Geologists need help classifying microscope rock images of sigma clasts; a type of mantled porphyroclasts widely used as kinematic indicators in rocks. Knowledge about the shear sense of sigma clast during formation (either CCW or CW shearing) gives insights into rock formation history. This work reports on early investigation of machine learning techniques for automatic detection and classification of sigma clasts and their rotation from photomicrographs. Convolutional Neural Networks (CNNs) are used to extract and leverage defining features of sigma clasts, such as shape, color, texture, and tail direction to improve accuracy. We leverage existing models that are pre-trained on very large collections of images, and use transfer learning techniques to apply them to microstructure images. We used YOLOv3 to identify different sigma clasts in a given image. We also experimented with other large pre-trained models such as ResNet50, VGG19, InceptionV3 with two additional layers trained specifically on our dataset. In order to facilitate exploration of different models with different settings, we are developing a computational experimentation environment to visualize different CNN network layers, classification heatmaps, and comparative metrics. Finally, since models perform better when more data are available, we are developing a web application to collect additional data from geoscientists and incentivize their participation in open science. The website allows researchers to upload images of rock microstructures, showing them the classification of the images based on the best models available, and allows them to correct any errors which can be used to improve the models.
