aj1365 / polsarformer Goto Github PK

This code is for the paper "Local Window Attention Transformer for Polarimetric SAR Image Classification" that is published in the IEEE Geoscience and Remote Sensing Letters journal.

License: Apache License 2.0

Jupyter Notebook 100.00%

classification cnn deep-learning image-classification polsar remote-sensing vision-transformer

polsarformer's Introduction

PolSARFormer: Local Window Attention Transformer for Polarimetric SAR Image Classification

Ali Jamali, Swalpa Kumar Roy, Avik Bhattacharya, and Pedram Ghamisi

This Keras code is for the paper A. Jamali, S. K. Roy, A. Bhattacharya and P. Ghamisi, "Local Window Attention Transformer for Polarimetric SAR Image Classification," in IEEE Geoscience and Remote Sensing Letters, doi: 10.1109/LGRS.2023.3239263 [https://ieeexplore.ieee.org/document/10024822].

Dataset

Flevoland dataset: NASA/JPL AIRSAR recorded the data of Flevoland, situated in the Netherlands, on August 16, 1989. The Flevoland image is $750\times1024$ pixels in size.

San Francisco dataset: The San Francisco illustrates a NASA/JPL AIRSAR L-band image of the San Francisco area. The resolution of the data of the San Francisco is $900\times1024$ pixels.

Appreciation from Geoscience and Remote Sensing Society (GRSS)

Citation

Please kindly cite the papers if this code is useful and helpful for your research.

@article{jamali2023local,
    title={Local window attention transformer for polarimetric SAR image classification},
    author={Jamali, Ali and Roy, Swalpa Kumar and Bhattacharya, Avik and Ghamisi, Pedram},
    journal={IEEE Geoscience and Remote Sensing Letters},
    volume={20},
    pages={1--5},
    year={2023},
    publisher={IEEE}
}

Acknowledgement

Part of the local window attention (LWA) block is implementated from Neighborhood Attention Transformer.

License

polsarformer's People

Contributors

Stargazers

Watchers

Forkers

westamine pomeloyou mikeywww tarequebashar rongtongxueya zxs1652 m89j loongzz lalitgis

polsarformer's Issues

About SanFrancisco_T3RF.mat download

Hello author, could you tell me where to download SanFrancisco_T3RF.mat download, thanks a lot.

Problems with the test_ratio

Thanks for your very nice code.
I have a doubt, is setting the test_ratio to 0.9 too large?

Thanks for your very nice code.
I met a problem with your code:
At the end of the PolSARFormer model, the code below just takes the x as the input, thus the output nn of the PolSARFormer is just the results of the output_block, and the attention operation is unused.
'nn = output_block(x, num_classes=num_classes, drop_rate=dropout, classifier_activation=classifier_activation)'

So I wonder if there is a mistake, and if the 'x' in the code should be takeplaced by other variable.

Code question

Thanks for your very nice code.I met a problem with your code:

According to my calculations, the shape of the patchesData returned from the ImageCube function should be (H*W, 12, 12, 9). However, I noticed that you are reshaping x to have a shape of (X1.shape[0], windowSize, windowSize, 12, 1) based on the setting of windowSize=12 in your paper, which translates to (H, 12, 12, 12, 1). This shape doesn't match my calculated shape of H*W, 12, 12, 9. Therefore, I have the following doubts:

Should the first dimension of the reshape operation be H*W? If my understanding is incorrect, I would like to know how you performed this step.
For the third dimension of my data, which represents the scattering dimension and has a size of 9, I am curious about the fourth dimension in your reshape operation, which is the last 12. Is this 12 related to the scattering dimension in your dataset? If not, I would like to understand the significance of this 12.
After running the code, I noticed that our data shapes do not seem to match. Could you please confirm the dimensionality of the scattering data in your San Francisco dataset? This information would be extremely helpful for my research.
Thank you very much.

Problems with datasets

Hello, I am very happy to see your research results, I am a graduate student, also working on polarized SAR Landcover classification problems。But I haven't been able to find these open source dataset to downloadcan you share the dataset you used in this article: Flevoland dataset, San Francisco dataset. If you can, thanks for sending me [email protected] email, thanks.