Comparing linear and nonlinear regression models in
Table 3, it was found that the most optimized regression
models were quadratic and cubic. However, there were only
exception for a few spectral indices, such as
RSI
(2259,1870) .
The best regression model was logarithmic function, and
NDWI
1940
, and the best regression model was power function.
The R
2
of published vegetation indices with
VWC
were under
0.5, and it had a poor correlation. In contrast, the R
2
of newly
developed vegetation indices were greater than 0.65, and
had a better correlation. Therefore, the newly developed
vegetation indices performed well in regression models.
According to the rule that the larger the R
2
, the smaller the
RMSE
, the greater the F-test, the higher the accuracy.
After the modeling analysis, the authors used the newly
developed vegetation indices as the three inverse factors to
establish an optimal equation for regression, the results were
as follows:
VWC
= 203.901*
DVI
(1712,1382)
+220.360*
NDSI
(2201,1870)
+
12.430*
RSI
(2259,1870)
+75.808
(R
2
= 0.812
RMSE
= 7.598 F-test = 37.416)
However, the result in this study demonstrated that,
compared with the estimations using only one vegetation
Figure 8. Coefficient of determination between
VWC
of
leaves and
DVI
,
NDSI
, and
RSI
.
Table 4. Comparison of precision test results of estimation models.
Type
Spectrum parameters Regression equation R
2
RMSE RPD F-test
Published
Spectral
indices
MSI
WI
SRWI
NDII
NDWI
1200
NDWI
1240
NDWI
1450
NDWI
1640
NDWI
1940
NDWI
2130
NMDI
GVMI
y =0.406x+34.57
y=0.395x+36.36
y=0.326x+40.46
y=0.297x+41.83
y=0.407x+34.85
y=0.411x+34.64
y=0.319x+40.26
y =0.377x+36.63
y=0.0793x+52.4
y=0.229x+45.58
y=0.350x+41.51
y=0.400x+35.11
0.375
0.421
0.350
0.327
0.414
0.413
0.306
0.358
0.039
0.211
0.220
0.374
13.348
12.848
13.609
13.843
12.919
13.925
14.065
13.522
16.552
14.995
14.907
13.349
1.047
1.023
0.979
0.949
1.067
1.069
0.956
1.018
0.757
0.877
0.986
1.050
16.799
20.349
15.096
13.649
19.824
19.777
12.348
15.650
1.133
7.499
7.919
16.792
Newly
developed
Spectral indices
RSI
(2259,1870)
NDSI
(2201,1870)
DVI
(1712,1382)
y=0.853x+9.10
y=0.819x+9.66
y=0.758x+13.20
0.777
0.786
0.759
7.968
7.804
8.287
2.020
2.029
1.861
97.701
103.039
88.211
Optimized
Spectral indices
RSI+NDSI+DVI
y = 0.812x+10.67 0.812 6.597 2.061 120.881
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2018
543
