Parameter-Efficient Adaptation of Vision Foundation Models for Seismic and DAS-VSP Data Denoising

Abstract

Deep learning approaches for seismic and Distributed Acoustic Sensing (DAS) data denoising often exhibit limited generalization under complex noise profiles and domain shifts. To address this, we propose a computationally efficient framework that repurposes general-purpose Vision Foundation Models (VFMs) for geophysical tasks using Parameter-Efficient Fine-Tuning (PEFT). Specifically, our architecture employs a pre-trained DINOv3 encoder adapted via Low-Rank Adaptation (LoRA), which adapts latent features with minimal memory overhead while mitigating catastrophic forgetting. Moreover, to improve performance on unseen data, we introduce a kurtosis-guided unsupervised Test-Time Adaptation (TTA) strategy that updates only LoRA parameters during inference. This allows the model to self-calibrate to site-specific noise characteristics by identifying information-rich regions based on kurtosis and performs self-training without access to ground truth. Experiments on exploration field seismograms and DAS-VSP data from the Utah FORGE site show that our framework matches or exceeds the performance of domain-specific models and other VFMs baselines. The evaluation on unseen cross-site data from Groß Schönebeck geothermal site further demonstrates the robust generalization capabilities of the framework, highlighting the potential of adapting pre-trained VFMs for data-intensive challenges in exploration seismology.