基于多尺度近端特征拼接网络的高光谱图像分类方法

doi:10.11959/j.issn.1000-436x.2021024

通信学报 ›› 2021, Vol. 42 ›› Issue (2): 92-102.doi: 10.11959/j.issn.1000-436x.2021024

基于多尺度近端特征拼接网络的高光谱图像分类方法

高红民, 曹雪莹, 陈忠昊, 花再军, 李臣明, 陈月

河海大学计算机与信息学院，江苏南京 211100

修回日期:2020-10-20 出版日期:2021-02-25 发布日期:2021-02-01
作者简介:高红民（1983- ），男，江苏仪征人，博士，河海大学教授，主要研究方向为深度学习等智能计算方法、图像大数据与人工智能、遥感影像处理等。
曹雪莹（1994- ），女，江苏睢宁人，河海大学硕士生，主要研究方向为深度学习与图像处理等。
陈忠昊（1997- ），男，安徽马鞍山人，河海大学硕士生，主要研究方向为深度学习与高光谱图像处理等。
花再军（1983- ），男，江苏姜堰人，河海大学实验师，主要研究方向为机器学习及图像处理等。
李臣明（1969- ），男，内蒙古通辽人，河海大学教授，主要研究方向为智能信息处理、遥感技术与系统等。
陈月（1995- ），女，江苏扬州人，河海大学硕士生，主要研究方向为高光谱遥感影像分类。
基金资助:
国家自然科学基金资助项目(62071168);国家重点研发计划基金资助项目(2018YFC1508106);中央高校基本科研业务费资金资助项目(B200202183);江苏省研究生科研与实践创新计划基金资助项目(SJCX20_0181)

Hyperspectral image classification method based on multi-scale proximal feature concatenate network

Hongmin GAO, Xueying CAO, Zhonghao CHEN, Zaijun HUA, Chenming LI, Yue CHEN

College of Computer and Information Engineering, Hohai University, Nanjing 211100, China

Revised:2020-10-20 Online:2021-02-25 Published:2021-02-01
Supported by:
The National Natural Science Foundation of China(62071168);The National Key Research and Develop-ment Program of China(2018YFC1508106);Fundamental Research Funds for the Central Universities(B200202183);Post-graduate Research ＆ Practice Innovation Program of Jiangsu Province(SJCX20_0181)

摘要/Abstract

摘要：

针对基于传统卷积神经网络模型的高光谱图像分类算法细节表现力不强及网络结构过于复杂的问题，设计了一种基于多尺度近端特征拼接网络的高光谱图像分类方法。通过引入多尺度滤波器和空洞卷积，在保持模型轻量化的同时可以获取更丰富的空间-光谱判别特征，并提出利用卷积神经网络近端特征间的相互联系进一步增强细节表现力。在3个基准高光谱图像数据集上的实验结果表明，所提方法优于其他分类模型。

关键词: 卷积神经网络, 高光谱图像分类, 特征拼接, 多尺度滤波器, 空洞卷积

Abstract:

Aiming at the phenomenon that the hyperspectral classification algorithm based on traditional CNN model was not expressive enough in detail and the network structure was too complex, a hyperspectral image classification method based on multi-scale proximal feature concatenate network (MPFCN) was designed.By introducing multi-scale filter and cavity convolution, the model could be kept light and the discriminative features of the space spectrum could be obtained, and the correlation between the proximal features of the CNN was proposed to further enhance the detail expression.Experimental results on three benchmark hyperspectral image data sets show that the proposed method is superior to other classification models.

Key words: convolutional neural network, hyperspectral image classification, feature concatenate, multi-scale filter, di-lated convolution

中图分类号:

TP302

高红民, 曹雪莹, 陈忠昊, 花再军, 李臣明, 陈月. 基于多尺度近端特征拼接网络的高光谱图像分类方法[J]. 通信学报, 2021, 42(2): 92-102.

Hongmin GAO, Xueying CAO, Zhonghao CHEN, Zaijun HUA, Chenming LI, Yue CHEN. Hyperspectral image classification method based on multi-scale proximal feature concatenate network[J]. Journal on Communications, 2021, 42(2): 92-102.

