SPARC-Net: Spectral-Preserving Amplitude-Phase and Reliable Correction Network for Long-Tailed Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification presents significant challenges due to the high dimensionality of spectral data and the long-tailed distribution of available samples. Current methods often employ principal component analysis or related preprocessing techniques to reduce spectral dimensionality, which may disrupt spectral continuity and weaken frequency-aware representation learning. Furthermore, most existing long-tailed HSI classification methods primarily focus on loss function design or classifier adjustment, while the interactions among representation learning, prototype modeling, training strategy, and output calibration remain insufficiently explored. To address these limitations, we propose SPARC-Net, a Spectral-Preserving Amplitude-Phase and Reliable Correction Network for long-tailed HSI classification. Its core component is a Spectral-Preserving Amplitude-Phase (SPAP) Backbone, which p'reserves the original spectral order, jointly models amplitude-phase representations, spatial high-frequency information, and spatial-spectral frequency interactions, and constrains feature distortion during representation learning. For long-tailed decision learning, a Main-anchored Reliable Prototype Correction (MRPC) Head retains a cosine classifier as the primary decision branch and employs reliability-aware dual-anchor prototypes solely for gated and bounded auxiliary correction. A main-branch-first staged training strategy and Reversible Tail-Prior Calibration (RTPC) further stabilize prototype learning and mitigate residual head-class bias during inference. Experiments on four datasets, including controlled comparisons with Mamba-based classifiers, comparisons between PCA and original-band inputs, head/medium/tail group evaluations, and sensitivity analyses under varying imbalance ratios, demonstrate competitive performance and improved tail-class reliability. These results provide complementary evidence for the effectiveness of the SPARC-Net for long-tailed HSI classification.