Random noise can be removed with traditional methods like median filter, f-x filter, etc., requiring manual parameters selection and hope for the best that one set of parameters works the whole seismic.
Deep learning supervised methods were also applied to battle the same challenge; however, it requires training on the pair of noise-free and contaminated images. We can get pairs of images by either using real field data and cleaning it with a traditional method, which hinders the meaning since we would not be able to considerably surpass the used method, or synthetic seismic generator, which might not be of a high standard to model various types of noises.
The paper "Birnie et al., 2021, Self-supervised learning for random noise suppression in seismic data" proposes an alternative way to suppress random noise using the self-supervised Noise2Void method. They train a 2-layer U-Net network on manually corrupted (blind-spot) and original noised data pairs. Blind-spot patches are constructed by replacing some pixels on the original image with random ones. The loss is calculated for the tampered pixels only.
Noise2Void is a self-supervised method that does not require learning through prepared sets of contaminated and clean data. The data for the training generates during the training on a particular seismic.
The method expects statically independent noise, suitable for random noise but has limited ability for correlated noise suppression.
The link for the LinkedIn post