SSCAConv

Source Code and Datasets for "SSCAConv: Self-guided Spatial-Channel Adaptive Convolution for Image Fusion"

Source Codes and Datasets for Hyperspectral Pansharpening in Deep Learning

Homepage: Liang-Jian Deng, Xiaoya Lu and Yu-Wei Zhuo
Code for the paper: "SSCAConv: Self-guided Spatial-Channel Adaptive Convolution for Image Fusion, IEEE GRSL, 2023" [paper]

Reference

Please cite the related paper:

@ARTICLE{2023lu,
  author={Lu, Xiaoya and Zhuo, Yu-Wei and Chen, Hongming and Deng, Liang-Jian and Hou, Junming},
  journal={IEEE Geoscience and Remote Sensing Letters}, 
  title={SSCAConv: Self-guided Spatial-Channel Adaptive Convolution for Image Fusion}, 
  year={2023},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/LGRS.2023.3344944}}

Dependencies and Installation

Python 3.8 (Recommend to use Anaconda)
Pytorch 1.7.0
NVIDIA GPU + CUDA
Python packages: pip install numpy scipy h5py
TensorBoard

Dataset

Please find the dataset shared by our team at [PanCollection]
More info. about these datasets can be found from the published paper.

Download by `Baidu Cloud`

Use [Baidu Cloud] to download these datasets.
Pansharpening datasets used in this paper are WV3 and GF2. Here we provide complete training and testing data for download.

Dataset Link Password

WV3 [download link] p6hx

GF2 [download link] vbj6
HISR task Harvard datasets are used for HISR task. Also, we provided the prepared data used in this paper.

Dataset Link Password

Harvard [download link] 8e2i

Dataset	Link	Password
WV3	[download link]	p6hx
GF2	[download link]	vbj6

Dataset	Link	Password
Harvard	[download link]	8e2i

Code

data.py: The dataloader of the training and testing data.
train.py: The main training function of our SSCANet.
SSCANet.py: The whole model of our SSCANet.
test.py: The main testing function of our SSCANet.

Get Started

For training, you need to set the file_path in the main function, adopt to your train set, validate set, and test set as well. Our code trains the .h5 file, you may change it through changing the code in data function.
After prepareing the dataset, you can modify the model and experiment hyperparameters as needed, such as epoch, learning rate, convergence function, etc.
At the same time, you also need to set the path where the model and log are saved.
Then you can start training, the code will automatically save the trained model in .pth format.
As for testing, you also need to set the path to open and load the data and trained .pth file, and get the test result in .mat format.

Method

Motivation
Proposed module
Overall Architecture

Visual Results
Quantitative Results
Please see the paper for other details.

Slupplementary experiment resluts

Diffusion model, as a newly developed generative technique, has really drawn much attention. For a more comprehensive and rigorous comparison, we have implemented a newly published diffusion model, i.e., PanDiff [R1], according to your suggestion. Specifically, we compare our SSCANet and PanDiff model on the reduced-resolution WV3 dataset, and the total timesteps T is set to 2000 for PanDiff. Table S1 shows the corresponding quantitative results, in which our model achieves the favorable performance with fewer network parameters and inference time compared to PanDiff. Besides, the qualitative visual results are presented in Fig. S1. Due to the 5-page limit for this letter, we have provided supplementary experiment results in our online repository.

[R1] Q. Meng, W. Shi, S. Li and L. Zhang, "PanDiff: A Novel Pansharpening Method Based on Denoising Diffusion Probabilistic Model," in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-17, 2023, Art no. 5611317, doi: 10.1109/TGRS.2023.3279864.

Contact

We are glad to hear from you. If you have any questions, please feel free to contact Yuuweii@yeah.net.

Pluto-wei/SSCAConv