Mamba Pyramid：

🔥 News

The paper Remote Sensing Change Detection by Pyramid Sequential Processing With Mamba is accepted by JSTARS 2025!

🚀 The progress

• Release all the codes.

😃 Start up

The repo is on the basis of the Changemamba. Thanks to the authors for their great work.

1 Set up

To start with, please follow the steps below:

git clone https://github.com/Band-127/Pyramid-Mamba-CD.git
conda create -n mambapyramid
cd MambaPyramid
conda activate mambapyramid

That is the way to create the env.

Then we install the requirements.

pip install -r requirements.txt
cd kernels/selective_scan && pip install .

Please organize the datasets as follows:

--workspace(MambaPyramid)
    --data
        --LEVIR-CD
            -- train
                -- A
                -- B
                -- label
            -- test
                (same)
            -- val
                (same)
        --DSIFN-CD
            -- train
                -- mask
                -- A
                -- B
            (same)
        --SYSU-CD
            -- train
                -- time1
                -- time2
                -- lable
            -- test
                (same)
        --WHU-CD
            -- A
                -- whucd-00001.png
                ......
            -- B
                -- whucd-00001.png
                ......
            -- label
                -- whucd-00001.png
                ......
            -- list
                -- train.txt
                -- test.txt
                -- val.txt

Note that the train/val/test sets in WHU-CD are splited by gen_train_val_test.py in list folder. You can find them in it. And please run it before your training and testing in case that raising error.

2 Train your model

For the 4 datasets above, we provide bash scripts for them in scripts folder.

    cd changedetection
    bash script/run/train_<dataset_name>.sh

Before you use them, please make sure that the data has been uploaded and the config path is correct.

Also, the evaluate the training process, you can see the matrics like train loss and corresponding metrics on Tensorboard.

3 Test your model

It is easy for you to get the test result of your model. We provide two forms, the binary change map and the RGB difference map. For the former, you only need to change the if_visible hyperparameter to 'gray' so you can get it. The white region stands for changed region, while the black stands for unchanged region. If you set the param as 'diff', you'll get the RGB maps. The red region stands for mistakes, while the green stands for misses.

    bash script/run/test_<dataset_name>.sh

The pre-trained model will be uploaded soon.

If the repository is useful for you, please tick a star and refer it in your paper.

@ARTICLE{11090033,
  author={Ma, Jiancong and Li, Bo and Li, Hanxi and Meng, Siying and Lu, Ruitao and Mei, Shaohui},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, 
  title={Remote Sensing Change Detection by Pyramid Sequential Processing With Mamba}, 
  year={2025},
  volume={18},
  number={},
  pages={19481-19495},
  keywords={Feature extraction;Transformers;Remote sensing;Biological system modeling;Convolutional neural networks;Computational modeling;Training;Computer architecture;Visualization;Deep learning;Change detection (CD);Mamba;multiscale;pyramid;state space model (SSM)},
  doi={10.1109/JSTARS.2025.3591834}}