Score-space knowledge graph propagation for generalized multi-label zero-shot classification of remote sensing images

Generalized multi-label zero-shot learning predicts seen and unseen labels from a predefined vocabulary, yet independent image-text matching leaves label relations unused. We propose Score-TEGAR, which refines frozen GeoRSCLIP scores through a typed label-relation graph. Seen-label statistics define co-occurrence, exclusion, and hierarchy; text affinity connects unseen labels. The Typed-Edge Graph Attention Refiner (TEGAR) combines relation-specific propagation, a non-positive exclusion residual, and uncertainty-aware scaling. Across MLRSNet, BigEarthNet, and China-MAS-50k, it achieved the highest mean harmonic mean of seen and unseen mean average precision (mAP-HM) among thirteen configurations: 0.614, 0.439, and 0.476. Parameter-matched controls support the contribution of typed graph structure, while label-level analysis reveals uneven benefits across unseen categories. Score-space refinement provides a route to structured prediction when candidate labels are known but training annotations are incomplete.