Pytorch implementation of the paper "Single-stream CNN with Learnable Architecture for Multi-source Remote Sensing Data".

Multi-stream CNNs are commonly used in multi-source remote sensing data fusion. In this work we propose an efficient strategy that enables single-stream CNNs to approximate multi-stream models using group convolution. The proposed method is applied to ResNet and UNet, and evaluated on Houston2018, Berlin, MUUFL data sets, obtaining promising results. An interesting finding is that regularization is playing an important role in these models.

Find our paper at: [IEEE Xplore] [arxiv]

• Requirements: python3, pytorch, gdal, sklearn.
• Simply run

python3 main.py

• To customize training/model arguments, modify common.py. Arguments are automatically loaded to main.py.

This repository also contains Pytorch implementation of the following models, which we use as baselines:

• Fusion-FCN: A three-branch CNN for MS-HSI-LiDAR data fusion. Award-winning model in 2018 IEEE DFC. [Paper]: "Multi-Source Remote Sensing Data Classification via Fully Convolutional Networks and Post-Classification Processing"

• Two-branch CNN (TB-CNN): A two-branch CNN architecture for feasture fusion with HSI and other remote scensing imagery. [Paper]: "Multisource Remote Sensing Data Classification Based on Convolutional Neural Network" [Official Tensorflow implementation]

Implementation of these models can be found at model/baseline/.

We made some modifications (merely tif→numpy, stacking) to the original data files. Our data files are available at this Google Drive site, which can be directly used in this code. Please note that we used channel-wise normalization AFTER loading these files, and this step is already implemented in our code.

Below are links to the original data sets:

[Houston2018]   [Berlin]   [MUUFL]  

If you find our work helpful, please kindly cite:

@ARTICLE{9761218,
  author={Yang, Yi and Zhu, Daoye and Qu, Tengteng and Wang, Qiangyu and Ren, Fuhu and Cheng, Chengqi},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Single-Stream CNN With Learnable Architecture for Multisource Remote Sensing Data}, 
  year={2022},
  volume={60},
  number={},
  pages={1-18},
  doi={10.1109/TGRS.2022.3169163}}