Conditional random field classification method based on hyperspectral-LiDAR fusion

Leiguang Wang; Ruozheng Geng; Qinling Dai; Jun Wang; Chen Zheng; Zhitao Fu

doi:10.3788/IRLA20210112

Journals >Infrared and Laser Engineering >Volume 50 >Issue 12 >Page 20210112 > Article

Infrared and Laser Engineering
Vol. 50, Issue 12, 20210112 (2021)

Conditional random field classification method based on hyperspectral-LiDAR fusion

Leiguang Wang^1、2, Ruozheng Geng³, Qinling Dai⁴, Jun Wang³, Chen Zheng^5、*, and Zhitao Fu⁶

Author Affiliations

¹Institutes of Big Data and Artificial Intelligence, Southwest Forestry University, Kunming 650224, China

²Key Laboratory of National Forestry and Grassland Administration on Forestry and Ecological Big Data, Southwest Forestry University, Kunming 650224, China

³Forestry College, Southwest Forestry University, Kunming 650224, China

⁴College of Art and Design, Southwest Forestry University, Kunming 650224, China

⁵College of Mathematics and Statistic, Henan University, Kaifeng 475004, China

⁶Faculty of Land and Resources Engineering, Kunming University of Science and Technology, Kunming 650093, China

show less

DOI: 10.3788/IRLA20210112 Cite this Article

Leiguang Wang, Ruozheng Geng, Qinling Dai, Jun Wang, Chen Zheng, Zhitao Fu. Conditional random field classification method based on hyperspectral-LiDAR fusion[J]. Infrared and Laser Engineering, 2021, 50(12): 20210112 Copy Citation Text

show less

Fig. 1. Hyperspectral and LiDAR co-classification by CRF integrating feature dissimilarity and class co-occurrence

Download full size | View in the Article

Fig. 2. Houston data set

Download full size | View in the Article

Fig. 3. Gaofeng forest farm data set

Download full size | View in the Article

Fig. 4. Classification results obtained from different feature fusion settings in the shaded area (Houston data set)

Download full size | View in the Article

Fig. 5. Initial classification map and results optimized by different CRF methods (Houston data set)

Download full size | View in the Article

(a) 休斯顿 (a) Houston			(b) 高峰林场 (b) Gaofeng forest farm
Class name	Number of training/testing samples/pixel	Sample color		Class name	Number of training/testing samples/pixel	Sample color
Healthy grass	198/1053			Eucalyptus	193/315
Stressed grass	190/1064			Road	74/106
Synthetic grass	192/505			Tilia tuan	40/52
Trees	188/1056			Cultivated land	95/127
Soil	186/1056			Acacia crassicarpa benth	208/308
Water	182/143			Wasteland	16/20
Residential	196/1072			Michelia macclurei dandy	69/95
Commercial	191/1053			Building	165/251
Road	193/1059			Other broad leaved forests	184/275
Highway	191/1036			Pinus massoniana lamb	214/300
Railway	181/1054			Cunninghamia lanceolata	34/47
Parking Lot 1	192/1041			Water	390/562
Parking Lot 2	184/285			Mixed shrub forest	53/84
Tennis court	181/247			Bamboo	21/34
Running track	187/473			Grassland	23/20

Table 1. Class names and their numbers

View in the Article

Precision	β
Precision	0.5	1	1.5	2	2.5	3	3.5	4	4.5
OA	93.99%	94.00%	93.93%	93.88%	93.89%	93.86%	93.85	93.84%	93.83%
Kappa	0.935	0.935	0.934	0.933	0.934	0.933	0.933	0.933	0.933
AA	93.47%	93.42%	93.19%	93.07%	91.30%	93.06%	93.04%	93.04%	93.03%

Table 2. Influence of different β values on the final classification accuracy (Houston)

