Evaluating Deep‐Learning‐Based Earthquake Detection in a Sparse Coastal Network: Insights From the 2017 Changdao Swarm, Eastern China

Constructing reliable earthquake catalogs in regions with sparse station coverage and strong noise remains challenging. We evaluate the performance of a deep‐learning–based detection and location workflow (LOC‐FLOW) using the 2017 Changdao offshore earthquake swarm in eastern China. Three catalogs are compared: an automatic deep‐learning catalog, a matched‐filter catalog, and a routine network catalog. For moderate‐magnitude events, the deep‐learning and matched‐filter catalogs show consistent temporal evolution and frequency–magnitude statistics, and both delineate a northwest–southeast–trending fault zone at 8–14 km depth. Differences are confined to small magnitudes. The deep‐learning catalog has a completeness magnitude of ∼1.1, similar to the routine catalog, whereas the matched‐filter catalog extends completeness to ∼0.8 and detects substantially more microearthquakes. Event correspondence analysis indicates that most deep‐learning detections are shared with the matched‐filter catalog, with missed events primarily occurring under low signal‐to‐noise and poor azimuthal coverage. These results demonstrate that deep‐learning workflows provide a stable backbone catalog under sparse‐network conditions, while matched filtering offers complementary sensitivity to weak events. Their joint application provides an effective strategy for constructing high‐resolution earthquake catalogs in coastal transition zones and other sparsely instrumented regions.