Unveiling Subsurface Pinnacle Reefs Using Seismic Attributes and Machine Learning-Based Image Segmentation of 3D Seismic Data

Abstract Pinnacle reefs on carbonate platforms are crucial geological features that function as both excellent reservoirs and vital paleo-ecological habitats. However, their subtle and complex geometries pose a significant challenge for accurate extraction and quantification as geobodies using seismic attributes alone. To overcome this limitation, this study introduces a novel machine learning-based image segmentation approach for detecting and extracting pinnacle reefs as geobodies from 3D seismic data, enabling more accurate and detailed statistical and geological analysis. The proposed workflow begins with seismic data conditioning, utilizing a 3D deep convolutional model to effectively suppress noise and enhance reflector continuity. Next, spectral decomposition techniques are applied to improve the visibility of carbonate reefs and buildups. A supervised image segmentation model is then trained on manually labeled reef and non-reef features, using RGB-blended spectral decomposition slices from the Poseidon 3D volume. Following training, the model is applied to the entire Oligocene-Miocene carbonate sequence, predicting similar features with high accuracy, which are extracted as geobodies that are subjected to statistical analysis based on their volume and vertical extent. Initial analysis using conventional seismic attributes, such as variance, edge evidence, 3D curvature, and dip illumination, revealed the presence of pinnacle reefs in the Poseidon 3D volume (Courtesy of Geosciences Australia). However, extracting these features individually proved to be time-consuming and error-prone process, particularly for smaller or less distinct features. In contrast, the image segmentation-based machine learning model successfully identified both major and minor pinnacle reefs, including those that were not clearly visible through attribute analysis alone. Trained on a limited number of time slices with labeled examples, the model achieved an accuracy of over 87% when predicting reef features across the entire seismic volume. Once predicted, the pinnacle reefs were extracted as discrete geobodies and analyzed for their height and volume, enabling a consistent and automated workflow for statistical quantification. Compared to traditional attribute-based methods, the segmentation approach significantly reduced manual effort while improving detection completeness. The extracted geobodies provided valuable insights for both stratigraphic interpretation and prospectivity assessment. These findings support the integration of image segmentation and machine learning into standard seismic interpretation workflows, particularly in carbonate settings where complex morphologies challenge conventional methods. This study introduces a novel approach to seismic interpretation, leveraging image segmentation to automate the detection and extraction of pinnacle reefs. By training a machine learning model on RGB-blended spectral decomposition attributes, the workflow achieves accurate identification of reef features and conversion into measurable geobodies. This method overcomes the limitations of traditional seismic attribute techniques, enabling efficient and repeatable extraction of subtle geological features in stratigraphically challenging environments.