A hybrid graph-spatial spectral transformer framework for hyperspectral image analysis

Hyperspectral Image Classification (HSIC) is proving challenging because of the data's huge dimensionality and the intricate link between spectral and spatial information. This study introduces a new approach called the Hybrid Graph-Spatial Spectral Transformer, which aims to address these issues. This approach combines the benefits of graph-based algorithms with spatial-spectral transformers. Regarding this matter, it regards HSI characteristics as a graph structure, where each node represents a single pixel and edges denote spectral similarity. This enables complete integration of the two dimensions. The sophisticated attention mechanism in this model enables the acquisition of comprehensive spatial-spectral representations, hence improving the accuracy of class differentiation. Furthermore, the model excels at capturing extensive connections between different spectral bands, allowing for a more profound understanding of the complex relationships within hyperspectral imaging (HSI) data. The capacity to flexibly modify attention weights across various spectral resolutions enhances its flexibility and accuracy. Empirical findings using established hyperspectral imaging (HSI) datasets consistently demonstrate that this model consistently surpasses current state-of-the-art methods, leading to a substantial increase in classification accuracy.