Cross-scene hyperspectral image classification combined spatial-spectral domain adaptation with XGBoost

Fig. 2. Spatial-spectral attention model

Fig. 3. Comparison of two different convolution procedure

Fig. 4. Schematic diagram of large convolution kernel decomposition

Fig. 5. False-color map and ground-truth map of Pavia dataset

Fig. 6. False-color map and ground-truth map of Indiana dataset

Fig. 7. Comparison of feature distribution before and after DA on Pavia dataset

Fig. 8. Comparison of feature distribution before and after DA on Indiana dataset

Fig. 9. Classification results of Pavia dataset

Fig. 10. Classification results of Indiana dataset

Table 1. Land class details in Pavia dataset
View table
View in Article
Table 1. Land class details in Pavia dataset
No. Category Source Target
C1 Trees 266 2 424
C2 Asphalt 266 1 704
C3 Parking lot 265 287
C4 Bitumen 206 685
C5 Meadow 273 1 251
C6 Soil 213 1 475

Table 2. Land class details in Indiana dataset

Table 2. Land class details in Indiana dataset

No.	Category	Source	Target
C1	Concrete/ Asphalt	4 867	2 942
C2	Corn cleanTill	9 822	6 029
C3	Corn cleanTill EW	11 414	7 999
C4	Orchard	5 106	1 562
C5	Soybeans cleanTill	4 731	4 792
C6	Soybeans cleanTill EW	2 996	1 638
C7	Wheat	3 223	10 739

Table 3. Parameters of spatial-spectral attention model

Table 3. Parameters of spatial-spectral attention model

Layer name	Output shape	Filter size	Padding	Dilation	Groups
Input	5×5×ch	-	-	-	-
DoConv	5×5×200	（1， 1）	（0， 0）	（1， 1）	1
BN	5×5×200	-	-	-	-
ReLu	5×5×200	-	-	-	-
Dropout	5×5×200	-	-	-	-
Avgpool	1×1×200	（5， 5）	（0， 0）	（1， 1）	-
DW-Conv	1×1×200	（5， 5）	（2， 2）	（1， 1）	200
DW-D-Conv	1×1×200	（7， 7）	（9， 9）	（3，3）	200
1×1 Conv	1×1×200	（1， 1）	（0， 0）	（1， 1）	1
FC	128	-	-	-	-

Table 4. Comparison of classification accuracy of different methods on Pavia dataset

Table 4. Comparison of classification accuracy of different methods on Pavia dataset

Method	C1	C2	C3	C4	C5	C6	OA/%	AA/%	K×100	Time/s
RBF-SVM	79.58	80.19	81.55	85.92	84.12	74.97	79.32	79.40	77.34	66.84
EMP-SVM	80.40	80.75	84.17	88.12	83.29	78.56	80.53	80.89	78.15	330.90
DCNN	86.47	84.40	86.66	90.45	83.04	82.56	84.20	83.94	81.17	277.76
ED-DMM-UDA	89.20	86.55	94.80	92.20	96.09	82.12	87.98	88.50	84.85	39.31
DCDA	92.14	94.36	100	81.35	95.78	81.99	90.63	90.08	88.00	42.74
SSDAX	94.17	90.58	99.13	92.15	94.22	84.81	91.62	92.51	89.39	36.93

Table 5. Comparison of classification accuracy of different methods on Indiana dataset

Table 5. Comparison of classification accuracy of different methods on Indiana dataset

Method	C1	C2	C3	C4	C5	C6	C7	OA/%	AA/%	K×100	Time/s
RBF-SVM	59.43	29.32	42.91	69.30	22.04	55.30	51.94	43.61	47.18	35.76	158.26
EMP-SVM	60.82	34.90	34.02	88.77	30.95	58.10	75.65	51.98	54.74	43.07	424.65
DCNN	65.22	39.14	41.98	82.96	36.12	61.15	77.87	56.10	57.78	46.26	294.50
ED-DMM-UDA	72.22	41.48	36.00	91.81	33.32	67.15	80.60	56.82	60.37	48.10	170.24
DCDA	53.18	25.01	41.15	93.90	42.29	80.72	82.16	61.60	61.79	53.00	153.72
SSDAX	77.44	54.20	44.63	91.19	54.78	64.38	86.98	65.98	67.66	58.57	151.05

Table 6. Ablation experimental results on Pavia dataset

Table 6. Ablation experimental results on Pavia dataset

Method	C1	C2	C3	C4	C5	C6	OA/%	AA/%	K×100	Time/s
MDDUWK	89.36	91.40	97.23	97.23	96.14	84.79	90.12	91.03	87.54	37.84
SSDA	89.81	91.55	96.97	87.94	96.43	84.74	90.38	91.24	87.86	38.47
SDAX	93.04	89.12	98.09	92.00	94.59	85.88	91.16	92.12	88.82	35.72
SSDAX	94.17	90.58	99.13	92.15	94.22	84.81	91.62	92.51	89.39	36.93

Table 7. Ablation experimental results on Indiana dataset