New framework improves remote sensing image fusion through frequency-based learning

Pan-sharpening, a vital technique in remote sensing, merges high-resolution panchromatic images with lower-resolution multispectral images to produce detailed, high-resolution multispectral outputs. It plays a crucial role in improving the balance between spatial and spectral resolution for optical remote sensing satellites. However, many existing methods struggle when applied to out-of-distribution data, as they rely on the assumption that training and testing datasets share identical distributions.

To overcome these limitations, the team developed a frequency decoupled domain-independent feature learning framework. This approach isolates domain-independent information within amplitude and phase components of images, employing frequency separation modules and learnable high-frequency filters to disentangle image details. Two specialized sub-networks then process this information, dynamically adjusting feature channels to optimize image fusion quality.

Extensive tests on multiple public datasets revealed the framework's exceptional ability to generalize across diverse data distributions. For instance, when trained on the WorldView-III dataset, the method retained high performance on the training set while outperforming other methods on generalization datasets. Visual evaluations confirmed its ability to consistently extract and apply information, ensuring robust performance despite variations in data distribution.

The team emphasized that this framework represents a significant advancement for applications requiring accurate, high-fidelity image data in diverse satellite imaging scenarios.

Research Report:Frequency decoupled domain-irrelevant feature learning for Pan-sharpening