Sedimentary Rock Contour Images Synthesis Using Generative Networks based on Optimal Transport
Gloria M. Castaneda Campos, José Ismael De la Rosa Vargas, Gamaliel Moreno, José de Jesús Villa Hernández, Efrén González · 2024
Morphology is a descriptor of sedimentary rocks useful in earth science and economic geology. In the contour of the rock, there is information on its origin, transport, and deposition. The two most important parameters are general form and roundness. These parameters are obtained by algorithms that process images of outcrops or samples of them. Capturing these images is a difficult task due to inaccessibility or risk. On the other hand, to evaluate an algorithm, it is necessary to capture several phenomena since usually, a single phenomenon produces only one type of contour. In this work, we train a Wasserstein Generative Adversarial Network (WGAN) that synthesizes contour images of rocks with a wide range of morphology. The WGAN was trained with 5200 contour images of various phenomena such as ash and debris flows, lahars, pyroclastic falls, and ballistic. The database was classified into five roundness categories to improve balance. Additionally, a data augmentation technique was used to enhance the learning process. The quality and realism of the generated images were evaluated using KDE and MMD metrics. The results indicated that WGAN generates good-quality images while maintaining class diversity.