A comparison of deep and shallow models for the detection of induced seismicity

Abstract

Can an interpretable logistic regression model perform comparably to a deep learning model in the task of earthquake detection? In spite of the recent focus in academic seismological research on deep learning, we find there is hope that it can. Using data from the Groningen Gas Field in the Netherlands, relating to low‐magnitude induced seismicity, we build on a recently presented four‐input logistic regression model by adding to it four further statistically derived features. We evaluate the performance of our feature‐enhanced model relative to both the original logistic regression model (shallow machine learning model) and a deep learning model proposed by the same research group. We discover that at the signal‐to‐noise ratio of this earlier work, our enhanced logistic regression model in fact overall outperforms the deep learning model and displays no false negative errors. At the lower signal‐to‐noise ratios also considered here, while the number of false positive errors made by the logistic regression model increases, the number of undetected earthquakes remains zero. Though the number of false positives is for the highest imbalance ratios currently prohibitive, the benefit of our four additional features, which increases as the signal‐to‐noise ratio decreases, suggests that an interpretable model might be made to perform comparably to a more complex deep learning model at real‐world class imbalance ratios if further useful inputs could be identified.