Seismic interpreters label faults by hand on a handful of 2D slices cut from a 3D survey that can hold billions of samples. Recent work shows that a 3D network trained only on those few slices can produce continuous fault surfaces across the whole survey. These models still share one blind spot. On every slice the interpreter touched, anything left unpicked is treated as confident "no fault." Interpreters routinely skip subtle or inconvenient faults, so the model is penalized for finding real faults on exactly the slices where supervision is densest. The more an expert annotates, the more of these omissions enter training as hard negatives.
This proposal uses the third dimension to detect those omissions. A fault is a surface. If the network, looking at the unlabeled slices on either side of an annotated slice, sees a coherent surface that runs straight through that slice where the interpreter drew nothing, that is evidence the interpreter missed it. We would down-weight the "no fault" label only at those places, and leave the rest of the background label intact. In two sentences: train a 3D fault segmenter from sparse human slices, but trust each unpicked pixel on an annotated slice only as much as the neighboring unlabeled volume agrees it is empty. A 2D or narrow-window model cannot make this correction, because the evidence comes from fault continuity across many slices.
Core Claim
The closest work is Adaptive Finetuning of 3D CNNs with Interpretation Uncertainty for Seismic Fault Prediction. It fine-tunes a synthetic-pretrained 3D U-Net on single annotated field slices in three steps. It samples training crops with a Poisson distribution centered on picks. It lets the positive-label weight decay exponentially with distance from each pick. It masks the loss to the annotated slices only. The paper models uncertainty only about where a picked fault lies, and its authors note that it treats everything far from a pick as confident background. That is the gap we target.
The classic 3D U-Net takes the opposite extreme. It sets every unannotated voxel to zero loss weight, but within an annotated slice it still trusts all labels. It reports IoU of 0.863 against 0.796 for a matched 2D network on sparse orthogonal slices of confocal microscopy volumes. That result motivates our volumetric design, but it says nothing about noisy negatives. Learning to label seismic structures with deconvolution networks and weak labels addresses label noise with a focal loss, which down-weights easy pixels uniformly and not by location.
Our claim is that interpreter omission is spatially structured and can be detected from volumetric fault continuity. Correcting only the negatives that this continuity contradicts should beat three alternatives under the same annotation budget: trusting all negatives, ignoring all of them, and down-weighting them uniformly. The gain should be largest in recall of the omitted faults and should grow with the omission rate.
Research Program
The base model follows the Mustafa et al. recipe. A 3D U-Net is pretrained on synthetic fault volumes. It is then fine-tuned on field sub-volumes with Poisson-centered sampling, distance-decay weights on positive picks, and loss masked to annotated slices. We change only the weight on unpicked voxels of annotated slices.
A slowly updated copy of the network, the "teacher" (an exponential moving average of the student's weights), predicts fault probability over each full training sub-volume. We take the teacher's fault probabilities on the unlabeled slices within a band of $\pm m$ slices around an annotated slice. We fit local surface patches by connected-component analysis, oriented along the dominant fault dip. For each unpicked voxel $x$ on the annotated slice, we define an omission score $q(x) \in [0,1]$. It is high only when two conditions hold. First, a teacher surface patch passes through $x$ with support on both sides of the slice, covering at least $m$ consecutive slices each way. Second, $x$ lies far from any human pick, so the case is not ordinary positional fuzziness, which the decay weights already handle. The negative-class loss at $x$ is multiplied by $1 - q(x)$. With $\ell$ the per-voxel binary cross-entropy, $w^+$ the Mustafa decay weight on picks, $P$ the picked voxels, and $N$ the unpicked voxels of annotated slices, the loss is
$$ \mathcal{L} = \sum_{x \in P} w^+(x)\,\ell(\hat{y}_x, 1) + \sum_{x \in N} \big(1 - q(x)\big)\,\ell(\hat{y}_x, 0). $$
Here $\hat{y}_x$ is the student's predicted fault probability at $x$. Requiring two-sided support stops the model from "explaining away" a negative using its own prediction on the slice itself. The evidence must come from slices the interpreter never saw. A warm-up period with $q = 0$ keeps the teacher from confirming its own early errors.
We will analyze three things. The first is how the benefit scales with the omission rate and with the annotation budget, meaning the number of labeled slices. The second is whether $q$ actually concentrates on true omissions, which we can measure where ground truth is known. The third is whether the band width $m$ must be large, which would show that volumetric context and not local smoothing is doing the work.
Data and Protocol
The main field benchmark is the Thebe gas field ($1803 \times 1537 \times 3174$ voxels; 900 training, 200 validation, and 703 test crosslines). We will simulate sparse interpretation by revealing labels on $k \in \{1, 3, 10, 30\}$ training crosslines. The remaining training volume stays in use as unlabeled context. We then simulate omissions by deleting whole connected fault segments from the revealed slices at rates of 0, 20, and 40 percent. Evaluation uses Optimal Image Scale and Optimal Dataset Scale F-scores (OIS and ODS) on the 703 test crosslines. We add an omission-recovery recall: the fraction of deleted training segments the model predicts on those slices. We also measure false-positive density in regions that are fault-free on every labeled slice.
