Simultaneous multi-class detection of interplanetary space weather events

Cataloging past space weather events, such as Interplanetary Coronal Mass Ejections (ICMEs) and Stream Interaction Regions (SIRs), is essential for both scientific research and operational applications. Firstly, it enables a comprehensive statistical analysis of their intrinsic physical and climatological properties. This, in turn, helps improve our understanding of their impact on the near-Earth space environment, including effects on human activities. It also enables the definition of event-driven space weather scenarii. Such studies benefit directly from the rapid and reproducible expansion of these catalogs, made possible by automatic event detection from in-situ time series data. Previous studies revealed the efficiency of deep-learning based methods for this task over traditional threshold-based techniques. Nevertheless, these methods have never been designed to simultaneously identify multiple event types. In this paper, we present a novel method for the multi-class automatic detection of ICMEs and SIRs. Our approach is inspired by the You Only Look Once (YOLO) family of algorithms, widely used in object detection. It works by directly identifying candidate time intervals, providing quick visual indicators of event occurrences that can be readily used by human observers in an operational space weather context. Thanks to its simple architecture, our method is easily implementable in data visualization tools and could be easily extended to additional event types. Tested on OMNI data between 1995 and 2024, the method detects at its best 644 out of 840 existing ICMEs and 1110 out of 1237 existing SIRs. 174 out of the 876 identified ICMEs and 192 out of the 1358 identified SIRs are actual False Positives for a maximal F1-score of 0.784 for ICMEs and 0.878 for SIRs. Our model performs slightly better at detecting events than other existing deep-learning based methods while achieving comparable results when estimating the events beginning and ending times. A detailed analysis of our method output shows that the great majority of detection errors are short events with a weak in-situ signature, which are expected to be among the least geoeffective event or misclassified events. We also show that the confusion made between ICMEs and SIRs is also shown to be comparable to the one actually made by human observers when manually establishing their catalogs.