Source-linked AI summary
A Machine Learning Benchmark for Facies Classification
Yazeed Alaudah, Patrycja Michalowicz, Motaz Alfarraj, Ghassan AlRegib
TL;DR
Seismic facies classification lacks large, publicly available annotated datasets and standardized quantitative evaluation, limiting objective comparison among models. This paper releases a fully annotated Netherlands F3 Block geological model, two deconvolution-network baselines, their code, and evaluation metrics; the section-based baseline outperformed the patch-based baseline, with a 15% MCA improvement.
Problem
Large public annotated datasets and quantitative evaluation are scarce in seismic facies classification, making comparisons between approaches subjective and difficult.
Method
The paper constructs a fully annotated 3D Netherlands F3 Block model, proposes patch-based and section-based deconvolution baselines, releases their code, and defines evaluation metrics.
Results
The section-based baseline outperformed the patch-based baseline overall, including a 15% MCA improvement and clearer advantages for smaller classes.
Takeaways & Limitations
The released dataset, code, baselines, and metrics establish an objective benchmark for comparing machine-learning approaches to facies classification.
Abstract
from arXiv · showhide
The recent interest in using deep learning for seismic interpretation tasks, such as facies classification, has been facing a significant obstacle, namely the absence of large publicly available annotated datasets for training and testing models. As a result, researchers have often resorted to annotating their own training and testing data. However, different researchers may annotate different classes, or use different train and test splits. In addition, it is common for papers that apply machine learning for facies classification to not contain quantitative results, and rather rely solely on visual inspection of the results. All of these practices have lead to subjective results and have greatly hindered the ability to compare different machine learning models against each other and understand the advantages and disadvantages of each approach. To address these issues, we open-source a fully-annotated 3D geological model of the Netherlands F3 Block. This model is based on the study of the 3D seismic data in addition to 26 well logs, and is grounded on the careful study of the geology of the region. Furthermore, we propose two baseline models for facies classification based on a deconvolution network architecture and make their codes publicly available. Finally, we propose a scheme for evaluating different models on this dataset, and we share the results of our baseline models. In addition to making the dataset and the code publicly available, this work helps advance research in this area by creating an objective benchmark for comparing the results of different machine learning approaches for facies classification.
INTRODUCTION
Seismic facies classification needs large, consistently annotated datasets and quantitative evaluation, but existing practices often rely on limited, researcher-specific annotations and subjective visual inspection. The paper addresses this gap by releasing a fully annotated Netherlands F3 Block model with baseline models, code, and evaluation procedures.
- Research motivation: Deep learning facies classification requires substantial annotated training data, but large publicly available seismic interpretation datasets remain scarce.Researchers have therefore annotated different numbers of F3 Block inlines, ranging from one to 40.
- Research motivation: Weakly supervised and unsupervised methods have been explored to reduce reliance on large manually annotated datasets.Examples include image-level labels, similarity-based label generation, principal component analysis, self-organizing maps, and deep convolutional autoencoders.
- Research motivation: The lack of shared annotated data and quantitative results makes facies-classification comparisons subjective and hinders understanding of model advantages and disadvantages.Many studies rely primarily on visual inspection rather than reported metrics.
- Contributions: The paper releases a fully annotated 3D geological model of the Netherlands F3 Block, based on 3D seismic data, 26 well logs, regional geology, and extracted fault planes.The fault data are released for possible fault-detection research but are not used in this work.
- Study area: The F3 Block is a 16 km x 24 km North Sea survey area whose publicly available seismic data and boreholes have supported extensive geological and seismic interpretation research.The survey was conducted in 1987 to identify geological structures and search for hydrocarbon reservoirs.
- Study area: The geological setting includes ten lithostratigraphic groups, spanning North Sea, Chalk, Rijnland, Jurassic, Triassic, Zechstein, and Limburg units.The listed groups are ordered from newest to oldest and include distinct rock types and geological ages.
- Study area: The F3 Block lies between the Step Graben and Dutch Central Graben, where tectonic activity and salt diapirs produce variable lithostratigraphic structures.Several groups occur only within the Dutch Central Graben.
The modeling process
The 3D geological model was prepared from both well-log and 3D seismic data, with the two data sources described as complementary inputs to the modeling process.
- The modeling process: The F3 Block geological model was prepared using both well logs and 3D seismic data.The following subsections describe how each data source contributed to model construction.
3D model building using well logs data
Well-log data from 26 boreholes were organized into geological horizons, correlated, and combined with seismic interpretation to build and refine the final 3D model. Time-depth conversion, fault interpretation, horizon tracking, and structural modeling were key steps.
- Well-log preparation: Well-log data for 26 boreholes were collected with coordinates, depth, inclination, and horizon information.The wells were located within or near the F3 Block, and their locations are shown in Figure 3.
- Well-log preparation: The 40 horizons originally present in the wells were assigned to adopted lithostratigraphic units and correlated across wells.These correlations supported a preliminary 3D model based on well-log data and guided subsequent seismic interpretation.
- Seismic-data preparation: Because the seismic data were migrated in time rather than depth, the workflow performed time-depth conversion using the supplied velocity model.OpendTect 5.0 was used for the conversion.
- Seismic-data preparation: Main fault surfaces were interpreted and fault networks were constructed, while horizons were tracked automatically with a seeded 3D autotracking tool that accounted for modeled faults.Petrel tools were used for polygon editing, fault-framework modeling, and horizon interpretation.
- Model refinement: The preliminary model was refined through fault re-modeling and horizon corrections before producing the final three-dimensional model.Volume-based horizon modeling and Petrel structural-modeling tools were used to create zones between individual horizons.
The 3D geological model
The F3 Block model contains seven lithostratigraphic groups spanning Cenozoic, Mesozoic, and Permian structural levels, with strong spatial variation caused by faults and salt diapirs.
