PhD scholarship on deep learning and constraint reasoning, France

PhD subject: Deep learning methods coupled with constraint-based reasoning for object recognition in remote sensing data Application to the detection of the Babacu palm tree in Amazonia

Supervisor
Carmen Gervet (Université de Montpellier)
Co-supervisors
Morgan Mangeas (IRD), Samira El-Yacoubi (Université de Perpignan)
Contact
Carmen Gervet carmen.gervet@umontpellier.fr

TOPIC. The most common approaches for object identification within satellite images are based on a pixel analysis, or image segmentation techniques, using expert information or GIS. Machine learning techniques have been successfully  used when the images are enriched with GIS (Huang et al., 1997).

This thesis considers a novel approach based on convolutional neural networks, enriched with structuring methods such as ontologies and constraints generation. Convolutional neural networks have brought a spark in the learning power of neural networks in the past five years, as a deep learning method for efficient classification using a substantial volume of data (1 million of images for thousand classes) (Krizhevsky et al., 2012). Their usage for remote sensing data seems obvious, even though novel to this date. In a context of remote sensing, the images are complex because they are produced by more and more refined spatial and spectral resolutions, and obtained using many complementary sensors.


This thesis has two main objectives:

•    To generate a set-based constraint model that will allow to reduce the search space to be explored (Gervet et Van Hentenryck, 2006). The resulting model will be used to adapt the core architecture of convolutional networks in order to take into account the specific constraints relative to satellite imaging: (i) multiple resolutions (spatial, spectral and temporal), (ii) multiple types of sensors (radar, lidar, optical), (iii) radiometric noises, (iv) high heterogeneity of the observed environment, (v) handling of shadows, clouds, reliefs. A particular attention will be given to the pre-processing that can improve the recognition and the concepts of sensitivity to noise within image used and the robustness of the convergence corresponding to the learning process.

•    To better understand the results by analyzing the mathematical and statistical properties of the parameters found. To do so, ontological models using set constraints will be developed to structure and try to explain the learning process. The idea is to enter the black box of the convolutional network, by seeking to extract its structure and the connections reached, once the learning process completed.
This work will benefit from numerous data, from different projects within the research centre ESPACE- DEV (Maison de la télédétection in Montpellier), applied to different fields such as evolution of land usage and monitoring of land degradation, or the detection of the palm trees in Brazil. A result of digitalization made by human experts can be used as learning basis. Finally, the developed algorithms will be implemented in an automatic processing chain (based on OTB and Python) in order to make available cartography from environmental areas through

Report Content

Share this Scholarship with your Friends

Scholarships