图/表 22

图1

图2

图3

图4

图5

表1

表2

表3

表4

图6

表5

表6

图7

图8

表7

表8

表9

表10

表11

图9

图10

图11

参考文献 23

[1]	ZHANG L F , ZHANG L P , TAO D C ,et al. Hyperspectral remote sensing image subpixel target detection based on supervised metric learning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2014,52(8): 4955-4965.
[2]	张亚光, 陈建平, 李诗 . 高光谱遥感在土壤重金属污染监测中的应用[J]. 地质学刊, 2019,43(3): 491-498.
	ZHANG Y G , CHEN J P , LI S . Application of hyper-spectral remote sensing in soil heavy metal pollution monitoring[J]. Journal of Geology, 2019,43(3): 491-498.
[3]	王新忠, 卢青, 张晓东 ,等. 基于高光谱图像的黄瓜种子活力无损检测[J]. 江苏农业学报, 2019,35(5): 1197-1202.
	WANG X Z , LU Q , ZHANG X D ,et al. Non-destructive detection of cucumber seeds vigor based on hyperspectral imaging[J]. Jiangsu Journal of Agricultural Sciences, 2019,35(5): 1197-1202.
[4]	SANDIDGE J C , HOLYER R J . Coastal bathymetry from hyperspectral observations of water radiance[J]. Remote Sensing of Environment, 1998,65(3): 341-352.
[5]	SAMANIEGO L , BARDOSSY A , SCHULZ K . Supervised classification of remotely sensed imagery using a modified K-NN technique[J]. IEEE Transactions on Geoscience and Remote Sensing, 2008,46(7): 2112-2125.
[6]	肖博林 . 基于支持向量机的高光谱遥感影像分类[J]. 科技创新与应用, 2020(4): 22-24.
	XIAO B L . Hyperspectral remote sensing image classification based on support vector machin[J]. Technology Innovation and Application, 2020(4): 22-24.
[7]	WANG Q , LIN J Z , YUAN Y . Salient band selection for hyperspectral image classification via manifold ranking[J]. IEEE Transactions on Neural Networks and Learning Systems, 2016,27(6): 1279-1289.
[8]	LICCIARDI G , MARPU P R , CHANUSSOT J ,et al. Linear versus nonlinear PCA for the classification of hyperspectral data based on the extended morphological profiles[J]. IEEE Geoscience and Remote Sensing Letters, 2012,9(3): 447-451.
[9]	GAO W , PENG Y . Ideal kernel-based multiple kernel learning for spectral-spatial classification of hyperspectral image[J]. IEEE Geoscience and Remote Sensing Letters, 2017,14(7): 1051-1055.
[10]	CHEN Y , NASRABADI N M , TRAN T D . Hyperspectral image classification via kernel sparse representation[J]. IEEE Transactions on Geoscience and Remote Sensing, 2013,51(1): 217-231.
[11]	LECUN Y , BENGIO Y , HINTON G . Deep learning[J]. Nature, 2015,521(7553): 436-444.
[12]	CHEN Y S , LIN Z H , ZHAO X ,et al. Deep learning-based classification of hyperspectral data[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2014,7(6): 2094-2107.
[13]	LIU P , ZHANG H , EOM K B . Active deep learning for classification of hyperspectral images[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017,10(2): 712-724.
[14]	CHEN Y S , JIANG H L , LI C Y ,et al. Deep feature extraction and classification of hyperspectral images based on convolutional neural networks[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016,54(10): 6232-6251.
[15]	LI Y , ZHANG H K , SHEN Q ,et al. Spectral-spatial classification of hyperspectral imagery with 3D convolutional neural network[J]. Remote Sensing, 2017,9(1): 67.
[16]	HE K M , ZHANG X Y , REN S Q ,et al. Deep residual learning for image recognition[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016: 770-778.
[17]	ZHONG Z L , LI J , LUO Z M ,et al. Spectral-spatial residual network for hyperspectral image classification:a 3-D deep learning framework[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018,56(2): 847-858.
[18]	SONG W W , LI S T , FANG L Y ,et al. Hyperspectral image classification with deep feature fusion network[J]. IEEE Transactions on Geoscience and Remote Sensing, 2018,56(6): 3173-3184.
[19]	WANG W J , DOU S G , JIANG Z M ,et al. A fast dense spectral-spatial convolution network framework for hyperspectral images classification[J]. Remote Sensing, 2018,10(7): 1068.
[20]	王莹 . 改进的基于 CNN 的高光谱遥感图像分类办法[J]. 现代商贸工业, 2019,40(35): 204-206.
	WANG Y . Improved hyperspectral remote sensing image classification method based on CNN[J]. Modern Business Trade Industry, 2019,40(35): 204-206.
[21]	SZEGEDY C , LIU W , JIA Y Q ,et al. Going deeper with convolutions[C]// 2015 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2015: 1-9.
[22]	LEE H , KWON H . Going deeper with contextual CNN for hyperspectral image classification[J]. IEEE Transactions on Image Processing, 2017,26(10): 4843-4855.
[23]	ZHANG C J , LI G D , DU S H . Multi-scale dense networks for hyperspectral remote sensing image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019,57(11): 9201-9222.

