From the Abundance Perspective: Multi-modal Scene Fusion-based Hyperspectral Image Synthesis

Introduction

This is the source code for our paper: From the Abundance Perspective: Multi-modal Scene Fusion-based Hyperspectral Image Synthesis.

Usage

Step 0: Data preparation

Download the Chikusei dataset (HSI) from https://naotoyokoya.com/Download.html, and divide and crop the HSI into several mat files of spatial size $256\times 256\times 59$ for training the unmixing network. Put them in ./dataset/trains/ and ./dataset/evals/

Download the HSRS-SC dataset (HSI) from http://www.cjig.cn/jig/ch/reader/view_abstract.aspx?file_no=20210805, and resample the HSIs into $256\times 256\times 59$ for validation of the unmixing network. Put them in ./dataset/tests/.

Download the AID dataset (RGB) from https://hyper.ai/datasets/5446, and resize the images into $256\times 256$. Put them in ./datasets/RGB/

Step 1: Scene-based Unmixing

For training the unmixing net, change the file path and run the following code.

python 1_scene-based-unmixing.py train

After training, run the following code to infer the abundance maps of external RGB datasets.

python 1_scene-based-unmixing.py infer

After that, we can obtain the inferred abundance of RGB dataset in ./datasets/inferred_abu/.

Step 2: Abundance-based Diffusion

For training the Abundance-based Diffusion, run the following code:

python 2_abundance-based-diffusion.py -p train

After training, modify the 'resume_state' in the ./config/*.json file, and run:

python 2_abundance-based-diffusion.py -p val

After that, we can obtain the synthesized abundance in ./experiments/ddpm/\*/mat_results/.

Step 3: Fusion-based generation

Change the train_path (path of synthesized abundances) and the model_name(the trained model of the unmixing net)

Run the following code to obtain the synthetic HSIs:

python 3_fusion-based_generation.py

After that, we can obtain the synthesized HSIs in ./experiments/fusion/HSI/ and its corresponding false-color images in ./experiments/fusion/RGB/.

Citation

If you find this work useful, please cite our paper:

@article{pan2024abundance,
  title={From the abundance perspective: Multi-modal scene fusion-based hyperspectral image synthesis},
  author={Pan, Erting and Yu, Yang and Mei, Xiaoguang and Huang, Jun and Ma, Jiayi},
  journal={Information Fusion},
  pages={102419},
  year={2024},
  publisher={Elsevier}
}

Contact

Feel free to open an issue if you have any question. You could also directly contact us through email at panerting@whu.edu.cn (Erting Pan)

EtPan/MSF-Diff