Artificial Intelligence for high performing inversion models
Generative Artificial Intelligence is a new branch of Machine Learning algorithms gaining popularity since less than five years. Generative Models differs from the classical Discriminative Models because they are able to mimic any distribution of data by progressively understand the main features and consequently generate additional samples.
This technology can be applied to healthcare, finance, engineering, distribution chain, automotive, music and also art. The most advanced application was presented late 2018 with the Chinese virtual anchorman experiment in the Xinhua TV channel. Aspecifically designed machine learning model (with human appearance) has been fed with real time news for days until it was able to announce and discuss news every day and 24-hours a day. The overall GANs design is very demanding in terms of computing resources and often they are hosted, executed and deployed in cloud and/or decentralized computing infrastructures. Experiments performed on portraits pictures reported training time of the order of days.
Novel applications based on machine learning require massive amounts of data for development and testing. Despite the large availability of Earth Observation data from satellites, data sets with consistent and complete imagery, ground truths and model simulations are not so easy to obtain because they are costly and complex. There are promising techniques in machine learning to generate data from a representative set called Generative adversarial networks (GANs) and they can be an alternative to expensive and costly simulations.
An activity with TDE and CSEM in Switzerland has assessed which GAN algorithms might be the most suitable for space applications and developed simulations for Earth Observation data. The activity developed and tested several GAN models. The most suitable were then selected for the two test cases: thermal-infrared data and atmospheric maps. An alternative algorithm for thermal-infrared retrieval of Land Surface Temperature has also been developed that can be used.
The tests demonstrated the capabilities to generate synthetic features for the two cases were valid and very quick to obtain. The very quick running time of trained GANs algorithms may simplify future processes for design and simulation and the algorithms for Earth Observation can be simplified in many numerical parts by including these machine learning components. Next, the activity plans to expand the algorithm tests to cover pre-operational phases., with an intention that the technology can be incorporated in user algorithms for Copernicus and Earth Watch in a cloud environment.
4000128225 completed in 2021.