Explainable Deep Neural Network (DNN) for model inversion: application to Synthetic Aperture Radar (SAR) Satellite Image Time Series (SITS)
The availability 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 satellite SAR data exploitation to a new data driven era, which provides scientists with both opportunities and challenges for operational monitoring of Earth deformation by exploiting SAR image time series.
This Ph.D subject proposes, for the first time, to tackle the major issue of neural networks based inversion and prediction from SAR displacement time series. We aim to propose a supervised neural networks based learning framework to predict the evolution of some key geophysical parameters (strongly related to natural hazards but cannot be observed directly, e.g. overpressure of a magma chamber) from SAR displacement time series. To solve a regression problem with a few training data, we will adopt a GAN strategy as in [E18]. To follow the temporal evolution, LSTM or alternatives (related ref. [T19][Z19]) will be exploited to capture both long and short range data dependencies in SAR displacement time series. Though deep neural networks achieve high level performances, their prediction rationale remain obscure to AI developers and related stakeholders such as companies or scientists. Works such as those surveyed in [G18] aim to open such black boxes and often propose to explain outcomes using saliency masks, i.e. by identifying the input data subsets utilized by the networks to establish their predictions [X15, R18]. We propose to combine this kind of approaches with a new one which is about building networks having neurons/concepts matching interpretable spatiotemporal datamining patterns such as those proposed in [M19].
The previously developed methods will be applied to targets of geophysical interest with different displacement behaviours: the Merapi volcano located in central Java in Indonesia (Sentinel-1 A/B), the Piton de la Fournaise on the eastern side of the Réunion Island (Sentinel-1 A/B) SAR images, the Grimsvotn volcano in southeast Iceland covered by daily GNSS (Global Navigation Satellite System) data, the Alpine glaciers covered by Sentinel-1 A/B images and by high resolution PAZ (first Spanish radar Earth Observation satellite).
Encadrements : Nicolas Méger (email@example.com), Alexandre Benoit (firstname.lastname@example.org) , Yajing Yan (email@example.com)
Laboratoire : LISTIC, Université Savoie Mont-Blanc
Lieu : Annecy
[E18] E. Laloy, R. Hérault, D. Jacques and N. Linde, (2018), Training-image based geostatistical inversion using a spatial generative adversarial neural network, Water Resources Research, vol.54, pp.381-406.
[G18] R. Guidotti, et al. 2018. A Survey of Methods for Explaining Black Box Models. ACM Comput. Surv. 51, 5, Article 93 (August 2018), 42 pages. DOI: https://doi.org/10.1145/3236009
[M19] N. Méger, C. Rigotti, C. Pothier, T. Nguyen, F. Lodge, L. Gueguen, R. Andréoli, M-P. Doin and M. Datcu. Ranking Evolution Maps for Satellite Image Time Series Exploration – Application to Crustal Deformation and Environmental Monitoring. Data Mining and Knowledge Discovery, volume 33, issue 1, pp. 131-167, January 2019. doi: 10.1007/s10618-018-0591-9
[R17] R. R. Selvaraju et al. Grad-cam: Why did you say that? visual explanations from deep networks via gradientbased localization. 2017 IEEE International Conference on Computer Vision (ICCV), Venice, 2017, pp. 618-626.
[T19] M. Titos, A. Bueno, L. Garcia, M.C. Benitez and J. Ibanez, (2019), Detection and classification of continuous volcano-seismic signals with recurrent neural networks, IEEE Transactions on Geosciences & Remote Sensing, 57(4), pp.1936-1948, doi: 10.1109/TGRS.2018.2870202
[X15] K. Xu et al. Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning, pages 2048–2057, 2015.
[Z19] R. Zhang, Z. Chen, S. Chen, J. Zheng, O. Buyukozturk and H. Sun, (2019), Deep long short-term memory networks for nonlinear structural seismic response prediction, Computers and Structures, vol.220, pp.55-68.
(c) GdR 720 ISIS - CNRS - 2011-2020.