Periodic Reporting for period 1 - ICELEARNING (Artificial Intelligence techniques for ice core analyses)
Période du rapport: 2020-01-15 au 2022-01-14
Earth System scientists can investigate the Earth past climate by studying natural archives where information can be frozen in time. Examples of such archives are the bed floor of the oceans or ice from polar regions and glaciers. Ice cores have provided us with some of the most pristine records of the Earth’s past climate. While the oldest Antarctic and Greenland cores date back to up to 800,000 and 125,000 year ago respectively and register variability of climate parameters at hemispheric scales, ice stored in glaciers and small ice caps located at lower latitudes typically contain a fingerprint or local to regional climate. Understanding climate variability is important to bring forward our knowledge of Earth system climate, understand the past and refine models to better constrain future scenarios.
Different types of climate information can be trapped in ice cores and can be retrieved using many methods and analytical techniques. The climate proxies that have been targeted in this project are insoluble particles, i.e. visible particles that do not dissolve in water. The broad category of insoluble particles include mineral dust, volcanic particles, biological matter like pollen grains and marine particulates like diatoms and foraminifera. The investigation of these particles in ice cores allow to produce past climate records which scientific importance are far reaching. Dust records allow to investigate the Earth past aridity, vegetation cover and atmospheric circulation. Volcanic particles tell us about past volcanic eruptions thereby allowing to investigate past volcanic activity. Biological matter and marine particles tell climate scientists crucial information about past vegetation and sea level.
At present, by far the most common technique deployed to find and classify insoluble particles in ice cores is manual microscopy. Despite bench microscopy can provide precise information to the operator, its use is extremely laborious, time consuming and a limited number of samples can be processed. The overall objective of the ICELEARNING project is to develop a methodology that would be able to support human researchers in finding and at the same time automatically classifying different types of insoluble particles in ice cores. We do so by deploying a Flow Imaging Microscope, called FlowCam, an instrument capable to continuously acquire images of particles in liquid ice core samples. The images are then processed by neural networks, specifically trained to recognize and count autonomously different types of particles at the same time. The developed methodology would not only benefit ice core scientists but would positively impact different branches of science where investigations by manual bench microscopy are still the standard approach.
The scientific project objectives include: characterization of the FlowCam instrument for ice core samples, selection and acquisition of the training datasets necessary to train classification models, development and training of the models, application of the models to real ice core samples and comparison with human microscopy to access the potentials and limitations of the developed methodology.
Different types of climate information can be trapped in ice cores and can be retrieved using many methods and analytical techniques. The climate proxies that have been targeted in this project are insoluble particles, i.e. visible particles that do not dissolve in water. The broad category of insoluble particles include mineral dust, volcanic particles, biological matter like pollen grains and marine particulates like diatoms and foraminifera. The investigation of these particles in ice cores allow to produce past climate records which scientific importance are far reaching. Dust records allow to investigate the Earth past aridity, vegetation cover and atmospheric circulation. Volcanic particles tell us about past volcanic eruptions thereby allowing to investigate past volcanic activity. Biological matter and marine particles tell climate scientists crucial information about past vegetation and sea level.
At present, by far the most common technique deployed to find and classify insoluble particles in ice cores is manual microscopy. Despite bench microscopy can provide precise information to the operator, its use is extremely laborious, time consuming and a limited number of samples can be processed. The overall objective of the ICELEARNING project is to develop a methodology that would be able to support human researchers in finding and at the same time automatically classifying different types of insoluble particles in ice cores. We do so by deploying a Flow Imaging Microscope, called FlowCam, an instrument capable to continuously acquire images of particles in liquid ice core samples. The images are then processed by neural networks, specifically trained to recognize and count autonomously different types of particles at the same time. The developed methodology would not only benefit ice core scientists but would positively impact different branches of science where investigations by manual bench microscopy are still the standard approach.
The scientific project objectives include: characterization of the FlowCam instrument for ice core samples, selection and acquisition of the training datasets necessary to train classification models, development and training of the models, application of the models to real ice core samples and comparison with human microscopy to access the potentials and limitations of the developed methodology.
- The FlowCam instrument: the characterization of the instrument was carried out at the University of Bergen. The characterization included studies on the blank background level, optimization of the settings for ice core measurements and selection of the necessary materials for trace level analyses. Two 3-month secondment periods at the University of Bergen were carried out in 2020 and 2021.
