Opening the black box: explainable AI for automated bioturbation analysis in cores and outcrops

The application of machine learning (ML) in geology has gained significant momentum over the past decade. Given the importance of visual interpretation in geological tasks such as lithological classification and trace fossil identification, automating these processes through ML has offered substantial benefits to both researchers and industry professionals, including reduced human error and increased efficiency. While machine learning methodologies commonly yield highly accurate results, the opacity of decision-making processes can present challenges, as they often provide "black box" responses. To address this issue and improve user trust and interpretability, the integration of Explainable Artificial Intelligence (XAI) has emerged as a promising solution. XAI techniques provide visual explanations, typically in the form of heat maps, indicating the specific regions within an image that the model prioritized in reaching its classification. Despite its growing relevance, the application of XAI within the geosciences remains in its infancy. This study evaluates the effectiveness of XAI in the context of bioturbation intensity classification using a pre-trained deep-learning model. The model classifies core and outcrop images into three categories: (1) unbioturbated, (2) moderately bioturbated, and (3) intensely bioturbated. XAI visualizations reveal strong alignment with an experienced ichnologist's interpretations. Our findings highlight the utility of XAI in enhancing the transparency, and reliability, as well as ability to conduct a consistent, rapid assessment of images to increase efficiency of ML-based geological interpretation. In addition to its potential for improving confidence, XAI may serve as a valuable educational tool, especially in specialized domains such as ichnology, wherein expert knowledge is typically limited in academic and industry settings.