The Importance of Blending Different Data Types to Train Machine Learning Classifiers for Sedimentary Structure Detection

Summary This study demonstrates that, by adding sketched interpretation data to photographic datasets of geological outcrops, we can improve the quality of sedimentary structure classification, even for smaller volume datasets. We blended raw outcrop photos with sketches of sedimentary structures to use as input into a Convolutional Neural Network (CNN) model which will predict and classify certain geological structures. The use of CNN can make geological classification easier for us by assisting in the collection of geological observations in seconds. Our work shows that the CNN model misclassified various geological features when trained only with one type of data (outcrop photos or geological sketches). The efficacy and novelty of the system described in this paper lies in the blending of two different data types (both outcrop photographs and geological sketches) when training our CNN model for geological feature detection. The use of the blended dataset in learning, at an optimal balance between sketches and outcrop photos (from 40% to 67% sketch proportion in the training dataset), results in fewer misclassifications and higher test accuracy of the model predictions of the sedimentary structures.