Marine seismic surveys can contain missing or unusable receiver traces. This paper tests how well those traces can be reconstructed. It compares zero fill, linear interpolation, a U-Net trained from random initialization on BP, SwinV2 with ImageNet weights, and SFM with seismic pretraining. BP 2007 supplies complete synthetic shots, so receiver traces can be removed from the input and scored against the target. Viking Line 12 is a field line; the test removes observed field traces and scores their reconstruction after the models are trained only on BP. All learned models use inputs computed from visible traces, predict a correction to a linear interpolation estimate, copy measured traces back, and are scored only on removed traces. On BP 2007, with 75% of receiver traces removed in groups of eight, the U-Net has the lowest RMSE on removed traces, 1.170 ± 0.467. SwinV2 with LoRA is the best pretrained method on BP, 2.247 ± 0.196. On Viking Line 12, using BP-trained weights without field retraining, the U-Net has the lowest mean RMSE, 16.38 ± 22.88, but repeat variation is large. Frozen SFM is the best pretrained method on Viking by RMSE, 19.98 ± 0.62, and has the highest SSIM. ... Read more

Seismic trace denoising is a long-standing problem in geophysical data processing, and recent work has shown that vision foundation models pre-trained on natural images can be adapted to it parameter-efficiently rather than trained from scratch. Such adaptation is typically applied uniformly across all transformer layers, but it is not known where along the network the adaptation effort is actually needed --- that is, where the representation gap between natural images and seismic data is concentrated. We investigate this question for two structurally distinct families of parameter-efficient fine-tuning (PEFT): Low-Rank Adaptation (LoRA), which injects a low-rank update into the attention projections, and Pfeiffer bottleneck adapters, which insert a residual MLP module after the feed-forward sub-layer. Using a DINOv3 ViT-S/16 backbone on active-source seismic image denoising and holding the per-layer parameter budget fixed across both mechanisms, we sweep adaptation placement across restricted subsets of the twelve transformer layers within the DINOv3 architecture and measure denoising quality at each placement. We find that adaptation is strongly concentrated in the early layers: placing modules on only the first four layers recovers most of the denoising quality of full adaptation at one third of the adapter parameters, and even a single early layer is already competitive with full adaptation, while a roughly monotonic early-to-late importance ordering holds across placements. Crucially, this profile is near-identical for the two mechanisms at matched budget, indicating that the effect is a property of where adaptation is applied rather than of the particular PEFT design. These results suggest that, for this task and backbone, the natural-image-to-seismic gap is primarily a low-level, input-stage shift, and that early-layer-heavy placement is an effective and economical default for PEFT-based adaptation of vision foundation models to seismic data. ... Read more

Seismic denoising is essential for subsurface imaging. Deterministic methods such as Radon transforms exist but struggle to separate signal from noise. Recent solutions like vision foundation models (VFMs) offer better signal separation, but generalize poorly across geological domains. Parameter-efficient fine-tuning with LoRA and kurtosis- guided test-time adaptation (TTA) improves cross-domain generalization, yet TTA performs weight updates during inference and collapses throughput from 114.29 to 1.09 patches per second, making real-time deployment impractical. This paper inves- tigates whether group relative query optimization (GRQO), a reinforcement-learning fine-tuning method applied before deployment, can reduce reliance on TTA. A LoRA- adapted DINOv3 backbone with a U-Net decoder is fine-tuned with a GRQO objective that combines a multi-head reward, KL-divergence regularization against a frozen ref- erence, and a head-diversity penalty. On a 4,000-patch unseen synthetic data set, SFT+GRQO alone improves signal retention (MS-SSIM-R 0.638→0.604) while pre- serving a similar inference speed. GRQO does not fully replace TTA, but it lowers the optimal TTA budget for signal retention: the best configuration for MS-SSIM- R pairs GRQO with only 25 TTA epochs (MS-SSIM 0.862, MS-SSIM-R 0.452), and longer adaptation degrades performance and decreases throughput. However, GRQO alone minimally deteriorates performance when evaluated on real seismic data (LS 0.373→0.381). These results indicate that pre-deployment reinforcement learning can shorten the costly real-time adaptation and increase signal retention, but may perform worse on unseen seismic domains. ... Read more

Active seismic imaging is used to infer subsurface structure from reflected wavefields, but acquisition and ambient noise can obscure weak reflectors and reduce interpretation reliability. Seismic denoising must remove noise while keeping geological structure intact. This thesis studies a parameter-efficient method to adapt pretrained vision foundation models to this task. The method treats each seismic section as a 2D grayscale image, maps it into a format compatible with vision backbones, and applies Low-Rank Adaptation (LoRA) to limit the number of trainable parameters. It then combines the denoised outputs of multiple adapted vision models through a learned soft spatial blender. This blender merges the expert predictions at the pixel level, allowing the final model to use complementary architectural strengths such as multiscale representation and long-range dependencies. The method is evaluated against a seismic foundation model baseline, using both quantitative metrics and qualitative inspection. Across 25 seed/split repetitions, the residual joint spatial blender achieves a mean absolute error of 0.0463, a peak signal-to-noise ratio of 33.98\,dB, and a structural similarity index of 0.9727, substantially outperforming the standalone adapted experts and the frozen baseline. These results show that jointly trained spatial fusion improves seismic denoising performance while keeping training parameter-efficient. ... Read more

Seismic reflection surveys image subsurface geological structures by recording waves reflected from interfaces between rock layers, which are then processed to form 3D seismic volumes. However, the acquired signals are often contaminated by noise that degrades interpretation quality. Existing denoising approaches adapt pretrained visual foundation models to seismic data but process each slice of the 3D seismic volume independently, discarding useful spatial context. To incorporate this context while retaining the efficiency of a 2D model, three input strategies are compared: 2D-1ch, in which a single slice is repeated across the input channels; 2.5D-3ch, which uses three consecutive slices from the same volume; and 2.5D-5ch, which uses five consecutive slices. DINOv3 is adapted with low-rank adaptation (LoRA), and a lightweight decoder is trained to predict the clean central slice. On 30 synthetic Image Impeccable volumes, mean test multi-scale structural similarity improves from $0.8624$ for 2D-1ch to $0.8947$ for 2.5D-3ch and $0.9039$ for 2.5D-5ch. 2.5D gains are largest on the slices with the most noise. Additional experiments show that these improvements arise primarily from the spatial context provided by neighbouring slices. On the real-field F3 dataset, 2.5D behaviour depends on slice orientation. In the horizontal time orientation the models were trained on, 2D-1ch leaves the least structured residual, while 2.5D-3ch and 2.5D-5ch over-smooth the output. However, in the inline/crossline F3 evaluation, 2.5D reduces the over-smoothing seen in 2D-1ch and retains more structure. Cross-backbone experiments with SFM-Base and SwinV2-T show that the 2D-to-2.5D trend is not specific to DINOv3, while full fine-tuning controls show that PEFT is sufficient to achieve the observed gains. These results support 2.5D input as an effective extension on synthetic data when neighbouring slices are aligned, while highlighting its sensitivity to field-data neighbour relationships. ... Read more