Global Distribution of Low Frequency Family Marsquakes From Deep-Learning-Based Polarization Estimation

Abstract

The seismometer has recorded thousands of marsquakes. Accurately locating these events is crucial for understanding Mars' internal structure and geological evolution. With only a single station, determining the location, especially the accurate back-azimuth, is more challenging than on Earth. Deep learning, being data-driven, can learn patterns of complex noise that are difficult for traditional methods to model, making it promising for improving back-azimuth estimation of marsquakes. However, challenges arise when applying deep learning to estimate marsquake polarization due to the limited quantity and low signal-to-noise ratios (SNR) of the data. In this study, we trained deep learning models for learning the noise patterns preceding marsquakes to address these challenges. By combining the proposed Sliding Window Inference and Featured-Training (SWIFT) to handle the high uncertainty in P phase picking, we are able to estimate polarizations of low frequency family marsquakes with improved accuracy. As a result, we have further improved the localization of marsquakes by relocating 56 events, including seven Quality C events with epicentral distances over 90°. For two Martian impact events with ground-truth locations, S1000a and S1094b, our deviations are only ∼5.65° and ∼2.72°. Our results reveal a new identified clustered seismicity zone around compressional structures in Hesperia Planum, including seven marsquakes with magnitudes from 2.7 to 3.6. Marsquakes are also widely distributed along the northern lowlands, the dichotomy boundary, and the higher-latitude southern highlands, suggesting a globally distributed pattern. Our renewed marsquake locations provide new insights into the tectonic interpretation of marsquakes.