edcstfn's Introduction

An Enhanced Deep Convolutional SpatioTemporal Network for Remote Sensing Images (EDCSTFN)

Notice:

The source code has been refactored. If this causes any runtime error, please feel free to contact me or make a pull request.
There is a small mistake in Figure. 2 regarding the original manuscript, the input and output channels should be the same for the AutoEncoder. (I wrote to the editor, but they said it cannot be corrected after the paper being released.)

Models:

The master branch is the implementation of the EDCSTFN model.
The autoencoder branch is the implementation of the hourglass AutoEncoder. The pretrained Encoder subnet is used in the EDCSTFN to calculate the feature loss.

Environment:

Tested on the following environment:

Python: >=3.6
PyTorch: >=0.4

If you find this code useful in your research, please consider citing our work:

@Article{rs11242898,
AUTHOR = {Tan, Zhenyu and Di, Liping and Zhang, Mingda and Guo, Liying and Gao, Meiling},
TITLE = {An Enhanced Deep Convolutional Model for Spatiotemporal Image Fusion},
JOURNAL = {Remote Sensing},
VOLUME = {11},
YEAR = {2019},
NUMBER = {24},
ARTICLE-NUMBER = {2898},
URL = {https://www.mdpi.com/2072-4292/11/24/2898},
ISSN = {2072-4292},
DOI = {10.3390/rs11242898}
}

https://www.mdpi.com/2072-4292/11/24/2898/htm

edcstfn's People

Contributors

Stargazers

edcstfn's Issues

您好，在尝试跑这个代码的时候。
读取pretrained_dict训练参数时，出现了这样的报错，尝试了很久都没有找到解决方法，方便解答一下吗？
出错地方：
File "F:\STF\new_way\edcstfn-master\utils.py", line 98, in load_pretrained
pretrained_dict = torch.load(pretrained)['state_dict']
报错：
'NoneType' object has no attribute 'seek'. You can only torch.load from a file that is seekable. Please pre-load the data into a buffer like io.BytesIO and try to load from it instead.

关于数据集的设置问题

您好，我是刚开始接触的遥感影像融合的研一学生，请问训练集和测试集数据的目录应该是怎样设置的呢？

autoencoder的训练结果都非常糟糕

你好！你们的论文写得很好有意复现
但是复现的过程中发现训练的结果很糟糕不知道问题出在哪（感觉是自己的数据集放置方式不对
下图是autoencoder的训练结果，我在data_pre/train的目录下放了20张Landsat图供训练，但是效果很糟糕调参也试过很多次，也没有改善，所以想询问一下你们的正确训练方式（数据集放置方式

训练过程中遇到的问题

作者您好，我尝试使用您的代码进行训练，但是效果似乎不是很好。我采用的参数是run.py文件开头给出的建议参数，不同之处在于根据我训练集图片大小改动了切块的大小。预训练模型的自编码器训练了30个epoch。在EDCSTFN的训练结果来看loss长时间稳定E-4，训练准确性长时间稳定E-6。验证结果来看，RMSE=6~10，ERGAS=3.4，SAM=1.3，这些似乎比论文中提到的大不少。我尝试修改学习率，做归一化处理，但是没有好转。所以我才向作者请教，作者若有空能给予一些帮助？谢谢