Class Incremental Learning Method for Hyperspectral Images Based on Real Data Playback Mechanism and Classification Network Optimization

In recent years, the hyperspectral image (HSI) classification has attracted great attention in the field of earth observation. With the expansion of application scenarios and the continuous improvement of application requirements, new classes of HSI continue to emerge. In the face of open and dynamic application scenarios, the model is required to be able to continuously learn new categories based on maintaining existing category knowledge. Class incremental learning algorithms have received extensive attention as key solutions to the problem above. In this paper, we use a regularization-based incremental learning algorithm, elastic weight consolidation (EWC) for class incremental learning for HSI classification. After data preprocessing to reduce data dimensions, we select new classes for research on class incremental learning. We combine 3D-2D convolutional structures and residual blocks as the classification network and use real data playback mechanism to improve class recognition accuracy and realize class incremental recognition of HSI data. We demonstrate the feasibility of the algorithm with extensive experiments based on three widely used hyperspectral datasets.