ISPRS P&PS 2025: CGSL: Commonality Graph Structure Learning for Unsupervised Multimodal Change Detection

This repository provides the official implementation of the paper:

Jianjian Xu, Tongfei Liu*, Tao Lei, Hongruixuan Chen, Naoto Yokoya, Zhiyong Lv, and Maoguo Gong[J]. CGSL: Commonality Graph Structure Learning for Unsupervised Multimodal Change Detection, ISPRS Journal of Photogrammetry and Remote Sensing, 2025, 229:92-106.

---

🚀 Features

• ✅ Unsupervised multimodal change detection
• ✅ Graph-based commonality structure learning
• ✅ Siamese graph encoder with probabilistic latent representation
• ✅ Composite loss (reconstruction, KL divergence, commonality)
• ✅ Support for optical, SAR, NDVI, and other multimodal combinations

---

📖 Abstract

Multimodal change detection (MCD) has attracted a great deal of attention due to its significant advantages in processing heterogeneous remote sensing images (RSIs) from different sensors (e.g., optical and synthetic aperture radar). The major challenge of MCD is that it is difficult to acquire the changed areas by directly comparing heterogeneous RSIs. Although many MCD methods have made important progress, they are still insufficient in capturing the modality-independence complex structural relationships in the feature space of heterogeneous RSIs. To this end, we propose a novel commonality graph structure learning (CGSL) for unsupervised MCD, which aims to extract potential commonality graph structural features between heterogeneous RSIs and directly compare them to detect changes. In this study, heterogeneous RSIs are first segmented and constructed as superpixel-based heterogeneous graph structural data consisting of nodes and edges. Then, the heterogeneous graphs are input into the proposed CGSL to capture the commonalities of graph structural features with modality-independence. The proposed CGSL consists of a Siamese graph encoder and two graph decoders. The Siamese graph encoder maps heterogeneous graphs into a shared space and effectively extracts potential commonality in graph structural features from heterogeneous graphs. The two graph decoders reconstruct the mapped node features as original node features to maintain consistency with the original graph features. Finally, the changes between heterogeneous RSIs can be detected by measuring the differences in commonality graph structural features using the mean squared error. In addition, we design a composite loss with regularization to guide CGSL in effectively excavating the potential commonality graph structural features between heterogeneous graphs in an unsupervised learning manner. Extensive experiments on seven MCD datasets show that the proposed CGSL outperforms the existing state-of-the-art methods, demonstrating its superior performance in MCD.

The framework of the proposed CGSL is presented as follows:

---

📊 Results Preview

Change Maps

Note: If you want to get the difference map and binary map in our paper, you can get them directly through the "Results of Our CGSL" folder.

📦 Installation

Requirements

• Python 3.8+
• PyTorch 1.9+
• scikit-image
• scikit-learn
• NumPy

Install Dependencies

pip install torch torchvision scikit-image scikit-learn numpy

---

📁 Dataset Preparation

Organize your datasets as follows:

root_dir/
├── dataset1/
│   ├── t1_image.tif
│   ├── t2_image.tif
│   └── ground_truth.tif
├── dataset2/
│   ├── ...

• Supported image types: .tif, .png, .jpg
• Each dataset should contain:
  • T1 image (pre-event)
  • T2 image (post-event)
  • Ground truth image (if available)

---

🛠 Usage

Training and Inference

Run the “main.py” script with appropriate arguments:

python main.py \
  --root_dir /path/to/root \
  --load_data_dir /path/to/dataset \
  --result_folder ./results \
  --img_t1_name t1_image.tif \
  --img_t2_name t2_image.tif \
  --ref_name ground_truth.tif \
  --type_a optical \
  --type_b sar \
  --n_seg 1900 \
  --com 32 \
  --item 1 \
  --epochs 50 \
  --lr 0.0001 \
  --weight_decay 0.0001 \
  --lambda_reg 0.001

Parameters

Argument	Type	Default	Description
root_dir	str	'your data path'	Root directory for datasets
load_data_dir	str	'your data path'	Subdirectory of the dataset
result_folder	str	'your data path'	Folder to save results
img_t1_name	str	'your data name'	Pre-event image filename
img_t2_name	str	'your data name'	Post-event image filename
ref_name	str	'your data name'	Ground truth image filename
type_a	str	'optical or sar'	Modality of first image (e.g., optical, sar)
type_b	str	'optical or sar'	Modality of second image (e.g., optical, sar)
n_seg	int	0	Number of superpixels for SLIC
com	float	0	Compactness parameter for SLIC
item	int	5	Dataset index (for logging)
epochs	int	50	Number of training epochs
lr	float	0.0001	Learning rate
weight_decay	float	0.0001	Weight decay for optimizer
lambda_reg	float	0.001	L2 regularization coefficient

---

📊 Recommended Hyperparameters

Based on the paper, here are the suggested parameters for each dataset:

Dataset	n_seg	com
#1	1900	32
#2	2800	0.4
#3	1100	30
#4	1100	42
#5	1400	30
#6	1800	0.4
#7	1500	32

---

📈 Results Output

The method outputs:

• Difference Image (DI)
• Binary Change Image (BCI) via Otsu thresholding
• Evaluation metrics: OA, Kappa, F1, AUR, AUP

CGSL outperforms multiple SOTA methods (e.g., SR-GCAE, CACD, PRBCD-Net, BGAAE).
For detailed results, please refer to Tables 2–3 and Figures 4–6 in our paper.

---

📜 Citation

If you use this code or our paper, please cite:

@article{ISPRS2025CGSL,
  title={CGSL: Commonality Graph Structure Learning for Unsupervised Multimodal Change Detection},
  author={Xu, Jianjian and Liu, Tongfei and Lei, Tao and Chen, Hongruixuan and Yokoya, Naoto and Lv, Zhiyong and Gong, Maoguo},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={229},
  pages={92--106},
  year={2025},
  publisher={Elsevier}
}

@article{TIP2025CFRL,
  author={Liu, Tongfei and Zhang, Mingyang and Gong, Maoguo and Zhang, Qingfu and Jiang, Fenlong and Zheng, Hanhong and Lu, Di},
  journal={IEEE Transactions on Image Processing}, 
  title={Commonality Feature Representation Learning for Unsupervised Multimodal Change Detection}, 
  year={2025},
  volume={34},
  number={},
  pages={1219-1233},
  keywords={Feature extraction;Image reconstruction;Training;Data mining;Autoencoders;Representation learning;Image sensors;Electronic mail;Decoding;Clustering algorithms;Multimodal change detection;unsupervised change detection;heterogeneous images;representation learning;commonality feature},
  doi={10.1109/TIP.2025.3539461}}

---

📮 Contact

For questions or suggestions, please open an issue or contact:

• Jianjian Xu: [email protected]
• Tongfei Liu: [email protected]

---