Self-Supervised Learning Tested on Real Seismic Data Denoising
A study evaluated the Noisy-as-Clean (NaC) self-supervised learning method for denoising real seismic data. Two independent seismic acquisitions, each with noisy and filtered data, formed four real datasets. The NaC method was adapted to add real noise controlled by a parameter. Ten experiments compared NaC SSL against supervised learning using identical network topology and hyperparameters. Results showed synthetic additive white Gaussian noise (AWGN) is inadequate for seismic data denoising.
Key facts
- Self-supervised learning (SSL) does not require clean reference data.
- Noisy-as-Clean (NaC) method was evaluated for real seismic data denoising.
- Two independent seismic acquisitions were organized into four real datasets.
- NaC SSL method was adapted to add real noise controlled by a parameter.
- Ten experiments compared different NaC SSL deployment strategies.
- Supervised learning baseline used identical network topology and hyperparameters.
- Models were evaluated on denoising performance, computational cost, and generalization.
- Synthetic additive white Gaussian noise (AWGN) is inadequate for seismic data denoising.
Entities
—