Better understanding of Generative Adversarial Networks (GAN) for space applications

Generative Artificial Intelligence is a new branch of machine learning algorithms which have been gaining popularity for the last five years. Inversion models are mathematical and scientific algorithms used to derive parameters of interest from indirect measurements. Both these techniques are widely used in everyday life, from healthcare and finance, to the automotive industry or in music and art.
A new activity with TDE and CSEM, Switzerland, has taken both of these principles and tested two algorithms over two use cases to have a better understanding of Generative Adversarial Networks (GAN) for space applications.The activity focused on applying GAN algorithms and neural networks to spatial images. 

The first use case looked at thermal-infrared inversion to see if it was possible to retrieve land surface temperature and emissivity data from an initial set of features such as airborne acquisitions and ground measurements.
First, a generative model was assessed to simplify the production of data, augment the variability of features and assess the quality of the augmented data sets. Secondly, a new approach based on neural networks was implemented which enabled the extraction of Land Surface Temperature and Emissivity information out of these thermal infrared images. The activity found that with this technique the measurements of these features were more precise and efficient than state-of-the-art methods.
Use case 2 used the computation of vertical Total Electron Content. Total Electron 

Content (TEC) is an important descriptive quantity for the ionosphere of the Earth. vTEC maps suffer the presence of outliers, borders effects and they become inaccurate far from the measuring stations. In addition they barely model off-nominal events and anomalies that often happen in between measuring stations/sampling, due to geomagnetic storms. To assess the performance of the network with spatial gaps, the filled vTEC maps generated from GANs are compared against vTEC maps filled using classical interpolation and computer vision methods. This gives an overall idea how good GANs are able to learn general distribution of values in vTEC maps.
Overall the activity was able to address the problem of expensive, time consuming data collection. These methods could mean it is possible to collect only a fragment of the data necessary for training a neural network which implements the scientific problem. CSEM plans to reuse this know-how in analysis of hyperspectral medical images.