Author Affiliations
1School of Geomatics, Liaoning Technical University, Fuxin 123000, China2Heilongjiang Institute of Geomatics Engineering, Harbin 150081, Chinashow less
Fig. 1. Titan MS-LiDAR point cloud of Area1 and its processing results. (a) C1 intensity; (b) C2 intensity; (c) C3 intensity; (d) Points colored by elevation; (e) Merged point cloud (R = C1 intensity, G = C2 intensity and B = C3 intensity);(f) 3D land cover classified point cloud
Fig. 2. Titan MS-LiDAR point cloud of Area2 and its processing results. (a) C1 intensity; (b) C2 intensity; (c) C3 intensity; (d) Points colored by elevation; (e) Merged point cloud (R = C1 intensity, G = C2 intensity and B = C3 intensity);(f) 3D land cover classified point cloud
Fig. 3. Flowchart of the proposed algorithm
Fig. 4. Relationships between wavelengths and reflectance properties of objects
[26] Fig. 5. NDRI statistical histogram for ground points of Area1
Fig. 6. NDRI statistics histogram for different objects of Area1. (a) Grass; (b) Road; (c) Tree; (d) Building
Fig. 7. Differentiability of different NDRI used for distinguishing from objects of Area1. (a) C1−C2; (b) C1−C3; (c) C2−C3
Parameter | Specification | Wavelengths | C1: 1550 nm, C2: 1064 nm, C3: 532 nm
| Forward angles | C1: 3.5°, C2: 0°, C3: 7°
| Beam divergence | C1: 0.35 mrad, C2: 0.35 mrad,
C3: 0.7 mrad
| Altitude | Topographic: 300-2000 m above ground
level (AGL), all channels
| Pulse repetition
frequency
| 50-300 kHz/channel; 900 kHz total | Scan angle (FOV) | Programmable; 0-60° max | Scan frequency | Programmable; 0-210 Hz | Swath width | 0-115% AGL | Accuracy | Horizontal: 1/7500×altitude, 1σ;
Vertical: < 5-10 cm, 1 σ | Laser range precision | <0.008 m, 1σ |
|
Table 1. Specifications of Titan
Experimental area | Type I error | Type II error | Total error | Area1 | 5.24% | 1.87% | 3.90% | Area2 | 2.61% | 4.25% | 3.26% |
|
Table 2. The errors of the filtering results
NDRI | Area1 | | Area2 | Ground points | Non-
ground points
| Ground points | Non-
ground points
| C2−C3
| 0.344 | 0.373 | | 0.296 | 0.367 | C1−C3
| 0.227 | 0.284 | 0.344 | 0.373 | C1−C2
| 0.405 | 0.262 | 0.231 | 0.258 |
|
Table 3. NDRI threshold t*
NDRI | Area1 | | Area2 | OA | Kappa | OA | Kappa | C2−C3
| 90.45% | 0.869% | | 89.89% | 0.853% | C1−C3
| 88.59% | 0.841% | 86.98% | 0.812% | C1−C2
| 80.68% | 0.738% | 83.07% | 0.752% |
|
Table 4. Classification accuracy of different NDRI indexes
Classification data | Reference data | Total row | User’s accuracy | Roads | Grass | Trees | Buildings | Roads | 24501 (A1)
| 1524 (A2)
| 939 (A3)
| 746 (A4)
| 27710 (Ta)
| 88.42%
(Ua=A1/Ta)
| Grass | 1386 (B1)
| 40260 (B2)
| 1137 (B3)
| 1128 (B4)
| 43911 (Tb)
| 91.69%
(Ub=B2/Tb)
| Trees | 234 (C1)
| 360 (C2)
| 24019 (C3)
| 1594 (C4)
| 26207 (Tc)
| 91.65%
(Uc=C3/Tc)
| Buildings | 40 (D1)
| 248 (D2)
| 2017 (D3)
| 18750 (D4)
| 21055 (Td)
| 89.05%
(Ud=D4/Td)
| Total column | 26161 (T1)
| 42392 (T2)
| 28112 (T3)
| 22218 (T4)
| 118883 | | Producer’s
accuracy
| 93.65%
(P1=A1/T1)
| 94.97%
P2=B2/T2 | 85.44%
(P3=C3/T3)
| 84.39%
(P4=D1/T4)
| | | Overall accuracy: 90.45%; Kappa statistic: 0.869 |
|
Table 5. Confusion matrix of the land cover classification result (Area1)
Classification data
| Reference data | Total row | User’s ccuracy | Roads | Grass | Trees | Buildings | Roads | 53 974 | 11 422 | 1 130 | 695 | 67 221 | 80.29% | Grass | 9 165 | 150 101 | 6 712 | 2 175 | 168 153 | 89.26% | Trees | 239 | 619 | 93 484 | 2 095 | 96 437 | 96.94% | Buildings | 282 | 249 | 5 222 | 35 435 | 41 188 | 86.03% | Total column | 63 660 | 162 391 | 106 548 | 40 400 | 372 999 | | Producer’s accuracy | 84.78% | 92.43% | 87.74% | 87.71% | | | Overall accuracy: 89.89%; Kappa statistic: 0.853 |
|
Table 6. Confusion matrix of the land cover classification result (Area2)
Authors | Algorithm principle | Features | OA | Kappa | Our paper | Step-by-step separation | Spatial;NDRI | 90.17% | 0.861 | Fernandez-Diaz et al[16] | Mahalanobis distance | Five structural; three intensity images | 82.33% | 0.77 | Chen et al[13] | SVM;k-NN re- classification | Spectral reflectance of four channels; five
vegetation indexes; neighborhood spatial
| 87.19% | N/A | Fernandez-
Diaz et al[16] | Maximum likelihood | Five structural; two intensity images | 90.22% | 0.870 | Zou et al[20] | Decision tree | Pseudo normalized difference vegetation index;
ratio of green; ratio of returns counts; difference of
elevation between maximum elevation of first
returns and minimum elevation of last returns
| 91.63% | 0.895 | Chen et al[13] | Random forest | 7 spectral features; 11 geometric features | 93.00% | N/A | Wang and Gu[9] | SVM | Spatial location; spectral; neighborhood geometric;
spectral structures; geometric-spectral
| 94.76% | 0.935 |
|
Table 7. Accuracy comparison between the proposed algorithm and the other classical algorithms