Spatial-Spectral Attention for Geological Mapping of Hyperspectral Sensor on Board Chandrayaan-2 Mission

Lunar hyperspectral remote sensing is one of the most important means to understand the mineralogy mapping of the lunar surface. Hyperspectral (HS) images are characterized by hundreds of channels of reflectance data from multiple bands across the Electro-magnetic spectrum, enabling the fine identification of materials by capturing subtle spectral discrepancies. To overcome the positional encoding issues that are inherent in transformer architecture hyperspectral data, we propose to use an encoder only based transformer network with a novel module - spatial positional encoding (SPE) layer and apply it on lunar hyperspectral image data i.e the Cuprite dataset. This work also compares the novel module with the state-of-the-art neural network models in the hyperspectral image classification domain. Then, we apply the novel architecture on the lunar surface data i.e Moon mineralogy mapper data and IIRS data.