HSACNet: Hierarchical Scale-Aware Consistency Regularized Semi-Supervised Change Detection

📖 Overview

This repository implements a semi-supervised change detection framework that leverages the powerful Segment Anything Model 2 (SAM2) as a foundation encoder. Our approach addresses the challenge of limited labeled data in remote sensing change detection tasks through:

• SAM2-Based Encoder: Utilizes the pre-trained Hiera backbone from SAM2 for robust feature extraction
• Scale-Aware Difference Attention Module (SADAM): Multi-scale difference feature extraction with attention mechanism
• Semi-Supervised Learning Framework: 1-weak-2-strong augmentation strategy with pseudo-labeling
• Parameter-Efficient Fine-Tuning: Both Adapter-based and LoRA-based implementations

Network Architecture

Figure 1: Overall architecture of the proposed change detection network, consisting of SAM2 encoder with adapters, SADAM modules, and decoder.

Figure 2: (Left) Structure of the proposed SADAM module with multi-branch group convolutions and spatial-channel attention. (Right) Overall semi-supervised learning framework with supervised and unsupervised components.

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/xqa001/SAM2-UNet-SSCD.git
cd SAM2-UNet-SSCD

# Create environment
conda create -n sam2-sscd python=3.10
conda activate sam2-sscd

# Install dependencies
pip install -r requirements.txt

Requirements

torch~=2.4.0
torchvision
torchaudio
hydra-core==1.3.2
tqdm==4.65.2
imageio
pysodmetrics
peft  # For LoRA implementation

Download Pretrained Models

Download SAM2 Hiera pretrained weights and place them in pretrained/:

• sam2_hiera_tiny.pt (recommended)
• sam2_hiera_small.pt
• sam2_hiera_base_plus.pt
• sam2_hiera_large.pt

📁 Project Structure

SAM2-UNet-SSCD/
├── sam2/                          # SAM2 backbone
│   ├── modeling/
│   │   ├── backbones/            # Hiera encoder
│   │   └── sam/                  # SAM components
│   └── build_sam.py
├── sam2_configs/                  # Model configs (T/S/B+/L)
├── splits/                        # Dataset splits
│   ├── whu/                      
│   │   ├── train.txt
│   │   ├── test.txt
│   │   ├── 5%/
│   │   ├── 10%/
│   │   └── 20%/
│   ├── levir/
│   └── ...
├── metrics/                       # Evaluation metrics
│   └── metric_tool.py
├── model.py                       # Network architectures
│   ├── SAM2UNetCD               # Adapter version
│   └── SAM2UNetCD_LoRA          # LoRA version
├── modules.py                     # Core modules
│   ├── ScaleAwareDiffAttnModule # SADAM
│   ├── AttentionalBlock         # Spatial attention
│   └── DifferenceModule         # Basic diff module
├── cddataset.py                  # Dataset loader
├── train-fully-ce.py             # Fully-supervised training
├── train-semi-1w2s-ce.py         # Semi-supervised (1w2s)
├── test.py                        # Testing
├── eval.py                        # Evaluation
└── *.sh                          # Shell scripts

📊 Dataset

Download Links

• WHU-CD: imageA, imageB, and label
• LEVIR-CD: imageA, imageB, and label
• GZ-CD: aistudio
• EGY-BCD: baidu drive, password: EGYD. google drive

Data Structure

The data directory should follow the structure below:

├── [Your WHU-CD/LEVIR-CD/GZ-CD/EGY-BCD Path]
    ├── A
    ├── B
    └── label

🏋️ Training

Fully-Supervised Training

Train with 100% labeled data:

CUDA_VISIBLE_DEVICES="0" \
python train-fully-ce.py \
  --hiera_path "pretrained/sam2_hiera_tiny.pt" \
  --dataset_name "whu" \
  --data_root "/path/to/WHU-CD256" \
  --labeled_id_path "splits/whu/train.txt" \
  --unlabeled_id_path "splits/whu/train.txt" \
  --save_path "exp/whu-fully" \
  --image_size 256 \
  --epoch 80 \
  --lr 3e-4 \
  --batch_size 8

Or simply:

bash train-fully.sh

Semi-Supervised Training

Train with limited labeled data (e.g., 5%):

CUDA_VISIBLE_DEVICES="0" \
python train-semi-1w2s-ce.py \
  --hiera_path "pretrained/sam2_hiera_tiny.pt" \
  --dataset_name "whu" \
  --data_root "/path/to/WHU-CD256" \
  --labeled_id_path "splits/whu/5%/labeled.txt" \
  --unlabeled_id_path "splits/whu/5%/unlabeled.txt" \
  --save_path "exp/whu-semi-5%" \
  --image_size 256 \
  --epoch 80 \
  --lr 3e-4 \
  --batch_size 4

Or:

bash train-semi.sh

🧪 Testing

Test a trained model:

CUDA_VISIBLE_DEVICES="0" \
python test.py \
  --checkpoint "exp/whu-semi-5%/best_model.pth" \
  --dataset_name "whu" \
  --data_root "/path/to/WHU-CD256" \
  --save_path "exp/whu-semi-5%/results"

Or:

bash test.sh

This will:

1. Load the trained model
2. Run inference on the test set
3. Save prediction masks to save_path

📈 Evaluation

Evaluate predictions against ground truth:

python eval.py \
  --dataset_name "whu" \
  --pred_path "exp/whu-semi-5%/results" \
  --gt_path "/path/to/WHU-CD256/label"

Or:

bash eval.sh

📈 Results

Our method is evaluated on four challenging change detection datasets (WHU-CD, LEVIR-CD, GZ-CD, and EGY-BCD) with different labeled ratios (5%, 10%, 20%, 40%).

Quantitative Results

Figure 3: Quantitative comparison on WHU-CD, LEVIR-CD, GZ-CD, and EGY-BCD datasets. The results show IoU and OA metrics across different labeled ratios (5%-40%), demonstrating consistent improvements over state-of-the-art semi-supervised methods.

Qualitative Results

Figure 4: Qualitative detection results of different methods on four CD datasets at 5% labeled ratio. White represents TP (true positives), black for TN (true negatives), red for FP (false positives), and green for FN (false negatives). Our method produces cleaner predictions with fewer false positives and false negatives.

📚 Citation

If you find this work useful for your research, please cite:

@inproceedings{xu2025hierarchical,
  title={HSACNet: Hierarchical Scale-Aware Consistency Regularized Semi-Supervised Change Detection},
  author={Xu, Qi'ao and Wang, Pengfei and Li, Yanjun and Qian, Tianwen and Wang, Xiaoling},
  booktitle={2025 IEEE International Conference on Multimedia and Expo (ICME)},
  year={2025},
  pages={1-6},
  doi={10.1109/ICME59968.2025.11209170},
  organization={IEEE}
}

🙏 Acknowledgements

Thanks to the following projects for their excellent work:

• Segment Anything 2 - Foundation model and Hiera backbone
• SAM2-UNet - SAM2 adaptation for medical image segmentation
• UniMatch - Semi-supervised learning framework

We thank the authors for open-sourcing their code.