Hydrocarbon Identification Using Supervised Machine Learning: Seismic Image Segmentation and Multi-Attribute Analysis Validated by Production and Log Data

Abstract Accurate identification of hydrocarbon pockets remains crucial for maximizing reservoir recovery, minimizing exploration risk, and improving field development strategies. Recently, artificial intelligence (AI) applications in seismic interpretation have shown promise for supplementing traditional geophysical workflows. Yet, challenges persist, especially in integrating data from multiple sources — such as seismic attributes, well logs, and production history to create cohesive, predictive models. Historically, seismic data have been underutilized for direct fluid identification due to resolution limitations, interpretational subjectivity, and difficulty in incorporating labeled datasets into supervised learning workflows. In this study, we propose a dual-method supervised machine learning framework that combines seismic grayscale image segmentation with tabular seismic attribute classification to locate hydrocarbon-bearing zones in two mature fields. The first method involves seismic-derived grayscale images generated from multiple attributes. Each pixel's intensity, reflecting variations in lithology and fluid properties, is manually labeled through petrophysical interpretation linked to well control. These labeled images are segmented using supervised learning techniques to differentiate hydrocarbon zones from water zones within the seismic volume. At the same time, the second approach uses tabular seismic attributes where selected records are labeled as Hydrocarbon, Water, or Dense based on core-calibrated well logs and validated with production data. Multiple supervised algorithms, including decision trees, support vector machines, and ensemble models, are evaluated and optimized to classify the larger dataset and predict fluid distributions in both fields. Results from both methods demonstrate strong agreement with known producing intervals and reveal the capacity to detect subtle heterogeneities, including potential bypassed pay zones that are not visible in conventional interpretation. The hybrid approach enhances subsurface understanding and provides a scalable, repeatable AI framework validated by petrophysical data. By integrating visual and numerical data through machine learning, this work illustrates a promising path forward for advancing seismic-based hydrocarbon detection, especially in data-limited or geologically complex environments. The innovative aspect lies in the combined use of image-based and attribute-driven learning within a production-calibrated setting, offering a practical AI workflow that extends traditional interpretation and reservoir modeling capabilities.