Keywords: remote sensing, change detection, deep learning, big data
Remote sensing data are growing fast, with improved observation frequency, spatial resolution, and data accessibility. Also, this recent context change lets benefit from the contributions of deep learning methods in the field of Earth observation images.
Revisiting times over a single site are getting shorter and shorter (a few days for satellites), which makes it possible to consider the rapid updating of land use products. Today, this update is a long and expensive process, essentially done by human expertise. Benefiting from an automatic method using satellite imagery would improve the frequency of updating these databases. In this field, much work has already been undertaken, both on radar images and optical images, as each of these modalities having different advantages and drawbacks:
- Radar images are exploitable whatever are the weather conditions of acquisition. Moreover, since radars are active systems and the acquisitions are done in very close geometric conditions, the images offer a very good temporal stability on the infrastructures, to the detriment however, of particular speckle statistics on the natural zones.
- Optical images are undeniably easier to interpret and to use by everyone, though sometimes the ground is masked by clouds. They allow to classify and detect a wide range of objects of interest, especially when using spectral bands of the non-visible domain (musltispectral imaging). By nature, they are intrinsically close to usual photos, and the transfer of deep network approaches (CNN, DBN, etc.) to these images is easy and has shown excellent performances. However, their specificities (top view, multispectral, geolocation, etc.) require the development of original methods to extract information.
The objective of this post-doc is to design change detection and semantic classification approaches for multimodal SAR / optical data. This raises several issues to tackle successively.
Fast and massive change detection
Recently, a method was developed at ONERA to rapidly detect changes in a temporal stack from open source Sentinel 1 radar images [Colin-Koeniguer et al, 2018a]. This method is very effective for quickly creating a large base of localized and dated examples of changes. [Colin-Koeniguer et al, 2018b] The idea of the first task is to use this algorithm to instantiate a potential database, for typical changes of intermediate scales of ten meters, such as construction sites, road works. The location of these changes and their dating obtained by radar could be coupled with other databases of land use, and an automatic search for high resolution Pléiades optical images, to obtain pairs of images before and after the date of the event.
The main idea is therefore to make maximum use of the radar data to build a multimodal learning dataset containing high resolution optical images.
Neural networks for change detection and semantic classification
By implementing machine learning strategies on this dataset, in particular on optical images, will allow to build change detectors in high-resolution, optical images. Indeed, datasets dedicated to this problem are rare and often suffer from imprecise and erroneous annotations [Daudt et al., 2018a] [Daudt et al., 2018b]. Of course, several problems will remain for the validation of the training dataset: removal of cloudy areas, characterization and relevance of changes detected in SAR, etc. Then, by combining these images with cartographic data (such as OpenStreetMap, cadastrer, etc.) it will be possible to develop neural network architectures for semantic land classification [Audebert et al., 2016 ] [Zhu et al., 2017] as well as for semantic classification of changes, an indispensable tool for studying and understanding the evolution of the use of urban and natural spaces.
PROFILE OF THE CANDIDATE
Education: PhD in image processing and machine learning
General knowledge of remote sensing is a plus.
Contacts :Elise KOENIGUER (firstname.lastname@example.org)
Bertrand LE SAUX (email@example.com)
Mihai DATCU (firstname.lastname@example.org)
Michel CRUCIANU (email@example.com)
Mihai DATCU is an internationally recognized researcher in the field of Earth Observation Data Science, whose research project focuses on new machine learning challenges for highly heterogeneous data. In 2017, he was awarded the "Blaise Pascal" International Chair of Excellence, funded by the Ile-de-France Region, awarded to world-renowned foreign researchers at a research or higher education institution in Ile-de-France. As such, he is hosted by the CEDRIC-CNAM and offers a collaboration with ONERA through the following post-doctoral subject.
In the CEDRIC lab of CNAM (http://cedric.cnam.fr), several research activities focus on machine learning for computer vision and in particular on zero-shot learning, deep learning from stream data and deep learning in remote sensing.
At ONERA, the IVA team (Image Vision LeArning) designs and develops methods for data processing, computer vision and machine learning. In particular, current research activities focus on SAR and optical Earth-observation, scene understanding and deep neural networks.
The post-doc is jointly financed by the "Blaise Pascal" International Chair (hosted at CNAM) and by the ONERA, so the student will benefit from a double affiliation with ONERA and the CNAM.
See this post-doc proposal also on http://spacedatascience.cnam.fr/
[Colin-Koeniguer et al, 2018a] Colin-Koeniguer, E., Boulch, A., Trouve-Peloux, P., & Janez, F. (2018, June). Colored visualization of multitemporal SAR data for change detection: issues and methods. In EUSAR 2018; 12th European Conference on Synthetic Aperture Radar (pp. 1-4). VDE.
[Colin-Koeniguer et al, 2018b] Elise Koeniguer, Jean-Marie Nicolas, Béatrice Pinel-Puyssegur, Jean-Michel Lagrange and Fabrice Janez, Visualisation des changements sur séries temporelles radar : méthode REACTIV évaluée à l’échelle mondiale sous Google Earth Engine. CFPT/RFIAP 2018, https://rfiap2018.ign.fr/sites/default/files/ARTICLES/CFPT2018/Oraux/CFPT2018_paper_koeniguer.pdf
[Daudt et al, 2018b] Fully Convolutional Siamese Networks for Change Detection, R.C. Daudt, B. Le Saux, A. Boulch, ICIP 2018
[Daudt et al., 2018a] Urban Change Detection for Multispectral Earth Observation Using Convolutional Neural Networks, R.C. Daudt, B. Le Saux, A. Boulch, Y. Gousseau, IGARSS 2018
[Audebert et al., 2016] Semantic Segmentation of Earth Observation Data Using Multimodal and Multi-scale Deep Networks", N. Audebert, B. Le Saux, S. Lefèvre, ACCV 2016
[Zhu et al., 2017] Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources, X.X. Zhu, D. Tuia, L. Mou, G.S. Xia, L. Zhang, F. Xu, F. Fraundorfer, in IEEE Geoscience and Remote Sensing Magazine, vol. 5, no. 4, pp. 8-36, Dec. 2017.
(c) GdR 720 ISIS - CNRS - 2011-2018.