DS0104 - Innovations scientifiques et technologiques pour anticiper ou remédier les risques environnementaux

FORAMINIFERA IMAGE RECOGNITION AND SORTING TOOL – FIRST

A working automated microfossil sorter

The FIRST project has enabled the development of a prototype for automatic sorting of microfossils.<br />This prototype is based on a classification by convolutional neural networks of microfossils. This machine launches high-throughput micro-paleontology for paleocanography and biostratigraphy.

Automation of microfossil analysis for high throughput paleoceanography

The objective of the FIRST project was to develop a prototype for the automatic analysis and processing of carbonate microfossils. These microfossils and in particular foraminifera have made it possible to discover the extent of climate change in the 20th century. These organisms, which range in size from one hundred metres to one to two millimetres, are found throughout the global ocean and secrete fossilized carbonate shells on the ocean floor. The chemical composition of these shells, the size and mass of these shells, as well as the different species (faunal assemblage) are indicators that are routinely used by a large scientific community. They make it possible to plot several of the parameters critical to the estimation of anthropogenic influence on the climate system: ocean surface temperature, hydrological structure and thus the amount of heat stored in the ocean, and salinity. Their faunal and geochemical analysis provides a unique way to quantify climate variability over the past thousands to millions of years. The objective of the FIRST project was to automate the processing of these calcareous tests.

The sorting prototype - named MiSo (Microfossil Sorter) - built within the consortium combines ATG Technologies' expertise in robotics and mechanics with the development of automatic classification methods based on convolutional neural networks at CEREGE. The principle adopted for the machine makes it possible to separate sedimentary particles individually with a success rate of about 90%,
despite the diversity of sedimentary particles. Each particle is imaged with a z-stack system, which allows high quality images to be obtained quickly (<1s). A topology adapted to sedimentary objects of convolutional neural networks has been developed and integrated into a software suite. The approach chosen in the project is to develop site-specific image data sets, sediment cores, data sets that allow learning about the neural network. The prototype, due to the stability of the imaging conditions, allows to obtain an accuracy of about 90% for the main foraminiferous species studied, in real conditions on sedimentary sequences.

The MiSo prototype makes it possible to sort around 8000 sedimentary particles 24 hours a day, and to sort up to 30 samples continuously for 4 different species. It allows to build massive data sets of microfossil images (> 2 Million images during the project), opening a new field of analysis in biometrics on sediment sequences. The tools developed in imaging (image purification, pretreatment, and neural network construction) have been applied to other images such as pollens, and living plankton, opening up new applications in palynology (fossil and modern), characterization of coastal ecosystems (marine station monitoring) and pelagic ecosystems.

The FIRST project makes it possible to automate the microfossil extraction processing chain for
geochemical analyses: the stitching of massive samples for clumped isotope or carbon14 analyses is now possible and saves significant operator time. On a larger scale, this prototype allows the development of automated biostratigraphy based on microfossil biometry. Finally, in principle, the prototype can be used as a basis for sorting other particles such as microplastics contained in sediments (media characterization), heavy minerals for geological dating, and waste shredder residues.

The separation principle implemented is currently being patented. Two software repositories at the APP correspond to the machine control software (MiSo), and one software for image annotation and tracking (ParticleSorter). The topology of the convolutional neural network and the validation of the dataset are published (Marchant et al, patent pending). Finally, this software has been successfully applied to images of modern and fossil pollens (Bourel et al, submitted).

The aim of the FIRST project is to develop an automated prototype for the processing and analysis of foraminifera, a group of calcareous microfossils. These marine organisms, whose size spans 100µm to 1 mm, are cosmopolitan, and secrete unique carbonate shells that accumulate on the ocean floor. The geochemical composition of these shells, the faunal assemblages, as well as the size and mass of foraminifera are environmental and biostratigraphic indicators used daily by a worldwide scientific community. These shells yield insights into many critical parameters used to quantify the anthropogenic impact on natural systems, and on climate: sea surface temperature (SST), the hydrological structure therefore oceanic heat content, and salinity. The study of faunal and geochemical indicators of foraminifera is thus a unique way to quantify past climatic variability during the last thousand years and beyond.

The protocol used to extract foraminifera from sediments has not evolved since the XIX century: (i) sediments are wet sieved to remove clays from the larger biogenic and clastic fraction. (ii) After this initial step, foraminifera are individually picked and sorted by species by a human operator, using a fine brush, under a stereomicroscope. (iii) Once separated, the shells are counted, measured and (iv) analyzed for geochemical compositions. The aim of this project is to restrict the manpower needed for steps (ii) and (iii) of physical sorting and recognition of shells, in a similar manner to cell sorting used in biomedical and oceanographic studies based on flow cytometry. Our ultimate goal is to use this automated method to document past climatic changes at high resolution for the last thousand years.

Automation is justified by the need to tackle three basic scientific requirements and one economic criterion: (1) an improvement of the accuracy of identification (2) the standardization of traceability of morphologic analyses (3) an improvement in reproducibility and last (4) a decrease in manpower time dedicated to tedious hand-picking.

To build this prototype, the consortium will have to resolve three technological bottle-necks : (i) the separation and segregation of foraminifera to achieve the handling of those microparticles, and eventually their extraction, (ii) the development of imaging systems relevant to pattern recognition and (iii) the development of classification algorithms allowing the computerto discriminate between different morphotypes of foraminifera.

This prototype will be developed by a consortium comprising one small business, ATG Group specialized in robotics and vision, and a group of scientists working at CEREGE (CNRS & Aix-Marseille Université), specializing in calcareous microfossils and image processing. The project will build upon existing experience at CEREGE with the SYRACO system, a software package capable of automatic recognition of coccolithophorids. This system, patent pending, is used in numerous research groups and in industry.

Project coordination

Thibault DE GARIDEL-THORON (Centre National de la Recherche Scientifique, délégation Provence et Corse _ Centre Européen de Recherche et d'Enseignement en Géosciences de l'Environnement)

The author of this summary is the project coordinator, who is responsible for the content of this summary. The ANR declines any responsibility as for its contents.

Partner

ATG ATG Technologie
CNRS DR12 _ CEREGE Centre National de la Recherche Scientifique, délégation Provence et Corse _ Centre Européen de Recherche et d'Enseignement en Géosciences de l'Environnement

Help of the ANR 313,024 euros
Beginning and duration of the scientific project: September 2015 - 36 Months

Useful links

Explorez notre base de projets financés

 

 

ANR makes available its datasets on funded projects, click here to find more.

Sign up for the latest news:
Subscribe to our newsletter