Boosting Hyperspectral Image Classification via Class-Difficulty-Aware and Spectral–Spatial Feature Representation

Abstract

Hyperspectral images (HSIs) provide rich spectral information, offering unique advantages for fine-grained land-cover classification. However, HSI classification remains challenged by insufficient spectral–spatial feature exploitation and significant variations in class difficulty under limited labeled samples. To address these issues, this letter proposes a classification framework that integrates spectral pixel-map feature extraction with class-difficulty-aware adaptive completion (CDAC). A dual-branch pyramid network (DBPN) is designed, where the spectral branch reshapes spectral vectors into pixel maps to capture multiscale spectral dependencies, while the spatial branch is deeply fused through a cross-attention mechanism. Furthermore, a CDAC mechanism is developed to adaptively allocate completion quotas according to class difficulty and select reliable pseudolabeled samples using multiple reliability indicators, thereby providing targeted compensation for underperforming classes. Experiments on three public hyperspectral datasets demonstrate the effectiveness of the proposed method and the complementary benefits of its components.