A systematic literature review of shallow and deep prior modeling methods for hyperspectral-multispectral image fusion

This review examines the evolution of prior modeling in hyperspectral–multispectral (HS–MS) image fusion and compares how shallow, deep, and hybrid priors address reconstruction accuracy, interpretability, and real-world deployment challenges. A PRISMA-guided review was conducted on 62 journal studies retrieved from Scopus, Taylor & Francis Online, and Wiley Online Library. The selected studies were classified according to methodological paradigm, historical trajectory, prior type, and reported limitations. The evidence shows a clear transition from Bayesian, matrix-factorization, sparse, and tensor-based priors to CNN-, GAN-, and Transformer-based deep models, followed by model-driven and physics-guided hybrid frameworks. Shallow priors remain effective for spectral fidelity, limited data, and interpretability, whereas deep priors enhance nonlinear feature learning and spatial detail recovery. Hybrid priors increasingly combine these strengths but remain constrained by sensor uncertainty, spectral variability, and domain shift. No single paradigm fully resolves the practical challenges of HS–MS fusion. Generalizable hybrid prior modeling represents the most promising direction. Future research should embed sensor physics, uncertainty awareness, and real-scene robustness into learning-based fusion to improve cross-sensor applicability and operational deployment.