Mapping of Karst Using Deep Learning Method: A Case Study in Central Luconia, Sarawak

Abstract This study focuses on leveraging deep learning method, specifically Convolutional Neural Network (CNN), to enhance seismic data interpretation for mapping karst bodies in Central Luconia carbonate platform, offshore Sarawak, Malaysia. The research aims to improve the accuracy of seismic data interpretation through detailed mapping of karst geobodies and evaluate the risks associated with karsts. The CNN architecture is designed to effectively capture spatial patterns and features pertinent to karst detection. The model undergoes supervised learning process, wherein it is trained using a labeled dataset. Throughout the training phase, the model learns to discern characteristics patterns and textures associated with karst features by iteratively adjusting its parameters to minimize prediction errors. The model is validated using a separate dataset to evaluate its performance and generalization ability. Hyperparameters of the model are then fine-tuned based on the validation results to enhance its accuracy and robustness. The CNN accurately identifies and maps sinkholes and dendritic karsts, significantly reducing manual interpretation time and effort. The results are validated in 3D cube but it is presented on map view in this paper. The maps provide insights into the distribution, density, and morphology of the karst features which are essential for assessing drilling risks. By highlighting the different types of karst within a carbonate platform, the CNN aids in planning and mitigating risks associated with drilling, such as collapses and fluid loss, ensuring safer and more efficient operations. This advanced approach enhances the understanding of the geological significance of karst formations and supports effective scanning for hydrocarbon reservoirs as well as future planning in CO2 storage selection, if needed.