基于自适应流形滤波的高光谱图像分类方法

廖建尚; 王立国; 郝思媛

doi:doi:10.3788/LOP55.041010

激光与光电子学进展, 2018, 55 (4): 041010, 网络出版: 2018-09-11

基于自适应流形滤波的高光谱图像分类方法下载： 861次

Hyperspectral Image Classification Method Based on Adaptive Manifold Filtering

廖建尚 ^*王立国郝思媛

作者单位

广东交通职业技术学院轨道交通学院, 广东广州 510650

图 & 表

图 1. 印第安农林数据集的原始光谱和滤波结果。(a)第10个波段;(b)第80个波段;(c)第120个波段;(d)第180个波段

Fig. 1. Original spectrum and filtering results of Indian Pines data sets. (a) 10th band; (b) 80th band; (c) 120th band; (d) 180th band

下载图片查看原文

图 2. 印第安农林数据集的流形滤波系数寻优。 (a)空间偏差系数σs; (b)范围偏差系数σr

Fig. 2. Optimization for manifold filtering coefficient of Indian Pines data sets. (a) Spatial deviation coefficient σs; (b) range deviation coefficient σr

下载图片查看原文

图 3. AMF-SVM流程

Fig. 3. Flow of AMF-SVM

下载图片查看原文

图 4. 印第安农林数据集分类情况。(a)真实图像;(b) SVM,OA为80.93%;(c) SVM-PCA,OA为80.46%;(d) GBF-SVM,OA为82.82%;(e) BF-SVM,OA为88.99%;(f) GDF-SVM,OA为91.08%;(g) EPF-B-g,OA为92.99%; (h) EPF-G-g,OA为92.83%;(i) IFRF,OA为93.64%;(j) AMF-SVM,OA为95.16%

Fig. 4. Classification of Indian Pines data sets. (a) Ground truth; (b) SVM, OA is 80.93%; (c) SVM-PCA, OA is 80.46%; (d) GBF-SVM, OA is 82.82%; (e) BF-SVM, OA is 88.99%; (f) GDF-SVM, OA is 91.08%; (g) EPF-B-g, OA is 92.99%; (h) EPF-G-g, OA is 92.83%; (i) IFRF, OA is 93.64%; (j) AMF-SVM, OA is 95.16%

下载图片查看原文

图 5. 帕威亚大学数据集分类情况。(a)真实图像;(b) SVM, OA为84.80%;(c) SVM-PCA,OA为83.95%;(d) GBF-SVM,OA为85.20%;(e) BF-SVM,OA为89.03%;(f) GDF-SVM,OA为94.20%;(g) EPF-B-g,OA为91.29%;(h) EPF-G-g,OA为91.68%;(i) IFRF,OA为95.31%;(j) AMF-SVM,OA为97.92%

Fig. 5. Classification for Pavia University. (a) Ground truth;(b) SVM, OA is 84.80%; (c) SVM-PCA, OA is 83.95%; (d) GBF-SVM, OA is 85.20%; (e) BF-SVM, OA is 89.03%; (f) GDF-SVM, OA is 94.20%; (g) EPF-B-g, OA is 91.29%; (h) EPF-G-g, OA is 91.68%; (i) IFRF, OA is 95.31%; (j) AMF-SVM, OA is 97.92%

下载图片查看原文

图 6. 不同训练样本比例分类后OA和Kappa系数折线图。 (a)印第安农林;(b)帕维亚大学

Fig. 6. Charts of OA and Kappa coefficient with different training samples. (a) Indian Pines; (b) Pavia University

下载图片查看原文

图 7. 分类算法的OA和Kappa系数。(a)印第安农林1%的训练样本;(b)帕维亚大学0.1%的训练样本

Fig. 7. OA and Kappa coefficient for different classification methods. (a) 1% training sample for Indian Pins; (b) 0.1% training sample for Pavia University

下载图片查看原文

图 8. 自适应流形滤波的高光谱分类寻优。 (a)印第安农林;(b)帕维亚大学

Fig. 8. Optimization for hyperspectral classification of adaptive manifold filtering. (a) Indian Pins; (b) Pavia University

