Repository Overview

Official implementation for the paper JL1-CD: A New Benchmark for Remote Sensing Change Detection and a Robust Multi-Teacher Knowledge Distillation Framework. This code is built upon the OpenCD toolbox.

News

• 6/30/2026 - JL1-CC & JL1-QA datasets have been open-sourced.
• 2/23/2025 - JL1-CD dataset has been open-sourced.

Dataset

JL1-CD

The JL1-CD dataset is now publicly available. You can download the dataset from:

• Google Drive
• Baidu Disk
• Hugging Face

JL1-CC & JL1-QA

Built upon JL1-CD, JL1-CC provides 17,021 quality-verified change captions and JL1-QA provides 20,060 question–answer pairs across eight question types. Download:

• Google Drive

Usage

Install

To set up the environment, follow the installation instructions provided in the OpenCD repository.

Training

The training process for the MTKD framework consists of three steps. Below, we use the Changer-MiT-b0 model as an example:

Step 1: Train the original model

Run the following command to train the original model:

python tools/train.py configs/changer/changer_ex_mit-b0_512x512_200k_cgwx.py --work-dir /path/to/save/models/Changer-mit-b0/initial

Step 2: Train teacher models for different CAR partitions (e.g., 3 partitions)

Train the teacher models for small, medium, and large CAR partitions as follows:

python tools/train.py configs/distill-changer/changer_ex_mit-b0_512x512_200k_cgwx-small.py --work-dir /path/to/save/models/Changer-mit-b0/small

python tools/train.py configs/distill-changer/changer_ex_mit-b0_512x512_200k_cgwx-medium.py --work-dir /path/to/save/models/Changer-mit-b0/medium

python tools/train.py configs/distill-changer/changer_ex_mit-b0_512x512_200k_cgwx-large.py --work-dir /path/to/save/models/Changer-mit-b0/large

In the above two steps, you will have four model versions for Changer-MiT-b0: the original model and three teacher models (small, medium, and large). At this point, the O-P strategy can already be applied.

Step 3: Train the student model

Initialize the checkpoint paths in configs/distill-changer/distill-changer_ex_mit-b0_512x512_200k_cgwx.py for the student model and teacher models as follows:

• checkpoint_student
• checkpoint_teacher_l
• checkpoint_teacher_m
• checkpoint_teacher_s

Then, run the following command to train the student model:

python tools/train.py configs/distill-changer/distill-changer_ex_mit-b0_512x512_200k_cgwx.py --work-dir /path/to/save/models/Changer-mit-b0/distill

After this step, you will have the student model trained within the MTKD framework.

Testing

Testing the student model trained with MTKD is simple. Run the following command:

python test.py <config-file> <checkpoint>

Testing the O-P strategy is more complex. You can refer to the script located at tools/test_pipline/single-partition-3-test.py for more details.

Checkpoints

You can download checkpoint files from:

• Baidu Disk
• Hugging Face

Citation

If you find JL1-CD, JL1-CC, or JL1-QA useful in your research, please consider citing our papers:

@article{liu2026jl1cd,
  title={JL1-CD: A New Benchmark and a Robust Multi-Teacher Knowledge Distillation Framework for Remote Sensing Change Detection},
  author={Liu, Ziyuan and Zhu, Ruifei and Gao, Long and Zhou, Yuanxiu and Ma, Jingyu and Gu, Yuantao},
  journal={IEEE Transactions on Instrumentation and Measurement},
  year={2026},
  publisher={IEEE}
}

@article{liu2026jl1ccqa,
  title={JL1-CC\&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering},
  author={Liu, Ziyuan and Zhu, Ruifei and Ma, Ouqiao and Gu, Yuantao},
  journal={arXiv preprint arXiv:2606.31745},
  year={2026}
}