Comparison of Pansharpening Algorithms: Developing a Data-Driven Evaluation Framework

Convolutional neural network (CNN)-based methods are increasingly crucial in image processing and computer vision. In remote sensing image pansharpening, CNNs have shown excellent fusion performance, attracting significant research interest. However, challenges remain, such as the lack of detailed pansharpening dataset simulations, standardized training-testing datasets, and simple, unified coding frameworks. Additionally, there is insufficient evaluation of recent deep learning pansharpening methods and evaluation of algorithms from the perspective of different subsequent applications, limiting comprehensive evaluation. This study constructs a comprehensive evaluation system for datasets, fusion algorithms, and evaluation methods. By statistically analyzing results, improve the accuracy and richness of algorithm evaluations, enabling more effective and objective method comparisons. Used WorldView-2 satellite imagery for visual assessments, supervised and unsupervised quantitative evaluations on 100 images, and land cover classification accuracy experiments. Traditional methods like BDSD and PRACS showed strong fusion results and land cover classification accuracy, deep learning methods generally outperformed traditional ones. This research provides detailed images and data resources for pansharpening, offering valuable references for future studies in this field.