Explainable blind-spot networks for self-supervised seismic coherent noise suppression

Summary Self-supervised denoising presents an alternative to supervised deep learning procedures, relieving the demand for pairs of noisy-clean training samples - an unobtainable requirement for field seismic data. For coherent noise suppression, a mask on the input to the network must be defined as part of the training process associated with a self-supervised denoising neural network. Such a mask however requires prior knowledge of the properties of the contaminating noise. Thus, we illustrate how, by incorporating explainable AI techniques in blind-spot denoising, namely Jacobian inspection, we can automatically identify the noise characteristics and design an optimal noise mask. The proposed workflow remains fully self-supervised, therefore no clean training labels are required. The effectiveness of our method is illustrated on a complex synthetic dataset contaminated by various coherent noise types: namely, coloured Gaussian noise, linear noise, and rig noise. The tailor-made noise masks are shown to almost completely suppress the different coherent noises, with minimal signal damage. Future work will focus on reducing the human interaction within the workflow, by automating the creation of the noise mask directly from the Jacobian matrix.