Band-wise Hyperspectral Image Pansharpening using CNN Model Propagation ( ArXiv ) introduces Rolling Pansharpening Neural Network (R-PNN), a Deep Learning method that use a simple CNN model propagation strategy for Hyperspectral Pansharpening. Starting from an initial weights configuration, each band is pansharpened refining the model tuned on the preceding one. In this way, R-PNN is able to work with hyperspectral images with an arbitrary number of bands. The proposed method has been tested on real hyperspectral images (PRISMA Dataset), both Full and Reduced resoution and then compared with several Pansharpening methods, both model-based and Deep-Learning.
If you use R-PNN in your research, please use the following BibTeX entry.
@ARTICLE{10341305,
author={Guarino, Giuseppe and Ciotola, Matteo and Vivone, Gemine and Scarpa, Giuseppe},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={Band-wise Hyperspectral Image Pansharpening using CNN Model Propagation},
year={2023},
volume={},
number={},
pages={1-1},
doi={10.1109/TGRS.2023.3339337}}
-
Giuseppe Guarino (giuseppe.guarino2@unina.it);
-
Matteo Ciotola (matteo.ciotola@unina.it);
-
Gemine Vivone;
-
Giuseppe Scarpa (giuseppe.scarpa@uniparthenope.it).
Copyright (c) 2023 Image Processing Research Group of University Federico II of Naples ('GRIP-UNINA'). All rights reserved. This software should be used, reproduced and modified only for informational and nonprofit purposes.
By downloading and/or using any of these files, you implicitly agree to all the
terms of the license, as specified in the document LICENSE
(included in this package)
All the functions and scripts were tested on Windows and Ubuntu O.S., with these constrains:
- Python 3.10.10
- PyTorch 2.0.0
- Cuda 11.7 or 11.8 (For GPU acceleration).
the operation is not guaranteed with other configurations.
- Install Anaconda and git
- Create a folder in which save the algorithm
- Download the algorithm and unzip it into the folder or, alternatively, from CLI:
git clone https://github.com/giu-guarino/R-PNN
- Create the virtual environment with the
r_pnn_env.yaml
conda env create -n r_pnn_env -f r_pnn_env.yaml
- Activate the Conda Environment
conda activate r_pnn_env
- Test it
python test.py -i example/PRISMA_example.mat -o ./Output_folder/
The easiest way for testing this algorithm is to create a .mat file. It must contain:
I_MS_LR: Original Hyperspectral Stack in channel-last configuration (Dimensions: H x W x B);I_MS: Upsampled version of original Hyperspectral Stack in channel-last configuration (Dimensions: HR x WR x B);I_PAN: Original Panchromatic band, without the third dimension (Dimensions: HR x WR).Wavelengths: Array of wavelengths (Dimensions: B x 1)
where R is the ratio of the sensor.
Please refer to --help for more details.
This project provide a set of weights obtained from a pre-training procedure described in the paper. You can directly start from this configuration using the command:
python test.py -i path/to/file.mat
where path/to/file.mat can be any dataset (with any number of bands) organized as described before.
Several options are possible. Please refer to the parser help for more details:
python test.py -h
If you want to train this network by yourself, you can do it providing in input the path of a Dataset folder that has to contain two subfolders: Training and Validation.
You can find the dataset used in this work at this link