Doctoral student, Polytech Annecy-Chambéry
Email:
lorenzo.lopez-uroz@univ-smb.fr
Website:
https://github.com/lolopezuroz
Office:
A117
Address:
LISTIC, 5 chemin de bellevue, CS80439, 74944 Annecy Cedex
Research team/group:
Thesis Title:
Neural network-based model inversion and hazard prediction with SAR displacement time series
"Inversion of physical models and prediction of natural hazards using neural networks based on SAR displacement time series"
Abstract:
The acquisition of Sentinel-1 A/B Synthetic Aperture Radar (SAR) images, covering Europe every 6 days (every 12 days elsewhere) and made available for free by the European Space Agency (ESA), brings the exploitation of SAR satellite data into a new era by enabling the creation of time series of SAR images whose exploitation for the operational monitoring of Earth’s deformation is a source of opportunities and challenges. This PhD topic proposes, for the first time, to address the major issue of inversion and prediction of geophysical parameters using Deep Neural Networks (DNN) exploiting displacement field time series obtained from SAR images. A GAN-type architecture (Generative Adversarial Network) DNN will first be proposed in order to enrich a learning data set with realistic simulations. This will then enable DNN to be trained on the inversion task, relying on LSTM-type network (Long Short-Term Memory) or causal convolution operators. This type of network can model the short- and long-term data dependencies in displacement time series. Particular attention will be paid to the explainability of the proposed network by using techniques such as salience masks, but also by ensuring that latent network features are matched with interpretable spatiotemporal data mining models/patterns.
_________________________________________________________________________________________________________________
The acquisition of Synthetic Aperture Radar (SAR) images Sentinel-1 A/B, covering Europe every 6 days (every 12 days elsewhere) and made available free of charge by the European Space Agency, ushers in a new era in the use of SAR satellite data by enabling the creation of time series of SAR images, the use of which for operational monitoring of Earth deformation presents both opportunities and challenges. This thesis topic proposes, for the first time, to address the major issue of inversion and prediction of geophysical parameters using Deep Neural Networks (DNNs) that exploit time series of displacement fields obtained from RSO images. A GAN-type DNN architecture will first be proposed to enrich existing real data with realistic simulations in order to enrich the training dataset. This will then optimize a DNN dedicated to the inversion problem. It will be capable of modeling short- and long-term data dependencies in displacement time series and will rely on Long Short-Term Memory or causal convolutional neurons. Particular attention will be paid to the explainability of the proposed network by using saliency mask techniques, but also by ensuring that representations of intermediate data, synthesized at the heart of the network, are matched with interpretable spatio-temporal data mining models/patterns.
Keywords:
inversion – prediction – neural networks – displacement – SAR
_________________________________________________________________________________________________________________
inversion – prediction – neural networks – displacement – SAR
Publications:
PhD supervisor / Thesis advisor:
PhD co-supervisors / Co-directors:
Start of the thesis:
October 1, 2022
Doctoral school:
Science, Engineering, Environment (SIE)