基于双通道空洞卷积神经网络的高光谱图像分类 下载: 1076次
Hyperspectral Image Classification Based on Dual-Channel Dilated Convolution Neural Network
1 燕山大学理学院, 河北 秦皇岛 066001
2 燕山大学电气工程学院, 河北 秦皇岛 066001
3 燕山大学机械学院, 河北 秦皇岛 066001
4 北京空间机电研究所, 北京 100094
图 & 表
图 1. 所提框架结构
Fig. 1. Structure diagram of the proposed framework
下载图片 查看原文
图 2. 一维空洞卷积示意图。(a)标准卷积;(b)空洞卷积
Fig. 2. Schematic of 1D dilated convolution. (a) Standard convolution; (b) dilated convolution
下载图片 查看原文
图 3. DCD-CNN结构
Fig. 3. DCD-CNN structure
下载图片 查看原文
图 4. 不同数据集在不同λ值的OA精度
Fig. 4. OA accuracy of different datasets at different λ values
下载图片 查看原文
图 5. 不同方法在Indian Pines数据集上的分类图。(a)真值图;(b) SVM;(c) AEAP;(d) DCNN; (e) FEFCN-ELM;(f)所提方法
Fig. 5. Classification maps of different methods on Indian Pines dataset. (a) Ground truth; (b) SVM; (c) AEAP; (d) DCNN; (e) FEFCN-ELM; (f) proposed method
下载图片 查看原文
图 6. 不同方法在Pavia University数据集上的分类图。(a)真值图;(b) SVM;(c) AEAP;(d) DCNN;(e) FEFCN-ELM;(f)所提方法
Fig. 6. Classification maps of different methods on Pavia University dataset. (a) Ground truth; (b) SVM; (c) AEAP; (d) DCNN; (e) FEFCN-ELM; (f) proposed method
下载图片 查看原文
表 1不同方法在Indian Pines数据集上的分类结果
Table1. Classification results of different methods on Indian Pines datasetunit: %
Class | SVM | AEAP | DCNN | FEFCN-ELM | Proposed |
---|
1 | 21.73 | 67.39 | 91.30 | 91.30 | 98.91 | 2 | 64.07 | 81.86 | 90.12 | 95.37 | 98.76 | 3 | 62.89 | 73.73 | 77.71 | 92.53 | 99.55 | 4 | 51.47 | 79.32 | 55.27 | 85.23 | 96.85 | 5 | 83.02 | 94.82 | 88.40 | 92.13 | 97.05 | 6 | 96.30 | 98.90 | 97.53 | 99.17 | 99.79 | 7 | 64.28 | 89.28 | 82.14 | 82.14 | 91.07 | 8 | 97.07 | 97.90 | 98.74 | 99.58 | 100.00 | 9 | 45.00 | 65.00 | 45.00 | 45.00 | 60.00 | 10 | 71.70 | 79.42 | 90.53 | 95.78 | 97.25 | 11 | 85.09 | 90.87 | 97.10 | 98.20 | 99.67 | 12 | 70.65 | 80.26 | 89.03 | 97.13 | 98.27 | 13 | 96.58 | 99.02 | 100.00 | 99.89 | 100.00 | 14 | 95.65 | 97.86 | 97.15 | 98.81 | 99.68 | 15 | 62.95 | 62.69 | 73.57 | 91.70 | 99.03 | 16 | 82.79 | 91.39 | 70.96 | 82.79 | 87.37 | OA | 79.00 | 87.15 | 90.97 | 96.16 | 98.83 | AA | 71.95 | 84.36 | 84.03 | 90.43 | 95.20 | Kappa | 75.94 | 85.28 | 89.65 | 95.61 | 98.66 |
|
查看原文
表 2不同方法在Pavia University数据集上的分类结果
Table2. Classification results of different methods on Pavia University datasetunit: %
Class | SVM | AEAP | DCNN | FEFCN-ELM | Proposed |
---|
1 | 84.05 | 92.51 | 94.40 | 96.24 | 99.95 | 2 | 94.90 | 97.75 | 98.52 | 98.12 | 99.99 | 3 | 55.93 | 82.46 | 80.56 | 81.65 | 100.00 | 4 | 66.74 | 94.64 | 90.79 | 94.81 | 98.21 | 5 | 88.02 | 98.43 | 98.88 | 97.47 | 100.00 | 6 | 70.07 | 84.21 | 86.41 | 80.53 | 100.00 | 7 | 86.01 | 83.08 | 82.78 | 91.95 | 100.00 | 8 | 90.76 | 83.43 | 88.91 | 93.07 | 99.45 | 9 | 99.68 | 100.00 | 99.89 | 99.68 | 100.00 | OA | 85.63 | 92.75 | 93.75 | 94.10 | 99.82 | AA | 81.80 | 90.72 | 91.24 | 92.61 | 99.73 | Kappa | 80.65 | 90.36 | 91.65 | 92.13 | 99.75 |
|
查看原文
胡丽, 单锐, 王芳, 江国乾, 赵静一, 张智. 基于双通道空洞卷积神经网络的高光谱图像分类[J]. 激光与光电子学进展, 2020, 57(12): 122803. Li Hu, Rui Shan, Fang Wang, Guoqian Jiang, Jingyi Zhao, Zhi Zhang. Hyperspectral Image Classification Based on Dual-Channel Dilated Convolution Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(12): 122803.