A novel multi-class land use/land cover classification using deep kernel attention transformer for hyperspectral images

Hyperspectral imaging is a prominent land use land cover (LULC)classification technology. However, due to fewer training samples, LULC classification using hyperspectral images remains complicated and labour-intensive. We have presented a Deep Kernel Attention Transformer (DKAT) to overcome these issues in classifying Land Use Land Cover classes. Before classifying the land cover, t-Distributed Stochastic Neighbouring Embedding (t-SNE) is exploited to extract the features from the LULC by applying the probability distribution function. To quantify the resemblance among the two points Kull Burk-Divergence (KL) is employed. Then, a searching-based band selection method is used to select the bands. The grey wolf optimization (GWO) technique is used in the searching-based band selection method to determine the informative bands. After choosing the informative bands from the hyperspectral data cube, we must classify the land cover. Experimental results are conducted by using five publicly available benchmark datasets. They are Indian Pines, Salinas, Pavia University, Botswana, and Kennedy Space Center. The classification accuracy is calculated using the overall accuracy, average accuracy, and kappa coefficient; we have achieved 99.19% overall accuracy, 99.32% average accuracy, and 99.14% kappa coefficient.