数据集	类别	名称	样本数/个	样本总数/个
	C1	Alfalfa	46
	C2	Corn-notill	1 428
	C3	Corn-mintill	830
	C4	Corn	237
	C5	Grass-pasture	483
	C6	Grass-trees	730
	C7	Grass-pasture-mowed	28
IP	C8	Hay-windrowed	478	10 249
	C9	Oats	20
	C10	Soybean-notill	972
	C11	Soybean-mintill	2 455
	C12	Soybean-clean	593
	C13	Wheat	205
	C14	Woods	1 265
	C15	Buildings-Grass-Trees_Drives	386
	C16	Stone-Steel-Towers	93
	C1	Asphalt	6 631
	C2	Meadows	18 649
	C3	Gravel	2 099
	C4	Trees	3 064
PU	C5	Painted metal sheets	1 345	42 776
	C6	Bare Soil	5 029
	C7	Bitumen	1 330
	C8	Self-Blocking Bricks	3 682
	C9	Shadows	947
	C1	Brocoli_green_weeds_1	2 009
	C2	Brocoli_green_weeds_2	3 726
	C3	Fallow	1 976
	C4	Fallow_rough_plow	1 394
	C5	Fallow_smooth	2 678
	C6	Stubble	3 959
	C7	Celery	3 579
SA	C8	Grapes_untrained	11 271	54 129
	C9	Soil_vinyard_develop	6 203
	C10	Corn_senesced_green_weeds	3 278
	C11	Lettuce_romaine_4wk	1 068
	C12	Lettuce_romaine_5wk	1 927
	C13	Lettuce_romaine_6wk	916
	C14	Lettuce_romaine_7wk	1 070
	C15	Vinyard_untrained	7 268
	C16	Vinyard_vertical_trellis	1 807

多尺度滤波器模块数/个	patch尺寸	OA	AA	KA
5	27×27	98.51%	98.50%	98.31%
4	27×27	98.48%	98.42%	98.27%
	13×13	95.89%	95.65%	95.31%
	27×27	98.43%	97.86%	98.22%
3	13×13	96.43%	96.11%	95.94%
	7×7	83.13%	82.51%	79.66%
	27×27	98.03%	96.31%	97.75%
2	13×13	95.85%	95.69%	92.26%
	7×7	85.86%	84.32%	83.93%

多尺度滤波器模块数/个	patch尺寸	OA	AA	KA
5	27×27	99.57%	99.14%	99.43%
4	27×27	99.50%	98.84%	99.34%
	13×13	98.29%	97.65%	97.74%
	27×27	99.40%	98.62%	99.21%
3	13×13	98.56%	97.70%	98.10%
	7×7	92.49%	90.46%	89.91%
	27×27	99.08%	97.83%	98.78%
2	13×13	98.30%	97.67%	97.74%
	7×7	94.28%	92.59%	92.36%

多尺度滤波器模块数/个	patch尺寸	OA	AA	KA
5	27×27	99.70%	99.71%	99.66%
4	27×27	99.78%	99.73%	99.76%
	13×13	95.97%	97.77%	95.51%
	27×27	99.70%	99.70%	99.67%
3	13×13	95.09%	98.03%	95.44%
	7×7	90.44%	93.58%	89.34%
	27×27	99.51%	99.49%	99.45%
2	13×13	95.05%	97.45%	94.48%
	7×7	90.77%	93.32%	89.75%

类别	SVM	DCNN	ResNet	MCNN	MPFCN
Alfalfa	37.21%	100%	96.34%	95.93%	97.93%
Corn-notill	77.60%	98.17%	97.20%	97.17%	98.52%
Corn-mintill	57.54%	93.04%	97.05%	94.24%	96.90%
Corn	60.29%	100%	99.77%	98.59%	99.84%
Grass-pasture	90.40%	97.01%	99.20%	96.55%	96.27%
Grass-trees	92.89%	97.49%	97.95%	99.49%	99.59%
Grass-pasture-mowed	86.36%	94.00%	100%	94.67%	99.8%
Hay-windrowed	97.90%	99.88%	97.90%	99.84%	100%
Oats	55.05%	96.875%	100%	97.92%	100%
Soybean-notill	50.45%	96.34%	96.80%	96.42%	97.75%
Soybean-mintill	72.16%	98.44%	99.30%	98.43%	98.30%
Soybean-clean	50.56%	93.16%	92.98%	95.94%	96.50%
Wheat	93.86%	100%	98.91%	95.83%	98.73%
Woods	89.00%	99.12	99.52%	98.60%	99.94%
Buildings-Grass-Trees-Drives	48.02%	99.14%	99.86%	99.90%	98.27%
Stone-Steel-Towers	91.67%	91.67%	88.71%	96.03%	97.62%

基于多尺度近端特征拼接网络的高光谱图像分类方法

Hyperspectral image classification method based on multi-scale proximal feature concatenate network

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 22

参考文献 23

相关文章 15

Metrics

推荐阅读 0

方法	OA	AA	KA	训练参数数量/个
SVM	75.07%	71.65%	71.52%	—
DCNN	97.50%	97.15%	97.15%	1 109 776
ResNet	97.97%	97.41%	97.69%	11 002 320
MCNN	97.62%	97.22%	97.29%	43 402 288
MPFCN	98.51%	98.50%	98.31%	12 593 104

