Research on Two-Branch Semi-Supervised Classification Method Based on Incremental Dictionary Learning for Hyperspectral Image

Hyperspectral image contains rich spectral and spatial information. The existing hyperspectral image classification methods, which rely solely on superpixel level features, suffer from the limitation of fixed kernel functions, failing to adequately account for the spectral similarity relationships between pixels and ignoring the correlations between feature dimensions. This leads to insufficient feature extraction and inefficient utilization of data. In the classification stage, existing hyperspectral image classification methods lack a dynamic optimization mechanism for the utilization of information from unlabeled samples. To address these challenges, a two-branch semi-supervised classification method based on incremental dictionary learning for hyperspectral image is proposed, where the two branches are designed to extract pixel level features and superpixel level features. In this method, the pixel spatial distance-enhanced multi-scale convolutional neural network is used to extract pixel level features, and the U-Net with superpixel segmentation based on adaptive composition is used to extract the superpixel level features under the graph structure of different scales. Then, the above two features are fused. Finally, the features are classified by sparse graph regularization based on incremental dictionary learning. Empirical studies conducted on four standard datasets verify the superior performance of our method.