2024 Volume 82 Issue 1-2 Pages 3-26
This study developed a methodology to detect sand particles automatically from scanned images of cross sections of sandstones using deep learning image analysis. First, we prepared 1589 manually traced images of 256×256 pixels taken from the sandstone specimens of the Cretaceous Izumi and Himenoura Groups as teacher images, and the supervised training of the convolutional neural network was conducted to produce an image segmentation model. The trained model was applied to the test image that was independent of training images, and achieved 90.7% in accuracy of pixel classification. In addition, the results of grain-size analysis of the segmented image agreed well with those of the test image. Next, the trained model was applied to 10 polished sections of turbidite sandstones obtained from the Upper Cretaceous Izumi Group distributed in the southern part of Awaji Island, Japan, to examine the diversity of parallel lamination in turbidites. All images were subdivided at every 0.5 mm vertical intervals, and 1 cm-thick analytical window was set in vertical profiles of mean and standard deviation calculated at every subdivision to obtain statistical values characterizing features of parallel lamination. As a result, statistical analysis of grain-size distribution in parallel laminated sandstones implied that the obvious (Type 1) and unobvious (Type 2) parallel lamination exhibit different characteristics. In Type 2, mean and standard deviation of grain-size distribution fluctuated periodically. In contrast, Type 1 is mainly composed of fine-grained sediments and irregularly intercalates coarse-grained layers. In the future studies, quantitative comparison of these characteristics with flume experiments or numerical calculation will provide understandings of the formation mechanisms of the parallel lamination.
