Extraction of stratigraphic exposures on visible images using a supervised machine learning technique

Abstract

As volcanic stratigraphy provides important information about volcanic activities, such as the eruption style, duration, magnitude, and their time sequences, its observation and description are fundamental tasks for volcanologists. Since outcrops are often obscured in nature, the first task would be identifying stratigraphic exposures in many cases. This identification/selection process has depended on humans and has difficulties in terms of time and effort consumption and in biases resulting from expertise levels. To address this issue, we present an approach that utilizes supervised machine learning with fine-tuning and forms the backbone to automatically extract the areas of stratigraphic exposures in visible images of volcanic outcrops. This study aimed to develop an automated method for identifying exposed stratigraphy. This method will aid in planning subsequent field investigations and quickly outputting results. We used U-Net and LinkNet, convolutional neural network architectures developed for image segmentation. Our dataset comprised 75 terrestrial outcrop images and their corresponding images with manually masked stratigraphic exposure areas. Aiming to recognize stratigraphic exposures in various situations, the original images include unnecessary objects such as sky and vegetation. Then, we compared 27 models with varying network architectures, hyperparameters, and training techniques. The highest validation accuracy was obtained by the model trained using the U-Net, fine-tuning, and ResNet50 backbone. Some of our trained U-Net and LinkNet models successfully excluded the sky and had difficulties in excluding vegetation, artifacts, and talus. Further surveys of reasonable training settings and network structures for obtaining higher prediction fidelities in lower time and effort costs are necessary. In this study, we demonstrated the usability of image segmentation algorithms in the observation and description of geological outcrops, which are often challenging for non-experts. Such approaches can contribute to passing accumulated knowledge on to future generations. The autonomous detection of stratigraphic exposures could enhance the output from the vast collection of remote sensing images obtained not only on Earth but also on other planetary bodies, such as Mars.