alanmorninglight / dcfsl-2021 Goto Github PK

Deep Cross-domain Few-shot Learning for Hyperspectral Image Classification

This is a code demo for the paper "Deep Cross-domain Few-shot Learning for Hyperspectral Image Classification"

Some of our code references the projects

• Learning to Compare: Relation Network for Few-Shot Learning

Requirements

CUDA = 10.0

Python = 3.7

Pytorch = 1.5

sklearn = 0.23.2

numpy = 1.19.2

dataset

1. target domain data set:

You can download the hyperspectral datasets in mat format at: http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes, and move the files to ./datasets folder.

2. source domain data set:

The source domain hyperspectral datasets (Chikusei) in mat format is available in:http://park.itc.utokyo.ac.jp/sal/hyperdata. You can download the preprocessed source domain data set (Chikusei_imdb_128.pickle) directly in pickle format, which is available in "link：https://pan.baidu.com/s/1JrCWJLmPFccfOrSh3P5QEA password：rnjv " , and move the files to ./datasets folder.

An example dataset folder has the following structure:

datasets
├── Chikusei_imdb_128.pickle
├── IP
│   ├── indian_pines_corrected.mat
│   ├── indian_pines_gt.mat
├── salinas
│   ├── salinas_corrected.mat
│   └── salinas_gt.mat
├── pavia
│   ├── pavia.mat
│   └── pavia_gt.mat
└── paviaU
    ├── paviaU_gt.mat
    └── paviaU.mat

Usage:

Take DCFSL method on the UP dataset as an example:

1. Download the required data set and move to folder./datasets.
2. If you down the source domain data set (Chikusei) in mat format,you need to run the script Chikusei_imdb_128.py to generate preprocessed source domain data.
3. Taking 5 labeled samples per class as an example, run DAFSC-UP.py --test_lsample_num_per_class 5 --tar_input_dim 103.

• --test_lsample_num_per_class denotes the number of labeled samples per class for the target domain data set.
• --tar_input_dim denotes the number of bands for the target domain data set.