14:30 〜 14:45
[HCG23-04] CNN画像判別モデルを用いたspaced stratificationの粒子配列解析
キーワード：トラクションカーペット、堆積構造、機械学習、CNN
Spaced stratification (SS) is composed of multiple bands typically exhibiting inverse grading with sharp and near-horizontal erosional surfaces. SS is observed in deposits of various processes, such as turbidites, hyperconcentrated flow deposits, and pyroclastic gravity flow deposits. Several studies have indicated that traction carpets or migration of low-amplitude bedforms can make SS. However, its formation mechanism has not been clarified because no experimental study succeeded in forming the structure. Hence, this study examined the diversity of individual bands of SS to understand its formation processes.
We studied the Upper Cretaceous Izumi Group in the northern part of Tokushima Prefecture and the southern part of Hyogo Prefecture, Southwest Japan. These areas are characterized by sand-dominated alternating beds of sandstone, mudstone, and conglomerate. SS are found in coarse-grained sandstone beds.
This study investigated SS using an image analysis model with the Convolutional Neural Network (CNN). The ResUnet50 was employed as an architecture of the CNN. The model was trained by teacher image datasets: cross-section images and labeled images artificially traced with grains and matrices. The trained CNN model was applied to the scanned cross-section images of SS to classify grains and matrices. Then, we measured and analyzed the characteristics of grains in mm-scale by using output images that were predicted by the trained CNN model. Using this model, it takes only a few tens of minutes to extract grains in an image of ten or more centimeters square.
Based on the grain data measured using CNN, several statistical values (e.g., the mean and standard deviation of grain size and grain long-axis orientations) characterizing the textures of the bands in SS were calculated. The degree of inverse- or normal-grading was also quantified. The k-means cluster analysis was performed based on those statistical values. BIC determined the adequate number of clusters. As a result, the bands in SS were classified into two clusters. Type 1 bands consisted of coarser grains and showed high (>30°) imbrication. Type 1 bands were also characterized by a large vertical variation in grain size and imbrication angle. In contrast, Type 2 bands consisted of finer grains and were represented by a small vertical variation in grain size and imbrication angle, although there was diversity in the mean imbrication angle. A single SS may contain both types of bands, but Type 1 bands were always below Type 2 bands.
Grain textural analysis using CNN revealed two types of bands in the SS. There are a few similarities between grain fabrics revealed in this study and fabrics of the experimental deposits of bedform or debris flow. However, the relationship between the formation processes of those deposits is uncertain, and whether these are formed by different processes or gradual changes in gravity flow conditions is still unclear; nonetheless, the result of this study implies that multiple processes can produce SS. It is expected to realize the formation process of the SS by comparing grain fabrics of this result and experimental deposits.
We studied the Upper Cretaceous Izumi Group in the northern part of Tokushima Prefecture and the southern part of Hyogo Prefecture, Southwest Japan. These areas are characterized by sand-dominated alternating beds of sandstone, mudstone, and conglomerate. SS are found in coarse-grained sandstone beds.
This study investigated SS using an image analysis model with the Convolutional Neural Network (CNN). The ResUnet50 was employed as an architecture of the CNN. The model was trained by teacher image datasets: cross-section images and labeled images artificially traced with grains and matrices. The trained CNN model was applied to the scanned cross-section images of SS to classify grains and matrices. Then, we measured and analyzed the characteristics of grains in mm-scale by using output images that were predicted by the trained CNN model. Using this model, it takes only a few tens of minutes to extract grains in an image of ten or more centimeters square.
Based on the grain data measured using CNN, several statistical values (e.g., the mean and standard deviation of grain size and grain long-axis orientations) characterizing the textures of the bands in SS were calculated. The degree of inverse- or normal-grading was also quantified. The k-means cluster analysis was performed based on those statistical values. BIC determined the adequate number of clusters. As a result, the bands in SS were classified into two clusters. Type 1 bands consisted of coarser grains and showed high (>30°) imbrication. Type 1 bands were also characterized by a large vertical variation in grain size and imbrication angle. In contrast, Type 2 bands consisted of finer grains and were represented by a small vertical variation in grain size and imbrication angle, although there was diversity in the mean imbrication angle. A single SS may contain both types of bands, but Type 1 bands were always below Type 2 bands.
Grain textural analysis using CNN revealed two types of bands in the SS. There are a few similarities between grain fabrics revealed in this study and fabrics of the experimental deposits of bedform or debris flow. However, the relationship between the formation processes of those deposits is uncertain, and whether these are formed by different processes or gradual changes in gravity flow conditions is still unclear; nonetheless, the result of this study implies that multiple processes can produce SS. It is expected to realize the formation process of the SS by comparing grain fabrics of this result and experimental deposits.