Search

Recent works on using Machine Learning methods for deriving estimates of the NO2 concentration at the Earth's surface from satellite observations exploit measurements taken from low Earth orbits, e.g. from the TROPOMI instrument on the Copernicus Sentinel-5P satellite. However, given geographic location, the time resolution is quite low, with a single measurement per day, which leads to rather small data sets. In order to increase the performance of Machine Learning methods, large data sets would be desirable.Launched in 2019, the Korean Geostationary Environmental Monitoring Spectrometer (GEMS) mission has been the first geostationary satellite mission for observing trace gas concentrations in the Earth's atmosphere over Asia. Geostationary orbits allows for hourly measurements, which leads to a much higher temporal resolution compared to measurements taken from low Earth orbits. Within the next years, two further geostationary missions will follow: NASA‘s TEMPO and ESA‘s Sentinel-4 mission, providing additional data with high temporal resolution over North America and Europe.One of the GEMS level-2 data products is the NO2 tropospheric vertical column density (VCD). In our research project we discuss and develop Deep Learning methods that use not only these NO2 VCDs, but also additional data such as meteorological and geographical data, to derive estimates of the NO2 surface concentration in high spatial as well as high temporal resolution, enabled by the geostationary GEMS measurements mentioned above. The validation of the network‘s prediction is realized by the consideration of in-situ NO2 observations from the air quality network of South Korea.

The recent development of scintillation crystals combined with γ-rays sources opens the way to an imaging concept based on Compton scattering, namely Compton scattering tomography. The associated inverse problem rises many challenges: non-linearity, multiple order-scattering and high level of noise. Already studied in the literature, these challenges lead unavoidably to uncertainty of the forward model. This work proposes to study exact and approximated forward models and develops two data-driven reconstruction algorithms able to tackle the inexactness of the forward model. The first one is based on the projective method called regularized sequential subspace optimization (RESESOP). We consider here a finite dimensional restriction of the semi-discrete forward model and show its well-posedness and regularization properties. The second one considers the unsupervised learning method, deep image prior, inspired by the construction of the model uncertainty in RESESOP. The methods are validated on Monte-Carlo data.