ZSBP: Zero-Shot Semi-Supervised Learning for Blind Pansharpening

Pansharpening aims to generate high-resolution multispectral images (HRMS) by effectively leveraging the complementary information in high-resolution panchromatic images (HRPan) and low-resolution multispectral images (LRMS). While recent deep learning-based methods have achieved promising results, they typically require large-scale training datasets and are often tailored to specific satellite sensors, limiting their generalization ability across different platforms. Moreover, at full resolution, accurate modeling of the degradation processes from HRMS to HRPan and from HRMS to LRMS remains a significant challenge, which directly affects the reliability of pansharpening in real-world applications. To address these issues, we propose a zero-shot semi-supervised learning framework for blind pansharpening, termed ZSBP. Our method performs both training and testing on a single pair of HRPan and LRMS, removing the dependency on large-scale datasets and mitigating the limitations of sensor-specific designs. Furthermore, we introduce two dedicated networks to model the degradation processes. By jointly training at both reduced-resolution and full-resolution, our method achieves stable and robust fusion. Extensive experiments demonstrate that our method outperforms state-of-the-art methods in both visual quality and quantitative metrics.