Francesco Pignatale, Telespazio Germany GmbH
Long Phil Chau, Telespazio Germany GmbH
Claudio Mammone, Telespazio Germany GmbH
Claudio Mammone, Head of Payload Operation Systems, Telespazio Germany GmbH
Satellite imagery and associated metadata are fundamental for providing insights for Earth-Observation studies, building up reliable models to improve our understanding of the evolution of the climate and for fostering applications for current and future space-borne missions.
In the current context, with many commercial missions being developed by new and more established companies, the time to reach the commercial operations is crucial to beat the competition. One of the key activities during the mission development cycle, assuming the use of COTS hardware, is the development of the data processing tools and associated algorithms. These are often very specific to the target applications and require reliable data to be developed. Synthetic or simulated data help in supporting these activities, speeding up the process of parallel development, testing, verification and validation of data-processors, retrieval algorithms and parameters configuration.
The standard approach to generate simulated data (based on physical models) requires effort and time. Moreover, the amount of produced data available is normally limited due to time and budget constraints.
Generative AI-based frameworks have been already widely used in the wider context of image processing, and since recent times have been applied to the remote sensing field. However, the scientific validity of the generated images is often neglected. The goal of this project is to review the state-of-the-art in the field and then build a Generative AI-based framework for building up synthetic satellite’s imagery tuned according to parameters (e.g. pixel resolution, aerosol optical thickness, solar zenith angles, sensing time, …) to provide a more scientifically valid product.
For this first prototype, the publicly available Copernicus Sentinel-2 data are used as reference dataset, as its characteristics match those of many commercial missions (orbit configuration, sensor spectral bands). Furthermore, the Copernicus Sentinel-2 data cover a large range of different observational parameters and conditions, are consistent for time-series analysis and scientifically verified.
In this work we present the proof of concept, its processing-chain, and preliminary results showing Generative-AI based Synthetic Copernicus Sentinel-2 data produced assuming different sets of initial conditions. Our approach can be then extended and the whole method scaled to include different types of sensors and datatypes.
This work is conducted with the contribution of the Technische Universität Darmstadt.