方法	OA	AA	KA	训练参数数量/个
SVM	87.52%	82.90%	83.28%	—
DCNN	98.54%	96.47%	98.06%	1 109 776
ResNet	98.66%	96.53%	98.22%	10 989 129
MCNN	98.98%	98.68%	98.65%	43 396 841
MPFCN	99.56%	99.14%	99.43%	12 583 497

方法	OA	AA	KA	训练参数数量/个
SVM	89.21%	93.65%	87.94%	—
DCNN	98.91%	99.22%	98.79%	1 109 776
ResNet	98.36%	98.96%	98.18%	10 994 640
MCNN	98.95%	99.25%	98.84%	43 402 288
MPFCN	99.70%	99.71%	99.66%	12 593 104

分类方法	训练时间/s	测试时间/s
DCNN	336.95	35.90
ResNet	335.24	35.83
MCNN	388.32	31.50
MPFCN	131.75	11.80

[1]	李昂, 陈建新, 魏昕, 周亮. 面向6G的跨模态信号重建技术[J]. 通信学报, 2022, 43(6): 28-40.
[2]	王晓丹, 李京泰, 宋亚飞. DDAC：面向卷积神经网络图像隐写分析模型的特征提取方法[J]. 通信学报, 2022, 43(5): 68-81.
[3]	廖育荣, 王海宁, 林存宝, 李阳, 方宇强, 倪淑燕. 基于深度学习的光学遥感图像目标检测研究进展[J]. 通信学报, 2022, 43(5): 190-203.
[4]	张帆, 黄赟, 方子茁, 郭威. 卷积神经网络的损失最小训练后参数量化方法[J]. 通信学报, 2022, 43(4): 114-122.
[5]	朱政宇, 侯庚旺, 黄崇文, 孙钢灿, 郝万明, 梁静. 基于并行CNN的RIS辅助D2D保密通信系统资源分配算法[J]. 通信学报, 2022, 43(3): 172-179.
[6]	安泽亮, 张天骐, 马宝泽, 邓盼, 徐雨晴. 基于一维CNN的多入多出OSTBC信号协作调制识别[J]. 通信学报, 2021, 42(7): 84-94.
[7]	麻文刚, 张亚东, 郭进. 基于LSTM与改进残差网络优化的异常流量检测方法[J]. 通信学报, 2021, 42(5): 23-40.
[8]	王洪雁, 张莉彬, 陈国强, 汪祖民, 管志远. 结合粒子滤波及度量学习的目标跟踪方法[J]. 通信学报, 2021, 42(5): 98-110.
[9]	司念文, 张文林, 屈丹, 常禾雨, 李盛祥, 牛铜. 基于对抗补丁的可泛化的Grad-CAM攻击方法[J]. 通信学报, 2021, 42(3): 23-35.
[10]	杨军,党吉圣. 基于上下文注意力CNN的三维点云语义分割[J]. 通信学报, 2020, 41(7): 195-203.
[11]	高红民,曹雪莹,杨耀,花再军,李臣明. 基于CNN的双边融合网络在高光谱图像分类中的应用[J]. 通信学报, 2020, 41(11): 132-140.
[12]	张猛,孙昊良,杨鹏. 基于改进卷积神经网络识别DNS隐蔽信道[J]. 通信学报, 2020, 41(1): 169-179.
[13]	周鑫,何晓新,郑昌文. 基于图像深度学习的无线电信号识别[J]. 通信学报, 2019, 40(7): 114-125.
[14]	查雄,彭华,秦鑫,李广,李天昀. 基于多端卷积神经网络的调制识别方法[J]. 通信学报, 2019, 40(11): 30-37.
[15]	李佳,云晓春,李书豪,张永铮,谢江,方方. 基于混合结构深度神经网络的HTTP恶意流量检测方法[J]. 通信学报, 2019, 40(1): 24-33.

类别	SVM	DCNN	ResNet	MCNN	MPFCN
Asphalt	91.66%	99.68%	99.99%	99.35%	99.64%
Meadows	94.91%	99.92%	99.93%	99.20%	99.92%
Gravel	71.37%	98.81%	97.94%	96.67%	99.06%
Trees	85.00%	90.38%	92.39%	98.22%	98.07%
Painted metal sheets	99.00%	96.39%	95.91%	99.77%	99.76%
Bare Soil	68.09%	100%	99.81%	99.15%	100%
Bitumen	49.41%	99.40%	99.78%	98.72%	99.10%
Self-Blocking Bricks	86.76%	97.90%	97.93%	98.92%	99.24%
Shadows	99.89%	85.77%	85.12%	98.17%	97.52%