基于多尺度残差网络的小样本高光谱图像分类 下载: 1119次
Classification of Small-Sized Sample Hyperspectral Images Based on Multi-Scale Residual Network
长安大学地质工程与测绘学院, 陕西 西安 710054
图 & 表
图 1. 残差学习模块
Fig. 1. Residual learning block
下载图片 查看原文
图 2. 多尺度光谱特征提取模块
Fig. 2. Multi-scale spectral feature extraction block
下载图片 查看原文
图 3. 多尺度空间特征提取模块
Fig. 3. Multi-scale spatial feature extraction block
下载图片 查看原文
图 4. 多尺度残差网络
Fig. 4. Multi-scale residual network
下载图片 查看原文
图 5. 不同卷积核数的模型总体精度
Fig. 5. Overall accuracy of models with different number of kernels
下载图片 查看原文
图 6. 不同空间尺寸输入的分类精度对比。(a) IN;(b) UP
Fig. 6. Comparison of classification accuracy of inputs with different spatial dimensions. (a) IN; (b) UP
下载图片 查看原文
图 7. IN 数据集分类图
Fig. 7. Classification maps of IN dataset
下载图片 查看原文
图 8. IN数据集分类图局部放大对比
Fig. 8. Partial enlargement comparison of classification maps of IN dataset
下载图片 查看原文
图 9. UP数据集分类图
Fig. 9. Classification maps of UP dataset
下载图片 查看原文
图 10. UP数据集分类图局部放大对比
Fig. 10. Partial enlargement comparison of classification maps of UP dataset
下载图片 查看原文
表 1IN数据集样本数量分布
Table1. Sample number distribution of IN dataset
Sample No. | Class | Train | Validation | Test |
---|
1 | Alfalfa | 5 | 4 | 37 | 2 | Corn-notill | 142 | 143 | 1143 | 3 | Corn-mintill | 83 | 83 | 664 | 4 | Corn | 23 | 24 | 190 | 5 | Grass-pasture | 48 | 48 | 387 | 6 | Grass-trees | 73 | 73 | 584 | 7 | Grass-pasture-mowed | 3 | 3 | 22 | 8 | Hay-windrow | 48 | 47 | 383 | 9 | Oats | 2 | 2 | 16 | 10 | Soybean-nottill | 97 | 97 | 778 | 11 | Soybean-mintill | 245 | 246 | 1964 | 12 | Soybean-clean | 59 | 59 | 475 | 13 | Wheat | 20 | 20 | 165 | 14 | Woods | 127 | 126 | 1012 | 15 | Building-Grass-Trees | 39 | 38 | 309 | 16 | Stone-Steel-Towers | 9 | 9 | 75 | Total | 1023 | 1022 | 8204 |
|
查看原文
表 2UP数据集样本数量分布
Table2. Sample number distribution of UP dataset
Sample No. | Class | Train | Validation | Test |
---|
1 | Asphalt | 331 | 332 | 5968 | 2 | Meadows | 932 | 933 | 16784 | 3 | Gravels | 105 | 105 | 1889 | 4 | Trees | 153 | 153 | 2758 | 5 | Painted-Metal-Sheets | 67 | 67 | 1211 | 6 | Bare-Soil | 251 | 252 | 4526 | 7 | Bitumen | 66 | 67 | 1197 | 8 | Self-Blocking-Bricks | 185 | 183 | 3314 | 9 | Shadows | 48 | 47 | 852 | Total | 2138 | 2139 | 38499 |
|
查看原文
表 3不同方法的分类精度对比
Table3. Comparison of classification accuracy of different methods%
Method | IN | UP |
---|
OA | AA | K | OA | AA | K |
---|
SVM | 78.82 | 74.66 | 76.43 | 86.22 | 85.65 | 85.76 | CNN | 91.45 | 89.87 | 90.36 | 96.69 | 96.20 | 95.98 | Res-3D-CNN | 95.63 | 91.02 | 92.35 | 97.65 | 97.24 | 96.85 | SSRN | 97.84 | 94.28 | 96.82 | 99.17 | 99.09 | 99.11 | ResDenNet | 97.98 | 96.48 | 96.89 | 99.33 | 99.14 | 99.21 | MSRN | 99.07 | 98.87 | 98.90 | 99.96 | 99.94 | 99.93 |
|
查看原文
表 4不同算法训练和测试时间对比
Table4. Comparison of training time and test time of different algorithmss
Dataset | Time | CNN | Res-3D-CNN | SSRN | ResDenNet | MSRN |
---|
IN | Training time | 509.50 | 596.40 | 628.60 | 256.52 | 229.61 | | Test time | 6.42 | 6.89 | 7.97 | 5.42 | 7.46 | UP | Training time | 1321.60 | 1256.20 | 1034.40 | 203.47 | 192.96 | | Test time | 8.73 | 18.69 | 16.54 | 17.85 | 22.57 |
|
查看原文
张祥东, 王腾军, 杨耘. 基于多尺度残差网络的小样本高光谱图像分类[J]. 激光与光电子学进展, 2020, 57(16): 162801. Xiangdong Zhang, Tengjun Wang, Yun Yang. Classification of Small-Sized Sample Hyperspectral Images Based on Multi-Scale Residual Network[J]. Laser & Optoelectronics Progress, 2020, 57(16): 162801.