Adopting Grad-CAM for Deep Learning-Based Rock Classification Model

Rock classification is often conducted by inspecting microscopic thin section images optically, but there can be some ambiguity involved. Recently, various studies have demonstrated the effectiveness of deep learning models for image classification. Thus, this study proposed a deep learning-based rock classification model with convolutional neural network (CNN) to classify six types of igneous rocks from their polarizing microscope images. Using the prepared image datasets, an accurate classification model was developed based on training and validation. Further, we assessed the model's performance by applying gradient-weighted class activation mapping (Grad-CAM) and visualized each pixel's weight using a heatmap. The result from Grad-CAM exhibited that the trained model identified adequate features within images such as mineral grain boundaries and textures to determine rock types. Using Grad-CAM as an explainable tool, the proposed model was proven to be reliable for rock classification.