- Training dataset acquisiton: we create a training dataset of the following types of particles found in ice core records: mineral dust, basaltic tephra, felsic tephra and three types of pollen grains: Corylus Avellana, Quercus Robur, Quercus Suber. The overall dataset consisted of more than 100,000 images.
- Model development: the training dataset was used to train a supervised-learning classification model. Research on the model was carried out throughout the project timeline and included investigations on the model architecture, preprocessing and data augmentation routines, training, validation and hyperparameter optimization.
- Investigation of ice core samples using the trained model. The trained model was deployed to investigate the content of particles inside GRIP (Greenland Ice Core Project) ice core samples, dating 13,000 to 17,000 years ago. The samples were selected as corresponding to known volcanic horizons during the Bølling–Allerød (GI-1) and glacial (GS-2) periods. The samples have been analyzed and the acquired images were classified by the model to infer the composition of particles contained within. The model found many volcanic shards. We have also investigated ice samples from the Quelccaya ice core (Peru).
- Dissemination and result exploitation: the project activities were disseminated at 2 conferences: the European Geophysical Union 2020 (virtual), American Geophysical Union 2021 (13-17 December 2021, New Orleans, USA). A website was created and updates were released during the project: www.icelearning.net. We have been in close collaboration with the instrument manufacturing company (Fluid Imaging Technologies Corporation). The project datasets were used as an Assignment for a MSc course “Applied Machine Learning” at the University of Copenhagen in 2021 and is currently being the topic of a 1 year Msc thesis in physics. A peer-reviewed publication is scheduled to be submitted within the first half of 2022 to The Cryosphere journal discussing all results of the project research and the developed methodology. All datasets and the model will be released publicly to foster further research.
- Results: we have successfully collected a training dataset for each class under investigation (mineral dust, Felsic tephra, Basaltic tephra, Corylus Avellana pollen, Quercus Robur pollen, Quercus Suber pollen) and developed a model capable of classifying the particles. With respect to dust, we have also developed a methodology that can quantify dust mass concentrations with similar performance and several advantages compared to state-of-the-art methods.
- Training dataset acquisiton: we create a training dataset of the following types of particles found in ice core records: mineral dust, basaltic tephra, felsic tephra and three types of pollen grains: Corylus Avellana, Quercus Robur, Quercus Suber. The overall dataset consisted of more than 100,000 images.
- Model development: the training dataset was used to train a supervised-learning classification model. Research on the model was carried out throughout the project timeline and included investigations on the model architecture, preprocessing and data augmentation routines, training, validation and hyperparameter optimization.
- Investigation of ice core samples using the trained model. The trained model was deployed to investigate the content of particles inside GRIP (Greenland Ice Core Project) ice core samples, dating 13,000 to 17,000 years ago. The samples were selected as corresponding to known volcanic horizons during the Bølling–Allerød (GI-1) and glacial (GS-2) periods. The samples have been analyzed and the acquired images were classified by the model to infer the composition of particles contained within. The model found many volcanic shards. We have also investigated ice samples from the Quelccaya ice core (Peru).
- Dissemination and result exploitation: the project activities were disseminated at 2 conferences: the European Geophysical Union 2020 (virtual), American Geophysical Union 2021 (13-17 December 2021, New Orleans, USA). A website was created and updates were released during the project: www.icelearning.net. We have been in close collaboration with the instrument manufacturing company (Fluid Imaging Technologies Corporation). The project datasets were used as an Assignment for a MSc course “Applied Machine Learning” at the University of Copenhagen in 2021 and is currently being the topic of a 1 year Msc thesis in physics. A peer-reviewed publication is scheduled to be submitted within the first half of 2022 to The Cryosphere journal discussing all results of the project research and the developed methodology. All datasets and the model will be released publicly to foster further research.
- Results: we have successfully collected a training dataset for each class under investigation (mineral dust, Felsic tephra, Basaltic tephra, Corylus Avellana pollen, Quercus Robur pollen, Quercus Suber pollen) and developed a model capable of classifying the particles. With respect to dust, we have also developed a methodology that can quantify dust mass concentrations with similar performance and several advantages compared to state-of-the-art methods.
The main result achieved with the project is the development a new methodology based on Flow Image Microscopy coupled to neural networks to investigate insoluble particles in ice core records. Before the project, such a method did not exist. The methodology can support the whole ice core community in analyzing ice core records using Continuous Flow Analysis setups. Additionally, the methodology can be a very important complementary tool for scientists involved in different research fields in which bench microscopy is the standard technique, such as palynology and multiple research lines within paleo science.