CSSM

Efficient Remote Sensing Change Detection with Change State Space Models

E.Ghazaei, E.Aptoula

Faculty of Engineering and Natural Sciences (VPALab), Sabanci University, Istanbul, Turkiye

[Paper Link]

🛎️Updates

• Notice🐍🐍: CSSM has been accepted by IEEE GRSL! We'd appreciate it if you could give this repo a ⭐️star⭐️ and stay tuned!!
• Nov 05th, 2025: The CSSM model and training code uploaded. You are welcome to use them!!

---

🚀 Overview

• CSSM serves as an efficient and state-of-the-art (SOTA) benchmark for binary change detection.

---

📦 Requirements

pip install torch torchvision torchaudio
pip install  pillow
pip install numpy scipy pandas
pip install matplotlib seaborn
pip install einops 
pip install torchinfo

---

📁 Dataset Preparation

This project supports three main change detection datasets:

• LEVIR-CD+ - Download
• SYSU-CD - Download
• WHU-CD - Download

---

Dataset Structure

For LEVIR-CD:

your_dataset/
├── train/
│   ├── A/          # Pre-change images
│   ├── B/          # Post-change images
│   └── label/      # Ground truth masks
├── test/
│   ├── A/
│   ├── B/
│   └── label/
└── val/
    ├── A/
    ├── B/
    └── label/

For SYSU-CD:

your_dataset/
├── train/
│   ├── time1/          # Pre-change images
│   ├── time2/          # Post-change images
│   └── label/      # Ground truth masks
├── test/
│   ├── time1/
│   ├── time2/
│   └── label/
└── val/
    ├── time1/
    ├── time2/
    └── label/

For WHU-CD:

WHU-CD/
├── A/              # Pre-change images
├── B/              # Post-change images
├── label/          # Ground truth masks
├── train_list.txt  # List of training samples
├── test_list.txt   # List of test samples
└── val_list.txt    # List of validation samples

The text files should contain image names (one per line): image_001.png image_002.png image_003.png ...

🚂 Training

LEVIR-CD Dataset

python main.py \
    --dataset levir \
    --train_path /path/to/LEVIR-CD/train \
    --test_path /path/to/LEVIR-CD/test \
    --val_path /path/to/LEVIR-CD/val \
    --batch_size 64 \
    --epochs 50 \
    --lr 0.001

SYSU-CD Dataset

python main.py \
    --dataset sysu \
    --train_path /path/to/SYSU-CD/train \
    --test_path /path/to/SYSU-CD/test \
    --val_path /path/to/SYSU-CD/val \
    --batch_size 32 \
    --epochs 100 \
    --lr 0.0001

WHU-CD Dataset

python main.py \
    --dataset whu \
    --train_path /path/to/WHU-CD \
    --train_txt /path/to/train_list.txt \
    --test_txt /path/to/test_list.txt \
    --val_txt /path/to/val_list.txt \
    --batch_size 64 \
    --epochs 50

---

⚙️ Arguments

Required Arguments

Argument	Description	Example
--dataset	Dataset type: levir, sysu, or whu	--dataset levir
--train_path	Path to training data	--train_path /data/train
--test_path	Path to test data (not for WHU)	--test_path /data/test
--val_path	Path to validation data (not for WHU)	--val_path /data/val

WHU-CD Specific Arguments

Argument	Description	Example
--train_txt	Training sample list file	--train_txt train_list.txt
--test_txt	Test sample list file	--test_txt test_list.txt
--val_txt	Validation sample list file	--val_txt val_list.txt

Optional Arguments

Argument	Default	Description
--batch_size	64	Batch size for training
--epochs	50	Number of training epochs
--lr	0.001	Learning rate
--step_size	10	Learning rate scheduler step size
--save_dir	./checkpoints	Directory to save model checkpoints
--model_name	best_model.pth	Filename for saved model
--seed	42	Random seed for reproducibility
--num_workers	4	Number of data loading workers

---

🔧 Advanced Usage Examples

Custom Save Directory and Model Name

python main.py \
    --dataset levir \
    --train_path /data/LEVIR-CD/train \
    --test_path /data/LEVIR-CD/test \
    --val_path /data/LEVIR-CD/val \
    --save_dir ./experiments/levir_exp1 \
    --model_name levir_model.pth \
    --epochs 100

Different Learning Rate Schedule

python main.py \
    --dataset sysu \
    --train_path /data/SYSU-CD/train \
    --test_path /data/SYSU-CD/test \
    --val_path /data/SYSU-CD/val \
    --lr 0.0005 \
    --step_size 20 \
    --epochs 150

Smaller Batch Size (for limited GPU memory)

python main.py \
    --dataset levir \
    --train_path /data/train \
    --test_path /data/test \
    --val_path /data/val \
    --batch_size 16 \
    --num_workers 2

---

📤 Output

During training, the script will:

• Display training loss for each batch
• Show validation metrics (IoU, confusion matrix) after each epoch
• Save the best model based on validation IoU
• Display learning rate and epoch time

Model Checkpoint

The best model is automatically saved to:

{save_dir}/{model_name}

Default: ./checkpoints/best_model.pth

---

🔍 Troubleshooting

Paths with Spaces

If your paths contain spaces, wrap them in quotes:

python main.py \
    --dataset levir \
    --train_path "/path/with spaces/train" \
    --test_path "/path/with spaces/test" \
    --val_path "/path/with spaces/val"

CUDA Out of Memory

Reduce batch size:

python main.py --dataset levir ... --batch_size 16

Missing WHU Text Files

For WHU dataset, ensure all three text files are provided:

python main.py \
    --dataset whu \
    --train_path /data/WHU-CD \
    --train_txt train_list.txt \
    --test_txt test_list.txt \
    --val_txt val_list.txt

---

💡 Getting Help

View all available arguments:

python main.py --help

---

📧 Contact

If you have any questions, please contact Elman Ghazaei at elman.ghazaei@sabanciuniv.edu

---

Qualitative Analysis:

---

Results