A mixed-attention deep autoencoder framework for denoising lunar hyperspectral images

In this paper, a deep learning (DL)-based denoising of lunar hyperpsectral image (HSI) framework is presented in which autoencoder (AE)-based mixed attention mechanism (MAAE) is proposed to effectively remove noise and achieve spectral fidelity. Our work used lunar hyperspectral data that was obtained by Chandrayaan-2 imaging infrared spectrometer (IIRS). Three architectures were constructed and compared that includes baseline AE, attention-enhanced AE, and proposed MAAE, which combined spatial and spectral correlation with both baseline as well as simple attention AE. The denoising process implies that the models learn to reproduce clean reflectance spectra using noisy inputs, and the mixed-attention component of the model is supposed to focus on emphasizing meaningful spectral-spatial features in a dynamic manner. The results of the proposed models were evaluated with respect to the traditional algorithms like block-matching 3D (BM3D) and deep image prior (DIP) using four commonly used quantitative metrics-the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), spectral angle mapper (SAM), and spectral information divergence (SID). The experimental results show that the MAAE has advantage over comparative methods on three hyperspectral data strips of lunar with the high PSNR (maximum of 75.07 dB), SSIM (0.63) and minimum SAM (0.11), SID (0.01) values indicating proposed MAAE outperforms other comparative techniques in noise reduction and spectral integrity. Our work confirms the effectiveness of mixed-attention process in hyperspectral denoising and provides a powerful baseline of DL to analyze and map minerals using a planetary surface in remote sensing applications.