Augmented Adversarial Training: Improving Robustness of Neural Network Based Geologic Interpretation

Summary Seismic images exhibit enormous diversity in structural complexity, resolution and signal to noise ratio across surveys. Consequently, convolutional neural network (CNN) based geologic interpretation often lack adequate generalization capabilities on such real data images. It is commonly observed that a CNN trained using data from one survey, exhibits significant degradation in interpretation accuracy when used on a new survey, never seen during the training stage. This makes production scale deployment of such models problematic and unreliable. In this paper we address the generalization issue by exploiting the presence of adversarial samples: defined as visually imperceptible, worst-case perturbations to an image that causes a CNN to misclassify the perturbed image with a high degree of confidence. We show that images from a new survey are likely close to adversarial points for a network optimally trained with legacy data. We then describe a training method which allows a CNN to develop robustness to such adversarial samples leading to significantly improved generalization capabilities. Using examples from salt interpretation during the model building stage on Gulf of Mexico (GOM) datasets, we demonstrate that our training strategy has very low generalization error and close to human accuracy on new, previously unseen surveys.