This repository contains the implementation of the Structural Dissimilarity Compensation Graph Autoencoder (SDCGA) for multimodal change detection, specifically designed for detecting land-cover changes using heterogeneous data such as optical and SAR images.

• Introduction
• Features
• Requirements
• Installation
• Usage
• Parameters
• Results
• Example
• License
• References

Multimodal change detection involves identifying changes between images captured at different times and using different sensors (e.g., optical and SAR). The SDCGA model utilizes a graph autoencoder with attention mechanisms to compensate for structural dissimilarities between modalities, improving change detection accuracy.

• Graph-based Modeling: Constructs graphs from superpixel segmentation to capture spatial relationships.
• Attention Mechanisms: Employs Graph Attention Networks (GAT) for better feature representation.
• Structural Compensation: Compensates for structural differences between modalities using learned parameters.
• Flexible Input: Supports multimodal data such as optical and SAR images.
• Performance Metrics: Calculates metrics like Overall Accuracy, Kappa coefficient, and F1 score.

• Python 3.6 or higher
• PyTorch 1.7 or higher
• Torch Geometric
• NumPy
• SciPy
• scikit-learn
• scikit-image
• matplotlib
• imageio
• OpenCV-Python

1. Clone the repository:

   git clone https://github.com/yourusername/SDCGA.git
   cd SDCGA

2. Set up a virtual environment (optional but recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows use `venv\Scripts\activate`

3. Install the required packages:

   pip install -r requirements.txt

   Note: For torch-geometric, you may need to follow the official installation guide, as it depends on your CUDA version and PyTorch installation.

   Example installation for torch-geometric:

   pip install torch
   pip install torch-geometric

   Make sure your PyTorch and torch-geometric versions are compatible.

1. Prepare your data:

   • First-time image (image_t1): e.g., SAR image.
   • Second-time image (image_t2): e.g., optical image.
   • Reference ground truth (Ref_gt): Ground truth change map for evaluation.

   Ensure that:

   • Images are in the same spatial resolution and aligned.
   • Images are in supported formats (e.g., TIFF).

2. Modify the script or provide command-line arguments:

   • Update the paths to your images in the script or pass them as arguments.

3. Run the script:

   python sdcga.py --image_t1_path path/to/your/image_t1.tif \
                   --image_t2_path path/to/your/image_t2.tif \
                   --ref_gt_path path/to/your/ref_gt.tif

4. Adjust parameters (optional):

   You can modify parameters such as learning rate, number of epochs, etc., via command-line arguments. See the Parameters section for details.

The script accepts several command-line arguments to adjust the model and processing:

• --lr: Learning rate (default: 0.01)
• --weight_decay: Weight decay (default: 0.0001)
• --dropout: Dropout rate (default: 0.0)
• --num_heads: Number of attention heads in GATConv (default: 4)
• --hidden_dim: Dimension of hidden layers (default: 16)
• --n_seg: Approximate number of superpixels (default: 10000)
• --cmp: Compactness parameter for superpixel segmentation (default: 15)
• --beta: Penalty coefficient for delta (default: 1)
• --regularization: Regularization term coefficient (default: 0.00001)
• --k_ratio: Neighbor ratio for graph construction (default: 0.1)
• --epoch: Number of training epochs (default: 300)

Example of adjusting parameters:

python sdcga.py --lr 0.005 --epoch 500 --num_heads 8

Results
After running the script, you will obtain:

Reconstructed Images: The model reconstructs input images which can be visualized.
Change Intensity Maps: Visual representations of change intensities.
Binary Change Maps: Thresholded maps showing detected changes.
Performance Metrics: Prints Overall Accuracy, Kappa coefficient, and F1 score.
Visualizations: Plots of images, change maps, and error maps.

Example
Here’s a step-by-step example to get you started:

Prepare the data:

1.Place your images in a directory, for example:
data/
├── image_t1.tif
├── image_t2.tif
└── ref_gt.tif
2.Run the script with your data:
python sdcga.py --image_t1_path data/image_t1.tif \
                --image_t2_path data/image_t2.tif \
                --ref_gt_path data/ref_gt.tif
3.View the results:

The script will display and save various plots.
Check the console output for performance metrics.

References
If you use this code in your research, please cite:

SDC-GAE: Structural Difference Compensation Graph Autoencoder for Unsupervised Multimodal Change Detection DOI:10.1109/TGRS.2024.3396141