Skip to content
Open access

Interpretable Dictionary Transfer-Based Spectral Super-Resolution for Similar-Scene Multispectral Images

2026 · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · Vol 19, pp. 32429-32443 · 0 citations · 47 references

Abstract

Deep learning-based transfer learning has been widely used in spectral super-resolution, which can transfer the general spectral mapping relationship learned from a pair of multispectral (MS) and hyperspectral (HS) images to other similar-scene MS images, to realize spectral super-resolution from one single MS image to the desired HS image. However, deep learning-based transfer learning requires large amounts of data for fine-tuning and suffers from poor interpretability, which may considerably reduce the effectiveness of spectral super-resolution. To alleviate this limitation, this article proposes an interpretable dictionary transfer-based spectral super-resolution method with closed-form solutions for similar-scene multispectral images, tries to adjust the prelearned spectral dictionary to fit with different task domain and to reconstruct a more accurate desired HS image. Specifically, first, taking typical spectral set in the source domain and similar spectral set in the task domain as a connection, a transfer matrix optimization strategy is proposed for the prelearned spectral dictionary. Second, an orthogonal matrix-based dictionary transfer scheme is established, to transfer each column of the spectral dictionary prelearned from the MS and HS image pairs, into the task domain of the observed similar-scene MS image for spectral super-resolution. Finally, the coefficient matrix of the transferred spectral dictionary is optimized as closed-form solutions, by using spatial information of the observed similar-scene MS image and the spectral similarity of the extracted spectral sets under sparse and low-rank constraints. Experimental results on both simulated and real datasets indicate that, our proposed method achieves superior performance and generalization ability across different scenarios, compared with the other state-of-the-art spectral super-resolution methods based on both deep learning and dictionary learning.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.