The weakly supervised learning method for 3D fault detection

In this work, we propose a weakly supervised learning method which could utilize sparse manual interpretation results as training data for 3D fault detection task. Following a weakly supervised learning setting, we design the masked training data, which are gathered from field seismic volumes, and a sparse loss function for training process. Synthetic seismic data and field seismic volumes are applied to testify the proposed method. While we make no claim that these results from weakly supervised learning method are better than results predicted by full supervised methods, we believe that weakly supervised method can provide at least competitive results with the supervised models in the literature and highlight the potential of the weakly supervised learning framework.