The Real-time Intelligent Seismic Processing system （RISP） is a system designed to address the practical demands for earthquake cataloguing in China. It employs the deep learning method to detect earthquakes and integrates functionalities including phase arrival picking, phase association, earthquake location and magnitude measurement. This system can process continuous seismic waveforms and dense earthquake sequences in real time and offline modes. In order to evaluate the practical application effectiveness of the RISP system in Guangdong Seismic Network and facilitate system optimization, this study investigates the detection performance of the RISP system for earthquake sequences in three characteristic seismogenic regions of Guangdong Province: Heyuan, Yangjiang, and Shantou, based on continuous waveforms recorded by Guangdong Seismic Network. Through comparative analysis with the manual catalog, we systematically assess the RISP system’s capabilities across multiple dimensions, such as phase picking accuracy, magnitude distribution, and hypocenter location precision. Furthermore, this paper analyzes three categories of earthquakes recorded by the RISP system: those more frequently detected and those weakly detected, as well as events whose seismic parameters deviate significantly from the manual catalog. The evaluation metrics adopted in the analysis process are the phase detection rate, event matching rate, and event authenticity rate.

The analysis results of this paper demonstrate that the RISP system exhibits high detection rates for Pg and Sg phases in Heyuan region, both at approximately 90%. 762 earthquakes were detected, which is 3.16 times as much as the number in the manual catalog. Most of the magnitudes of the more detected events are M_L＜0. Among them, 225 earthquakes in the automatic catalog successfully matched those in the manual catalog, with high event matching rates across all magnitude ranges. Whereas the detection rates for Pg and Sg phases in Yangjiang and Shantou regions are relatively low, both below 60%. The RISP system detected 163 and 130 earthquakes in these two regions respectively, with both totals being fewer than the corresponding numbers documented in the manual catalog. Most of the magnitudes of the weakly detected events are M_L＜1.0. Within the automatic catalog, 134 and 103 earthquakes were successfully matched with the manual catalog, respectively. The event matching rates for earthquakes below M_L1.0 were low. The more detected events were manually verified, and there were eight false detection events in the automatic catalog, resulting in an event authenticity rate of 99.2%. Compared with the manual catalog, the automatic catalog produced by the RISP system in Heyuan region has smaller deviations in the origin time, epicentral location, focal depth, and magnitude. In contrast, the discrepancies between the automatic and manual catalogs in Yangjiang and Shantou regions are significantly larger, particularly in terms of epicentral location and focal depth, for which deviations are most pronounced.

Additionally, the RISP system displays a tendency to slightly underestimate magnitudes for larger earthquakes and overestimate magnitudes for smaller ones in Yangjiang and Shantou regions. On the whole, the RISP system demonstrates strong performance in Heyuan region and can produce an earthquake catalogue with high completeness and strong consistency. However, its performance is mediocre in Yangjiang and Shantou regions due to limitations in the detection model and associated algorithms, necessitating subsequent optimization and improvement.

To accelerate the operational application of the RISP system in Guangdong region and further enhance the modernization of earthquake monitoring capabilities, this paper proposes recommendations for optimizing and improving the RISP system based on the aforementioned research findings and the practical requirements of earthquake cataloguing in Guangdong. The RISP system should retrain its models using historical seismic waveforms from Guangdong seismic network or adopt transfer learning strategies to refine the models, thereby improving the phase detection rate and arrival picking accuracy. Following the completion of the National Earthquake Intensity Rapid Reporting and Early Warning Project, the RISP system must urgently enhance functions such as phase arrival picking from acceleration waveform records and earthquake event type discrimination to meet emerging earthquake cataloguing requirements and improve operational efficiency.