A new kernel method for hyperspectral image feature extraction

Figures & data

Figure 1. Flowchart of the KM-KMNF algorithm.

Figure 2. Experiment results on Pavia University data (a) Original Pavia University image. (b) Ground reference map containing nine land-cover classes. (c) Original K-means clustering result (same-colored region belongs to the same subregion). (d) The result of the image segmentation using the OK-means clustering method (each closed region is a subregion). The results of the ML classification after using different dimensionality reduction methods (number of features = 8), (e) OMNF method, (f) KMNF method, (g) OKMNF method, and (h) KM-KMNF method.

Table 1. Pavia University samples used for training and testing.

Classes	Training	Testing
Asphalt	3315	3316
Meadows	9324	9325
Gravel	1049	1050
Trees	1532	1532
Metal sheets	672	673
Bare soil	2514	2515
Bitumen	665	665
Bricks	1841	1841
Shadows	473	474
Total	21,385	21,391

Table 2. Overall accuracy of the ML classification on Pavia University image using different dimensionality reduction methods.

Number of features	KMNF, %	OMNF, %	OKMNF, %	KM-KMNF, %
3	67.63	73.70	74.13	78.23
4	70.99	78.53	77.71	80.45
5	73.85	79.66	79.24	84.25
6	78.28	81.01	83.03	85.59
7	81.45	82.79	86.11	88.06
8	83.06	84.05	87.24	88.94
9	83.37	85.96	89.31	90.60
10	84.04	87.96	90.02	91.42
11	85.31	90.71	91.10	92.56
12	86.15	91.54	91.46	93.74
13	86.85	91.67	92.39	94.26
14	87.79	92.01	92.44	94.81
15	88.61	92.24	92.46	94.45
16	88.88	92.21	92.37	94.52
17	88.97	92.61	92.56	94.17
18	88.58	92.15	92.19	94.20
19	88.59	92.45	92.57	93.49
20	88.25	92.13	92.55	93.52

Figure 3. Comparison of the accuracies of the ML classification on Pavia University image using different number of features.

Figure 4. Experiment results on Indian Pines data (a) Original Indian Pines image. (b) Ground reference map containing nine land-cover classes. (c) Original K-means clustering result (same-colored region belongs to the same subregion). (d) The result of image segmentation using the optimal K-means clustering method (each closed region is a subregion). The results of the ML classification after using different dimensionality reduction methods (number of features = 8), (e) OMNF method, (f) KMNF method, (g) OKMNF method, and (h) KM-KMNF method.

Table 3. Samples from Indian Pines used for training and testing.

Classes	Training	Testing
Corn-no till	359	1075
Corn-min till	209	625
Grass/Pasture	124	373
Grass/Trees	187	560
Hay-windrowed	122	367
Soybean-no till	242	726
Soybean-min till	617	1851
Soybean-clean till	154	460
Woods	324	970
Total	2338	7007

Table 4. Overall accuracies of the ML classification on Indian Pines image using different dimensionality reduction methods.

Number of features	KMNF(Std)	OMNF(Std)	OKMNF(Std)	KM-KMNF(Std)
3	59.79% (0.012 3)	66.56% (0.004 8)	64.41% (0.005 2)	67.65% (0.004 3)
4	66.62% (0.011 6)	72.83% (0.004 7)	67.53% (0.007 7)	72.75% (0.003 0)
5	69.72% (0.017 2)	74.80% (0.004 4)	76.60% (0.004 1)	77.50% (0.004 6)
6	73.55% (0.008 5)	76.54% (0.004 2)	78.48% (0.002 6)	80.39% (0.003 1)
7	74.77% (0.006 7)	77.81% (0.004 5)	79.94% (0.003 3)	81.98% (0.003 7)
8	76.27% (0.003 8)	77.96% (0.004 2)	82.84% (0.005 3)	84.31% (0.004 0)
9	76.84% (0.001 7)	82.02% (0.004 3)	84.84% (0.006 2)	84.36% (0.007 1)
10	77.90% (0.004 2)	83.92% (0.003 9)	85.48% (0.003 9)	85.54% (0.004 8)
11	78.62% (0.006 5)	84.89% (0.003 7)	85.92% (0.003 7)	86.43% (0.004 5)
12	79.17% (0.017 6)	86.27% (0.003 3)	86.33% (0.003 3)	87.17% (0.004 5)
13	79.79% (0.015 0)	86.68% (0.005 7)	86.61% (0.005 7)	88.00% (0.002 4)
14	80.43% (0.013 7)	87.70% (0.004 0)	86.89% (0.001 6)	89.01% (0.004 6)
15	80.74% (0.002 0)	88.07% (0.002 7)	87.40% (0.001 7)	89.30% (0.003 4)
16	80.89% (0.004 5)	88.54% (0.003 9)	87.75% (0.001 0)	89.70% (0.003 1)
17	80.93% (0.007 3)	88.44% (0.003 4)	87.94% (0.002 2)	89.65% (0.003 4)
18	81.07% (0.006 2)	88.94% (0.004 4)	87.91% (0.001 2)	89.64% (0.003 7)
19	81.20% (0.005 4)	88.88% (0.004 2)	88.17% (0.002 0)	89.61% (0.003 3)
20	81.18% (0.005 0)	88.94% (0.004 2)	88.51% (0.003 3)	89.75% (0.003 2)

Table 5. Overall accuracies of the SVM classification on Indian Pines image using different dimensionality reduction methods.

Number of features	KMNF, %	OMNF, %	OKMNF, %	KM-KMNF, %
3	55.65	66.00	63.52	71.39
4	64.38	72.66	71.26	77.55
5	68.23	76.64	73.78	81.15
6	70.89	79.18	77.99	82.16
7	72.82	80.33	80.21	84.10
8	74.23	81.27	82.72	85.96
9	75.64	83.07	83.70	87.78
10	76.26	83.72	84.32	87.85
11	77.12	84.33	84.78	88.00
12	77.91	84.83	85.15	87.85
13	78.52	86.14	85.30	88.30
14	79.25	87.49	86.10	88.98
15	79.77	87.46	86.47	88.32
16	80.00	87.61	86.66	88.38
17	79.90	87.67	86.85	88.12
18	80.14	87.18	86.74	87.71
19	80.24	87.56	86.92	87.52
20	80.30	86.82	87.06	87.33

Figure 5. Comparison of the accuracies of the SVM classification on Indian Pines image after using different dimensionality reduction methods.

Table 6. Overall classification accuracies by using different number of training samples.

Methods	Training samples/class (Pavia University)				Training samples/class (Indian Pines)
	50	100	150	200	50	100	150	200
KMNF	80.95%	82.82%	82.93%	83.55%	74.72%	75.29%	75.82%	76.44%
KM-KMNF	86.94%	87.72%	88.46%	88.26%	83.11%	84.23%	84.28%	84.20%