| Summary: | Hyperspectral image classification using deep learning techniques has received great attention in recent years, considering the powerful spatial feature mining ability of deep learning. Fully convolutional network is an effective deep learning architecture that exploits spatial contextual information through a hierarchical convolutional structure. However, it often ignores the relationships between samples of the same category and different categories within the global context. Therefore, a fully convolutional spectral–spatial fusion network based on supervised contrastive learning (FCSCL) is proposed for hyperspectral image classification to enhance the separability between different categories and class aggregation among the same category. In the FCSCL framework, the spectral–spatial fusion classification network is developed to capture subtle spectral variations and spatial patterns by adaptively fusing the features extracted by the spectral branch and spatial branch. To improve intraclass compactness and interclass separability, the SCL module is integrated into the FCSCL framework. The positive and negative sample pairs are constructed by the designed hard example pairs sampling strategy. These constructed sample pairs are used to guide the network to learn more discriminative feature representations that pixels of the same category are closer to each other and pixels of different categories are pushed further apart in the feature space. The experiments using three public hyperspectral datasets verify the effectiveness of the FCSCL algorithm, and the FCSCL method achieves better classification performance.
|
|---|