Cascadic Multi-Receptive Learning for Multispectral Pansharpening

How to use?

• Directly run: test.py for the single WV3 example
• Directly run: test_mulExm.py for the multiple WV3 examples

Citation

@ARTICLE{10308614,
  author={Wang, Jun-Da and Deng, Liang-Jian and Zhao, Chen-Yu and Wu, Xiao and Chen, Hong-Ming and Vivone, Gemine},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Cascadic Multi-Receptive Learning for Multispectral Pansharpening}, 
  year={2023},
  doi={10.1109/TGRS.2023.3329881}}

Motivations

1. CML-resblock

A CML-resblock is proposed to extract information from different scales in a step-by-step manner. Specifically, every pixel of the output is able to perceive multi-scale information through a cascade-like connection strategy, which is an efficient and effective multi-receptive learning process.

2. Multiplication network structure

Inspired by the traditional multiplicative injection model for pansharpening, we design the novel multiplication network structure to learn the coefficients of the restoration mapping.

Datasets

1. training datasets

PanCollection

2. testing datasets

Pan.baidu

Results

1. Quantitative results

a. Single example

b. Multiple examples

2. Visual results

a. Reduced resolution

b. Full resolution