SemReNet: Semantic Relational Supervision for Domain Generalization in Hyperspectral Image Classification

Abstract

Remote sensing domain generalization aims to generalize hyperspectral image classifiers to unseen domains affected by sensor, imaging, and environmental shifts. Existing semantic-guided methods mainly rely on class-to-anchor alignment, independently matching each category to a textual anchor, which may depend on manually crafted descriptions and overlook interclass structure. We propose SemReNet, a semantic relation-preserving framework that replaces independent anchor alignment with class-level semantic relational supervision. SemReNet comprises a relational semantic encoder that constructs interclass semantic geometry and a relation-aware visual encoder that learns structured hyperspectral representations. A relational-semantic consistency objective regularizes the visual relation structure toward the semantic one during source-domain training, without enforcing absolute visual–text matching. Experiments on the Houston, Pavia, and HyRANK benchmarks show that SemReNet achieves overall accuracies of 84.21%, 88.78%, and 84.87%, outperforming the best competing methods by 1.88, 2.02, and 2.28 percentage points, respectively. These results demonstrate the effectiveness of semantic relational supervision for cross-domain hyperspectral classification.