Attention-Infused 3D-SRCNN for Hyperspectral Image Super Resolution

In this paper, a new approach for Hyperspectral Image Super Resolution (HSI-SR) using a combination of 3D Convolutional Neural Networks (3DCNNs) and an attention mechanism, mainly Squeeze-and-Excitation (SE), is proposed. The devised method aims to generate High Resolution HSI (HRHSI) from a single Low Resolution HSI (LR-HSI), an approach known as Single Image SR (SISR), which is a challenging task in remote sensing applications. 3D-SRCNN is utilized to extract and learn the spatial and spectral features of the input image, while the SE mechanism is employed to enhance the network's ability to model the interdependencies between the spectral bands. The proposed model is evaluated on ROSIS Pavia University and AVIRIS Botswana HSI datasets. Experimental results demonstrate that the proposed 3D-SE-SRCNN outperforms other methods in terms of both quantitative metrics and visual quality. The implementation of the proposed model is provided in this repository: https://github.com/NourO93/SISR_Library