Hyperspectral Remote Sensing Image Classification Based on Three-Dimensional Convolution Neural Network Combined with Conditional Random Field Optimization

Zhuqiang Li; Ruifei Zhu; Fang Gao; Xiangyu Meng; Yuan An; Xing Zhong

doi:10.3788/AOS201838.0828001

Journals >Acta Optica Sinica >Volume 38 >Issue 8 >Page 0828001 > Article

Acta Optica Sinica
Vol. 38, Issue 8, 0828001 (2018)

Hyperspectral Remote Sensing Image Classification Based on Three-Dimensional Convolution Neural Network Combined with Conditional Random Field Optimization

Zhuqiang Li^1、*, Ruifei Zhu^1、2, Fang Gao¹, Xiangyu Meng³, Yuan An¹, and Xing Zhong¹

Author Affiliations

¹ Chang Guang Satellite Technology Co.Ltd., Key Laboratory of Satellite Remote Sensing Application Technology of Jilin Province, Changchun, Jilin 130000, China

² Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun, Jilin 130033, China

³ Jilin Provincial Land Survey & Planning Institute, Changchun, Jilin 130061, China;

show less

DOI: 10.3788/AOS201838.0828001 Cite this Article Set citation alerts

Zhuqiang Li, Ruifei Zhu, Fang Gao, Xiangyu Meng, Yuan An, Xing Zhong. Hyperspectral Remote Sensing Image Classification Based on Three-Dimensional Convolution Neural Network Combined with Conditional Random Field Optimization[J]. Acta Optica Sinica, 2018, 38(8): 0828001 Copy Citation Text

show less

Fig. 1. Flow chart of experimental method

Download full size

Fig. 2. Hyperspectral image classification framework based on spatial-spectral 3D-CNN model

Download full size

Fig. 3. Illustration of hyperspectral pixel adjacent sparse matrix. (a) Pixel adjacent sparse matrix; (b) image four-neighbor model (K=4); (c) image eight-neighbor model (K=8)

Download full size

Fig. 4. Classification results comparison of different algorithms on Indian Pines dataset (16 categories). (a) Pseud color image; (b) true image; (c) LDM-FL; (d) 2D-CNN; (e) 3D-CNN; (f) 3D-CNN-CRF

Download full size

Fig. 5. Classification results comparison of different algorithms on Pavia University dataset (9 categories). (a) Pseud color image; (b) true image; (c) LDM-FL; (d) 2D-CNN; (e) 3D-CNN; (f) 3D-CNN-CRF

Download full size

Fig. 6. Classification and unknown region generalization result on Pavia Center dataset (9 categories). (a) Pseud color image; (b) true image;(c) 3D-CNN-CRF(feature dimension: 34); (d) 3D-CNN-CRF (feature dimension: 68); (e) 3D-CNN-CRF (feature dimension: 102); (f) unknown region result

Download full size

Fig. 7. Influence of spectral features with different dimensions on classification accuracy

Download full size

Algorithm	Parameter
Algorithm	k_c1	s_c1	p_c1	k_c2	s_c2	p_c2	f_c1	f_c2	l_r	k_crf	λ_crf
2D-CNN	3×3	[1,1]	2×2	3×3	[1,2]	2×2	400	200	0.005	-	-
3D-CNN	3×3×6	[1,1,4]	3×3×3	3×3×6	[1,1,2]	3×3×3	400	200	0.005	-	-
3D-CNN-CRF	3×3×6	[1,1,4]	3×3×3	3×3×6	[1,1,2]	3×3×3	400	200	0.005	8	0.375

Table 1. Related parameter settings of different algorithms

Accuracy indicator	Category	Algorithm
Accuracy indicator	Category	LDM-FL	p-CNN^*	2D-CNN	3D-CNN	3D-CNN-CRF
	C1	97.87	83.39	100	100	100
	C2	89.67	85.41	96.37	89.60	97.25
	C3	89.64	82.76	78.52	96.04	99.88
	C4	93.60	82.14	89.43	87.40	95.16
	C5	96.47	95.24	93.28	97.20	99.14
	C6	100	99.25	96.24	97.96	99.05
	C7	77.78	91.47	95.45	87.50	100
CA	C8	100	99.81	100	100	100
	C9	100	90.44	92.31	100	100
	C10	87.60	82.39	93.13	88.58	92.62
	C11	98.61	90.20	92.15	96.97	99.62
	C12	91.21	89.81	87.0	92.93	98.10
	C13	91.93	87.60	98.56	99.01	100
	C14	98.98	96.20	96.03	99.36	99.76
	C15	96.92	91.54	88.37	92.04	94.54
	C16	94.90	93.86	90.29	98.91	98.92
OA		94.6	90.16	92.27	94.85	98.18
Kappa		93.88	89.91	91.21	94.14	97.92

Table 2. Results of accuracy comparison of different algorithms on Indian Pines dataset (16 categories) %

Accuracy indicator	Category	Algorithm
Accuracy indicator	Category	LDM-FL	p-CNN^*	2D-CNN	3D-CNN	3D-CNN-CRF
	C1	96.03	87.34	97.12	98.22	99.01
	C2	99.03	94.63	99.45	99.34	99.67
	C3	90.42	86.47	89.84	90.46	94.17
	C4	91.99	96.29	98.48	99.28	99.74
CA	C5	97.8	99.65	100	99.48	99.78
	C6	89.61	93.23	86.85	90.57	95.45
	C7	71.75	93.19	86.12	95.09	98.21
	C8	87.17	86.42	94.46	96.32	98.18
	C9	92.68	100	98.85	98.44	98.23
OA	94	92.56	95.6	97.2	98.6
Kappa	92.1	91.7	94.7	96.3	98.1

Table 3. Results of accuracy comparison of different algorithms on Pavia University dataset (9 categories)%

Dataset	Size of dataset	Algorithm	Time
Dataset	Size of dataset	Algorithm		Feature extraction /min				2D/3D-CNN training /min	2D/3D-CNN testing /min	CRF /s
Indian Pines	145×145×220	2D-CNN3D-CNN-CRF	0.60.9		2.63.7	0.851.2	-26.3
Pavia University	610×340×103	2D-CNN3D-CNN-CRF	1.31.8		2.94.3	1.62.2	-32.5
Pavia Center	1096×715×102	2D-CNN3D-CNN-CRF	1.72.1		3.45.3	1.72.6	-42.5

Table 4. 2D/3D-CNN training, testing and optimization time for three datasets

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information