SAM-CD

Pytorch codes of Adapting Segment Anything Model for Change Detection in HR Remote Sensing Images [paper]

The SAM-CD adopts FastSAM as the visual encoder with some modifications.

How to Use

1. Installation

   • Install FastSAM following the instructions.
   • Modify the Ultralytics source files following the instructions at: 'SAM-CD/models/FastSAM/README.md'.

2. Dataset preparation.

   • Please split the data into training, validation and test sets and organize them as follows:

      YOUR_DATA_DIR
      ├── ...
      ├── train
      │   ├── A
      │   ├── B
      │   ├── label
      ├── val
      │   ├── A
      │   ├── B
      │   ├── label
      ├── test
      │   ├── A
      │   ├── B
      │   ├── label

• Find change line 13 in SAM-CD/datasets/Levir_CD.py (or other data-loading .py files), change /YOUR_DATA_ROOT/ to your local dataset directory.

3. Training

   classic CD training: python train_CD.py

   training CD with the proposed task-agnostic semantic learning: python train_SAM_CD.py

   line 16-45 are the major training args, which can be changed to load different datasets, models and adjust the training settings.

4. Inference and evaluation

   inference on test sets: set the chkpt_path and run

   python pred_CD.py

   evaluation of accuracy: set the prediction dir and GT dir, and run

   python eval_CD.py

(More details to be added...)

Dataset Download

In the following, we summarize links to some frequently used CD datasets:

• LEVIR-CD
• WHU-CD (baidu)
• CLCD (Baidu)
• S2Looking
• SYSU-CD

Pretrained Models

For readers to easily evaluate the accuracy, we provide the trained weights of the SAM-CD.

Drive
Baidu (pswd: SMCD)

Cite SAM-CD

If you find this work useful or interesting, please consider citing the following BibTeX entry.

@article{ding2024adapting,
title={Adapting Segment Anything Model for Change Detection in HR Remote Sensing Images},
author={Ding, Lei and Zhu, Kun and Peng, Daifeng and Tang, Hao and Yang, Kuiwu and Bruzzone, Lorenzo},
journal={IEEE Transactions on Geoscience and Remote Sensing}, 
year={2024},
volume={62},
pages={1-11},
doi={10.1109/TGRS.2024.3368168}
}