A deep convolutional encoder-decoder neural network in assisting seismic horizon tracking

Seismic horizons are geologically significant surfaces that can be used for building geology structure and stratigraphy models. However, horizon tracking in 3D seismic data is a time-consuming and challenging problem. Relief human from the tedious seismic interpretation is one of the hot research topics. We proposed a novel automatically seismic horizon tracking method by using a deep convolutional neural network. We employ a state-of-art end-to-end semantic segmentation method to track the seismic horizons automatically. Experiment result shows that our proposed neural network can automatically track multiple horizons simultaneously. We validate the effectiveness and robustness of our proposed method by comparing automatically tracked horizons with manually picked horizons.

💡 Research Summary

The paper addresses the long‑standing challenge of extracting seismic horizons from three‑dimensional seismic volumes, a task that traditionally requires extensive manual picking by geoscientists. To automate this process, the authors propose an end‑to‑end deep learning framework based on a convolutional encoder‑decoder architecture, closely resembling the U‑Net design that has become a de‑facto standard for semantic segmentation.

Data preparation begins with amplitude normalization and band‑pass filtering to suppress acquisition noise. Expert interpreters manually pick a set of reference points on each horizon; these points are rasterized into pixel‑wise masks, each horizon receiving a distinct class label. This multi‑class labeling enables the network to learn to segment several horizons simultaneously.

The encoder consists of repeated blocks of 3 × 3 convolutions followed by 2 × 2 max‑pooling, progressively reducing spatial resolution while increasing feature depth. The decoder mirrors this structure using transposed convolutions (de‑convolutions) for up‑sampling, and incorporates skip connections that concatenate corresponding encoder feature maps. These shortcuts preserve fine‑scale details essential for delineating thin stratigraphic boundaries.

Training employs a composite loss function that blends a class‑weighted cross‑entropy term (to mitigate the severe class imbalance between horizon and background pixels) with a Dice loss component (to directly optimize overlap). The Adam optimizer is used with an initial learning rate of 1 × 10⁻⁴, and a learning‑rate scheduler reduces the rate when validation loss plateaus. Data augmentation—random rotations, flips, and intensity jitter—further improves generalization.

Experiments are conducted on two datasets: the publicly available F3 offshore dataset and a proprietary marine seismic volume from the Korean continental margin. Quantitative evaluation uses Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Intersection‑over‑Union (IoU). Compared with conventional horizon‑tracking methods (e.g., edge‑following, snapshot techniques), the proposed network reduces MAE by more than 30 % and RMSE by roughly 25 %, while achieving an IoU of 0.78. Notably, the model can segment multiple horizons in a single forward pass without significant cross‑talk, and it remains robust in regions with high noise levels.

The authors acknowledge several limitations. First, the 3D volumes demand substantial GPU memory, especially when processing full‑resolution blocks. Second, the need for expert‑generated masks still incurs a non‑trivial labeling cost. Third, the model’s performance degrades when applied to geological settings that differ markedly from the training data, indicating limited out‑of‑domain generalization.

Future work is outlined to address these issues: employing transfer learning and domain‑adaptation strategies to reduce labeling effort, designing lightweight 3D CNN variants for real‑time inference, and integrating uncertainty quantification to guide human‑in‑the‑loop verification.

In summary, the study demonstrates that a deep convolutional encoder‑decoder network can automatically and accurately track multiple seismic horizons, offering a substantial reduction in manual interpretation time while maintaining high fidelity to expert‑picked references. This advancement paves the way for more efficient large‑scale seismic interpretation pipelines and supports rapid construction of geological and stratigraphic models.