Research on hyperspectral image classification method based on deep learning

Fig. 4. Feature classification results for the IP dataset (a) pseudo-colour composite image, (b) surface real data, (c) DPyResNet method, (d) SSRN method, (e) ContextNet method, (f) method proposed in this paper, (g) classification legend

Download full size

View in Article

Fig. 5. Feature classification results for the UP dataset (a) pseudo-colour composite image, (b) surface real data, (c) DPyResNet method, (d) SSRN method, (e) ContextNet method, (f) method proposed in this paper, (g) classification legend

Download full size

View in Article

Table 1. SST module network parameters

View table

View in Article

Table 1. SST module network parameters

Layer	Layer Type
$W^{Q}$	Conv3d（i=1，o=24，k=（1×1×7），s=（1，1，2），p=（1，1，0））
$W^{K}$	Conv3d（i=1，o=24，k=（3×3×7），s=（1，1，2），p=（1，1，0））
$W^{V}$	Conv3d（i=1，o=24，k=（3×3×7），s=（1，1，2），p=（1，1，0））
FFN	Conv3d（i=24，o=24，k=（1×1×7），s=1，p=（0，0，3）） BatchNorm3d（24） ReLU Conv3d（i=24，o=24，k=（1×1×7），s=1，p=（0，0，3）） BatchNorm3d（24）

Table 2. Pooled residual network parameters

View table

View in Article

Table 2. Pooled residual network parameters

Model Type	Module Type	Layer Type
PoolRes1	Pool	Maxpool Avgpool Conv2d（i=1，o=1，k=（3×3），s=1，p=（1，1）） Conv2d（i=1，o=1，k=（3×3），s=1，p=（1，1）） Sigmod
PoolRes1	ResBlock1	Conv3d（i=24，o=24，k=（1×1×7），s=1，p=（0，0，3）） BatchNorm3d（24） ReLU Conv3d（i=24，o=24，k=（1×1×7），s=1，p=（0，0，3）） BatchNorm3d（24）
PoolRes2	Pool	Maxpool Avgpool Conv2d（i=1，o=1，k=（3×3），s=1，p=（1，1）） Conv2d（i=1，o=1，k=（3×3），s=1，p=（1，1）） Sigmod
PoolRes2	ResBlock2	Conv3d（i=24，o=24，k=（3×3×1），s=1，p=（1，1，0）） BatchNorm3d（24） ReLU Conv3d（i=24，o=24，k=（3×3×1），s=1，p=（1，1，0）） BatchNorm3d（24）
PoolRes3	Pool	Maxpool Avgpool Conv2d（i=1，o=1，k=（3×3），s=1，p=（1，1）） Conv2d（i=1，o=1，k=（3×3），s=1，p=（1，1）） Sigmod
PoolRes3	ResBlock3	Conv3d（i=24，o=24，k=（3×3×1），s=1，p=（1，1，0）） BatchNorm3d（24） ReLU Conv3d（i=24，o=24，k=（3×3×1），s=1，p=（1，1，0）） BatchNorm3d（24）

Table 3. Comparison results of classification performance metrics OA (%), AA (%) and Kappa coefficient (κ) of different classification methods on IP dataset

View table

View in Article

Table 3. Comparison results of classification performance metrics OA (%), AA (%) and Kappa coefficient (κ) of different classification methods on IP dataset

