Hyperspectral Remote Sensing Image Classification Model Based on S2AF-GCN

Hailin Song; Xili Wang

doi:10.3788/LOP220612

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 2 >Page 0228010 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 2, 0228010 (2023)

Hyperspectral Remote Sensing Image Classification Model Based on S²AF-GCN

Hailin Song and Xili Wang^*

Author Affiliations

School of Computer Science, Shaanxi Normal University, Xi'an 710119, Shaanxi, China

show less

DOI: 10.3788/LOP220612 Cite this Article Set citation alerts

Hailin Song, Xili Wang. Hyperspectral Remote Sensing Image Classification Model Based on S²AF-GCN[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0228010 Copy Citation Text

show less

Classification model of hyperspectral remote sensing images based on S2AF-GCN

Fig. 1. Classification model of hyperspectral remote sensing images based on S²AF-GCN

Download full size

Renderings of spatial neighborhood feature aggregation. (a) European distance calculated by original features; (b) European distance calculated by aggregation features

Fig. 2. Renderings of spatial neighborhood feature aggregation. (a) European distance calculated by original features; (b) European distance calculated by aggregation features

Download full size

Fig. 3. Connections between nodes of different classes

Download full size

Fig. 4. Comparison of composition based on original features and aggregation features. (a) Random samples; (b) 5 nearest neighbors from original features; (c) 5 nearest neighbors from aggregation features

Download full size

Fig. 5. Algorithm flow of S²AF-GCN

Download full size

Fig. 6. Classification results on Indian Pines dataset. (a) False color; (b) ground truth; (c) FuNet-C; (d) S²GCN; (e) GCN(OF); (f) GCN(AF); (g) S²AF-GCN(OF); (h) S²AF-GCN(AF)

Download full size

Fig. 7. Classification results on Pavia University dataset. (a) False color; (b) ground truth; (c) FuNet-C; (d) S²GCN; (e) GCN(OF); (f) GCN(AF); (g) S²AF-GCN(OF); (h) S²AF-GCN(AF)

Download full size

Fig. 8. Classification results on Kennedy Space Center dataset. (a) False color; (b) ground truth; (c) FuNet-C; (d) S²GCN; (e) GCN(OF); (f) GCN(AF); (g) S²AF-GCN(OF); (h) S²AF-GCN(AF)

Download full size

Fig. 9. Influence of nearest neighbor number K on the overall accuracy on different datasets. (a) Indian Pines; (b) Pavia University; (c) Kennedy Space Center

Download full size

Fig. 10. OA of different models under different proportions of training samples. (a) Indian Pines; (b) Pavia University; (c) Kennedy Space Center

Download full size

Class No.	Class name	Number of samples
Class No.	Class name	Train	Test	Total
		105（1%）	10261	10366
1	Corn Notill	14	1420	1434
2	Corn Mintill	8	826	834
3	Corn	3	231	234
4	Grass Pasture	5	492	497
5	Grass Trees	7	740	747
6	Hay Windrowed	5	484	489
7	Soybean Notill	9	959	968
8	Soybean Mintill	24	2444	2468
9	Soybean Clean	6	608	614
10	Wheat	2	210	212
11	Woods	12	1282	1294
12	Buildings Grass Trees Drives	4	376	380
13	Stone Steel Towers	2	93	95
14	Alfalfa	2	52	54
15	Grass Pasture Mowed	1	25	26
16	Oats	1	19	20

Table 1. Land cover category and dataset division on Indian Pines dataset

Class No.	Class name	Number of samples
Class No.	Class name	Train	Test	Total
		427（1%）	42349	42776
1	Asphalt	66	6565	6631
2	Meadows	186	18463	18649
3	Gravel	21	2078	2099
4	Trees	31	3033	3064
5	Metal Sheets	13	1332	1345
6	Bare Soil	50	4979	5029
7	Bitumen	13	1317	1330
8	Bricks	37	3645	3682
9	Shadows	10	937	947

Table 2. Land cover category and dataset division on Pavia University dataset

Class No.	Class name	Number of samples
Class No.	Class name	Train	Test	Total
		52（1%）	5159	5211
1	Scrub	8	753	761
2	Willow swamp	2	241	243
3	CP hammock	3	253	256
4	Slash pine	3	249	252
5	Oak/Broadleaf	2	159	161
6	Hardwood	2	227	229
7	Swap	1	104	105
8	Graminoid marsh	4	427	431
9	Spartina marsh	5	515	520
10	Cattail marsh	4	400	404
11	Salt marsh	4	415	419
12	Mud flats	5	498	503
13	Water	9	918	927

Table 3. Land cover category and dataset division on Kennedy Space Center dataset

Model	CNN feature extraction	Information used	Features used in constructing graph	Adjacency matrix
FuNet-C	2DCNN	Spatial and spectral information	Original spectral features	Inaccurate
S²GCN	No	Spatial and spectral information	Original spectral features	Inaccurate
S²AF-GCN	No	Spatial and spectral information	Aggregation features	Accurate

Table 4. Comparison of each method

Model	Classify using original features	Classify using aggregation features	Construct adjacency matrix using original features	Construct adjacency matrix using aggregation features	Accurate adjacency matrix
GCN（OF）	√		√		×
GCN（AF）		√	√		×
S²AF-GCN（OF）	√			√	√
S²AF-GCN（AF）		√		√	√

