Alexander Barth
Speaking Sessions
Generative Deep Learning for Satellite Data Reconstruction
One of the major strengths of satellite data is its ability to provide near-global coverage. However, satellites operating in the visible or infrared range are affected by clouds (among other factors), which can significantly reduce the spatial and temporal coverage.
For many applications, complete coverage is either desired or even necessary. In addition, it is increasingly important to accurately reflect the underlying uncertainty of the reconstructed observations. In this study, we show the use of neural networks to fulfil these goals. We present DINCAE (Data-INterpolating Convolutional Auto-Encoder), a neural network that can be trained on incomplete images and is able to provide an error estimate of the reconstructed field. The reconstructed data and their associated error estimates are validated with in situ observations and withheld satellite data. DINCAE has been applied to various parameters, including sea surface temperature, chlorophyll-a concentration, total suspended matter and along-track sea-level altimetry.
Generative deep learning techniques (denoising diffusion models and conditional flow matching) can naturally provide an ensemble of reconstructions, where each member is spatially coherent with the scales of variability and with the available data. Rather than providing a single reconstruction, an ensemble of possible reconstructions can be generated, and the ensemble spread reflects the underlying uncertainty. We show how this method can be trained from a collection of satellite data without requiring prior interpolation of missing data. The reconstruction method is tested with chlorophyll-a concentration from the Ocean and Land Colour Instrument (OLCI) sensor aboard the Sentinel-3A and Sentinel-3B satellites.
The spatial scales of the reconstructed data are assessed via a variogram, and the accuracy and statistical validity of the reconstructed ensemble are quantified using the continuous ranked probability score and its decomposition into reliability, resolution, and uncertainty.
Biography
Alexander Barth is an associate professor at the GeoHydrodynamics and Environment Research (GHER) group, of the University of Liège (Belgium). He obtained his PhD in 2004 at the University of Liège on data assimilation in nested ocean models. He worked on variational interpolation, and he is the principal investigator of several European projects aiming to apply and improve the generation of climatologies based from in situ observations. Recently, he also investigated the use of deep learning with ocean remote sensing data and model results. His research focuses on probabilistic approaches and uncertainty quantification.