RgB2CASI dataset for traning and test

RgB2CASI dataset comes from the data set of 2018 IGARSS Data Fusion Contest acquired by CASI, which includes a hyperspectral image with 48 bands and a corresponding RGB image. Similar to RgB2CAVE dataset, the Green channel of the original RGB image is also spatially downsampled in RgB2CASI data set. Thus, for this data set, the goal of generalized spectral super-resolution is to restore a 48-band hyperspectral image from the degraded RGB images consist of a Red channel, a Blue channel, and a low-resolution Green channel.

The file named "RgB2CASI_train.h5" contains training samples where the high-resolution Red and Blue channel is with size of 128×128 while the size of Green channel is 64×64. Note that, data augmentation has also been implemented to enlarge the dataset eightfold. So there are 1904 samples. Moreover, we also give some test images in the "RgB2CASI_test.mat". Details about generalized spectral super-resolution can be found in the following paper.

***Notes*** The file named "RgB2CAI_train.h5" is the same as "RgB2CASI_train.h5", which is uplaoded twice by mistake and cannot be deleted in Zenodo!!!

Paper: J. He, Q. Yuan, J. Li, and L. Zhang, "PoNet: A universal physical optimization-based spectral super-resolution network for arbitrary multispectral images," Information Fusion, vol. 80, pp. 205–225, 2022.

More information about the author can be found at https://jianghe96.github.io/

If this dataset is helpful please cite as:

@article{He2022PoNet,
title={PoNet: A universal physical optimization-based spectral super-resolution network for arbitrary multispectral images},
author={He, Jiang and Yuan, Qiangqiang and Li, Jie and Zhang, Liangpei},
journal={Information Fusion},
volume={80},
pages={205--225},
year={2022},
}