Spectral Spatial Classification of Hyperspectral Image by Exploiting Extended Threshold-Free Attribute Profile and Transformer-Encoder Based Network

Classification of hyperspectral images is a challenging problem. Classification methods based on both spectral and spatial information has advantage over those based on spectral values alone. Multiple methods have been proposed in the literature in this direction among which transformer-based approaches are in recent highlight. The state-of-the-art method which have used both spectral and spatial information have shown remarkable performance but is time consuming. This paper presents a two-phase classification framework in which an extended threshold-free attribute profile is constructed for the hyperspectral image and then in second phase the constructed spectral-spatial features are classified with help of a neural network which has convolutional layers as well as a series of transformer encoders. Experiments are carried out on publicly available HSI dataset considering only 30 samples per class for training. The proposed method outperformed all the state-of-the-art-methods. Therefore, the proposed method has high potential for accurate classification of HSI.