This is a code demo for the paper "Cross-Domain Few-shot Hyperspectral Image Classification with Cross-Modal Alignment and Supervised Contrastive Learning".

• CUDA = 12.2

• python = 3.9.18

• torch = 1.11.0+cu113

• transformers = 4.30.2

• sklearn = 0.0.post9

• numpy = 1.26.0

• source domain dataset

  • Chikusei

• target domain datasets

  • Indian Pines
  • Houston
  • Salinas
  • WHU-Hi-LongKou

You can download the source and target datasets mentioned above at https://pan.baidu.com/s/1wo9xj85YaT3JGogVyJKZTQ?pwd=5lkl, and move to folder datasets. In particular, for the source dataset Chikusei, you can choose to download it in mat format, and then use the utils/chikusei_imdb_128.py file to process it to get the patch size you want, or directly use the preprocessed source dataset Chikusei_imdb_128_7_7.pickle with a patch size of 7 $\times$ 7. In addition, for the Houston dataset in the target domain, since it has officially divided training and test sets, the data loading part is slightly different from the other three datasets. We only take k samples of each class from the training set for training, and the samples of the test set are used for testing. For this reason, we specially uploaded train-HT.py and related data loading code.

An example datasets folder has the following structure:

datasets
├── Chikusei_imdb_128_7_7.pickle
├── Chikusei_raw_mat
│   ├── HyperspecVNIR_Chikusei_20140729.mat
│   └── HyperspecVNIR_Chikusei_20140729_Ground_Truth.mat
├── IP
│   ├── indian_pines_corrected.mat
│   └── indian_pines_gt.mat
├── Houston
│   ├── data.mat
│   ├── mask_train.mat
│   └── mask_test.mat
├── salinas
│   ├── salinas_corrected.mat
│   └── salinas_gt.mat
└── WHU-Hi-LongKou
    ├── WHU_Hi_LongKou.mat
    └── WHU_Hi_LongKou_gt.mat

You can download the pre-trained model of Base Bert, bert-base-uncased, at https://pan.baidu.com/s/1C6qExEcVd3foNtLcn7PKFw?pwd=enda, and move to folder pretrain-model.

An example pretrain-model folder has the following structure:

pretrain-model
└── bert-base-uncased
    ├── config.json
    ├── pytorch_model.bin
    ├── tokenizer.json
    ├── tokenizer_config.json
    └── vocab.txt

1. Download the required source and target datasets and move to folder datasets.

• If you down the source domain dataset (Chikusei) in mat format, you need to run the script Chikusei_imdb_128.py to generate preprocessed source domain data.
• If you downloaded Chikusei_imdb_128_7_7.pickle, move it directly to the corresponding dataset directory.

2. Download the required Base Bert pre-trained model and move to folder pretrain-model.
3. Run train.py.