Deep feature dendrite with weak mapping for small-sample hyperspectral image classification

The article introduces a novel neural network model called Deep Feature Dendrite (DFD) for small-sample Hyperspectral Image (HSI) classification. HSI classification is crucial for analyzing and interpreting rich spatial and spectral information in HSI data, but it faces challenges such as high dimensionality, redundancy, and the scarcity of labeled samples. Traditional methods and deep learning approaches have limitations in fully exploiting spectral-spatial features and require substantial labeled samples. The DFD model addresses these issues by combining deep feature extraction with a residual dendrite network for weak mapping. The deep feature extraction part uses convolutional layers, tokenizers, and attention encoders to effectively extract spatial-spectral features and semantic tokens from HSI data. The controllable mapping part employs a residual dendrite network to perform weak mapping, enhancing the model’s generalization ability. Experiments on four standard datasets demonstrate that the DFD model achieves higher classification accuracy and better generalization compared to other methods, highlighting its potential for practical applications in fields such as environmental monitoring and remote sensing.