17:15 〜 18:30
[SCG50-P05] 機械学習による鉛直重力偏差データを用いた断層の分類
キーワード：機械学習、重力偏差、断層
Changes in the density structure of the subsurface due to fault movement have been observed as gravity anomalies or gravity gradients. Gravity gradients are more sensitive to the change in the subsurface density structure compare to gravity anomalies, and are therefore widely used to estimate subsurface structures such as faults (e.g., Thompson, 1982; Copper and Cowan, 2006). Subsurface normal faults can be seen because the low-density layer is the upper plate. On the other hand, subsurface reverse faults are difficult to determine their structures such as their dips because the high-density layer is the upper plate.
Machine learning is one of the artificial intelligence (AI) techniques, which is used to filter out spam mail and automated driving technology, for example. It can classify data or pictures which are difficult to judge visually. In this study, we applied the machine learning method to classify faults, which have been done by researchers’ visual judgements or roadside surveys, from vertical gravity gradients data obtained from fault models. For machine learning, we used Convolutional Neural Network (CNN), which is useful for image classification. The gravity gradient data was obtained by assuming fault models using the Talwani’s method (Talwani, 1965), and the data were subjected to random noise with 0-50 Eötvös. The data was imaged and the image size was unified as 256×256 pixel for supervised learning. Here, we classified normal and reverse faults with different fault dips. At first, the normal faults were classified by fault dips of 45°, 60° and 75°. 200 images with a fault dip of 60° were classified. The value of the loss function was approximately 0.10, which yields the accuracy of 97%. Next, the reverse faults were classified by fault dips of 15°, 30°, 45° and 60°. 200 images with a fault dip of 30° were classified. The value of the loss function was approximately 0.15, which yields the accuracy of 88.5%. The values of the loss functions were approximately 0.10 and 0.15, indicating that classifications of the fault dip were done with high accuracy. This indicates that the classification of normal and reverse fault using the machine learning method is possible. However, the accuracy of the reverse faults was 10% lower than that of the normal faults. This indicates that the classification of reverse faults is more difficult than that of normal faults. We will apply this method to actually observed data in the future study. However, since our method is based on linear fault model, we need to use actual data that can be approximated to be linear faults.
Machine learning is one of the artificial intelligence (AI) techniques, which is used to filter out spam mail and automated driving technology, for example. It can classify data or pictures which are difficult to judge visually. In this study, we applied the machine learning method to classify faults, which have been done by researchers’ visual judgements or roadside surveys, from vertical gravity gradients data obtained from fault models. For machine learning, we used Convolutional Neural Network (CNN), which is useful for image classification. The gravity gradient data was obtained by assuming fault models using the Talwani’s method (Talwani, 1965), and the data were subjected to random noise with 0-50 Eötvös. The data was imaged and the image size was unified as 256×256 pixel for supervised learning. Here, we classified normal and reverse faults with different fault dips. At first, the normal faults were classified by fault dips of 45°, 60° and 75°. 200 images with a fault dip of 60° were classified. The value of the loss function was approximately 0.10, which yields the accuracy of 97%. Next, the reverse faults were classified by fault dips of 15°, 30°, 45° and 60°. 200 images with a fault dip of 30° were classified. The value of the loss function was approximately 0.15, which yields the accuracy of 88.5%. The values of the loss functions were approximately 0.10 and 0.15, indicating that classifications of the fault dip were done with high accuracy. This indicates that the classification of normal and reverse fault using the machine learning method is possible. However, the accuracy of the reverse faults was 10% lower than that of the normal faults. This indicates that the classification of reverse faults is more difficult than that of normal faults. We will apply this method to actually observed data in the future study. However, since our method is based on linear fault model, we need to use actual data that can be approximated to be linear faults.