Physically Constrained Transfer Learning Through Shared Abundance Space for Hyperspectral Image Classification

Hyperspectral image (HSI) classification is one of the most active research topics and has achieved promising results boosted by recent development deep learning. However, state-of-the-art approaches tend to perform poorly when training testing images are on different domains, e.g., source domain target domain, respectively, due spectral variability caused acquisition conditions. Transfer learning-based methods address this problem pretraining in fine-tuning domain. Nonetheless, a considerable amount data be labeled nonnegligible computational resources required retrain whole network. In article, we propose new transfer learning scheme bridge gap between domains projecting HSI from into shared abundance space based their own physical characteristics. way, discrepancy would largely reduced such that model trained could applied without extra efforts for labeling or network retraining. The proposed method referred as physically constrained through (PCTL-SAS). Extensive experimental demonstrate superiority compared state art. success endeavor facilitate deployment real-world sensing scenarios.