The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
265
among input variables, and input variables are not only related
to outputs, but also relevant to one another. Under this condition,
an input variable of model possibly shields the effects of the
other variables on results. In short, stepwise regression, as a
fixed processing, has risks for users to a certain extent, because
regressive results are closely related to the initial model as well
as the selective strategy for variables.
Considering the shortcomings of stepwise regression analysis,
this study firstly used single-correlation analysis, then selected
the wavebands with high correlation and long intervals (weakly
relevant to one another) as regressive input variables. Thus, on
one hand, under the premise of remaining sensitive wavebands,
the amount of input wavebands was reduced and overfitting
phenomenon was avoided; on the other hand, the selection of
input wavebands were not entirely based on the magnitude of
correlation coefficients but selecting 1 extremum (or 2
extremum for differential transforms with positive and negative
into three sections after elimination of water-absorption peaks.
The longer the intervals among selected wavebands are, the
weaker the correlation among them. Accordingly, high
correlation among selected variables (wavebands) was
effectively preventing by this way.
Table 2 shows the regression equations used for prediction of
SOM from various transforms of reflectance. And all equations
were performed F-value test with significance level 0.001. As
shown in table 2, among all transforms, the first order
differentiate of logarithm of reflectance ((lg/?)') has the
strongest ability to predict SOM content, and the validation
coefficient of its regression equation consisting of three
wavebands is 0.89, the maximum among all regression
equations. The effect of SOM content predictions from different
regression equations are as shown in figure 4, 5, 6, 7.
Transforms
Regression Equations
R 2
Adjusted
R 2
RMSE
X = R
f = 1.990-20.949^2137 +22.318^50!
0.684
0.679
0.608
X = y[R
Y = 2.386 - 24.938^137 + 24.689* 149 9
0.802
0.799
0.480
X = l/R
Y = 0.053 + 0.307^2277
0.789
0.787
0.486
X-lgR
y = 1.029 -12.359*2149 + 10.878*1504
0.851
0.849
0.417
x = (\/r)
y = 0.581-362.003* 863 +64.680*n45
0.840
0.837
0.432
*
II
\
Y = 0.626 + 1308.365*2222 + 2027.007*i 740 -135.885* 672
0.706
0.699
0.586
x = {igR)
7 = 1.772 + 1004.071*2187 +2893.272*849 -1682.915*168,
0.888
0.885
0.360
x={ig r)'
7 = 2.451 + 21952.91*587 “ 47995.4* 905
-4577.994*2219 + 13138.89*1726
0.839
0.833
0.431
X = Jr
7 = 1.971 + 1399.130*2180 + 4260.033* 846 - 2459.097* 16 85
0.861
0.858
0.403
* = Væ
7 = 2.661 + 38552.87*587 - 40731.4* 905
+ 33733.12*1725 -6072.362*2199
0.842
0.837
0.432
X = R'
7 = 1.891 + 5024.556*845 - 941.121* 203 7
+ 576.462*2180 + 615.901*1521
0.789
0.782
0.500
X = R
7 = 2.305 + 13830.69*587 + 52867.66*i 725 - 35305.5* 529
0.754
0.748
0.537
Table 2 Regression analytical result between different reflectance transforms and SOM Content
A