly improved classification performance. Combining synthetic data of rare crops only (oats and alfalfa) with real EO data yielded the highest accuracy, highlighting the effectiveness of targeted synthetic data in addressing classification challenges for multi-class tasks.

2. Categorization of human settlements. The integration improved overall performance in detecting built-up areas. Promising results using only synthetic data warrant further investigation, particularly when real EO data is hard to obtain.

3. Monitoring of photovoltaic panels. Model performance generally improved, with the best results achieved when both physically- and AI-based datasets were used. Findings suggest that an appropriate amount of synthetic data can enhance model performance, with the ideal quantity varying based on data distribution and volume.

   These promising results lay the foundation for further investigation, including refinement of synthetic data and additional experiments to better understand their advantages in EO applications. The simulated datasets and tools have been openly released to encourage further research and development:

• Generated synthetic data is available on Zenodo.

   

• Code for generating the physically-based data and the AI-based data is available on GitHub.