A spectral-spatial deep learning for secondary electron hyperspectral image super-resolution

The secondary electron hyperspectral imaging (SEHI) is a promising surface analysis method, especially for nano/micro chemical materials. However, SEHI poses challenges for achieving a trade-off between low-resolution imaging over a large field of view and high-resolution imaging over a limited field of view, and avoiding potential sample alteration caused by electron beam exposure. To address this limitation, we introduce, for the first time, a spectral-spatial deep learning model (SSDL) that produces a super-resolution SEHI image from a low resolution SEHI image, while substantially reducing the risk of sample alteration. Specifically, the SSDL model incorporates a spectral squeeze-and-excitation (SE) module and a spectral spatial fusion block, enhancing its flexibility in extracting both spatial and spectral features. Further, inception-residual modules are embedded within the proposed SSDL to capture multi-scale structural features. A custom spatial–spectral loss function is introduced, integrating a mean squared error (MSE), a spatial gradient loss, and a spectral angle mapper loss, which are essential for simultaneously preserving spectral and spatial integrity. The proposed model shows high performance compared to state-of-the-art methods during testing on two real SEHI datasets. Ablation experiments further confirm the effectiveness of the spectral SE block, inception-residual modules, and the custom loss function. This work demonstrates that deep learning can greatly enhance SEHI by considerably increasing the scanning throughput while preserving high-resolution, thereby improving the practical utility of SEHI for materials science.