下载图片查看原文

表 1印第安农林数据集图像分类数据统计

Table1. Classification data statistics of Indian Pines data sets

Groundtruth	SumsampleNo.	TrainsampleNo. /%	TestsampleNo. /%	SVM /%	SVM-PCA /%	GBF-SVM /%	BF-SVM /%	GDF-SVM /%	EPF-B-g /%	EPF-G-g /%	IFRF /%	AMF-SVM /%
Alfalfa	54	7	93	83.57	78.89	88.83	91.86	91.10	95.58	94.78	91.34	92.24
Corn-no till	1434	7	93	71.50	71.08	76.46	84.28	87.25	91.57	91.39	91.23	96.95
Corn-min till	834	7	93	70.63	72.02	70.38	88.93	91.67	87.34	87.54	84.64	97.90
Corn	234	7	93	44.19	41.48	51.61	57.98	66.21	62.57	61.39	86.22	87.16
Grass/pasture	497	7	93	89.90	89.12	88.96	92.29	93.60	95.48	94.82	93.96	93.58
Grass/trees	747	7	93	94.79	94.63	95.15	96.79	96.86	99.79	99.50	98.10	97.40
Grass/pasture-mowed	26	7	93	53.91	53.27	66.50	62.45	64.31	54.19	62.80	88.13	76.67
Hay-windrowed	489	7	93	97.16	96.18	99.56	98.33	97.47	100.0	100.0	99.58	99.16
Oats	20	7	93	46.99	47.52	75.94	57.46	62.42	22.69	39.34	89.47	94.07
Soybeans-no till	968	7	93	69.29	68.28	67.89	83.05	84.41	87.59	86.21	87.14	92.83
Soybeans-min till	2468	7	93	85.12	84.43	86.88	91.70	93.68	97.71	97.51	95.96	98.51
Soybeans-clean till	614	7	93	79.40	78.41	74.90	87.71	90.03	95.67	95.88	95.24	96.58
Wheat	212	7	93	95.98	96.53	97.00	97.38	97.58	99.95	99.60	99.34	98.38
Woods	1294	7	93	97.67	97.97	98.19	98.08	98.59	99.94	99.81	98.84	99.01
Bldg-Grass-Tree	380	7	93	45.94	43.59	68.51	64.42	74.86	60.16	61.26	91.16	78.77
Stone-steeltowers	95	7	93	76.42	76.16	71.16	76.42	81.82	93.35	97.73	83.37	82.04
OA /%	-	-	-	80.93	80.46	82.82	88.99	91.08	92.99	92.83	93.62	96.16
Kappa	-	-	-	78.12	77.58	80.28	87.41	89.81	91.96	91.78	92.11	95.62

查看原文

表 2帕维亚大学数据集图像分类数据统计

Table2. Classification statistics of Pavia University data sets

Groundtruth	Sum	Train /%	Test /%	SVM /%	SVM-PCA /%	GBF-SVM /%	BF-SVM /%	GDF-SVM /%	EPF-B-g /%	EPF-G-g /%	IFRF /%	AMF-SVM /%
Asphalt	6641	2	98	87.84	86.19	88.74	88.23	94.98	98.07	97.49	97.70	98.68
Meadows	18649	2	98	95.81	95.99	96.13	97.03	98.32	99.98	99.91	99.34	99.79
Gravel	2099	2	98	57.87	48.76	54.51	65.01	76.07	72.60	69.39	86.68	90.63
Trees	3064	2	98	88.17	85.01	89.21	91.98	96.19	91.84	92.26	92.78	96.56
Metalsheets	1345	2	98	98.34	98.72	98.84	97.54	98.38	99.85	99.94	99.02	99.40
Soil	5029	2	98	54.33	54.96	56.21	77.91	88.34	60.74	60.32	99.86	97.59
Bitumen	1330	2	98	64.64	64.79	65.89	70.50	82.89	81.27	86.38	96.37	95.12
Bricks	3682	2	98	78.97	79.41	77.88	80.18	91.43	98.47	95.95	73.13	97.05
Shadows	947	2	98	89.33	84.29	90.64	87.82	93.37	95.13	93.20	83.10	94.49
OA /%	-	-	-	84.80	83.96	85.20	89.03	94.20	92.32	91.92	95.31	98.17
Kappa	-	-	-	79.47	78.31	80.00	85.34	92.29	89.57	89.04	93.67	97.57

