Effectively Distinguishing Blast and Earthquake Sources in Eastern Canada

Eastern Canada lies within the stable interior of the North American Plate, yet several regions, including the Western Québec, Charlevoix, Lower St. Lawrence, and Northern Appalachians seismic zones, have experienced damaging intraplate earthquakes. Building reliable earthquake catalogs in these regions is challenging because tectonic earthquakes must be distinguished from industrial blasts, often under low signal-to-noise ratio (SNR) conditions. Here we develop a spectrogram-based source discrimination framework using pretrained convolutional neural network (CNN) image classifiers. Approximately 100,000 three-component waveform records of labeled earthquakes and blasts from the Canadian National Earthquake Database (2000-2024) were converted into standardized three-channel spectrogram inputs, and local event origin time was incorporated as an auxiliary feature. Among the tested models, an EfficientNet-based architecture achieved the best overall performance, yielding the lowest weighted classification cost under the imposed minimum-recall constraint. To account for uneven station coverage, we further developed an event-level framework that combines station-level CNN predictions using reliability-based weights derived from source-station distance and denoised SNR. Applied to an enhanced catalog in the Western Québec Seismic Zone, the framework separated blast-like from earthquake-like events while also identifying noise-dominated detections. This reduces false positives introduced by machine-learning based phase picker and helps produce a more reliable regional earthquake catalog. The proposed framework is efficient, transferable, and well suited for catalog enhancement in Eastern Canada, with clear application potential for other intraplate regions.