3D seismic multi-horizon reconstruction via dual-branch network with geological constraints

Abstract

Accurate 3D seismic multi-horizon reconstruction is pivotal for structural interpretation and reservoir characterization. Existing deep learning methods predominantly rely on convolutional neural networks (CNNs) to predict relative geological time (RGT) volumes from seismic data. However, CNN-based approaches typically face challenges in capturing long-range stratigraphic relationships and often struggle to accurately resolve complex geological structures, such as faults, especially when labeled data are sparse. To address these challenges, we propose a dual-branch deep neural network (DB-Net) that synergistically integrates CNN and transformer architectures for geologically constrained RGT reconstruction. The CNN branch extracts local spatial features through hierarchical 3D convolutions, while the transformer branch models global stratigraphic dependencies via multi-head self-attention mechanisms. A spatially adaptive gated fusion module is designed to dynamically combine complementary features from both branches based on local geological complexity. Unlike purely data-driven methods, our approach incorporates explicit geological knowledge through a multi-component loss function that enforces horizon continuity in conformable regions while preserving discontinuities at fault boundaries. Sparse horizon labels and fault attributes are utilized as structural priors to guide network training. Furthermore, an auxiliary feature-level domain adaptation module mitigates the distribution gap between synthetic training data and field seismic volumes, facilitating effective knowledge transfer despite limited field annotations. Extensive experiments on both synthetic and field datasets demonstrate that DB-Net delivers state-of-the-art reconstruction accuracy, significantly improving performance in geologically complex regions.