Assisted interpretation of thin sections and core samples with Computer Vision approaches: appraisal of Deep Learning workflows on operational datasets

@inproceedings{BOUZIAT_RM2021,
 abstract = { Computer Vision, a subfield of Artificial Intelligence dedicated to image processing, has known a major development in the last decade. Convolutional Neural Networks (CNN) and Deep Learning methodologies are increasingly used to facilitate image analysis in various domains, with applications such as automated medical diagnosis, facial recognition software or self-driving cars. These technologies are also raising growing interests in geology, as this discipline heavily relies on visual interpretation. However only few practical use cases on operational datasets are yet documented. Thus, in this work we appraise reference Deep Learning workflows to automate the interpretation of thin sections and core photographs from real-life geological studies. From this experience, we highlight good practices and recommendations for further use of these technologies in similar contexts. In a first appraisal, we use a dataset of 145 carbonate thin sections to assess object detection models with limited training. The thin sections come from the Alveolina Limestone formation in the Graus-Tremp basin (Spain) and contain 9 different species of microfossils with an unbalanced distribution. We train four state-of-the-art Deep Learning models to automatically spot, delineate and characterize these microfossils, using only 15 sections manually interpreted. The results on 130 other thin section images are then qualitatively assessed by expert geologists, and precisions and inference times quantitatively measured. The four models show good capabilities in detecting and categorising the microfossils. However differences in accuracy and performance are underlined and discussed. In a second appraisal, we use actual core images to evaluate the potential of supervised classification models in extrapolating human interpretation from a few segments to the entire wells. The dataset comes from an expedition of the International Ocean Drilling Program in the Gulf of Corinth and corresponds to 3 drilling sites in the Gulf. The images were interpreted by an expert in terms of 17 facies associations. We carry out a Transfer Learning methodology to generate and compare multiple models with different CNN architectures and training strategies. After fine tuning, the best model trained on 20% of the cores length is applied on the remaining 80% with a Top-1 accuracy of 70% and a Top-3 one of 90%. From this study, we draw guidelines for comparable projects and open discussions on the geological relevance of some differences between human and machine interpretation. On the whole, this work illustrates the promises of Computer Vision and Deep Learning approaches in geology, but also the challenges to face in their application to operational datasets. },
 author = { Bouziat, Antoine AND Lechevallier, Antoine AND Koroko, Abdoulaye AND Lecomte, Jean-Claude AND Feraille, Mathieu AND Desroziers, Sylvain },
 booktitle = { 2021 RING Meeting },
 publisher = { ASGA },
 title = { Assisted interpretation of thin sections and core samples with Computer Vision approaches: appraisal of Deep Learning workflows on operational datasets },
 year = { 2021 }
}