IAP-24-082

Unlocking Earth’s volcanic record with novel computer vision approaches

Explosive volcanic eruptions can have profound environmental and societal consequences. The most powerful volcanic events inject gas and ash into the stratosphere which can lead to significant global cooling. However, even relatively small volcanic eruptions (e.g. the 2010 Eyjafjallajökull eruption in Iceland) can have major impacts on our modern, globalised world.

A detailed record of past volcanic events is critical for improving eruption forecasts and preparing society for future eruptions. The discovery of microscopic shards of volcanic ash (tephra) in diverse environmental archives located far from volcanic areas, such as ice-cores, lake sediments, marine cores and peat deposits, has revolutionised our understanding of volcanic hazards, climate change and environmental change.

However, the full potential of volcanic ash to inform us about past, present and future volcanism has yet to be realised. This is because the techniques used to search for, identify and quantify the number of volcanic ash shards have not fundamentally changed for decades. These techniques rely on the laborious manual identification of ash shards using optical microscopy and from looking at scanning-electron microscope images. This severely limits how much of the existing environmental archives we check for the presence of volcanic ash. For example, of the 800’000 year ice-core archive from the polar ice sheets, only a tiny fraction of the available material has been searched in detail. For the vast majority (99 %) of volcanic sulfur deposits identified in these records, we have no idea which volcano was responsible and therefore which regions tend to produce the largest volcanic eruptions.

In addition, fundamental questions about the preservation of volcanic ash in distal settings remain unanswered as its difficult to process the volume of material required. Without addressing these issues, it is difficult to fully interpret the significance of the presence or absence of volcanic ash which has implications, e.g. for understanding the likelihood of volcanic ash closing airspace in north-western Europe.

This project will apply innovative new approaches in computer vision to directly address these constraints and to massively increase our capacity to look for volcanic ash. Computer vision techniques such as deep classification and segmentation neural networks are now routinely used in healthcare to provide rapid and reliable identification of features, for instance in cancer detection and localisation. We will apply similar techniques to identify volcanic ash in optical and SEM-EDX (Scanning Electron Microscope and Energy Dispersive X-ray spectroscopy) images (Figure 1). Volcanic ash is the ideal candidate for computer vision techniques as it has distinctive morphological and chemical properties which distinguish it from other common particles found in sediments and ice-cores, such as dust and pollen. Developing such a tool can only be achieved through new collaboration and this PhD project unites supervisors with expertise in volcanic ash identification and analysis, with experts in computer vision and deep learning.

The automation of identification of volcanic ash has the potential for generating a step change in the way researchers around the world search for tephra in diverse environmental archives. The increase in research speed this approach promises will allow tephra studies to fully realise its potential in providing critical insights into the hazards generated by volcanic activity and for greatly enhancing our understanding of environmental change.

This project addresses three main challenges:

Component 1: We need to design a protocol for rapidly collecting SEM-BDX images of sufficient quality and in a way that doesn’t affect the geochemistry of tephra shards
Component 2: We need to design and implement a computer vision tool which can reliably identify volcanic ash particles SEM-BDX images, and we need to test that tool against established methods of identification to ensure it is accurate and reliable
Component 3: To demonstrate that this new approach can generate new insights by speeding up identification of tephra in diverse environmental archives (peat and lake sediments, ice cores)

Component 1
Our existing workflows and analytical equipment can capture images and chemical data of high quality, but refinement of the analytical conditions used on the JEOL-IT200 are required to speed up image acquisition. Detailed geochemical analysis of samples using EPMA (Electron Probe Micro Analyser) will be required to quantify the degree of geochemical alteration of volcanic ash shards because of the BDX imaging.

Component 2
We lack the computing tools to use the SEM-BDX images to automatically and robustly identify volcanic ash shards as distinct from other material. We will use established computer vision techniques to develop such a tool for the identification of tephra shards. For deep learning we need a large training dataset. We are starting with a significant number of existing images, but we will augment these data creating synthetic dummy samples of a wide range of tephra particles with differing properties (e.g. with different chemistries and morphologies) to ensure our training dataset captures the full range of likely particles. Once the analysis tool is created, we will test its efficacy by comparing it against established techniques, such as optical microscopy counting of shards.

Component 3
To demonstrate the wider utility of the approach and how it can be used across a range of preservation environments, we will use the tool to identify and quantify the tephra within unanalysed environmental records. These would include unanalysed ice-core records, as well as the collection of new peat and lake sediment cores from sites in Scotland.

Year 1

Literature review. Creating a reference set of synthetic training material from a wide range of volcanic material (e.g. variations in tephra geochemistry, morphology) and mount this material in resin stubs. Collate existing research imagery. Experimenting with SEM-BDX analysis conditions to produce images of required quality and use this to generate a large training dataset from the synthetic material.

Year 2

Development and training of a software tool based on the synthetic images. Collecting peat and lake cores which likely contain cryptotephra, processing the cores and record tephra abundance using both traditional and new software approaches, attendance at international conference (IAVCEI)

Year 3

Continuation of tephra analysis, including geochemical analysis using EPMA; attendance
at major international conference (e.g., INQUA); writing up.

Year 3.5

Writing up, manuscript production.

The candidate will receive training in the following areas: Collection of SEM-BDX images, image recognition software development (using TensorFlow or PyTorch), field sampling of tephra (sediment coring and core processing techniques), EPMA geochemical analysis, statistical training in R and Python, creation of resin stubs for geochemical analysis.