Thebe's own labels surely contain omissions, which would make our gains look smaller than they are. We therefore repeat the protocol on the FaultSeg3D synthetic volumes, where the labels are complete. Because these volumes are only $128^3$, we stitch several into larger blocks and use a smaller $m$. On the Netherlands F3 block, which has no dense fault labels, we run the single-inline setting of Mustafa et al. Expert interpreters will review a blinded sample of detections that appear only under our method, so we can check whether those detections are real faults.
Baselines
The naive reference is the synthetic-pretrained 3D U-Net applied with no field fine-tuning. It establishes how much the domain gap costs. Both FaultSeg Swin-UNETR and Mustafa et al. report that such models fragment on field data, and this baseline shows that failure on our splits.
The direct competitor is the full Mustafa et al. recipe. Its negatives are trusted, and it uses the same backbone, synthetic pretraining, crop sampling, optimizer, and number of steps as our method. Any gain over it is attributable to the loss term alone. On Lacuna we found no reported quantitative result for this method on Thebe or F3, since its evaluation is qualitative. We will therefore report it ourselves at every budget.
The strongest published Thebe number is Seismic Fault SAM, with OIS 0.881 and ODS 0.879. It adapts SAM with lightweight adapters and stacks 5 adjacent slices as input channels. FaultSeg Swin-UNETR reports OIS 0.875 and ODS 0.870. Both results use all 900 training crosslines, so they are a dense-label ceiling and not a like-for-like comparison. We will retrain Seismic Fault SAM in our sparse-slice and omission settings with matched GPU-hours and the same hyperparameter-search budget, about 20 trials per method on the validation split. Only Seismic Fault SAM's sparse-label performance can be compared directly to ours. We found no reported sparse-slice result for any method on Lacuna.
Three ablations isolate the key ingredient. The first is "uniform negative down-weighting," with $1 - q$ replaced by a constant tuned on validation. The second is "focal loss on negatives," following Alaudah et al. The third is "same-slice $q$," in which omission evidence comes from the student's prediction on the annotated slice itself, with no neighboring slices. Beating the Mustafa recipe by a margin that grows with the omission rate is the minimum result. Beating the uniform and same-slice ablations on omission recall, at no worse false-positive density, is what would show that volumetric continuity is the mechanism.
Why not just…?
Why not just zero-weight everything unlabeled, including unpicked pixels on annotated slices? This is the 3D U-Net logic pushed one step further. It removes omission penalties but throws away the interpreter's real negative information, which is most of what an annotated slice contains. With faults this sparse and the class imbalance this severe, we expect false positives to rise sharply. We include this as a baseline and will report the false-positive density alongside recall.
Why not just down-weight all negatives uniformly, or use a focal loss? This is the real threat. If omissions were spread evenly, a well-tuned constant would capture most of the benefit. Our bet is that omissions cluster on subtle, low-throw faults that are continuous in 3D, so targeted relabeling should recover them without loosening the background elsewhere. The uniform and focal ablations, tuned on the same budget, settle the question. If they match us on recall at equal false-positive density, the targeted mechanism is not worth its complexity, and we will say so.
Why not just self-train with pseudo-labels on the unlabeled slices? Standard teacher-student pseudo-labeling adds supervision where the human gave none. It never questions human negatives, so it inherits the same omission penalty on annotated slices. We will run a mean-teacher pseudo-labeling baseline both alone and combined with our loss. If combining them adds little over pseudo-labeling alone, that suggests the extra unlabeled supervision already outweighs the bad negatives.
Why not just use the 2.5D SAM model, which holds the best Thebe score? Its 5-slice window is too narrow to supply the two-sided, many-slice surface evidence our omission score needs, and its dense-label advantage may not survive at $k = 1$ or $3$ slices. The sparse-budget comparison and the sweep over $m$ test this directly. If performance saturates at $m \le 2$, a 2.5D model could carry the same correction, and the volumetric argument weakens.
Decisive Checks
The idea succeeds if three things hold on Thebe at the 20 and 40 percent omission rates and at small budgets. Omission recall must rise clearly over the Mustafa recipe. False-positive density must not rise. The uniform and same-slice ablations must recover substantially less of that gain. The FaultSeg3D replication with complete labels must agree, so the result is not an artifact of Thebe's own missing labels. Benefit should increase with $m$ beyond the 5-slice regime. Expert review on F3 must confirm that a meaningful share of the new detections are real faults. If uniform down-weighting matches us, or the benefit vanishes beyond a few slices of context, the core claim fails.
References
- Adaptive Finetuning of 3D CNNs with Interpretation Uncertainty for Seismic Fault Prediction
- 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation
- Learning to label seismic structures with deconvolution networks and weak labels
- FaultSeg Swin-UNETR: Transformer-Based Self-Supervised Pretraining Model for Fault Recognition
- Seismic Fault SAM: Adapting SAM with Lightweight Modules and 2.5D Strategy for Fault Detection