Frequency-Relation Guided Historical Inversion for Raw-Exemplar-Free Continual Hyperspectral Image Classification

Raw-exemplar-free continual hyperspectral image (HSI) classification aims to learn newly introduced land-cover classes while retaining old-class discrimination without storing historical HSI patches. In this work, raw-exemplar-free denotes a compact-memory-assisted setting in which no historical HSI patches, pixel-level spectral-spatial tensors, spatial coordinates, or persistent generated samples are retained; only compact non input statistics and normalized feature references are used for reconstruction guidance and memory calibration. The main challenge is to construct historical evidence that is structurally faithful, discriminative, and consistent with the evolving feature space. To address this challenge, this paper proposes a Frequency-Relation Guided Structured Historical Inversion framework, termed FR-SHI. Frequency-Relation Guided Historical Reconstruction (FRHR) reconstructs old-class auxiliary samples by combining frequency structural constraints with relation-memory guidance. Old-New Collaborative Relational Distillation (OCRD) preserves both output responses and channel-spatial feature relations of the previous model, while Temporal Drift Suppression and Calibration (TDSC) suppresses representation drift and calibrates compact historical memories after each task. Experiments on Houston, Salinas, and WHU-Hi LongKou show that FR-SHI provides more stable old-class retention and better final classification performance than representative raw-exemplar-free continual learning methods. Compared with the closest raw-exemplar-free HSI-specific inversion baseline, HyperSC, FR-SHI improves final OA by 3.58, 0.88, and 3.91 percentage points on Houston, Salinas, and WHU Hi-LongKou, respectively.