Skip to content

README.md

Latest commit

 

History

History
175 lines (119 loc) · 6.78 KB

README.md

File metadata and controls

175 lines (119 loc) · 6.78 KB

RSCNet 🌍

PyTorch implementation of RSCNet: Representative Spectral Correlation Network for Multi-source Remote Sensing Image Classification.

This project corresponds to the paper "Representative Spectral Correlation Network for Multi-source Remote Sensing Image Classification", published in IEEE Transactions on Geoscience and Remote Sensing (IEEE TGRS), 2026.

Paper Button Dataset Button Weights Button

🔎 Overview

RSCNet is proposed for multi-source remote sensing image classification, especially for joint interpretation of HSI with SAR/LiDAR. Although HSI provides rich spectral signatures, its hundreds of contiguous bands also introduce strong spectral redundancy. At the same time, HSI and SAR/LiDAR are generated by very different imaging mechanisms, which makes direct feature fusion difficult due to the large cross-source semantic gap.

To address these two challenges, the paper introduces Representative Spectral Correlation Network (RSCNet), which tightly couples spectral band selection with cross-source feature interaction. Instead of treating spectral reduction as an isolated preprocessing step such as PCA, RSCNet allows representative spectral bands to be dynamically selected under multi-source guidance, so that discriminative class-related spectral information can be preserved during fusion.

The framework is built around two key components:

  • 🧠 Key Band Selection Module (KBSM): adaptively evaluates the importance of original HSI bands using cross-source fused features as guidance, suppressing redundant bands while preserving representative and task-relevant spectral responses.
  • 🔗 Cross-source Adaptive Fusion Module (CAFM): enhances heterogeneous feature interaction through adaptive cross-source weighting together with local-global contextual refinement, producing more discriminative fused representations for classification.

By progressively alternating representative band selection and cross-source fusion, RSCNet learns compact yet expressive multi-source features and achieves strong classification performance with relatively low computational complexity. frame

📊 Supported Datasets

The code currently supports the following dataset settings:

datasetType Dataset Modalities Classes
0 Houston2013 HSI + LiDAR 15
1 Houston2018 HSI + LiDAR 20
2 Trento HSI + LiDAR 6
3 Berlin HSI + SAR 8
4 Augsburg HSI + SAR 7

📁 Repository Structure

RSCNet/
├── code/
│   ├── task.py
│   ├── train.py
│   ├── test.py
│   ├── dataset.py
│   ├── net_RSCNet.py
│   ├── report.py
│   ├── visualization.py
│   └── parameter.py
├── model/
├── log/
├── report/
└── pic/

📦 Data Preparation

The datasets can be downloaded from:

The current code expects the data directory to be placed alongside the repository, for example:

workspace/
├── RSCNet/
└── data/
    ├── Houston2013/
    ├── Houston2018/
    ├── Trento/
    ├── Berlin/
    └── Augsburg/

Please keep the original .mat filenames used by the provided dataset package, since the paths are hard-coded in code/dataset.py.

🧩 Pretrained Weights

Pretrained model weights can be downloaded from:

Place the downloaded weights under the model/ directory, or update the paths in code/parameter.py if you want to use a different location.

🚀 Training and Testing

The main entry script is code/task.py.

  1. Move into the code directory:
cd code
  1. Edit the configuration if needed:
  • 🗂️ Set the dataset by changing datasetType in task.py
  • 💻 Set the device by changing dev, for example cuda:0
  • ⚙️ Adjust training hyperparameters in parameter.py or in myTask(...)
  1. Run training and testing:
python task.py

Current default call in task.py:

myTask(0.0001, 1, 3, dev)

This means:

  • learning rate = 1e-4
  • epochs = 1
  • dataset = Berlin (datasetType = 3)

For a full experiment, please update the epoch number and dataset setting before running.

📝 Outputs

By default, the project saves:

  • 💾 trained models to model/
  • 📜 training logs to log/
  • 📈 evaluation reports to report/
  • 🖼️ visualization maps to pic/ when visualization is enabled

💡 Notes

  • The dataset paths in the current implementation are defined with relative paths in code/dataset.py.
  • The parameter file contains several demo/default output filenames.
  • If you only want evaluation with pretrained weights, you can call myTest(...) in code/test.py after setting the correct dataset index and model path.

📚 Citation

If you find this repository useful, please cite:

@ARTICLE{11483234,
  author={Gong, Chuanzheng and Gao, Feng and Lin, Junyan and Dong, Junyu and Du, Qian},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Representative Spectral Correlation Network for Multisource Remote Sensing Image Classification}, 
  year={2026},
  volume={64},
  number={},
  pages={1-14},
  keywords={Sentinel-1;Earth Observing System;Apertures;Feeds;Antennas;Filtering;Filters;Semiconductor lasers;Semiconductor optical amplifiers;Circuits and systems;Hyperspectral image (HSI);multisource data fusion;sparse attention;spectral band;synthetic aperture radar (SAR)},
  doi={10.1109/TGRS.2026.3684877}}

📬 Contact

If you have any questions, feel free to contact us via email:

We hope RSCNet helps your research ⭐ If you find our work useful, please consider citing the paper and starring this repository.