Graph--Based Event Fingerprints for Classifying Geomagnetic Storm--Driven Forbush Decreases

Abstract

Forbush decreases (FDs) are transient depressions in the galactic cosmic-ray flux observed by global neutron-monitor networks and are commonly associated with interplanetary disturbances driven by coronal mass ejections and related shocks. Despite extensive observational work, quantitatively comparing FD morphology across events and linking it to storm severity remains challenging due to heterogeneous station responses, coverage gaps, and the multivariate nature of the network. This work introduces a graph-based event representation in which each FD is mapped to an event network constructed from pairwise dissimilarities between station response time series. A controlled sparse backbone is obtained via the minimum spanning tree, enabling comparable event graphs across cases. From each graph, a compact set of geometric/topological fingerprints is computed, including global integration measures, spectral summaries, mesoscopic structure, centrality aggregates, and complexity descriptors. Predictive skill is assessed using strict leave-one-event-out validation over a pre-defined grid of distance metrics and distance-domain transformations, with selection criteria fixed \emph{a priori}. The proposed fingerprints exhibit measurable signal for three tasks: (i) multi-class classification of geomagnetic storm intensity (G3/G4/G5) with moderate but consistent performance and errors dominated by adjacent categories; (ii) stronger binary severity screening ($\ge$G4 vs.\ G3) with high sensitivity to severe events; and (iii) drop regression with partial least squares achieving positive explained variance relative to a fold-wise mean baseline.

Figures (4)

• Figure 1: Illustrative two event graph for 2012-03-08 (G3) and 2023-04-23 (G2) from a representative Forbush decrease. Nodes are stations colored by cutoff rigidity (GV). Edges correspond to the MST computed from the event-specific dissimilarity matrix, yielding a sparse, connected backbone with $N-1$ edges that minimizes total distance. Left: Manhattan ($p=1$); right: Minkowski/Euclidean ($p=3$). While the main station groupings persist, the MST topology (branching patterns and connector nodes) changes with the distance metric.
• Figure 2: Rigidity-conditioned distributions of node-role summaries across events. For each FD event graph, node-level metrics are aggregated within three cutoff-rigidity bands defined by fixed thresholds: low ($R_c < 3$ GV), medium ($3 \le R_c < 6$ GV), and high ($R_c \ge 6$ GV). Boxplots summarize the distributions across events. Low-rigidity stations show systematically larger avg. katz, avg. closeness, and avg. betweenness, indicating more central and bridging roles in event backbones, while the Laplacian-based summary peaks for the medium-rigidity group.
• Figure 3: LDA projection (LD1--LD2) for the best multi-class intensity pipeline (log distance transform, no normalization and Minkowski adjacency). Colors indicate storm classes (G3/G4/G5).
• Figure 4: PLS (5 components) LOEO regression: predicted vs. observed Forbush drop (%) for the best pipeline (no transform; no normalization; Minkowski adjacency). The dashed line indicates perfect agreement.