Class	Training	Test	Classical Models				Deep Neural Networks
Class	Training	Test	MLR	RF	SVM	LSTM	ResNet	ContextNet	Ms-3DNet	ENL-FCN	DPyResNet	SSRN	SST
1	4	42	15.45±0.023	28.46±0.061	51.22±0.190	69.11±0.090	98.66±0.018	88.78±0.080	66.67±0.471	97.56±0.000	94.69±0.076	57.78±0.423	99.07±0.013
2	142	1286	73.77±0.006	56.63±0.024	81.22±0.037	74.22±0.016	87.85±0.020	98.19±0.005	75.94±0.080	93.15±0.000	93.83±0.040	98.37±0.012	98.73±0 .006
3	83	747	51.14±0.027	48.42±0.013	65.82±0.013	71.49±0.030	92.71±0.007	95.37±0.028	81.39±0.007	97.59±0.000	89.30±0.003	97.47±0.010	98.83±0.006
4	23	214	43.97± 0.051	33.49±0.025	57.75±0.041	60.72±0.041	95.43±0.046	97.04±0.021	88.63±0.063	91.55±0.000	93.51±0.055	99.12±0.0099	98.63±0.020
5	48	435	83.52±0.034	85.21±0.025	90.04±0.014	87.51±0.015	98.23±0.015	97.78±0.015	95.61±0.054	97.47±0.000	99.26±0.004	97.79±0.013	99.03±0.006
6	73	657	94.82±0.009	92.64±0.027	96.25±0.006	94.77±0.015	97.98±0.011	98.60±0.008	96.78±0.026	99.24±0.000	98.52±0.007	98.50±0.010	98.55±0.009
7	2	26	41.33±0.186	2.67±0.038	73.33±0.019	85.33±0.094	92.98±0.099	90.35±0.098	100.00±0.000	100.00±0.000	83.08±0.178	66.67±0.471	90.70±0.193
8	47	431	98.53±0.006	97.67±0.015	97.98±0.006	97.83±0.009	95.06±0.014	97.76±0.026	89.51±0.091	97.44±0.000	97.63±0.022	96.45±0.029	100.00±0.000
9	2	18	5.56±0.045	9.26±0.094	50.00±0.045	53.70±0.139	60.83±0.283	86.90±0.102	66.67±0.471	72.22±0.000	66.66±0.471	56.25±0.418	78.34±0.091
10	97	875	65.41±0.041	60.91±0.047	73.87±0.018	73.68±0.025	96.05±0.013	96.08±0.018	87.41±0.070	94.74±0.000	93.77±0.029	98.33±0.009	97.49±0.002
11	245	2210	80.37±0.015	87.88±0.019	82.90±0.012	84.93±0.024	93.32±0.041	97.35±0.004	76.69±0.096	95.61±0.000	89.78±0.040	99.08±0.005	99.26±0.002
12	59	534	55.68±0.007	41.26±0.030	74.91±0.043	73.68±0.025	86.65±0.077	94.00±0.012	88.65±0.036	97.00±0.000	83.43±0.107	98.46±0.009	98.03±0.002
13	20	185	97.66±0.005	90.09±0.040	96.94±0.021	98.74±0.005	82.16±0.076	95.01±0.03	99.78±0.003	97.83±0.000	98.19±0.021	100.00±0.000	99.59±0.005
14	126	1139	95.61±0.004	95.46±0.014	93.82±0.010	96.22±0.004	95.39±0.016	98.49±0.014	90.06±0.087	99.12±0.000	96.00±0.021	98.63±0，010	99.38±0.005
15	38	348	56.00±0.045	41.11±0.029	60.42±0.044	60.04±0.029	90.96±0.127	94.10±0.031	88.21±0.044	92.80±0.000	91.22±0.040	99.24±0.005	98.08±0.009
16	9	84	84.92±0.020	79.37±0.030	91.27±0.054	90.87±0.022	94.73±0.038	93.57±0.046	98.53±0.021	100.00±0.000	70.90±0.388	95.63±0.062	96.94±0.023
OA	1018	9231	76.23±0.008	72.98±0.006	82.00±0.006	82.13±0.004	92.44±0.006	96.98±0.006	83.44±0.060	96.15±0.054	91.47±0.029	98.38±0.004	98.67±0.001
AA			65.23±0.019	59.41±0.005	77.36±0.019	79.53±0.005	91.19±0.025	94.96±0.003	86.91±0.084	95.21±0.028	94.14±0.006	91.11±0.080	96.63±0.014
$κ$			0.7266±0.010	0.6862±0.007	0.7941±0.007	0.7954±0.004	0.9137±0.006	0.9655±0.007	0.8082±0.070	0.9560±0.030	0.9020±0.034	0.9815±0.005	0.9849±0.002

