Operational application of artificial intelligence-based seismic phase identification and arrival-picking models requires rigorous testing to define their capabilities and limitations and guide further model improvements. In this study, we evaluate the PhaseNet model monitoring performance for distant M≥5.0 earthquakes under sparse-network conditions. To this end, we establish an end-to-end operational chain encompassing seismic phase detection, association, and localization, and compare the results against those obtained using the conventional STA/LTA method.

We selected 1 507 global M≥5.0 earthquakes from 2024 and 309 well-distributed seismic stations worldwide. We downloaded seismic data from the EarthScope for each event, covering a time window of 3 minutes before and 30 minutes after the origin time. We picked arrival times using both methods, then associated seismic phases using GLASS3.

Our results indicate that the PhaseNet model performs optimally for seismic data sampled at 20 Hz and generalizes well to distant M≥5.0 events. Specifically, it can effectively identify signals from earthquakes with epicentral distances up to 10 000 km, while simultaneously suppressing false picks induced by high-frequency noise and reducing the false alarm rate for small events—making it well suited for automatic monitoring tasks. In terms of association performance, PhaseNet achieves an overall success rate of 63.50%, outperforming the STA/LTA method, which yields 51.76%.

We further analyzed the temporal and spatial error distributions of the associated events. The results show that PhaseNet produces more concentrated error distributions, indicating greater stability. When evaluating detection success rates by magnitude, PhaseNet significantly outperforms STA/LTA for earthquakes below M5.7, whereas STA/LTA exhibits a higher overall success rate for events between M5.7 and M6.5. Case studies of typical events show that PhaseNet’s primary advantage over STA/LTA is its superior phase-picking accuracy, which underpins its markedly better detection performance for lower-magnitude events.

To further evaluate PhaseNet’s ability to identify seismic phases from distant M≥5.0 earthquakes, we extracted the maximum epicentral distances of all successfully associated events for both methods. We then calculated, within the range of these maximum epicentral distances, the ratio of successfully associated stations to the total number of stations. Overall, the maximum associated epicentral distances for earthquakes of various magnitudes can reach over 100° for both methods, with medians exceeding 89°. This distance far exceeds the near-field epicentral distances that dominate the PhaseNet training dataset, demonstrating the model’s strong generalization capability. In terms of the ratio of successfully associated stations to total stations, the PhaseNet model performs comparably to the STA/LTA method, with ratios mainly between 0.05 and 1, and a slightly higher median than STA/LTA. These results indicate that the PhaseNet model achieves performance comparable to the STA/LTA method in identifying distant earthquakes of magnitude 5.0 and above, while also highlighting its better applicability and potential in this scenario.

To assess the two methods’ ability to detect aftershock sequences, we defined events as potentially subject to mainshock interference if an earthquake of equal or greater magnitude occurred within 15 minutes and 100 km prior. We identified 75 such events. For these, both PhaseNet and STA/LTA detected only 10 events each, corresponding to a success rate of just 13.33%—substantially lower than the overall rates. Although STA/LTA generally underperforms PhaseNet overall, it appears to perform better for this specific category. After excluding these 75 events, STA/LTA detected 770 out of the remaining 1 432 events （53.77%）, while PhaseNet detected 947 （66.13%), with both methods showing improved success rates relative to the original results. Improvements were observed across all magnitude bins; however, PhaseNet still underperforms STA/LTA at magnitudes 5.9 and 6.2. PhaseNet’s relatively poor performance in aftershock settings may be attributed to its training dataset, which is predominantly composed of isolated earthquake events and lacks specific training and optimization for scenarios involving superimposed waveforms from multiple events.

Based on epicentral locations, we used the 10 km crustal thickness contour from the Crust1.0 model as a geographic divider to further classify the 1 432 events （after excluding aftershock-interfered events） into two groups: offshore events and continental/nearshore events. This yielded 548 offshore and 884 continental/nearshore events. Separate success-rate calculations for each method and category show that both methods perform significantly better for continental and nearshore earthquakes than for offshore ones. Cross-comparison reveals that PhaseNet outperforms STA/LTA by 15.72% for continental and nearshore events, a margin considerably larger than the 6.94% observed for offshore events. The magnitude-dependent success-rate distributions further confirm notable improvements for both methods across all magnitude ranges.

Notably, after excluding aftershock-interfered events, the PhaseNet model success rate for continental and nearshore earthquakes rises markedly to 82.58%. Further analysis suggests that the relatively low monitoring success rates for distant oceanic and aftershock-sequence events are mainly due to two factors: ① the vast majority of small-event samples used in PhaseNet training originate from onshore regions, and ② the model lacks sufficient dedicated training for scenarios involving overlapping seismic signals. As a result, the model has not adequately learned the seismic signal characteristics of these two earthquake types. To address this, future work should incorporate more representative samples of such events to enhance the model’s overall monitoring performance.