查看原文

表 3自适应流形滤波的高光谱分类数据统计

Table3. Hyperspectral classification data statistics of adaptive manifold filtering

Index	n	0	1	2	3	4	5	6	7	8
	Tree height	2	3	4	5	6	7	8	9	10
	Tree node	3	7	15	31	63	127	255	511	1023
Indian Pines	OA /%	95.61	95.67	95.97	95.61	95.80	96.15	96.04	96.16	96.13
Indian Pines	Kappa	94.98	95.05	95.39	94.99	95.20	95.61	95.47	95.62	95.58
Pavia	OA /%	98.02	98.19	98.14	98.20	98.08	98.17	98.40	98.29	98.17
Pavia	Kappa	97.37	97.60	97.53	97.62	97.45	97.58	97.88	97.73	97.57

查看原文

廖建尚, 王立国, 郝思媛. 基于自适应流形滤波的高光谱图像分类方法[J]. 激光与光电子学进展, 2018, 55(4): 041010. Jianshang Liao, Liguo Wang, Siyuan Hao. Hyperspectral Image Classification Method Based on Adaptive Manifold Filtering[J]. Laser & Optoelectronics Progress, 2018, 55(4): 041010.

基于自适应流形滤波的高光谱图像分类方法下载： 861次

图 1. 印第安农林数据集的原始光谱和滤波结果。(a)第10个波段;(b)第80个波段;(c)第120个波段;(d)第180个波段

Fig. 1. Original spectrum and filtering results of Indian Pines data sets. (a) 10th band; (b) 80th band; (c) 120th band; (d) 180th band

图 2. 印第安农林数据集的流形滤波系数寻优。 (a)空间偏差系数σs; (b)范围偏差系数σr

Fig. 2. Optimization for manifold filtering coefficient of Indian Pines data sets. (a) Spatial deviation coefficient σs; (b) range deviation coefficient σr

图 3. AMF-SVM流程

Fig. 3. Flow of AMF-SVM

图 4. 印第安农林数据集分类情况。(a)真实图像;(b) SVM,OA为80.93%;(c) SVM-PCA,OA为80.46%;(d) GBF-SVM,OA为82.82%;(e) BF-SVM,OA为88.99%;(f) GDF-SVM,OA为91.08%;(g) EPF-B-g,OA为92.99%; (h) EPF-G-g,OA为92.83%;(i) IFRF,OA为93.64%;(j) AMF-SVM,OA为95.16%

图 5. 帕威亚大学数据集分类情况。(a)真实图像;(b) SVM, OA为84.80%;(c) SVM-PCA,OA为83.95%;(d) GBF-SVM,OA为85.20%;(e) BF-SVM,OA为89.03%;(f) GDF-SVM,OA为94.20%;(g) EPF-B-g,OA为91.29%;(h) EPF-G-g,OA为91.68%;(i) IFRF,OA为95.31%;(j) AMF-SVM,OA为97.92%

Fig. 5. Classification for Pavia University. (a) Ground truth;(b) SVM, OA is 84.80%; (c) SVM-PCA, OA is 83.95%; (d) GBF-SVM, OA is 85.20%; (e) BF-SVM, OA is 89.03%; (f) GDF-SVM, OA is 94.20%; (g) EPF-B-g, OA is 91.29%; (h) EPF-G-g, OA is 91.68%; (i) IFRF, OA is 95.31%; (j) AMF-SVM, OA is 97.92%

图 6. 不同训练样本比例分类后OA和Kappa系数折线图。 (a)印第安农林;(b)帕维亚大学

Fig. 6. Charts of OA and Kappa coefficient with different training samples. (a) Indian Pines; (b) Pavia University

图 7. 分类算法的OA和Kappa系数。(a)印第安农林1%的训练样本;(b)帕维亚大学0.1%的训练样本

Fig. 7. OA and Kappa coefficient for different classification methods. (a) 1% training sample for Indian Pins; (b) 0.1% training sample for Pavia University

