Answers are grounded in this paper and cite their sources.
Hyperspectral–multispectral image fusion aims to reconstruct high-fidelity high spatial-resolution hyperspectral images by integrating the complementary information from both modalities. However, most existing methods directly perform coupled spatial–spectral modeling in the spatial domain, which easily causes mutual interference among complementary features. Moreover, unified network architectures lack the capability to adaptively model the complementary characteristics between spatial and spectral features, thereby limiting the fidelity of reconstructed results. To address these issues, we propose a wavelet-decoupled complementarity-guided expert network (WCGE-Net), whose core idea is to combine wavelet-based frequency disentanglement with mixture-of-experts (MoE) to progressively decouple, calibrate, and integrate complementary features for high-fidelity reconstruction. Specifically, the feature decomposition based on wavelet transform module first decomposes features into high-frequency and low-frequency components. Subsequently, the cross-modal feature-guided MoE module constructs spectral and spatial expert branches to calibrate high-frequency and low-frequency components under cross-modal complementary feature guidance. Furthermore, the cross-frequency interaction MoE module dynamically adjusts the aggregation weights of high-frequency and low-frequency experts through cross-frequency interaction, thereby enabling adaptive and high-fidelity reconstruction of spatial and spectral information. Extensive experiments conducted on four public datasets demonstrate that the proposed WCGE-Net consistently outperforms existing state-of-the-art methods in both quantitative metrics and visual quality.
No takes yet. Share an insight, caveat, or question.
Mou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: