Learning Sensor-Specific Spatial-Spectral Features of Hyperspectral Images via Convolutional Neural Networks

By Shaohui Mei, Jingyu Ji, Junhui Hou, Xu Li, Qian Du.

Introduction

The C-CNN is an unified framework for hyperspectral image classification with a single network. You can use the code to train/evaluate a network for hsi classification. For more details, please refer to our paper and our slide.

Proposed Framework

The Feature of AVIRIS sensor are as follow.

the Feature of AVIRIS sensor

Some classification results (overall accuracy) are listed below:

Datasets	SVM-RBF	Proposed C-CNN (1x1)	3x3 mean	5x5 mean	3x3 mean + std	5x5 mean + std
Indian	84.11	85.75	94.62	96.38	94.71	96.76
Salina	81.55	92.19	96.40	97.73	95.85	97.42

Citing our work

Please cite our work in your publications if it helps your research:

@ARTICLE{7919223,
    author={S. Mei and J. Ji and J. Hou and X. Li and Q. Du},
    journal={IEEE Transactions on Geoscience and Remote Sensing},
    title={Learning Sensor-Specific Spatial-Spectral Features of Hyperspectral Images via Convolutional Neural Networks},
    year={2017},
    volume={PP},
    number={99},
    pages={1-14},
    keywords={Feature extraction;Hyperspectral imaging;Image sensors;Machine learning;Principal component analysis;Sensors;Classification;convolutional neural network (CNN);feature learning;hyperspectral;spatial-spectral.},
    doi={10.1109/TGRS.2017.2693346},
    ISSN={0196-2892},
    month={},
}

Installation

Install Caffe

Please follow Caffe instruction to install all necessary packages and build it.

git clone https://github.com/BVLC/caffe.git
# Modify Makefile.config according to your Caffe installation.
cp Makefile.config.example Makefile.config
make -j8
# Make sure to include $CAFFE_ROOT/python to your PYTHONPATH.
make py
make test -j8
# (Optional)
make runtest -j8

Clone this repository.
Note: We currently only support Python 2.7

Preparation

Modify Caffe root

if necessary, modify caffe root in the code:

gedit ./src/net/find_caffe.py

Download datasets

The Indian Pines and Salina Valley will be download by the following stript:

bash ./src/get_HSI_data.sh

Preprocess

In our paper, amount of spatial information about hyperspectral image was explored. In our code, 1x1, 3x3 mean, 5x5 mean, 3x3 mean+std, 5x5 mean+std was listed to compare.

cd ./src
python data_analysis/pre_process.py

Traineval

To train CNN using the train script simply specify the parameters listed in train.py as a flag or manually change them.
```
cd ./src
python train_cnn.py
```

Training Parameter Options:

parser = argparse.ArgumentParser(description="train bn net", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--spatial_info', type=str, default='5x5_mean_std', help="'1x1_mean' | '3x3_mean' | '3x3_mean_std' | '5x5_mean'| '5x5_mean_std'")
parser.add_argument('--gpu', type=int, default=0, help='the number of gpu id, only one number is required')
parser.add_argument('--dst_dir', type=str, default='bn_net_200', help='the destination dir for the experiments')
parser.add_argument('--data_set', type=str, default='indian_pines', help='indian_pines, salina')
parser.add_argument('--max_iter', type=int, default=1000000, help='how many iters')
parser.add_argument('--train_nums', type=float, default=200, help='how many samples or how much percents for training, 200 or 0.1')
args = parser.parse_args()

Feature extraction

Use the following script to extract the supervised deep feature or sensor-specific feature of datasets acquired by AVIRIS sensor.
```
cd ./src
python feature_extract.py
```