5:15 PM - 7:15 PM
[MGI28-P01] Automatic Classification of the Sediment Lithology Using Microscopic Images of Smear Slides of Siliceous Sediment Cores from the Japan Sea
Keywords:smear slide, lithological classification, artificial intelligence (AI), siliceous sediment, the Japan Sea
With the recent development of artificial intelligence (AI), the introduction of AI into geological research has been expanded. In particular, automatic classification of grains using AI is almost on a practical level. Such technology can compensate for limited human resources and realize much faster data accumulation than ever before.
To enlarge the potential of AI, we tried applying the AI automatic classification technology to “an assemblage of grains”, whose variation is much larger than “a single grain”. Our purpose is to automate the lithological classification using smear slides, which is essential work in describing unconsolidated sediments. The manual observation of smear slides takes a lot of time, and the criteria of classification, to some extent, may differ among observers. Automatic classification using AI can lead to a cut down on human resources and the establishment of unified criteria for classification.
In this study, we used the Miocene deep-sea sediment cores in the Japan Sea. They show a simple lithological variation from clay to diatom ooze, which is suitable to the first step for the AI-based lithological classification. Through the Repository Core Re-Discovery Program (ReCoRD) ReC23-03: the Japan Sea Paleoceanography and Paleoclimatology during the Miocene, 536 smear slides were made using the sediment cores from International Ocean Discovery Program (IODP) and Ocean Drilling Program (ODP) Sites U1430, U1425, 794, and 795. The slides were scanned by a digital slide scanner, NanoZoomer S360 (Hamamatsu Photonics), and their digital images cut into ca. 0.9-square-mm tiles were used for construction of and classification by the AI models. By using 106 smear slides (38, 288 tiles) from U1430 Hole A, we constructed the AI models through a software, GSJ Particle Analyzer (Miyakawa et al., 2024). The models were trained to classify five lithological types: clay, diatomaceous clay, clayey diatom ooze, diatom ooze, and tephra.
The accuracy score of the AI classification model exceeded 90% for the validation data set from U1430 Hole A. Moreover, we constructed a method to predict the lithology of a smear slide based on the classification result of its tiles. The slide-level lithological classification at U1430 Hole A, which was used for the training data set, succeeded at 100%. Furthermore, the percentage of correct answers to U1430 Hole B, which was not used for the training data set, also reached 90%. The depth profiles of the AI classification result from two holes of U1430 show the possibility of inter-hole lithological correlation. After this, we will check the AI model using the other sites (U1425, 794, and 795) and examine the possibility of inter-site lithological correlation.
Through this study, we made the automatic lithological classification model specializing in siliceous sediments or the Japan Sea during the Miocene. In the future, the AI classification model covering the world’s ocean sediments might be realized if we can prepare a comprehensive training data set encompassing various lithological types, various sea areas, and various ages.
To enlarge the potential of AI, we tried applying the AI automatic classification technology to “an assemblage of grains”, whose variation is much larger than “a single grain”. Our purpose is to automate the lithological classification using smear slides, which is essential work in describing unconsolidated sediments. The manual observation of smear slides takes a lot of time, and the criteria of classification, to some extent, may differ among observers. Automatic classification using AI can lead to a cut down on human resources and the establishment of unified criteria for classification.
In this study, we used the Miocene deep-sea sediment cores in the Japan Sea. They show a simple lithological variation from clay to diatom ooze, which is suitable to the first step for the AI-based lithological classification. Through the Repository Core Re-Discovery Program (ReCoRD) ReC23-03: the Japan Sea Paleoceanography and Paleoclimatology during the Miocene, 536 smear slides were made using the sediment cores from International Ocean Discovery Program (IODP) and Ocean Drilling Program (ODP) Sites U1430, U1425, 794, and 795. The slides were scanned by a digital slide scanner, NanoZoomer S360 (Hamamatsu Photonics), and their digital images cut into ca. 0.9-square-mm tiles were used for construction of and classification by the AI models. By using 106 smear slides (38, 288 tiles) from U1430 Hole A, we constructed the AI models through a software, GSJ Particle Analyzer (Miyakawa et al., 2024). The models were trained to classify five lithological types: clay, diatomaceous clay, clayey diatom ooze, diatom ooze, and tephra.
The accuracy score of the AI classification model exceeded 90% for the validation data set from U1430 Hole A. Moreover, we constructed a method to predict the lithology of a smear slide based on the classification result of its tiles. The slide-level lithological classification at U1430 Hole A, which was used for the training data set, succeeded at 100%. Furthermore, the percentage of correct answers to U1430 Hole B, which was not used for the training data set, also reached 90%. The depth profiles of the AI classification result from two holes of U1430 show the possibility of inter-hole lithological correlation. After this, we will check the AI model using the other sites (U1425, 794, and 795) and examine the possibility of inter-site lithological correlation.
Through this study, we made the automatic lithological classification model specializing in siliceous sediments or the Japan Sea during the Miocene. In the future, the AI classification model covering the world’s ocean sediments might be realized if we can prepare a comprehensive training data set encompassing various lithological types, various sea areas, and various ages.