MACHINE LEARNING-BASED TAXONOMIC CLASSIFICATION OF CARBONATE SKELETAL GRAINS

Assessing the types and abundances of skeletal grains in carbonate rocks across geologic timescales provides important information to reconstruct past climatic, depositional, and ocean chemistry conditions. However, analysis of carbonate skeletal grains is often obscured by their relatively poor preservation due to fragmentation, dissolution, and recrystallization during diagenesis, making them hard to identify with the human eye. This is compounded by the costly and time-consuming petrographic analysis of carbonate microfacies and microfossil identification. Previous studies have shown the power of machine learning algorithms, specifically Deep Convolutional Neural Networks (DCNN), as an ideal interdisciplinary solution to the problem of petrographic analysis. This study aims to develop an efficient and cost-effective AI-driven tool to classify and predict types and abundances of skeletal grains from petrographic images. The experimental design compares the performances of three different machine learning algorithms (support vector machine (SVM), DCNN, and deep transfer learning using ResNet152 architecture) when tasked with identifying 15 target classes of skeletal grains in carbonate rocks. Our dataset included approximately 1500 training images, with each target class being equally represented. Out of the tested algorithms, DCNN performs better than traditional machine learning (SVM) with the transfer learning method achieving the highest accuracy (98%). Overall, the skeletal grain classes with the highest and lowest accuracies were sponges and cephalopods, respectively. Furthermore, when tested against human identification (experienced geologists and undergraduate geoscience students) the DCNN transfer learning method proved to be both more accurate and faster. This study shows that deep learning is capable of matching and surpassing human accuracy and speed when identifying carbonate microfacies. Future work will ascertain the accuracies of DCNN-based object detection using the same 15 target classes.