PaperlayerView exact revision on arXiv

Source-linked AI summary

Voxel-based 3D Facies Segmentation from Seismic Data: A Comparative Study

Duc-Thanh Pham, Minh-Tan Pham, Anh Nguyen, Van Nguyen

arXiv:2608.14058v1cs.CVcs.AI

TL;DR

Existing 2D slice-based approaches do not preserve the volumetric continuity of seismic data. This paper benchmarks voxel-based 3D architectures on public datasets under standardized settings, finding that 3D U-Net is the strongest overall and most computationally efficient baseline.

  • Problem

    Existing 2D slice-based seismic segmentation studies process volumetric data independently, ignoring inter-slice spatial correlations and volumetric continuity.

  • Method

    The paper benchmarks CNN-, transformer-, and state-space-based voxel architectures on the Netherlands F3 and Parihaka datasets using standardized experimental settings.

  • Results

    3D U-Net achieved the best overall performance on both datasets while also having the lowest computational and memory cost among evaluated methods.

  • Takeaways & Limitations

    CNN-based 3D U-Net is a strong, reliable baseline for volumetric seismic facies segmentation with limited labeled data.

  • Takeaways & Limitations

    The Parihaka benchmark uses only training regions because official labels for some test regions are unavailable.

Abstract

from arXiv · show

Seismic facies segmentation has emerged as a significant challenge in geophysics, requiring robust methods and systems to effectively identify geologically analogous facies with limited labeled data. Although existing studies have shown promising results in 2D facies segmentation, they often preprocess the original 3D seismic volumes into sets of 2D slices, typically the inline and crossline directions, and treat this problem as a purely 2D segmentation task. This simplification introduces discontinuities across slices and fails to preserve the spatial and structural continuity in 3D seismic data, thus limiting the model's ability to learn coherent geological patterns. In this work, we present a comparative and reproducible benchmark for voxel-based 3D seismic facies segmentation, built upon publicly available seismic volumes including the Netherlands F3 and the Parihaka datasets, with standardized data splits and evaluation metrics. By evaluating the three representative families of modern 3D segmentation architectures, we establish strong baseline results that highlight the potential and remaining challenges for future research in this domain.

I. INTRODUCTION

Seismic facies segmentation is important for seismic interpretation but remains constrained by reliance on 2D models for inherently 3D data. This paper addresses that gap with a comparative, reproducible benchmark of voxel-based 3D architectures on public seismic datasets.

  • Motivation: Seismic facies segmentation supports resource exploration, reservoir characterization, and subsurface infrastructure development.Facies are characterized using seismic attributes including amplitude, frequency content, waveform shape, and reflector geometry and continuity.
  • Problem: Existing studies commonly use 2D deep learning models despite the inherently 3D volumetric nature of seismic data.CNN-based architectures such as SegNet, U-Net, and DeepLabv3 are used mainly because of practical constraints and limited data availability.
  • Contribution: The benchmark compares voxel-based 3D-UNet, transformer-based UNETR, and state-space-model-based SegMamba.These represent three families of voxel-based deep-learning architectures.
  • Contribution: Experiments use the public Netherlands F3 and Parihaka seismic datasets with standardized data splits to promote reproducibility and comparability.The authors release code, data splits, and trained models to establish a reproducible and extensible baseline.

II. METHODOLOGY

The methodology defines a reproducible and comparable benchmark, specifying the datasets, model architectures, evaluation metrics, and implementation details used in the study.

  • II. METHODOLOGY: The benchmark specifies datasets, model architectures, evaluation metrics, and implementation details to support reproducibility and comparability.These experimental settings are intended to enable future studies to build directly upon the benchmark.

A. Datasets and Metrics

The benchmark uses the publicly available Netherlands F3 and Parihaka 3D seismic datasets, preserving full seismic cubes with facies labels. Evaluation uses standardized five-fold cross-validation with an 80/20 partition and the Dice coefficient for volumetric segmentation accuracy.

  • Datasets: The study evaluates facies segmentation on the publicly available Netherlands F3 Block and Parihaka 3D seismic datasets, both containing full seismic cubes with facies labels.The labels derive from interpreted horizons or well-based stratigraphic units.
  • Datasets: Parihaka contains complex geology, including varied faults and distinct gas hydrate deposits, and provides six geological facies classes.The survey was conducted offshore Taranaki, New Zealand, with labels provided by Chevron USA.
  • Data preparation: The benchmark adopts five-fold cross-validation with an 80/20 partition to support fair evaluation and statistical significance testing across independent folds.This protocol follows a standardized evaluation strategy established in prior seismic facies interpretation work.
  • Data preparation: Seismic volumes are uniformly sampled into fixed-size 128×128×128-voxel sub-volumes, using mirror padding and a stride of 64 during splitting.These choices address the limited spatial extent and irregular shape of the annotated seismic datasets.
  • Evaluation Metrics: The Dice coefficient quantifies 3D facies segmentation accuracy by measuring agreement between predicted and ground-truth voxel labels for each semantic class.Its range is 0 to 1, with higher values indicating better agreement.

B. Model architectures

The study evaluates three voxel-based 3D segmentation architectures for seismic facies segmentation: CNN-based 3D-Unet, transformer-based UNETR, and state-space-model-based SegMamba. Their designs combine hierarchical convolutional processing, global-context modeling, or Mamba-based multi-scale representations with convolutional decoding.

