基于注意力机制的多目标优化高光谱波段选择 下载: 1076次
Multi-Objective Optimization of Hyperspectral Band Selection Based on Attention Mechanism
1 中国人民解放军空军航空大学航空作战勤务学院, 吉林 长春 130022
2 东北师范大学地理科学学院, 吉林 长春 130024
图 & 表
图 1. SENet结构
Fig. 1. SENet structure
下载图片 查看原文
图 2. 波段选择模型结构
Fig. 2. Band selection model structure
下载图片 查看原文
图 3. Botswana数据集的真彩色图像和地物真值图。(a)真彩色图像;(b)地物真值图
Fig. 3. True color image and ground truth map of Botswana data set. (a) True color image; (b) ground truth map
下载图片 查看原文
图 4. Indian Pines数据集的真彩色图像和地物真值图。(a)真彩色图像;(b)地物真值图
Fig. 4. True color image and ground truth map of Indian Pines data set. (a) True color image; (b) ground truth map
下载图片 查看原文
图 5. 实验中使用的SENet结构
Fig. 5. SENet structure in the experiment
下载图片 查看原文
图 6. Botswana数据集上的总体分类精度、训练损失和波段权重变化。(a)总体分类精度;(b)训练损失;(c)波段权重热力图
Fig. 6. Overall classification accuracy, training loss, and band weight changes in the Botswana data set. (a) Overall classification accuracy; (b) training loss; (c) band weight thermal map
下载图片 查看原文
图 7. Indian Pines数据集上的总体分类精度、训练损失和波段权重变化。(a)总体分类精度;(b)训练损失;(c)波段权重热力图
Fig. 7. Overall classification accuracy, training loss and band weight changes on the Indian Pines data set. (a) Overall classification accuracy; (b) training loss; (c) band weight thermal map
下载图片 查看原文
图 8. 各算法在Botswana数据集上的总体分类精度、平均分类精度和Kappa系数。(a)总体分类精度;(b)平均分类精度;(c) Kappa系数
Fig. 8. Overall classification accuracy, average classification accuracy and Kappa coefficient of each algorithm in the Botswana data set. (a) Overall classification accuracy; (b) average classification accuracy; (c) Kappa coefficient
下载图片 查看原文
图 9. 各算法在Botswana数据集上平均光谱散度
Fig. 9. Average spectral divergence of each algorithm on the Botswana data set
下载图片 查看原文
图 10. 各算法在Indian Pines数据集上的总体分类精度、平均分类精度和Kappa系数。(a)总体分类精度;(b)平均分类精度;(c) Kappa系数
Fig. 10. Overall classification accuracy, average classification accuracy and Kappa coefficient of each algorithm in the Indian Pines data set. (a) Overall classification accuracy; (b) average classification accuracy; (c) Kappa coefficient
下载图片 查看原文
图 11. 各算法在Indian Pines数据集上的平均光谱散度
Fig. 11. Average spectral divergence of each algorithm on the Indian Pines data set
下载图片 查看原文
表 1模型中的数据尺寸及激活函数变化
Table1. Data size and activation function change in the model
Module | Layer | Input size | Output size | Activation |
---|
| Input | | 1×1×b | | Attention module | FC-1(fully connected layer) | 1×1×b | 1×1×(b/16) | ReLU | | FC-2(fully connected layer) | 1×1×(b/16) | 1×1×b | Sigmoid | | Encoder-1(autoencoder) | 1×1×b | 1×1×256 | | | BN-1(batch normalization) | 1×1×256 | 1×1×256 | ReLU | | Encoder-2(autoencoder) | 1×1×256 | 1×1×128 | | | BN-2(batch normalization) | 1×1×128 | 1×1×128 | ReLU | | Encoder-3(autoencoder) | 1×1×128 | 1×1×64 | | | BN-3(batch normalization) | 1×1×64 | 1×1×64 | ReLU | Reconstruction module | Encoder-4(autoencoder) | 1×1×64 | 1×1×64 | | | BN-4(batch normalization) | 1×1×64 | 1×1×64 | ReLU | | Decoder-1(autoencoder) | 1×1×64 | 1×1×128 | | | BN-5(batch normalization) | 1×1×128 | 1×1×128 | ReLU | | Decoder-2(autoencoder) | 1×1×128 | 1×1×256 | | | BN-6(batch normalization) | 1×1×256 | 1×1×256 | ReLU | | Decoder-3(autoencoder) | 1×1×256 | 1×1×b | Sigmoid | Classification module | Latent vector | 1×1×64 | 1×1×64 | | | FC-3(fully connected layer) | 1×1×64 | Number of class | Softmax |
|
查看原文
表 2高光谱图像数据集
Table2. Hyperspectral image data set
Item | Botswana | Indian Pines |
---|
Shooting area | Okavango Delta, Botswana | Indiana, USA | Imaging spectrometer | Hyperion | AVIRIS | Spectral range /nm | 400-2500 | 400-2500 | Number of wavelengths (remove strong noise and water vapor band) | 145 | 200 | Image size /(pixel×pixel) | 1476×256 | 145×145 | Spatial resolution /m | 30 | 20 | Sample size | 3248 | 10249 | Object types | 14 | 16 |
|
查看原文
表 3不同权重系数下两个数据集的实验结果
Table3. Experimental results of two data sets with different weight coefficients
γ | Botswana | Indian Pines | |
---|
OA /% | AA /% | Kappa | OA /% | AA /% | Kappa |
---|
0.1 | 88.9 | 89.5 | 0.873 | 73.1 | 71.4 | 0.708 | 0.3 | 89.3 | 89.8 | 0.886 | 73.6 | 71.5 | 0.706 | 0.5 | 88.6 | 87.1 | 0.869 | 74.3 | 70.4 | 0.712 | 0.7 | 87.2 | 86.8 | 0.853 | 72.1 | 69.5 | 0.698 | 0.9 | 85.3 | 86.7 | 0.839 | 69.7 | 66.1 | 0.664 |
|
查看原文
关世豪, 杨桄, 卢珊, 付严宇. 基于注意力机制的多目标优化高光谱波段选择[J]. 光学学报, 2020, 40(21): 2128002. Shihao Guan, Guang Yang, Shan Lu, Yanyu Fu. Multi-Objective Optimization of Hyperspectral Band Selection Based on Attention Mechanism[J]. Acta Optica Sinica, 2020, 40(21): 2128002.