A Light-Weighted Spectral-Spatial Transformer Model for Hyperspectral Image Classification

Classifying hyperspectral images in remote sensing applications is challenging due to limited training samples and high dimensionality of data. Deep-learning-based methods have recently demonstrated promising results in the classification of HSI. This article presents a proposed methodology for extracting local features and high-level semantic features from HSI input data using a light-weighted spectral-spatial transformer. This approach will allow us to comprehensively examine the spatial and spectral characteristics while reducing the computing expenses. The proposed model integrates lightweight multihead self-attention and residual feedforward modules in order to effectively capture long-range dependencies and address the computational challenges associated with this model. In order to assess the efficiency of the proposed model, we conducted experiments on four publicly available datasets. The obtained experimental results were then compared with those of the existing state-of-the-art models. The proposed model obtains the best classification results in terms of classification accuracy and computational complexity under limited training samples. The overall accuracy of the proposed model achieved 99.91, 98.06, 99.43 and 99.01 on four datasets.