<p>Hyperspectral image (HSI) classification plays a vital role in remote sensing by enabling precise material identification and land-cover mapping through rich spectral information. However, its performance is hindered by high spectral dimensionality causing redundancy, spatial complexity caused by diverse landform textures, spectral mixing within pixels due to limited spatial resolution, and insufficient labeling samples that restrict model generalization. Existing hypergraph-based methods extract features from HSI but are often shallow and lack mechanisms to refine hypergraph structures, limiting their representation ability. To achieve high-accuracy classification under these conditions, we propose a framework called HyCoReNet that models the HSI as superpixels, which are generated using a semantic-aware segmentation network. A multi-strategy hypergraph generation approach is further designed to integrate spatial adjacency, spectral similarity, and spectral unmixing information. HyCoReNet consists of a four-layer hypergraph convolutional network, during which the hypergraph is compressed and progressively reconstructed through the key vertices selection and the discarded vertices reintroduction mechanism. To preserve pixel-level details, a parallel convolutional neural network branch extracts multi-scale features. Features from both branches are fused and fed into a classifier for final prediction. Experiments on benchmark datasets demonstrate that the proposed framework achieves superior accuracy, effectively modeling complex hyperspectral images.</p>