Comparative analysis of traditional machine learning and deep learning for seismic facies classification using F3 data from the Dutch North Sea

Seismic facies classification (SFC) plays a critical role in subsurface interpretation, and both traditional machine learning (ML) and deep learning (DL) approaches have been increasingly applied to automate this task. This study presents a comparative analysis of ML and DL methods using the publicly available F3 seismic dataset from the Dutch North Sea. A single inline was manually labeled for both binary and multiclass (12-class) classification scenarios. Fifty seismic attributes were extracted and refined using feature selection techniques—mutual information (MI), analysis of variance (ANOVA), recursive feature elimination (RFE), and random forest (RF)—to train nine ML classifiers. In parallel, a convolutional neural network (CNN) was trained directly on raw seismic data using the same labeled inline. The results show that RFE and RF produced the most compact and effective feature subsets (19–20 attributes), with spectral and geometric attributes contributing most to classification performance. Most traditional ML models achieved over 85% accuracy in the binary classification task, although performance declined in the multiclass scenario. Notably, k-nearest neighbors (KNN), random forest (RF), multilayer perceptron (MLP), and support vector machine (SVM) maintained relatively strong performance across both scenarios. The CNN outperformed all ML models, achieving near-perfect accuracy (0.99 binary and 0.98 multiclass) and demonstrating superior generalization on unseen data, despite minor overfitting in the 12-class case. While traditional ML approaches remain suitable for resource-limited environments, the results highlight the strong potential of DL methods for handling complex SFC tasks when sufficient labeled data are available.