Hybrid Context-Fusion Attention (CFA) U-Net and Clustering for Robust Seismic Horizon Interpretation

TL;DR

The paper tackles the challenge of robust seismic horizon interpretation under complex geology and sparse annotations by introducing a Context-Fusion Attention (CFA) U-Net that integrates semantic, spatial, and edge information into its attention gates. It systematically compares five U-Net variants and a CFA U-Net across two datasets, using a hybrid workflow that couples DBSCAN-based post-processing with orthogonal (inline and crossline) prediction fusion. The CFA U-Net, especially with all three heads (semantic, spatial, and Sobel), achieves state-of-the-art IoU and low MAE on the Mexilhão field, and high horizon surface coverage under sparse sampling on the F3 block, illustrating a strong precision–recall balance. The results demonstrate that context-aware attention and density-based refinement substantially improve horizon continuity and geological plausibility, making the approach practical for challenging seismic interpretation tasks.

Abstract

Interpreting seismic horizons is a critical task for characterizing subsurface structures in hydrocarbon exploration. Recent advances in deep learning, particularly U-Net-based architectures, have significantly improved automated horizon tracking. However, challenges remain in accurately segmenting complex geological features and interpolating horizons from sparse annotations. To address these issues, a hybrid framework is presented that integrates advanced U-Net variants with spatial clustering to enhance horizon continuity and geometric fidelity. The core contribution is the Context Fusion Attention (CFA) U-Net, a novel architecture that fuses spatial and Sobel-derived geometric features within attention gates to improve both precision and surface completeness. The performance of five architectures, the U-Net (Standard and compressed), U-Net++, Attention U-Net, and CFA U-Net, was systematically evaluated across various data sparsity regimes (10-, 20-, and 40-line spacing). This approach outperformed existing baselines, achieving state-of-the-art results on the Mexilhao field (Santos Basin, Brazil) dataset with a validation IoU of 0.881 and MAE of 2.49ms, and excellent surface coverage of 97.6% on the F3 Block of the North Sea dataset under sparse conditions. The framework further refines merged horizon predictions (inline and cross-line) using Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to produce geologically plausible surfaces. The results demonstrate the advantages of hybrid methodologies and attention-based architectures enhanced with geometric context, providing a robust and generalizable solution for seismic interpretation in structurally complex and data-scarce environments.

Figures (12)

• Figure 1: The two 3D seismic cubes utilized in the experiments are (a) the F3-Block in the Dutch North Sea and (b) the Mexilhão field in the Santos Basin, Brazil.
• Figure 2: Geological complexity of the Mexilhão field. (a) A representative seismic amplitude slice. (b) The pseudo-relief attribute highlights structural features. (c) A manual fault interpretation overlaid on the pseudo-relief. (d) The seismic amplitude slice with both the fault interpretation and the target horizon highlighted.
• Figure 3: Schematic of the Attention U-Net architecture. The model processes a 2D seismic image patch (left) through a symmetric encoder-decoder network to produce a probability mask (right). The encoder (contracting path) uses convolutional (orange), dropout (green), and max-pooling (red) blocks to extract hierarchical features. The decoder (expansive path) uses upsampling (purple) and convolutional blocks to restore spatial resolution. The key components are the attention gates ($\Omega$) on the skip connections, which adaptively re-weight features before they are fused via concatenation ($||$).
• Figure 4: Diagram of Context-Fusion Attention gate mechanism
• Figure 5: The DBSCAN filtering workflow. (a) The raw 3D probability volume is predicted by the neural network, with points colored by their two-way travel time (TWT). (b) DBSCAN groups the data into distinct clusters, identified by different categorical colors. (c) The final result after filtering for the largest point cluster, with the TWT-based color map reapplied to the cleaned horizon.