Hyperspectral Image Classification Based on Modified DenseNet and Spatial Spectrum Attention Mechanism

Xin Wang; Yanguo Fan

doi:10.3788/LOP202259.0210014

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 2 >Page 0210014 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 2, 0210014 (2022)

Hyperspectral Image Classification Based on Modified DenseNet and Spatial Spectrum Attention Mechanism

Xin Wang^* and Yanguo Fan

Author Affiliations

College of Oceanography and Spatial Information, China University of Petroleum (East China), Qingdao , Shandong 266500, China

show less

DOI: 10.3788/LOP202259.0210014 Cite this Article Set citation alerts

Xin Wang, Yanguo Fan. Hyperspectral Image Classification Based on Modified DenseNet and Spatial Spectrum Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0210014 Copy Citation Text

show less

Fig. 1. Modified three-dimensional convolution module

Download full size

Fig. 2. Structure of the modified Dense_Layer

Download full size

Fig. 3. Model of the spatial spectrum attention mechanism. (a) Channel attention mechanism; (b) spatial attention mechanism

Download full size

Fig. 4. Structure of the Dense_Layer

Download full size

Fig. 5. Structure of the MDSSAN model

Download full size

Fig. 6. Indian Pines data set and label. (a) Indian Pines data set; (b) label

Download full size

Fig. 7. Pavia University data set and label. (a) Pavia University data set; (b) label

Download full size

Fig. 8. KSC data set and label. (a) KSC data set; (b) label

Download full size

Fig. 9. Classification results of the Indian Pines data set. (a) Indian Pines data set; (b) label; (c) 2D_CNN; (d) 3D_CNN; (e) M3RCNN; (f) 3D_DenseNet; (g) MDSSAN

Download full size

Fig. 10. Classification results of the Pavia University data set. (a) Pavia University data set; (b) label; (c) 2D_CNN;(d) 3D_CNN; (e) M3RCNN; (f) 3D_DenseNet; (g) MDSSAN

Download full size

Fig. 11. Classification result diagram of the KSC data set. (a) KSC data set; (b) label; (c) 2D_CNN; (d) 3D_CNN; (e) M3RCNN; (f) 3D_DenseNet; (g) MDSSAN

Download full size

No.	Category	Training set	Verification set	Test set
Total		2164	1182	7311
1	alfalfa	11	4	34
2	corn-notill	284	141	1009
3	corn-min	167	100	572
4	corn	58	28	163
5	grass/pasture	103	75	320
6	grass/trees	148	70	530
7	grass/pasture-mowed	13	11	25
8	hay-windrowed	106	51	345
9	oats	12	12	23
10	soybeans-notill	202	111	689
11	soybeans-min	509	242	1737
12	soybeans-clean	126	73	430
13	wheat	55	26	163
14	woods	250	155	902
15	building-grass-trees-drives	85	55	291
16	stone-steel towers	35	28	78

Table 1. Number of samples of training set, verification set and test set selected from Indian Pines data set

No.	Category	Training set	Verification set	Test set
Total		8514	4408	29989
1	asphalt	1306	691	4637
2	meadows	3730	1902	13023
3	gravel	400	225	1483
4	trees	626	336	2114
5	sheets	244	155	961
6	bare soil	1033	474	3540
7	bitumen	282	128	941
8	bricks	714	405	2587
9	shadows	179	92	703

Table 2. Number of samples of training set, verification set and test set selected from Pavia University data set

No.	Category	Training set	Verification set	Test set
Total		1122	623	3739
1	scrub	172	70	522
2	willow swamp	50	32	167
3	camping hammock	52	28	185
4	slash pine	66	26	172
5	oak/broadleaf	39	19	118
6	hardwood	49	32	166
7	swap	23	24	79
8	graminoid marsh	81	62	312
9	spartina marsh	106	55	386
10	cattail marsh	88	44	302
11	salt marsh	90	58	304
12	mud flats	103	66	370
13	water	203	107	656

Table 3. Number of samples of training set, verification set and test set selected from KSC data set

