Are seismic attributes still helpful for deep learning?

Abstract

Seismic classification is a machine learning process of mapping seismic data to classes, and this process is traditionally performed using methods that rely exclusively on seismic attributes. With the recent evolution of deep learning, seismic classification can now be performed directly on seismic data without the need to carefully select seismic attributes as inputs. Recent studies show deep learning methods trained directly on the seismic data can perform better than traditional methods that must train on seismic attributes. While this is true, what remains on open question is if deep learning methods could perform even better if trained on seismic data and appropriate seismic attributes. We investigate this idea using a synthetic seismic classification problem. The algorithm we choose is a recurrent neural network (RNN), and we test different input scenarios (i.e., with and without seismic attributes) and vary the amount of training date. Our findings suggest that including seismic attributes as inputs to deep learning algorithms may still be beneficial under certain conditions.