Detecting moonquakes using convolutional neural networks, a non-local training set, and transfer learning

Abstract

The costly power requirements of delivering seismic data back to Earth from planetary missions requires the development of algorithms for lander-side signal analysis for telemetry prioritization. This is difficult to explicitly program, especially if no prior seismic data are available from the planetary body. Deep learning computer vision has been used to generalize seismic signals on Earth for earthquake early warning problems but such techniques have not yet been expanded to planetary science. We demonstrate that Convolutional Neural Networks can be used to accurately catalogue planetary seismicity without local training data by building binary noise/signal classifiers from a single Earth seismic station and applying the models to moonquakes from the Apollo Passive Seismic Experiment (PSE) and the Lunar Seismic Profiling Experiment (LSPE). In order to promote generality and reduce the amount of training data, the algorithms use spectral images instead of time-series. Two- to five-layer convolution models are tested against a subset of 200 Grade-A events from the PSE and obtained station accuracy averages of 89-96 per cent. As the model was applied to an hour trace of data (30 min before and after the Grade-A event), additional detections besides the Grade-A event are unavoidable. In order to comprehensively address algorithm accuracy, additional seismic detections corresponding to valid signals such as other moonquakes or multiples within a particularly long event needed to be compared with those caused by algorithm error or instrument glitches. We developed an 'extra-arrival accuracy' metric to quantify how many of the additional detections were due to valid seismic events and used it to select the three-layer model as the best fit. The three-layer model was applied to the entire LSPE record and matched the lunar day-night cycle driving thermal moonquake generation with fewer false detections than a recent study using Hidden Markov Models. We anticipate that these methods for lander-side signal detection can be easily expanded to non-seismological data and may provide even stronger results when supplemented with synthetic training data.