Feature Extraction with Bidirectional Encoder Representations from Transformers in Hyperspectral Images

The spectral information alone is not sufficient to achieve successful results in the classification of hyperspectral images. In studies on increasing the success of classification results, spatial information is also exploited. In this study, BERT and ALBERT models, which were recently introduced to the literature in natural language processing, have been used as feature extraction models in spatial information acquisition. These models, which try to understand the context by learning the relationship between tokens in the field of natural language processing, are used to learn the relationship between the pixel itself and its neighbors and rearrange the spectral signature. Then, spectral signatures that have been rearranged by using spatial information are trained in classifiers and models are created. The results of the BERT and ALBERT models are presented comparatively. According to the results obtained, although the size of the model is small, ALBERT produced better results in twelve of twenty tests.