Table 4. Comparison results of classification performance metrics OA (%), AA (%) and Kappa coefficient (κ) of different classification methods on UP dataset

View table

View in Article

Table 4. Comparison results of classification performance metrics OA (%), AA (%) and Kappa coefficient (κ) of different classification methods on UP dataset

Class	Training	Test	Classical Models				Deep Neural Networks
Class	Training	Test	MLR	RF	SVM	LSTM	ResNet	ContextNet	Ms-3DNet	ENL-FCN	DPyResNet	SSRN	SST
1	663	5968	92.30±0.004	91.11±0.007	94.30±0.008	95.47±0.005	96.82±0.023	99.56±0.002	99.36±0.001	99.98±0.000	98.35±0.017	99.85±0.001	99.91±0.001
2	1864	16785	96.18±0.003	98.11±0.003	97.65±0.002	96.91±0.002	98.59±0.008	99.85±0.002	99.80±0.000	100.00±0.000	98.76±0.008	99.98±0.000	99.98±0.001
3	209	1890	72.75±0.013	67.71±0.014	81.26±0.018	78.01±0.011	90.01±0.061	99.19±0.001	98.02±0.017	99.68±0.000	94.22±0.034	99.68±0.003	99.75±0.002
4	306	2758	89.28±0.002	88.20±0.006	94.63±0.004	94.92±0.007	99.32±0.003	99.80±0.002	99.71±0.001	98.94±0.000	99.20±0.005	99.92±0.001	99.97±0.004
5	134	1211	99.42±0.003	98.93±0.002	99.20±0.002	99.26±0.003	99.81±0.000	99.91±0.001	99.94±0.000	100.00±0.000	99.72±0.003	99.94±0.000	99.9±0.003
6	502	4527	77.45±0.005	72.14±0.022	90.58±0.008	87.85±0.012	99.41±0.002	99.75±0.003	99.43±0.003	99.87±0.000	98.52±0.006	99.95±0.001	99.95±0.005
7	133	1197	55.69±0.043	75.68±0.017	85.71±0.011	80.23±0.007	96.90±0.017	98.37±0.022	99.18±0.005	100.00±0.000	97.37±0.004	100.00±0.000	100.00±0.000
8	368	3314	87.04±0.004	89.64±0.013	88.20±0.003	88.49±0.008	92.00±0.044	98.48±0.008	97.13±0.005	99.69±0.000	97.37±0.004	98.28±0.015	99.02±0.005
9	94	853	99.77±0.001	99.77±0.002	99.84±0.001	99.88±0.001	98.88±0.012	99.26±0.005	99.74±0.002	100.00±0.000	84.51±0.071	99.39±0.003	99.94±0.006
OA	4273	38503	89.87±0.001	90.41±0.001	94.19±0.002	93.45±0.001	97.38±0.007	99.57±0.001	99.35±0.001	99.76±0.002	99.60±0.001	99.77±0.001	99.87±0.004
AA			85.54±0.004	86.81±0.002	92.38±0.003	91.23±0.001	96.86±0.005	99.35±0.002	99.15±0.002	99.70±0.002	97.05±0.010	99.66±0.01	99.83±0.005
$κ$			0.8646±0.001	0.8710±0.002	0.9229±0.002	0.9130±0.001	0.9652±0.009	0.9943±0.001	0.9913±0.002	0.9972±0.001	96.69±0.006	0.9969±0.01	0.9983±0.006

Tools

Get Citation

Copy Citation Text

Bin ZHANG, Liang LIU, Xiao-Jie LI, Wei ZHOU. Research on hyperspectral image classification method based on deep learning[J]. Journal of Infrared and Millimeter Waves, 2023, 42(6): 824

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Jan. 6, 2023

Accepted: --

Published Online: Dec. 26, 2023

The Author Email: Wei ZHOU (yeaweam@163.com)

DOI:10.11972/j.issn.1001-9014.2023.06.016

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology