Hyperspectral Image Classification under PDE-based Cross-sensor Transfer Learning

The application of deep neural networks in image classification often demands a substantial amount of supervised data for training to yield relatively satisfactory outcomes. However, in the task of hyperspectral image classification, given the highly specialized nature of data annotation, the existing datasets are predominantly small-sized. The dearth of a large quantity of labeled image datasets for training poses a challenge in enhancing performance. Transfer learning mitigates the requirement for effective training with large-scale labeled samples by reusing the parameters of trained models and adapting them to related new tasks. Nevertheless, from the perspective of application scenarios, the characteristic that a single type of sensor only serves a specific region at a specific time results in the extraction of only one type of feature, which performs suboptimally on diverse datasets. Hence, the research on cross-sensor transfer learning is extremely pressing. Currently, the research on cross-sensor transfer learning mainly centers on extracting consistent features through more efficient models, yet this gives rise to the issue of an excessive number of parameters. PDE enhances the model generalization ability in cross-sensor hyperspectral image classification transfer learning through its stronger capacity for extracting consistent features, significantly reducing the number of parameters.