Spatial-Spectral-Associative Contrastive Learning for Satellite Hyperspectral Image Classification with Transformers

Albeit hyperspectral image (HSI) classification methods based on deep learning have presented high accuracy in supervised classification, these traditional required quite a few labeled samples for parameter optimization. When processing HSIs, however, artificially are always insufficient, and class imbalance limited is inevitable. This study proposed Transformer-based framework of spatial–spectral–associative contrastive to extract both spatial spectral features HSIs by the self-supervised method. Firstly, label information generated spatial–spectral augmentation transform entropy. Then, Transformer modules used learn high-level semantic domain domain, respectively, from which cross-domain fused associative Finally, we design classifier Transformer. The invariant distinguished properties satellite further discriminant between different pixels, intersection over union imported into loss function avoid collapse caused imbalance. Conducting experiments two HSI datasets, this verified performance model. results showed that model can effective more accurate compared with model, especially average various classifications.