Abstract: Although deep learning-based pansharpening methods for remote sensing images have made certain progress, most rely on supervised training with downsampled data, which is susceptible to scale bias and struggles to maintain stable performance at full resolution. In contrast, unsupervised methods optimize directly on full-resolution images, avoiding issues caused by downsampling, but generally exhibit weaker robustness due to the lack of explicit supervisory signals. To address this, a pansharpening Network based on Two-stage Collaborative Optimization (TCONet) was proposed. The first stage involves supervised training on downsampled data, combining an improved multi-resolution analysis method with an attention mechanism to improve spatial detail and spectral preservation capabilities. The second stage constructed an unsupervised information compensation network, optimizing directly on full-resolution images to compensate for information loss caused by scale inconsistency. Experimental results indicate that TCONet outperforms other comparative methods in terms of both visual quality and evaluation metrics on three remote sensing datasets.

Key words: pansharpening, supervised learning, unsupervised learning, image fusion, remote sensing image processing

摘要： 尽管基于深度学习的遥感图像全色锐化方法取得了一定进展，但多数依赖于降采样数据进行监督训练，易受到尺度偏差影响，难以在全分辨率条件下保持性能稳定；相比之下，无监督方法可直接在全分辨率图像上优化，规避了降采样引发的问题，但由于缺乏明确的监督信号，其鲁棒性普遍较弱。为此，本文提出一种两阶段协同优化的全色锐化网络(TCONet)。第一阶段通过在降采样数据上进行监督训练，结合改进的多分辨率分析方法与注意力机制，提升空间细节与光谱保持能力；第二阶段构建无监督信息补偿网络，直接在全分辨率图像上优化，弥补尺度不一致带来的信息缺失。实验结果表明，TCONet在三个遥感数据集上的视觉效果与评价指标方面均优于其他对比方法。

关键词: 全色锐化, 监督学习, 无监督学习, 图像融合, 遥感图像处理