Supervised Seismic Facies Classification in Ataga Field of Niger Delta Basin using a Thalweg Tracker and Deep Learning Approach

Seismic facies classification remains a critical yet challenging task in subsurface characterization due to the complexity of depositional systems and limitations of manual interpretation. This study presents a supervised deep learning workflow for seismic facies prediction using a dip-steered median filtered (DSMF) seismic volume. Ten facies label sets, representing minima and maxima responses, were derived from unsupervised vector quantization (UVQ) clustering and subsequently refined using Thalweg tracking, with 1,500 samples per class. These labels, together with the DSMF volume, were used to extract 3D cubelets (41×41×3) for training convolutional neural networks (CNNs). Three architectures, LeNet, ResNet18, and SimpleNet, were implemented and evaluated based on classification performance and geological consistency of predicted facies volumes. Among the tested models, LeNet demonstrated superior performance, producing laterally continuous and stratigraphically consistent facies distributions that closely follow the geometry of seismic reflectors. The predicted facies exhibit clear differentiation between sand-prone, shale-dominated, and heterolithic units. Validation using gamma-ray (GR) and density (RHOB) logs confirms a strong correspondence between predicted facies and lithological variations, with sand intervals corresponding to low GR responses and shale intervals to high GR values. Minor limitations include partial merging of polarity-based facies classes; however, overall depositional patterns remain well preserved. The results highlight the effectiveness of a DSMF-driven LeNet approach for reliable and geologically meaningful seismic facies classification.