Spectral-Spatial Classification of Hyperspectral Images Using BERT-Based Methods With HyperSLIC Segment Embeddings

The classification performance is highly affected because hyperspectral images include many bands, have high dimensions, and few labeled training samples. This challenge reduced by using rich spatial information an effective classifier. classifiers in this study are BERT-based (Bidirectional Encoder Representations from Transformers) models, which recently been applied natural language processing. BERT model its performance-improved version, the ALBERT (A Lite BERT) model, utilized as transformer-based models. Because of their structure, these models can also accept via ‘segment embeddings’. Segmentation algorithms commonly used literature to get information. Superpixel methods shown superior results segmentation due utility working at superpixel level rather than conventional pixel level. HyperSLIC, a modified version SLIC method for images, employed input study. In addition, HyperSLIC merged with DBSCAN algorithm similar superpixels increase size spatially areas called HyperSLIC-DBSCAN. effects segment embedding on accuracy studied experimentally. Experimental show that outperform deep learning-based 1D/2D convolutional neural network when help