Large Scale Interferometric Processing of Sentinel-1 Data over the Atacama Desert - a Contribution to the TecVolSA Project

Kurzfassung

The project TecVolSA (Tectonics and Volcanoes in South America) aims at developing an intelligent Earth Observation (EO) data processing system for monitoring the earthquake cycle and volcanic events in South America. The Remote Sensing Technology Institute of DLR participates to this project together with GFZ (German Research Centre for Geosciences). The project is partially financed by Helmholtz. So far we have processed about 40 Sentinel-1 slices covering the Atacama Desert with mixed Permanent Scatterer and Distributed Scatterer (PS/DS) techniques. The area is very dry and the spatio-temporal coverage is excellent. Tropospheric correction have been applied using ECMWF ERA5 data, hence improving the performance in observing both topography related and large scale deformation signals. The current results reveal, as expected, plenty of interesting signals to be interpreted (see attached figure for an overview of the velocity field). Preliminary GPS cross-validation, thanks to data freely available from the Geodetic Nevada Laboratory, confirm that the InSAR relative error in the estimated velocities is in the order of 1 mm/yr at large scale (>100 km) and confirms the large scale signal related to the subduction of the Nazca plate (see attached figure). More GNSS validation will be possible with additional GPS stations. The challenge of the project is the separation of different contributions to the InSAR measurements: apart from the tectonic effects, there are contributions coming from volcanic unrest, atmospheric delays, moisture effects, snow, flank instability (likely downhill creep or solifluction related to permafrost, see attached figure), salt lake growth, mining, and likely more. We are dealing with this complexity with a diversity of tools: physical modeling and statistical analysis, deep neural networks, and expert knowledge. GFZ contributes process knowledge, historic seismic data, in-situ motion measurements and observations and 4D geophysical modelling codes for producing a diverse database for the training of neural networks in order to autonomously discover significant events in noisy data. We tackle the problem as a semi-supervised multi-class classification approach where the labeling of the known deformation phenomena is provided by GFZ. Signals for which the source of deformation is unknown are identified and clustered automatically using advanced unsupervised machine-learning techniques. Therefore, we leverage from the advantages of both supervised and unsupervised learning and improve the accuracy for detection and classification of different deformation sources. The networks and AI-based methods are developed at DLR. This new approach (InSAR + Artificial Intelligence) should be able to process the massive data stream of the Copernicus Sentinel-1 SAR mission. South America was selected because manifold geophysical signals can be expected there in short time scales and plenty of in-situ data are available. This project will complement the current model-based geophysical research by a data-driven AI-based approach. Training and applying this intelligent system shall improve our understanding of geophysical processes related to natural and anthropogenic hazards. At a later stage the system shall be scalable to global processing capacity. Future developments on the InSAR processing will include ionospheric corrections based on split-spectrum and mosaicking of the velocity and displacement series. Some issues with the L1 processor are hindering the deployment of the split-spectrum technique. Stacks from the ascending geometry are already being processed and will help the geophysical interpretation.