Geo-spatial Information Science (Dec 2024)
Latent spectral-spatial diffusion model for single hyperspectral super-resolution
Abstract
In recent years, significant advances have been achieved in addressing super-resolution (SR) tasks for hyperspectral images, primarily through deep learning-based methodologies. Nevertheless, methods oriented toward optimizing peak signal-to-noise ratio (PSNR) often tend to drive the SR image to an average of several possible SR predictions, resulting in visually over-smoothed outputs. Furthermore, the current landscape of hyperspectral SR techniques exhibits a notable deficiency in accounting for the inherent noise complexities within the realistic data, which hinders their efficacy in real-world scenarios. To address these issues, we propose a novel latent spectral-spatial diffusion model (LSDiff) for single hyperspectral SR. The diffusion model is chosen for its remarkable generative capabilities and noise robustness. However, hyperspectral images are characterized by their exceptionally high spectral dimensions and complex spectral-spatial properties. To address this complexity, we leverage a large-scale pre-trained autoencoder to map the data into a low-dimensional space conducive to diffusion. Additionally, the incorporation of the 3DConvNext module, complemented by meticulously designed loss constraints, empowers the extraction of rich spectral-spatial features that are inherent in hyperspectral data. Subsequently, the diffusion model undergoes efficient optimization through a variant of the variational bound on the data likelihood. During the reverse transformation, LSDiff systematically converts Gaussian noise into SR images, conditioned on the low-resolution input. Comprehensive experiments provide empirical evidence of LSDiff’s clear superiority over existing state-of-the-art SR techniques. It generates images with markedly improved spatial and spectral fidelity while concurrently showcasing robustness to noise.
Keywords
