Xiangyu Rui, Xiangyong Cao, Li Pang, Zeyu Zhu, Zongsheng Yue, Deyu Meng
Pretrained diffusion model can be downloaded from
https://github.com/wgcban/ddpm-cd#arrow_forwardpre-trained-models--trainvaltest-logs
This model is not required to be retrained.
All datasets used in this work can be found in Google Driver or BaiduCloud.
These datasets can be directly used to reproduce the results.
Here are the links to the original datasets. You can download them, crop the data and generate your own datasets if interested.
Chikusei: https://naotoyokoya.com/Download.html
Houston: https://hyperspectral.ee.uh.edu/?page id=459
Pavia: https://github.com/liangjiandeng/HyperPanCollection
We generate our Chikusei and Houston datasets by 'data/generate_data.m'. Pavia can be directly downloaded from the above link for use.
run python3 demo_syn.py -res opt
Before you running the script, please first download the pre-trained diffusion model, put it to your file and change the --resume in demo_syn.py.
there are several options you can set:
-gpu: int
-dn: dataname,str. e.g. 'Chikusei_test'. The dataset should contain "HRMS", "LRMS" and "PAN".
-krtype: int. Set 0 for the first time in order to estimate kernel and srf. Set 1 if you have already save them in './estKR'.
-res: str. Set 'opt' for estimating the residual and 'no' for R=0.
Other options include eta1, eta2, scale, ks, step, accstep. Please refer to demo_syn.py.