Two-branch-CNN-Multisource-RS-classification

This example implements the paper Multisource Remote Sensing Data Classification Based on Convolutional Neural Network

A two-branch CNN architecture for feasture fusion with HSI and other remote scensing imagery. Reach a quite high classification accuracy. Evaluated on the dataset of Houston, Trento, Salinas and Pavia.

Prerequisites

• System Ubuntu 14.04 or upper
• Python 2.7 or 3.6
• Packages

pip install -r requirements.txt

Usage

dataset utilization

Please modify line 10-22 in data_util.py for the dataset details.

Training

1. Train HSI

python main.py --train hsi --epochs 20 --modelname ./logs/weights/hsi.h5

2. Train LiDAR/VIS

python main.py --train lidar --epochs 20 --modelname ./logs/weights/lidar.h5

3. Train two branches

python main.py --train finetune --epochs 20 --modelname ./logs/weights/model.h5

Results

All the results are cited from original paper. More details can be found in the paper.

dataset	Kappa	OA
Houston	0.8698	87.98%
Trento	0.9681	97.92%
Pavia	0.9883	99.13%
Salinas	0.9745	97.72%

Citation

@article{xu2017multisource,
  title={Multisource Remote Sensing Data Classification Based on Convolutional Neural Network},
  author={Xu, Xiaodong and Li, Wei and Ran, Qiong and Du, Qian and Gao, Lianru and Zhang, Bing},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2017},
  publisher={IEEE}
}

TODO

1. pytorch version.
2. more flexiable dataset utilization