The dual seismological developments of large-scale offshore observational networks and high-quality machine-learning earthquake detection and picking allow unprecedented seismicity detection in high hazard subduction zones. We develop a high resolution machine learning enhanced seismicity catalog for the Alaska Peninsula stretch of the Alaska-Aleutian Subduction Zone, which was instrumented with both land and ocean bottom seismometers during the Alaska Amphibious Community Seismic Experiment (AACSE) in 2018–2019. We first use an analyst-checked AACSE catalog to develop new local land and ocean-bottom earthquake waveform training data sets, which we then use to fine-tune the EQTransformer machine learning picking model. In the process, we test multiple training strategies for best picking performance on ocean-bottom waveform data. We evaluate the effect of choice of pre-trained starting model, addition of waveform bandpass filtering, and choice of random metadata split. We find no training strategy to be substantially better than the others. Instead, simply having a local waveform data set to fine-tune the picking model substantially improves performance on both the ocean-bottom and land seismic data from AACSE. We finally apply the separately trained land and ocean-bottom models to continuous seismic data for AACSE and use an iterative catalog refinement method to detect new small-magnitude seismicity. The machine learning enhanced catalog has roughly 3x more earthquakes than the analyst-checked catalog, and the magnitude of completeness is regionally variable but decreases 0.1–0.4 magnitude units. With more small seismicity detected, seismicity patterns mirror prior observations in the region. Detection is especially improved far offshore and in the outer rise.
