Adaptive threshold-guided multi-scale convolution and multi-order graph feature interaction for hyperspectral image classification

With the continuous advancement of deep learning techniques, hyperspectral image (HSI) classification has achieved significant progress. In particular, the ongoing optimization of convolutional neural networks (CNNs) has greatly enhanced their performance. However, conventional CNNs typically rely on fixed-size convolutional kernels, which may struggle to capture features with varying scales or shapes effectively. This limitation can hinder the overall accuracy and robustness of HSI classification. To address this issue, we propose an adaptive threshold-guided multi-scale convolution and multi-order graph feature interaction network, referred to as AT-MCGFI Net, to further enhance HSI classification performance. The framework extracts discriminative features at both pixel and superpixel levels. At the superpixel level, a low-to-high-order spectral graph convolution module is introduced to learn non-Euclidean representations across multiple scales and structural variations. At the pixel level, a multi-scale differential filtering module (MDFM) is designed to perform dual-path feature extraction, capturing complementary spatial–spectral patterns. To refine and emphasize informative features, a multi-head grouping attention module is applied. Finally, an adaptive threshold-guided feature cross-attention module (ATFCM) enables effective interaction between pixel-level and superpixel-level features, facilitating comprehensive feature fusion. Through this hierarchical and multi-level feature learning strategy, AT-MCGFI Net is able to capture more informative and representative features, thereby significantly benefiting HSI classification. Experimental results on benchmark datasets demonstrate that the proposed AT-MCGFI Net outperforms existing state-of-the-art methods.