ICELEARNING

Outlier detection.

We have trained a classification model to classify ice core particles based on a number of classes. However, we should expect that real ice core samples contain a larger variety of particles than those we have trained the model with. It is therefore desirable to develop a method that is capable to alert us in case an unexpected particle is imaged by the FlowCam. Commonly referred to as outlier or anomaly detection, this fundamental task that can be tackled with machine learning techniques. The image shows a freshwater lake diatom found in the Peruvian Quelccaya ice core.

Update 17

Super-Resolution using Generative Adversarial Networks.

Super-Resolution imaging refers to all those techniques that aim at increasing the quality (resolution) of images. Recently, neural networks, and in particular Generative Adversarial Networks have shown great potential in this task, achieving extremely good results. In our case the images obtained from the FlowCam are at times very blurry, and an increase in the image quality is certainly desirable. The images show a raw tephra image acquired with the FlowCam and the same image at 4x resolution obtained using the Real-ESRGAN model developed by Xintao Wang et al. (2021). The question as to whether the high resolution image would help the classification model is certainly well posed.

Update 16

Ancient volcanic tephra from the last glacial cycle !

We have applied the model to real ice core samples from the GRIP ice core, Central Greenland. The samples we analyzed are known to contain volcanic tephra and they were chosen to test whether our model can successfully flag some of the particles as either basaltic or felsic tephra. And the answer is yes, some of the images classified as tephra look indeed like realistic tephra shards. This is a super result as it shows the potential of this system (FlowCam coupled to deep learning) to autonomously find particles that carry valuable information to understand the Earth past climate.

Update 15

Model training – accuracy up, loss down.

While training the model, in each training loop a mini batch of the training samples is evaluated and the loss is quantified. The gradients of the loss function with respect to all model weights are calculated and the weights are updated following a gradient descent algorithm. At the same time the model accuracy is evaluated on both the training and validation datasets. This is repeated until all items in the training dataset are scanned. In summary, during a well-behaved training time, the loss function decreases and the accuracy increases.

Update 14

A hybrid model.

From the FlowCam we can collect two types of information of each particle that passes through the instrument. The particle image and a number of numerical features, often called metadata, that are calculated from the instrument software. The majority of these features are related to the particle geometry, e.g. diameter, angle, circulatiry, while others relate to the image itself, e.g. pixel intensity, sharpness. To classify every particle we aim for a model that is be capable to integrate and use the information from both the images and the metadata (the more info the better). A Convolutional Neural Network (CNN) is the most natural choice to process images, while a Fully Connected Neural Networks (also referred to as Multi Layer Perceptron, MLP) can process numerical features. Our model uses and connects both types of models into one single hybrid architecture.

Update 13

Validation: hyperparameter optimization.

Training a neural network consists in finding the best model parameters. Some of these parameters are called hyperparameters and are fine tuned by evaluating the model performance on a subset of the training set, called validation set. The hyperparameters are found by maximizing the performance on the validation dataset. In our case we randomly collect 500 images per class from the training set to build the validation dataset. These images are not used for training so the model is has no a-priori knowledge of them and the performance evaluation has no bias. One of the most crucial hyperparameter is the learning rate. Image by JPuckett Design.

Update 12

Training dataset representativeness.

For a classification algorithm to be effective and general, its training datasets must be a very good general representation of the real data the model will be applied to. In particular, I am interested in investigating the representativeness of the pollen classes in the training set. Image by andrasgs from Pixabay

Update 11

Training dataset: Felsic Tephra.

Opposite to the dark looking basaltic ash, some types of tephra appear translucent or transparent. We therefore build an additional tephra class and use samples of the Campanian Ignimbrite eruption of the Phlegrean Fields, that occurred 40,000 years ago and the most explosive in Europe in the last 200,000 years. The photo shows a batch of felsic tephra particles which is part of the training dataset for this class. The particles are between 10 and 40 microns.

Update 10

Training dataset - Basaltic tephra.

Basaltic tephra layers are very important in both ice and sediment records. Therefore, we would like to have a model that is capable to tell us when a basaltic tephra particle is imaged in the FlowCam. Here, I am preparing some basaltic tephra powder to inject in the FlowCam in order to build a training set for this class. I am using ash from the Icelandic Grímsvötn volcan. This powder was collected during the 2011 eruption.

Update 9

Training dataset: Pollen.

Ice cores contain different types of biological matter, the type and the amount depending on the location of the ice core in relation to the vegetation source. We explore the possibility to automatically detect pollen grains, which are important for palynological studies. The photo shows a sample containing Corylus Avellana pollen grains diluted in ultra pure water. This sample will be analyzed to build a training dataset of images of this class.

Update 8

Training Dataset: Dust.

Mineral dust is are the most abundant particle type in ice core records. This particle has been historically extremely important for ice core studies and has led to many discoveries in the field. Dust is rather stable in the ice and its variability in concentration is linked to many climate parameters, like aridity, wind strength and circulation. To build a training dataset of dust particles, we use a very fine grained aluminium oxide powder, with a size distribution similar to that of dust found in Antarctica or Greenland ice cores. On the left is a photo of a sample obtained with scanning electron microscopy (Photo by Thomas P.J.Linsinger et al., 2019).

Update 7

The FlowCam instrument.

The goal of this project is to develop a system that can automate as much as possible the recognition of different types of particles in ice cores, a process that today is often carried out by bench microscopy, which is very time consuming. We will use the FlowCam, a microscope that can acquire images of particles flowing inside a liquid sample that is pumped inside the instrument. The FlowCam, a VS-I-B-model, is located at the University of Bergen, in Norway. Once the images are obtained, the particles can be classified by neural networks.

Update 6

Chasing clean water.

Since we are working with particles in ice cores, the concentrations are very low. Therefore, the background level of impurities in all materials we are using has also to be very low, including water. After a few days of investigations on the cause of high blank levels, I managed to understand the impurities were in the pure (not ultrapure water) I was using. We now got some very ultrapure water to work with.

Update 5

Image aquisition, analysis and training sets

The image aquisitions phase has started at the University of Bergen. The goal is to first characterize the instrument. Then I'll try to reproduce reference dust concentrations. Next, I will build training datasets for different particle classes, which include volcanic tephra.

One of the challenges is represented by the low concentration of particles in polar ice.

One the right are photos of particles in ice from the Adamello glacier (Alps, Italy)

Update 4

Convolutional Neural Network layers

During these months of lockdown I have been coding foward and backward passes for all of the most used layers in Convolutional Neural Network architectures. The backward passes are the trickiest part and deserve some careful differentiation and abstraction, which I enjoy doing first with pen and paper. The backproprgation techique was derived by several researchers in the early 60's and its implementation to train neural networks suggested in the 70's, but became widely used only in the 2010's thanks to powerful GPU-based computing systems. In the photo is the Group Normalization layer, which rationale to normalize ConvNets is similar to Layer Normalization. This technique was just recently suggested by Yuxin Wu and Kaiming He (2018)

Update 3

Dust detection

Eolian mineral dust in ice cores is particularly important to understand the past climate because its concentration is linked to the Earth aridity, atmospheric moisture content, atmospheric circulation patterns and wind strength. The figure on the left is a photo of an ice core sample obtained by optical microscope. The single dust particles have been identified, counted and color-coded singularly by automatic image analysis.

Update 2

15 Jan 2020 - START

The two year ICELEARNING project (2020-2022) starts today 15 january 2020 at the Ca' Foscari University of Venice (morning photo of the 'Canal Grande').

The first step is to select which class of particles to focus on. Dust and tephra particles are good candidates.

Update 1