- The model identifies seven lithostratigraphic groups divided into Cenozoic, Mesozoic, and Permian structural levels.
- The F3 Block has highly variable geological structures, including differential unit thicknesses and an expanded fault network related to salt tectonics.
- The Upper North Sea group is the youngest and flattest unit, with thicknesses ranging from about 1000 m to 1320 m.
- The Middle North Sea group ranges from 20 to 150 m thick, while the Lower North Sea group ranges from about 180 to 750 m.
- The Rijnland, Scruff, and Zechstein groups vary substantially in thickness and distribution, particularly near Zechstein diapirs.
- Three fault generations were identified, and diagonal cross sections follow the SW-NE and NW-SE axes of the model.
Deconvolution networks
Deconvolution networks use mirrored encoder-decoder architectures to combine feature recognition with accurate spatial localization for semantic labeling tasks such as facies classification.
- Fully convolutional architectures enabled semantic labeling with convolutional networks without pre- or post-processing steps.
- A deconvolution network uses convolution and pooling layers in an encoder mirrored by deconvolution and unpooling layers in a decoder.
- The encoder performs object detection and classification, while the decoder supports accurate localization.
- The network architecture was applied to seismic interpretation in prior studies and used for both baseline models in this work.
- Pooling layers select maximum filter responses, whose indices are shared with unpooling layers to recover higher-resolution outputs.
Baseline Models
The study compares patch-based and section-based baseline models that share an architecture and training settings but differ in data handling and labeling procedures.
- The two baseline models use the same architecture, optimizer, and hyperparameters but differ in training and seismic-volume labeling procedures.
- The patch-based model trains on randomly sampled patches and averages overlapping test-patch predictions in inline and crossline directions.
- The section-based model trains on entire inline and crossline sections, allowing relationships between lithostratigraphic units and depth information to be used.
- On an Nvidia Titan X GPU, the section-based network was trained eight sections at a time in about 70 minutes.
- Additional variations tested data augmentation and skip connections as potential model improvements.
The geological model
The final geological model uses a restricted, processed volume selected to avoid artifacts and missing data, while retaining labels, seismic data, horizons, and fault planes for research use.
- The training and testing volume is restricted to inlines 100–701, crosslines 300–1201, and depths 1005–1877 meters because of processing artifacts and missing side data.
- Inline and crossline pixels both have 25m×25m resolution, although their dimensions need not match.
- The training set contains different lithostratigraphic classes whose percentages are reported in Table 1.
- The released data include original horizons for all lithostratigraphic units and extracted fault planes from three generations.
The train/test split
The benchmark separates correlated seismic regions into one training set and two spatially distinct test sets. This design aims to reduce leakage and preserve an independent evaluation boundary.
- Rationale: Randomly selecting neighboring seismic sections can produce artificially high test results because those sections are highly correlated.The split therefore seeks to minimize correlation between training and testing data.
- Data partitions: The training set covers inlines [300,700] and crosslines [300,1000].
- Data partitions: Test set #1 covers inlines [100,299] and crosslines [300,1000].
- Data partitions: Test set #2 covers inlines [100,700] and crosslines [1001,1200] and contains a large Zechstein diapir unseen during training.
- Evaluation boundary: The test sets should be used only for final evaluation, not repeated validation during model retraining.Reusing them for parameter selection defeats their purpose as test sets.
Evaluation metrics
The study evaluates models with multiple objective metrics and compares patch- and section-based approaches across two spatial test sets. Section-based models perform better overall, while augmentation and skip connections improve results, especially for smaller classes.
- Metrics: The evaluation uses pixel accuracy, per-class accuracy, mean class accuracy, and frequency-weighted intersection over union.
- Test design: The two test sets are spatially separated from training, with test set #2 containing an unseen Zechstein diapir.
- Baseline comparison: Section-based models better classify classes that patch-based models often confuse with the Lower North Sea group.
- Baseline comparison: Section-based models achieve a 15% improvement in MCA over the patch-based baseline and show a clear advantage on Scruff and Zechstein.
- Baseline comparison: Section-based models are superior overall because they incorporate spatial and contextual information and train and test faster.
- Class imbalance: The dataset is highly imbalanced, favoring Lower and Upper North Sea groups and producing poor performance for smaller classes.The baseline models do not apply class-imbalance corrections, although such techniques could improve smaller-class results.
- Data augmentation: Data augmentation increases patch-based FWIU and MCA scores by more than 10% and improves performance on Zechstein and Scruff.
- Skip connections: Skip connections improve patch-based PA by about 1% and MCA and FWIU by about 1.5%, while Scruff accuracy increases by 6.5% and 25% in patch- and section-based models.
EVALUATION METRICS
The paper evaluates facies-classification models with computer-vision metrics based on correctly classified pixels, class overlap, and class frequency. These metrics provide complementary measures of overall accuracy, per-class performance, and intersection-over-union quality.
- The evaluation uses metrics commonly applied in computer vision to objectively assess model performance.The framework defines ground-truth pixels Gi, predicted pixels Fi, their intersection, and set cardinality.
- Pixel Accuracy (PA) measures the percentage of correctly classified pixels across all classes.
- Class Accuracy for class i (CAi) measures the percentage of pixels correctly classified within class i.Mean Class Accuracy (MCA) averages CAi over all classes, with nc denoting the number of classes.
- Intersection over Union (IUi) measures the overlap between ground-truth and predicted pixels for class i using their intersection divided by their union.The metric equals 1 if and only if all pixels are correctly classified; averaging IUi over classes gives Mean Intersection over Union (Mean IU).
- Frequency-Weighted Intersection over Union (FWIU) weights each class by its size to reduce sensitivity to small classes.