Interpretable Cross-Sensor Hyperspectral Image Super-Resolution

doi:10.1109/tgrs.2026.3727427

Interpretable Cross-Sensor Hyperspectral Image Super-Resolution

2026-08-26

0

PRE

AI

DOI：10.1109/tgrs.2026.3727427

Original

Original request for help

Save

Abstract

En 中文

Hyperspectral image super-resolution (HSI SR) has attracted increasing research attention. Despite recent advances, existing HSI SR methods face two major limitations: 1) differences across sensors hinder the generalization of models trained on source-sensor data to unseen sensor data and 2) data-driven methods generally lack interpretability and do not sufficiently exploit the intrinsic spectral properties of HSIs. To address these limitations, we propose an interpretable cross-sensor HSI SR framework that integrates explicit spectral priors into an optimization-driven unfolding network. This method enables the trained model to be applied to unseen target-sensor data without additional retraining. Specifically, an endmember-aware dynamic modulation network (EDMN) is designed to encode compact transferable material-related priors from endmembers into adaptive bandwise weights to modulate iterative HSI SR reconstruction, improving cross-sensor generalization. Furthermore, a local spectral smoothness and spatial-gradient constraint (LSSSC) is incorporated into the unfolding network. By combining LSSSC with adaptive attention weights, the proposed network promotes spectral smoothness in homogeneous regions while mitigating excessive smoothing near boundaries, thereby improving spectral preservation. Cross-sensor experiments across three datasets show that the proposed method achieves competitive overall performance. In addition, downstream HSI classification is performed to verify model performance.

Keywords:

Cross-sensor

hyperspectral image (HSI)

interpretable network

spectral prior

super-resolution (SR)

Interpretable Cross-Sensor Hyperspectral Image Super-Resolution

Abstract

Journal

Organization

Cited Papers