Classification of down-core foraminifera image sets using convolutional neural networks.

Abstract

Manual identification of foraminifera species or morphotypes under stereoscopic microscopes is time-consuming for the taxonomist, and a long-time goal has been automating this process to improve efficiency and repeatability. Recent advances in computation hardware have seen deep convolutional neural networks emerge as the state-of-the-art technique for imagebased automated classification. Here, we describe a method for classifying large down-core foraminifera image set using convolutional neural networks. Construction of the classifier is demonstrated on the publically available endless forams image set with an best accuracy of approximately 90%. A complete down-core analysis is performed for benthic species in the Holocene period for core MD02-2518 from the North Eastern Pacific, and the relative abundances compare favourably with manual counting, showing the same signal dynamics. Using our workflow opens the way to automated paleo-reconstruction based on computer image analysis, and can be employed using our labelling and classification software ParticleTrieur.

Innovative microfossil (radiolarian) analysis using a system for automated image collection and AI-based classification of species

“…Hsiang et al 5 constructed the Endless Forams (http://endle ssfor ams.org/) online portal, which hosts a large number of planktonic foraminiferal images that have been identified by experts, to compare the results of a CNN-based classification with the classification by humans. Furthermore, Marchant et al 13 reported that changes in the relative abundance of benthic foraminiferal assemblages estimated using a CNN-based classification showed good agreement with manual counts performed by humans. These recent studies have shown the effectiveness of deep learning as a method for microfossils classification.…”

Technical note: A new automated radiolarian image acquisition, stacking, processing, segmentation and identification workflow

“…ParticleTrieur is a dedicated software program developed at CEREGE (Marchant et al, 2020) that enables the operator to visualize and assign vignettes to manually defined classes; the program uses the k-NN (k-nearest-neighbours) algorithm to aid in identification by self-learning and progressively suggesting identification once enough radiolarian pictures are identified (the reader is referred to Marchant et al (2020) for more information). Using this software, a large dataset of radiolarian taxa images (called the AutoRadio Database) was progressively built (the current version of the database used in this study can be downloaded from http://microautomate.cerege.fr/dat, last access: 17 November 2020, and is freely accessible online as a catalogue at https://autoradio.cerege.fr/database/, last access: 17 November 2020).…”

“…ResNet50 topology (residual nets with a depth of 50 layers; He et al, 2015) with added cyclic (Dieleman, 2016) and gain layers (resnet50_cyclic_gain_tl; see Marchant et al, 2020, for a detailed description of the network), greyscale images resized to 256 px × 256 px, a batch size (number of images presented per training iteration) of 64, 30 epochs and four drops for the adaptive learning rate (ALR) system, and augmentation (Marchant et al, 2020). This training process lasts about 30 min and generates two files that can then be used for automated recognition (network_info.xml and frozen_model.pb files).…”

“…Several workflows inspired by SYRACO and now using CNNs have been successively developed at CEREGE and applied to microfossil taxa (e.g. Marchant et al, 2020;Bourel et al, 2020). Regarding radiolarians, previous attempts have mainly focused on the identification step.…”