Solar filament fragments stop being stitched by distance alone: a graph network treats each fragment as a node and reaches AUC 0.867 at kNN 10
Synopsis
Solar filaments in Hα images break into fragments as they evolve, and projection effects can separate pieces of one filament while bringing pieces of different filaments close together, so distance and orientation alone cannot decide which fragments belong together. The study treats each fragment as a graph node and each candidate pair as an edge, predicts at the edge level with a graph neural network, and replaces conventional additive aggregation with multiplicative aggregation so that inconsistent neighbors contribute less. On a dataset built from about 100 Kanzelhöhe Observatory Hα images and about 1,700 fragments, the method reaches AUC 0.867 and AP 0.792 at kNN 10, above GCN at 0.832/0.747 and GAT at 0.842/0.766.
Interpretation
Fragment membership becomes an edge-prediction problem on a graph: each fragment is a node, and the probability on the edge between two fragments states whether they belong to the same filament. Earlier automated practice mainly merged fragments by a fixed distance threshold, later adding geometric constraints such as orientation, and one line of work recast fragment association as an instance-segmentation task. On a dataset of about 100 Kanzelhöhe Observatory Hα images and about 1,700 fragments, the method reaches AUC 0.867 and AP 0.792 at kNN 10, against 0.832/0.747 for GCN and 0.842/0.766 for GAT.
Node features concatenate geometric scale, normalized centroid position, skeleton length and eccentricity, and intensity normalized by the image-wide mean and standard deviation into a 13-dimensional vector, while edge features record normalized distance, coordinate differences and absolute differences in area, perimeter, skeleton length and eccentricity, seven dimensions in all. Candidate edges come from a spatial kNN with a distance constraint that removes long-range connections and cuts computation; edge labels are binary, positive within the same group. Experiments report means and standard deviations over 5-fold cross-validation at kNN 5, 10 and 15, and at kNN 10 the method's AUC standard deviation is 0.014.
Aggregation in message passing changes from summation to weighted geometric aggregation, computed in the log domain and mapped back by exponentiation, so one inconsistent neighbor lowers the overall representation instead of being cancelled by the others. GCN uses fixed normalized weights and GAT uses attention weights, but both combine neighbor information by summation, and on denser graphs irrelevant connections weaken the useful signal. Raising kNN from 5 to 10 lifts the method's AUC from 0.854 to 0.867, and at kNN 15 it falls back to 0.859 with AP 0.785, still above GCN's 0.762 and GAT's 0.747.
Perspective
The demonstrated scope is about 1,700 fragments from about 100 full-disk Hα images at Kanzelhöhe Observatory, with fragments extracted by an off-the-shelf segmentation method and manually corrected, and AUC and AP as the metrics; under the temporal split the method reaches AUC 0.819 and AP 0.751.
All methods lose performance under the temporal split, the method's AP standard deviation rises to 0.071, and the variation comes mainly from changes in solar activity and observing conditions across periods. Graph construction uses only geometric kNN and does not include solar magnetic-field constraints; multiplicative aggregation may reduce sensitivity to subtle but informative differences; node and edge features rely on hand design.