View in the Article

(a) 休斯顿数据集 (a) Houston data set
Category	Pixel level classification method				CRF classification optimization method
Category	F^Spe	F^DSM	F^Spe+F^Spa	F^Spe+F^DSM	GGF	GGF_CRF1	GGF-CRF
Healthy grass	82.34	24.88	55.69	55.89	81.67	82.43	83.1
Stressed grass	83.36	55.92	84.40	84.49	99.34	99.62	99.81
Synthetic grass	100	91.88	100	100	100	100	100
Trees	93.37	67.23	91.57	98.11	99.24	99.24	99.62
Soil	98.30	76.80	100	99.15	100	100	100
Water	91.61	80.42	99.30	96.50	95.10	95.10	94.41
Residential	76.59	71.74	82.84	91.32	92.35	92.26	93.47
Commercial	56.51	61.92	53.09	52.42	94.59	94.78	95.73
Road	66.57	51.37	79.04	83.95	86.02	85.93	85.74
Highway	72.39	53.86	68.15	79.92	93.24	93.63	94.98
Railway	92.88	83.97	97.34	87.76	90.70	90.80	90.61
Parking Lot 1	78.58	60.71	97.70	79.63	94.24	94.43	97.41
Parking Lot 2	72.98	57.19	81.05	74.04	72.28	71.93	66.67
Tennis Court	98.79	97.17	100	98.79	100	100	100
Running Track	98.31	28.96	98.52	97.67	99.37	99.37	99.79
OA	81.98%	60.48%	85.12%	85.14%	93.34%	93.47%	94.00%
AA	84.17%	64.27%	85.91%	85.31%	93.21%	93.30%	93.42%
Kappa	0.805	0.597	0.839	0.839	0.928	0.929	0.935
(b)高峰林场数据集 (b) Gaofeng forest farm data set
Category	Pixel level classification method				CRF classification optimization method
Category	F^Spe	F^DSM	F^Spe+F^Spa	F^Spe+F^DSM	GGF	GGF_CRF1	GGF-CRF
Eucalyptus	73.65	60.63	90.79	77.46	86.67	96.82	97.14
Road	48.11	52.83	90.57	66.98	74.53	73.50	73.58
Tilia tuan	5.77	46.15	59.62	53.85	25	32.69	32.69
Cultivated land	83.46	98.43	100	96.85	100	100	100
Acacia crassicarpa benth	71.75	88.31	97.08	87.66	90.91	97.73	97.73
Wasteland	80	55	95	90	95	100	100
Michelia macclurei dandy	31.58	55.79	75.79	70.53	67.37	83.16	84.24
Building	83.27	84.06	96.41	92.83	98.01	97.21	97.21
Other broad leaved forests	70.91	66.55	96.36	65.82	83.27	85.45	85.82
Pinus massoniana lamb	73.67	92.67	92.00	85.67	89.00	96.67	97.00
Cunninghamia lanceolata	12.77	68.09	95.74	65.96	78.72	95.74	95.74
Water	99.82	98.22	100	99.64	100	100	100
Mixed shrub forest	2.38	51.19	73.81	67.86	73.81	88.1	88.1
Bamboo	0	2.94	17.65	17.65	0	0	0
Grassland	10.00	40.00	55.00	85.00	95.00	100	100
OA	71.46%	78.58%	92.41%	83.32%	87.71%	92.37%	92.84%
AA	49.81%	64.06%	82.39%	74.92%	77.15%	83.14%	83.28%
Kappa	0.674	0.756	0.914	0.811	0.860	0.913	0.919

Table 3. Producer's accuracy comparison of seven classification methods for different data sets

View in the Article

Precision	Deep fusion^[15]	HyMCKs^[8]	Multi level fusion method^[4]	EC-CRF^[11]	GGF-CRF
OA	91.32%	90.33%	93.22%	91.70%	94.00%
Kappa	0.9057	0.8949	0.930	0.907	0.935

Table 4. Comparison of classification accuracy of different methods on Houston data set

Download Citation

Save the article for my favorites

Paper Information