图 8. 自适应流形滤波的高光谱分类寻优。 (a)印第安农林;(b)帕维亚大学

Fig. 8. Optimization for hyperspectral classification of adaptive manifold filtering. (a) Indian Pins; (b) Pavia University

表 1印第安农林数据集图像分类数据统计

Table1. Classification data statistics of Indian Pines data sets

表 2帕维亚大学数据集图像分类数据统计

Table2. Classification statistics of Pavia University data sets

表 3自适应流形滤波的高光谱分类数据统计

Table3. Hyperspectral classification data statistics of adaptive manifold filtering

关于本站 Cookie 的使用提示

全站搜索

基于自适应流形滤波的高光谱图像分类方法 下载： 861次

图 1. 印第安农林数据集的原始光谱和滤波结果。(a)第10个波段;(b)第80个波段;(c)第120个波段;(d)第180个波段

Fig. 1. Original spectrum and filtering results of Indian Pines data sets. (a) 10th band; (b) 80th band; (c) 120th band; (d) 180th band

图 2. 印第安农林数据集的流形滤波系数寻优。 (a)空间偏差系数σs; (b)范围偏差系数σr

Fig. 2. Optimization for manifold filtering coefficient of Indian Pines data sets. (a) Spatial deviation coefficient σs; (b) range deviation coefficient σr

图 3. AMF-SVM流程

Fig. 3. Flow of AMF-SVM

图 4. 印第安农林数据集分类情况。(a)真实图像;(b) SVM,OA为80.93%;(c) SVM-PCA,OA为80.46%;(d) GBF-SVM,OA为82.82%;(e) BF-SVM,OA为88.99%;(f) GDF-SVM,OA为91.08%;(g) EPF-B-g,OA为92.99%; (h) EPF-G-g,OA为92.83%;(i) IFRF,OA为93.64%;(j) AMF-SVM,OA为95.16%

图 5. 帕威亚大学数据集分类情况。(a)真实图像;(b) SVM, OA为84.80%;(c) SVM-PCA,OA为83.95%;(d) GBF-SVM,OA为85.20%;(e) BF-SVM,OA为89.03%;(f) GDF-SVM,OA为94.20%;(g) EPF-B-g,OA为91.29%;(h) EPF-G-g,OA为91.68%;(i) IFRF,OA为95.31%;(j) AMF-SVM,OA为97.92%

Fig. 5. Classification for Pavia University. (a) Ground truth;(b) SVM, OA is 84.80%; (c) SVM-PCA, OA is 83.95%; (d) GBF-SVM, OA is 85.20%; (e) BF-SVM, OA is 89.03%; (f) GDF-SVM, OA is 94.20%; (g) EPF-B-g, OA is 91.29%; (h) EPF-G-g, OA is 91.68%; (i) IFRF, OA is 95.31%; (j) AMF-SVM, OA is 97.92%

图 6. 不同训练样本比例分类后OA和Kappa系数折线图。 (a)印第安农林;(b)帕维亚大学

Fig. 6. Charts of OA and Kappa coefficient with different training samples. (a) Indian Pines; (b) Pavia University

图 7. 分类算法的OA和Kappa系数。(a)印第安农林1%的训练样本;(b)帕维亚大学0.1%的训练样本

Fig. 7. OA and Kappa coefficient for different classification methods. (a) 1% training sample for Indian Pins; (b) 0.1% training sample for Pavia University

图 8. 自适应流形滤波的高光谱分类寻优。 (a)印第安农林;(b)帕维亚大学

Fig. 8. Optimization for hyperspectral classification of adaptive manifold filtering. (a) Indian Pins; (b) Pavia University

表 1印第安农林数据集图像分类数据统计

Table1. Classification data statistics of Indian Pines data sets

表 2帕维亚大学数据集图像分类数据统计

Table2. Classification statistics of Pavia University data sets

表 3自适应流形滤波的高光谱分类数据统计

Table3. Hyperspectral classification data statistics of adaptive manifold filtering

相关论文

相关资讯

关于本站 Cookie 的使用提示

全站搜索

基于自适应流形滤波的高光谱图像分类方法下载： 861次