  • Architecture families: The benchmark compares 3D-Unet, UNETR, and SegMamba as representative CNN-based, transformer-based, and state-space-model-based architectures.Each architecture reflects a distinct design philosophy for voxel-based 3D seismic facies segmentation.
  • 3D-Unet: 3D-Unet uses five encoder stages with paired 3×3×3 convolutions, batch normalization, ReLU activation, max-pooling, and progressively wider feature channels.Feature channels increase from 16 to 128, with the bottleneck also operating at the widest channel count.
  • 3D-Unet: 3D-Unet restores the input spatial dimensions and produces a complete 128 × 128 × 128 voxel prediction aligned with the input.The implementation avoids downsampling below the final layer while retaining hierarchical multi-scale encoder features.
  • UNETR: UNETR embeds nonoverlapping 3D patches with a transformer encoder and uses convolutional and deconvolutional decoding to produce voxel-wise semantic predictions.Its design combines global context from transformers with local detail from convolutional operations.
  • SegMamba: SegMamba combines Gated Spatial Convolution, Mamba-based 3D encoder blocks, convolutional decoding, and feature-level uncertainty estimation for feature reuse.GSC preserves voxel spatial relationships before flattening, while TSMamba downsampling maintains multi-scale representations and limits sequence length.

III. EXPERIMENTS AND RESULTS

The experiments compare three 3D segmentation methods on two datasets under identical training protocols, while also analyzing model complexity and sensitivity to input sub-volume size.

  • Comparative evaluation: The study evaluates three 3D segmentation methods on each of two datasets using identical training protocols for a fair comparison.The experiments include comparative evaluation across both datasets.
  • Additional analyses: The experiments analyze model complexity and assess sensitivity to the size of the input sub-volumes.Sensitivity analysis focuses on the sub-volumes used as model inputs.

A. Experimental Setup

The study compares each proposed method on two benchmarks, evaluating each fold on its corresponding validation set. Experiments use an NVIDIA A100 and hyperparameters drawn from prior standard 3D segmentation frameworks to support comparability and reproducibility.

  • Experimental Setup: Each proposed method is compared on two studied benchmarks, with evaluation performed after every fold on its corresponding validation set.Reported measures include mean accuracy for each class and the dataset-wide average accuracy.
  • Training Configuration: Experiments were performed on an NVIDIA A100 with 80 GB of VRAM and Ubuntu Server 18.04.3 LTS.
  • Training Configuration: Hyperparameters primarily followed prior studies and standard 3D segmentation frameworks, including learning-rate schedules, optimizer configurations, and batch sizes.This approach was intended to facilitate direct comparison, stable convergence, and reliable performance.

B. Facies Segmentation Performance

Performance is higher on the Netherlands F3 dataset than on Parihaka, while 3D U-Net achieves the strongest average Dice scores across both datasets. SegMamba remains competitive, whereas UNERT performs comparatively worse.

  • Cross-dataset performance: All evaluated methods achieve higher accuracy on F3 than on Parihaka, with Parihaka performance approximately 10% lower across models.The passage attributes this difference to F3’s relatively clean signal and well-defined geological structures.
  • Model comparison: 87.69% for F3 and 75.01% for Parihaka are the highest average Dice scores, achieved by 3D U-Net.The passage links this result to convolutional architectures capturing local volumetric features with limited labeled seismic data.
  • Model comparison: SegMamba performs competitively but remains approximately 3% lower than 3D U-Net on both datasets, while UNERT shows comparatively weaker performance.These comparisons concern the two models other than 3D U-Net.

C. Model Complexity and Computational Cost

3D U-Net is the most computationally efficient of the three evaluated methods, combining competitive performance with the fewest parameters, lowest training memory, and lowest reported computational cost.

  • Model Complexity and Computational Cost: 3D U-Net outperforms the other two methods in computational efficiency, requiring 2.51M parameters, 5.79 GB of training memory, and 95.41 GFLOPs.The analysis reports these as the lowest values among the compared methods.

D. Sensitivity Analysis on Sub-volume Size

The sensitivity analysis shows that input sub-volume size affects 3D U-Net segmentation performance on the Netherlands F3 dataset. Among the tested configurations, 128 × 128 × 128 voxels achieves the best average Dice score, while smaller and larger sizes perform worse.

  • D. Sensitivity Analysis on Sub-volume Size: 87.69% average Dice score is achieved with a 128 × 128 × 128-voxel sub-volume, outperforming the tested 96×96×96 and 144×144×144 sizes.The comparison uses 3D U-Net on the F3 dataset with five-fold cross-validation.
  • D. Sensitivity Analysis on Sub-volume Size: Smaller sub-volumes degrade performance because they fail to capture sufficient spatial context.

IV. CONCLUSION

The study presents the first comparative benchmark for voxel-based 3D seismic facies segmentation using state-of-the-art architectures. It finds that 3D U-Net is a strong, efficient baseline, supporting convolutional operations for seismic volumes with local continuity and limited training data.

  • Benchmark contribution: The work presents the first comparative benchmark for 3D seismic facies segmentation using state-of-the-art voxel-based deep learning architectures.The benchmark evaluates representative CNN-based, transformer-based, and state-space-based models.
  • Limitations of prior approaches: Existing manual and 2D slice-based facies analysis methods fail to capture volumetric continuity and spatial coherence in geological structures.The conclusion contrasts these approaches with voxel-based 3D segmentation of seismic data.
  • Benchmark findings: 3D U-Net consistently achieved the best overall performance on both datasets while exhibiting the lowest computational and memory cost among evaluated methods.The result identifies 3D U-Net as a strong and competitive baseline for seismic facies segmentation.
  • Architectural suitability: Convolutional operations are well suited for seismic volumes with strong local continuity and limited training data.This conclusion follows from the observed performance and resource efficiency of CNN-based architectures.
Loading 2608.14058v1…