PhD Thesis Proposal
Topic: Machine learning, image processing and chemometrics to develop new approaches for hyperspectral imaging and multimodal spectroscopy data analysis
Joint Supervision - Double PhD Title from Doctoral schools:
· PhD Course in Models and Methods for Materials and Environmental Sciences
University of Modena and Reggio Emilia, Modena (ITALY)
PhD Course in "Models an Methods for Material and Environmental Sciences" (Administrative Reference)
· Ecole Doctorale 104 Sciences de la Matière, du Rayonnement et de l’Environnement
Université Lille Nord de France (FRANCE)
Tutors:
Prof Marina Cocchi (Modena), Prof. Cyril Ruckebusch, Prof. Ludovic Duponchel (Lille)
Timing: Call Opening June 6th 2021 - Call Deadline July 6th 2021
Duration: 3 years (18 months in Modena and 18 months in Lille) starting from 1st November 2021
The student should have achieved his Master Diploma (officially released by university, dissertation accomplished is not enough) before 30 October 2021.
We would like a skilled candidate with some chemometrics and matlab or other language programming experience, being already familiar with hyperspectral images is a plus but not a prerequisite. Also it is not necessary having a chemistry background.
Contact person: This email address is being protected from spambots. You need JavaScript enabled to view it.
Project’s ABSTRACT
The main Focus will be on: Developing New chemometric approaches to investigate the spatial-spectral interplay in chemical imaging
Hyperspectral imaging and multimodal spectroscopic are mature non-destructive tool for the analyst, as well as suitable to monitor chemical systems in time, enabling characterization of composition/structure and their spatial and temporal evolution. From the data analysis standpoint, they match the paradigm of Big Data and pose new challenges to state of art chemometrics methodology.
The research project is aimed at exploiting the integration of machine learning, image processing and chemometrics to develop new approaches suitable to be tuned on the basis of sought information while balancing modeling capability and chemical interpretation.
Main focus will be on hyperspectral imaging data, where two different approaches may be currently envisioned. On the one hand, multivariate classification methods aiming at assigning a class label to every pixel. On the other hand, spectral unmixing approaches aiming at identifying individual sources of spectral variations, with each measured pixel now described as a linear mixture of the pure spectra characteristic of those unknown individual sources. For classification, the underlying assumption is that the spectral signature measured at one pixel is characteristic of one type of object (class) only. Thus, mixed pixels represent the major problem for classification approaches. By contrast, spectral unmixing is more difficult due to the inverse nature of the problem and can benefit from the knowledge of the presence of unmixed pixels. However, classification requires (some) pixels to be labelled with class membership for model training. This work will evaluate pros and cons and both approaches, in connection with the structure of the data hand, investigated in both spectral and spatial modes.
Machine learning methods so far have been used in the classification task at pixel level, either starting from spectral of spatial dimension, without unraveling their simultaneous interplay. Novelty is the exploitation of this interplay, as well by matching spectral unmixing and machine learning tools, which should lead to new development.
