5:15 PM - 7:15 PM
[MIS14-P11] Automated volcanic glass shards counting system for cryptotephra analysis using open-source hardware and software and its application to paleoenvironmental research
Keywords:tephrochronology, volcanic glass shards, cryptotephra, machine learning, YOLO (You Only Look Once)
Tephrochronology, which uses pyroclastic deposits as time marker horizons, is one of the most robust tools for synchronizing geologic records. Recently, a continuous analysis to obtain the concentration of glass shards in geological records such as marine, lacustrine, eolian sediments, and ice core records has been widely used. This method is called cryptotephra analysis (e.g., Davies, 2015). Cryptotephra detection allows the extension of geographic distribution and the contribution of a high-resolution tephrostratigraphic framework. Many cryptotephra studies have been conducted worldwide. However, cryptotephra analysis is time-consuming, so the development of a time-saving method is needed. Here, machine learning-based image analysis using a convolutional neural network was applied to cryptotephra analysis, and we established the Automated Cryptotephra Image Analysis System (ACIAS: Watanabe and Yoshida, 2025; GR of TMU). This system is easy to install due to its accessibility and low cost.
This system consists of the full-slide image acquisition component and the glass shards content detection component. In the former component, the PiAutoStage (Steiner and Roonie, 2021), consisting of Raspberry Pi, Arduino Uno, electrical parts, and 3D printer parts, was used to acquire full-slide images. In the latter component, the You Only Look Once (YOLO) v5 algorithm was used to detect glass shards and determine the concentration of glass shards. Training data consisted of 15 tephra samples from the proximal outcrops and six from the YNH-P2 core (Watanabe et al., in prep.) from Yanohara Bog, Fukushima Prefecture. Thin sections were prepared using particles between 62 and 125 µm and a light-curing agent (refractive index of 1.55 after curing). Photo images were taken with a Raspberry Pi HQ camera attached to a polarizing microscope under open Nicol conditions. The grains were classified into 11 categories: volcanic glass shards were divided into five categories according to morphology, minerals were three categories, and others were three categories. Based on this model, the overall accuracy for detecting glass shards was 84.9%.
Next, we evaluated the reliability of ACIAS counting using the YNH-P2 core. The comparison of manual counting and ACIAS counting showed that ACIAS counting could be an alternative method to manual counting. The two visible tephras and five cryptotephras were detected in the YNH-P2 core (Watanabe et al., in prep.). The variation in concentration of glass shards using both methods was very similar, and ACIAS was able to detect cryptotephras.
The concentration of glass shards has mainly been used to detect the cryptotephra horizon. On the other hand, it is often difficult to distinguish between the primary cryptotephra horizon (fall deposit) and secondary deposition. In Lake Suigetsu, many secondary depositional peaks of Aira Tn tephra (AT) glass shards were observed (Albert et al., 2024). These peaks were interpreted as deposition during higher energy events (e.g. floods). Based on this interpretation, the glass shards content is expected to identify "cryptic" events and reveal the frequency of small events. Thus, ACIAS can be a useful tool not only for tephrochronology but also for paleoenvironmental research.
This system consists of the full-slide image acquisition component and the glass shards content detection component. In the former component, the PiAutoStage (Steiner and Roonie, 2021), consisting of Raspberry Pi, Arduino Uno, electrical parts, and 3D printer parts, was used to acquire full-slide images. In the latter component, the You Only Look Once (YOLO) v5 algorithm was used to detect glass shards and determine the concentration of glass shards. Training data consisted of 15 tephra samples from the proximal outcrops and six from the YNH-P2 core (Watanabe et al., in prep.) from Yanohara Bog, Fukushima Prefecture. Thin sections were prepared using particles between 62 and 125 µm and a light-curing agent (refractive index of 1.55 after curing). Photo images were taken with a Raspberry Pi HQ camera attached to a polarizing microscope under open Nicol conditions. The grains were classified into 11 categories: volcanic glass shards were divided into five categories according to morphology, minerals were three categories, and others were three categories. Based on this model, the overall accuracy for detecting glass shards was 84.9%.
Next, we evaluated the reliability of ACIAS counting using the YNH-P2 core. The comparison of manual counting and ACIAS counting showed that ACIAS counting could be an alternative method to manual counting. The two visible tephras and five cryptotephras were detected in the YNH-P2 core (Watanabe et al., in prep.). The variation in concentration of glass shards using both methods was very similar, and ACIAS was able to detect cryptotephras.
The concentration of glass shards has mainly been used to detect the cryptotephra horizon. On the other hand, it is often difficult to distinguish between the primary cryptotephra horizon (fall deposit) and secondary deposition. In Lake Suigetsu, many secondary depositional peaks of Aira Tn tephra (AT) glass shards were observed (Albert et al., 2024). These peaks were interpreted as deposition during higher energy events (e.g. floods). Based on this interpretation, the glass shards content is expected to identify "cryptic" events and reveal the frequency of small events. Thus, ACIAS can be a useful tool not only for tephrochronology but also for paleoenvironmental research.