Hyperspectral Imaging for high resolution, non-destructive and fast analysis of sediment cores : application to Lake Le Bourget and Black Sea sediment cores.
Résumé
Sedimentary archives are used to infer past climate and environment thanks to their physical and chemical properties. Sampling methods (millimetric or centimetric) and routine analysis are time and material consuming. The use of some specific spectroscopic methods and data analysis, allow to develop and perform some robust methods capable of (i) fast high resolution (ii) performed at low costs (iii) non-destructive and (iv) monitor concentration variations of major sediment compounds. X-ray fluorescence spectroscopy is one of these techniques, but it is able to detect just mineral geochemistrys. Whereas hyperspectral imaging (VNIR 400-1000 nm, SWIR 1000-2500 nm) allow, behind each voxel (a pixel with several wavelengths), to define spectral fingerprint of organic or mineral chemical compounds. This type of data can be analyzed by statistical techniques. Many (pseudo-) univariate coefficients are available for the quantification of some molecules (RABD845 for BPhe a, RABD660-670 for chlr-a and chlorins). But in this study we choose to applied multivariate methods that take into account all spectra variations. To achieve such study we can use technic that usually applied in classical spectroscopy or for satellite data that can be unsupervised or supervised. For unsupervised techniques, without any prior knowledge of the sample, exploratory algorithms are used to determine groups in the data. Then, these groups can be interpreted with the comparison to other analytical methods. It is possible to find pure signal that corresponds to one or several organic or mineral sedimentary compounds by (i) endmembers techniques, (ii) spectral unmixing, or (iii) clustering. For supervised techniques, we can use the knowledge of sample chemical and physical properties to create prediction models, then it is possible to observe variations of a specific property along the core. To develop qualitative and quantitative model for focused spectral properties we can applied classification and regression techniques. They allow to discriminate spectral domains or some wavelengths for some interest property. In the present study, Lake Le Bourget (Savoie, France) and Black Sea (Northwest margin) sediment cores are used. From this two different environmental systems we could create and test several prediction models. The high-resolution acquisition is done with two hyperspectral cameras: VNIR (400-1000 nm) and SWIR (1000-2500 nm) with spatial resolution of several dozen micrometers. Both sensors are well designed to create predictive models for either physical or chemical properties. In order to improve prediction models and make them more robust, we can pair these two cameras and add XRF core scanner data. For the black sea sediment, we use unsupervised techniques to determine groups and define interesting spatial areas to take samples for analytic analysis. Whereas for the Lake le Bourget sediment, several previous studies allow us to have many available data, thus supervised techniques are used to observe along core variations. For some properties, we could try to use models of the Lake Le Bourget in the Black Sea data, for example if we create a grain-size model, chemical elements ratio or organic compounds.