Development of a MEMS-based Earthquake Dataset using the Raspberry Shake Network in New Zealand

Abstract

Traditional seismic monitoring is often limited by the high cost of instrumentation and logistical barriers, hindering the expansion of earthquake monitoring networks especially in under-resourced regions. Low-cost MEMS-based sensors offer a scalable alternative, but require specialized datasets to train machine learning models adapted to their unique noise characteristics and sensitivity profiles. To address this, we systematically collected waveforms from approximately 4,000 earthquakes (magnitude 2.7 to the highest recorded) recorded across 89 Raspberry Shake stations in New Zealand from 2020–2025. Events were matched to nearby stations based on epicentral distance criteria (100 km for M 2.7–5.5, 150 km for M>5.5). A staged filtering pipeline using PhaseNet, EQTransformer, and GPD models, cross-validated with theoretical TauP arrivals, was applied to ensure phase pick quality across three confidence tiers. The final curated dataset comprises 918 high-confidence waveforms with validated P and S wave arrivals, alongside approximately 16,035 total waveform records spanning all quality tiers. This dataset addresses the critical scarcity of labeled training data for low-cost seismic instrumentation, enabling the development of phase pickers specifically calibrated for MEMS sensors.