whuzhusihan96 / scsn Goto Github PK

A deep learning method for remote sensing image cross-scene generalization.

• python 3
• pytorch 1.10 or above

Four scene classification datasets are used in our experiments: AID, CLRS, MLRSN, and RSSCN7. Since they are all public benchmarks, they will not be provided here. Just get them by yoursef.

For more information on the scene classification (or more tasks) dataset, check out this paper.

./data/xxx.txt is to show samples in our experiment. This is not the way our code reads data.

If you want to use your own dataset, please organize your data in the following structure.

RootDir
└───Domain1Name
│   └───Class1Name
│       │   file1.jpg
│       │   file2.jpg
│       │   ...
│   ...
└───Domain2Name
|   ...    

And then, modifty util/util.py to contain the dataset.

The specific image suffixes supported can be referred to torchvision.datasets.ImageFolder.

1. Modify the file in the scripts.

2. The main script file is train.py, which can be runned by using run.sh from scripts/run.sh.

It is easy to design your own method following the steps:

1. Add your method (a Python file) to alg/algs, and add the reference to it in the alg/alg.py.

2. Modify utils/util.py to make it adapt your own parameters.

3. Modify scripts/run.sh and execuate it.

Here we provide two results, more can be found in our paper.

IMP and RSP denote ImageNet pre-training and Remote Sensing pre-training.

Great thanks to DomainBed and DeepDG. And our code is mainly based on DeepDG.

@article{ZHU20231,
title = {Style and content separation network for remote sensing image cross-scene generalization},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
volume = {201},
pages = {1-11},
year = {2023},
issn = {0924-2716},
doi = {https://doi.org/10.1016/j.isprsjprs.2023.05.007},
url = {https://www.sciencedirect.com/science/article/pii/S0924271623001247},
author = {Sihan Zhu and Chen Wu and Bo Du and Liangpei Zhang}
}