Table 5. Comparison of similarities and differences of various methods in ablation experiment

Class No.	FuNet-C	S²GCN	GCN（OF）	GCN（AF）	S²AF-GCN（OF）	S²AF-GCN（AF）
1	81.20	83.10	61.27	67.11	81.83	82.18
2	76.15	81.48	24.46	57.51	55.57	85.84
3	89.61	98.27	39.83	47.62	51.52	60.17
4	78.66	78.05	73.37	74.19	82.11	78.46
5	71.76	92.43	85.54	87.43	89.05	99.05
6	99.38	100.00	94.42	97.11	99.79	100.00
7	73.30	82.59	61.94	87.17	78.83	88.22
8	61.46	75.16	57.90	74.02	64.57	81.18
9	46.88	78.13	31.41	39.80	59.05	70.07
10	99.52	100.00	91.90	99.05	100.00	100.00
11	96.80	99.22	95.16	94.38	97.58	98.75
12	57.18	66.76	38.83	71.81	66.22	67.82
13	83.87	86.02	80.65	80.65	77.42	88.17
14	92.31	84.62	32.69	82.69	76.92	94.23
15	72.00	92.00	36.00	92.00	80.00	100.00
16	42.11	57.89	47.37	73.68	63.16	68.42
OA /%	74.99	84.08	63.19	75.53	76.48	85.51
AA /%	76.39	84.73	59.55	76.64	76.34	85.16
Kappa	0.7181	0.8196	0.5823	0.7226	0.7317	0.8353

Table 6. Classification results on Indian Pines dataset

Class No.	FuNet-C	S²GCN	GCN（OF）	GCN（AF）	S²AF-GCN（OF）	S²AF-GCN（AF）
1	85.73	95.55	88.35	91.32	91.93	94.58
2	99.20	98.26	95.79	95.72	98.60	99.13
3	77.53	79.40	68.62	62.03	77.53	82.24
4	57.73	86.55	90.74	92.35	89.85	89.99
5	99.47	100.00	95.35	99.40	99.55	100.00
6	75.34	89.54	74.11	89.27	90.72	98.77
7	96.13	56.42	61.73	88.84	92.10	99.92
8	85.27	91.85	82.36	97.31	95.58	98.86
9	99.89	100.00	99.15	100.00	100.00	100.00
OA /%	89.00	93.29	88.24	92.52	94.58	96.95
AA /%	86.25	88.62	84.02	90.69	92.87	95.43
Kappa	0.8511	0.9103	0.8429	0.9009	0.9278	0.9561

Table 7. Classification results on Pavia University dataset

Class No.	FuNet-C	S²GCN	GCN（OF）	GCN（AF）	S²AF-GCN（OF）	S²AF-GCN（AF）
1	98.41	100.00	92.96	96.95	95.22	98.94
2	100.00	100.00	87.97	94.19	98.76	98.34
3	45.06	100.00	82.21	90.91	70.36	100.00
4	6.83	59.84	17.27	58.63	64.26	72.69
5	90.57	100.00	61.01	62.26	92.45	97.48
6	22.47	62.56	50.66	59.03	64.32	92.95
7	40.38	69.23	37.50	46.15	40.38	76.92
8	83.84	100.00	88.52	91.57	94.38	100.00
9	83.69	83.69	84.85	83.11	89.90	83.69
10	96.50	94.00	71.00	93.75	80.25	94.00
11	100.00	100.00	80.24	87.47	87.23	97.11
12	96.50	96.39	94.58	95.78	95.38	96.39
13	100.00	100.00	99.78	99.46	100.00	100.00
OA /%	84.09	93.35	82.05	88.41	88.58	94.92
AA /%	74.16	89.67	72.97	81.48	82.53	92.96
Kappa	0.8228	0.9259	0.7999	0.8707	0.8727	0.9435

Table 8. Classification results on Kennedy Space Center dataset

Dataset	FuNet-C	S²GCN	S²AF-GCN
Indian Pines	420	117	182
Pavia University	963	133	240
Kennedy Space Center	280	40	56

Table 9. Running time of FuNet-C, S²GCN, S²AF-GCN models on three datasets

m+1	Indian Pines			Pavia University			Kennedy Space Center
m+1	OA /%	AA /%	Kappa	OA /%	AA /%	Kappa	OA /%	AA /%	Kappa
2	77.52	80.36	0.7451	93.98	92.34	0.9200	90.83	85.29	0.8978
4	81.56	82.24	0.7912	95.89	94.57	0.9455	92.61	87.54	0.9177
6	84.81	84.01	0.8276	96.42	95.52	0.9526	93.47	89.09	0.9272
8	85.51	85.16	0.8353	96.65	95.43	0.9561	94.92	92.96	0.9435
10	85.05	85.19	0.8304	96.69	95.94	0.9595	92.44	87.65	0.9157
12	83.74	85.14	0.8156	95.96	93.41	0.9462	93.14	89.12	0.9236

Table 10. Classification results of S²AF-GCN on different datasets under different aggregation times

Hailin Song, Xili Wang. Hyperspectral Remote Sensing Image Classification Model Based on S²AF-GCN[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0228010

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享