The presence of insoluble particles in ice cores carry fingerprints of multiple aspects of Earths past climate. Mineral dust records allow the investigation of dust source emissions, atmospheric transport and wind strength variability. Volcanic ash (cryptotephra) particles are emitted during eruptions and deposited as individual layers in the ice. Their detection and characterization is fundamental for reconstructions of past volcanism and as a method to date and synchronize different sedimentary records, such as marine or terrestrial cores. Pollen grains and biological matter are often found in alpine glacial ice records at mid latitudes and are proxies for ecosystem changes and vegetation dynamics. To date, the analytical detection of these particles is often based on intensive manual microscopic investigations and require multiple laborious and often destructive extraction steps. Here, we present an analytical framework that can overcome these limitations, based on flow imaging microscopy coupled to deep learning neural networks for the autonomous detection and quantification of dust, volcanic tephra and pollen grain particles. The network architecture structure joins a Resnet-backbone Convolutional Neural Network and a Fully Connected Net and is trained in supervised mode. We present the developed methodology and the results applied to real ice samples. The network performs particle image classification, thus allowing the simultaneous calculation of particle number concentrations of all classes. Using information on particle size, the framework also allows the quantification of mass concentrations. The network can efficiently identify dust particles with a detection limit of 10 ppb and can thus be deployed as a dust detector in ice core analyses. The network is also able to identify tephra shards, based on trials with known volcanic horizons in the Greenlandic GRIP ice core and is therefore suitable to produce time series of past volcanic activity from ice core records. The analytical routine is non-destructive and can operate in continuous mode, thus applicable in ice core continuous-flow-analysis setups. We believe the proposed framework provides an innovative tool to support human experts across multiple research areas of ice core science.

Autonomous detection of ice core particles via deep learning

Maffezzoli Niccolò
;
Burgay Francois;Spolaor Andrea;Vascon Sebastiano;Pelillo Marcello;Ferretti Patrizia;Steffensen Joergen Peder;Dahl-Jensen Dorthe;Barbante Carlo
2021

Abstract

The presence of insoluble particles in ice cores carry fingerprints of multiple aspects of Earths past climate. Mineral dust records allow the investigation of dust source emissions, atmospheric transport and wind strength variability. Volcanic ash (cryptotephra) particles are emitted during eruptions and deposited as individual layers in the ice. Their detection and characterization is fundamental for reconstructions of past volcanism and as a method to date and synchronize different sedimentary records, such as marine or terrestrial cores. Pollen grains and biological matter are often found in alpine glacial ice records at mid latitudes and are proxies for ecosystem changes and vegetation dynamics. To date, the analytical detection of these particles is often based on intensive manual microscopic investigations and require multiple laborious and often destructive extraction steps. Here, we present an analytical framework that can overcome these limitations, based on flow imaging microscopy coupled to deep learning neural networks for the autonomous detection and quantification of dust, volcanic tephra and pollen grain particles. The network architecture structure joins a Resnet-backbone Convolutional Neural Network and a Fully Connected Net and is trained in supervised mode. We present the developed methodology and the results applied to real ice samples. The network performs particle image classification, thus allowing the simultaneous calculation of particle number concentrations of all classes. Using information on particle size, the framework also allows the quantification of mass concentrations. The network can efficiently identify dust particles with a detection limit of 10 ppb and can thus be deployed as a dust detector in ice core analyses. The network is also able to identify tephra shards, based on trials with known volcanic horizons in the Greenlandic GRIP ice core and is therefore suitable to produce time series of past volcanic activity from ice core records. The analytical routine is non-destructive and can operate in continuous mode, thus applicable in ice core continuous-flow-analysis setups. We believe the proposed framework provides an innovative tool to support human experts across multiple research areas of ice core science.
AGU Fall Meeting Abstracts
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in ARCA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/10278/5004827
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact