lironui/double-branch-dual-attention-mechanism-network

Hyperspectral Image Classification

Hyperspectral Image Classification

⭐ Welcome to my HomePage ⭐

This repository implementates 6 frameworks for hyperspectral image classification based on PyTorch and sklearn.

The detailed results can be seen in the Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network.

Some of our code references the projects

• Dual-Attention-Network
• Remote sensing image classification
• A Fast Dense Spectral-Spatial Convolution Network Framework for Hyperspectral Images Classification

If our code is helpful to you, please cite
Li R, Zheng S, Duan C, et al. Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network[J]. Remote Sensing, 2020, 12(3): 582.

Requirements：

numpy >= 1.16.5
PyTorch >= 1.3.1
sklearn >= 0.20.4

Datasets:

You can download the hyperspectral datasets in mat format at: http://www.ehu.eus/ccwintco/index.php/Hyperspectral_Remote_Sensing_Scenes, and move the files to ./datasets folder.

Usage:

1. Set the percentage of training and validation samples by the load_dataset function in the file ./global_module/generate_pic.py.
2. Taking the DBDA framework as an example, run ./DBDA/main.py and type the name of dataset.
3. The classfication maps are obtained in ./DBDA/classification_maps folder, and accuracy result is generated in ./DBDA/records folder.

Network:

• DBDA
• DBMA
• FDSSC
• SSRN
• CDCNN
• SVM

Figure 1. The structure of the DBDA network. The upper Spectral Branch composed of the dense spectral block and channel attention block is designed to capture spectral features. The lower Spatial Branch constituted by dense spatial block, and spatial attention block is designed to exploit spatial features.

Results:

Figure 2. Classification maps for the IP dataset using 3% training samples. (a) False-color image. (b) Ground-truth (GT). (c)–(h) The classification maps with disparate algorithms. Figure 3. Classification maps for the UP dataset using 0.5% training samples. (a) False-color image. (b) Ground-truth (GT). (c)–(h) The classification maps with disparate algorithms. Figure 4. Classification maps for the UP dataset using 0.5% training samples. (a) False-color image. (b) Ground-truth (GT). (c)–(h) The classification maps with disparate algorithms. Figure 5. Classification maps for the BS dataset using 1.2% training samples. (a) False-color image. (b) Ground-truth (GT). (c)–(h) The classification maps with disparate algorithms.

double-branch-dual-attention-mechanism-network's Issues

IP数据集精度问题

请问使用原始的DBDA模型在Indian Pines 数据集上得到的平均精度AA与论文中有较大差异可能是由什么导致的？

3d convolution kernel parameter order

Thanks， I got it

depth size

hello
According to your paper
kernel_3d = math.floor((band - 6) / 2) should be kernel_3d = math.ceil((band - 6) / 2) , the model can output as expected, because of using self.global_pooling = nn.AdaptiveAvgPool3d(1) ,and the band is 200, and 194 is divisible by 2

why u do element-wist product between spatial/ spectral attention output with its input?

in your code "class DBDA_network_MISH" 's forward function :
"x1 = self.attention_spectral(x16)
x1 = torch.mul(x1, x16)"
and
"x2 = self.attention_spatial(x25)
x2 = torch.mul(x2, x25)"
above all don't exist in your paper, and lack of physical meaning.