Classification	2D_CNN	3D_CNN	M3RCNN	3D_DenseNet	MDSSAN
Alfalfa /%	93.48	100.00	71.74	97.83	97.83
Corn-notill /%	95.87	99.86	85.92	95.59	98.32
Corn-min /%	93.13	90.84	93.86	100.00	99.76
Corn /%	92.83	94.09	92.41	99.58	100.00
Grass/pasture /%	81.78	98.55	93.37	95.65	99.79
Grass/trees /%	98.49	98.77	98.49	99.73	99.86
Grass/pasture-mowed /%	89.29	100.00	100.00	100.00	100.00
Hay-windrowed /%	99.58	100.00	100.00	100.00	100.00
Oats /%	90.00	85.00	55.00	40.00	80.00
Soybeans-notill /%	95.88	89.51	83.54	99.59	99.38
Soybeans-min /%	98.21	83.95	97.84	99.96	99.76
Soybeans-clean /%	96.46	98.48	93.76	97.64	99.49
Wheat /%	98.05	100.00	99.51	94.15	100.00
Woods /%	99.60	99.60	98.97	99.45	99.92
Building-grass-trees-drives /%	97.93	98.19	97.15	99.22	98.45
Stone-steel towers /%	97.85	98.92	82.80	96.77	96.77
OA /%	96.43	93.86	93.89	98.59	99.43
AA /%	94.90	95.99	90.27	94.70	98.08
Kappa coefficient	95.93	93.03	93.01	98.40	99.35
Time /min	1.89	5.38	32.27	18.07	12.43

Table 4. Classification results of the Indian Pines data set

Classification	2D_CNN	3D_CNN	M3RCNN	3D_DenseNet	MDSSAN
Asphalt /%	98.97	99.40	99.26	98.66	99.82
Meadows /%	99.43	99.65	99.98	99.92	99.96
Gravel /%	97.28	94.33	97.14	99.29	97.76
Trees /%	99.71	97.78	99.61	98.47	99.45
Sheets /%	99.70	98.88	100.00	100.00	99.93
Baresoil /%	99.64	99.98	87.09	100.00	99.96
Bitumen /%	84.21	98.87	98.87	99.02	99.32
Bricks /%	97.91	99.35	99.57	100.00	99.78
Shadows /%	91.66	99.89	100.00	97.99	99.37
OA /%	98.53	99.18	98.12	99.54	99.74
AA /%	96.50	98.68	97.95	99.26	99.48
Kappa coefficient	98.05	98.92	97.49	99.39	99.66
Time /min	6.61	7.21	73.92	76.10	19.58

Table 5. Classification results of the Pavia University data set

Classification	2D_CNN	3D_CNN	M3RCNN	3D_DenseNet	MDSSAN
Scrub /%	97.77	99.74	99.74	98.16	100.00
Willow swamp /%	78.60	99.18	93.42	88.48	95.06
CP hammock /%	87.89	93.75	98.83	98.83	100.00
Slash pine /%	70.24	58.33	86.51	88.89	95.24
Oak/broadleaf /%	78.88	80.12	85.09	93.17	93.79
Hardwood /%	93.89	80.34	93.01	92.14	100.00
Swap /%	98.10	78.10	97.14	100.00	100.00
Graminoid marsh /%	89.33	99.54	98.84	99.07	98.84
Spartina marsh /%	90.77	98.65	100.00	99.81	99.62
Cattail marsh /%	90.84	99.26	100.00	100.00	98.51
Salt marsh /%	99.76	100.00	99.52	99.76	100.00
Mud flats /%	99.40	97.02	94.63	97.22	98.81
Water /%	99.89	100.00	99.89	100.00	100.00
OA /%	93.07	95.16	97.45	97.66	98.98
AA /%	90.41	91.08	95.89	96.58	98.45
Kappa coefficient	92.28	94.61	97.16	97.39	98.87
Time /min	1.11	1.74	18.04	11.25	8.19

Table 6. Classification results of the KSC data set

Data set	Model parameter		OA/%
Data set	Before improvement	After improvement	Before improvement	After improvement
Indian Pines	1186044	565167	98.70	99.43
Pavia University	1184749	563872	99.17	99.74
KSC	1185489	564612	95.93	98.98

Table 7. Effect of three-dimensional convolution module before and after the improvement

Xin Wang, Yanguo Fan. Hyperspectral Image Classification Based on Modified DenseNet and Spatial Spectrum Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(2): 0210014

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information