In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
173
Figure 1. Measured and simulated canopy reflectance spectra of
two sample plots.
Generally the simulated reflectances were in relatively good
agreement with the measured reflectances for canopies with
different LAI values. A more concise analysis reveals that most
spectral bands were modelled with average absolute error
(AAE) lower than 0.02 reflectance units. As Figure 2 shows the
AAE in some regions is relatively high (greater than 0.02),
especially close to the water vapour absorption regions (1135
nm to 1400 nm, and 1820 nm to 1940 nm). We considered the
bands with an AAE greater or equal to 0.02 as wavelengths
being either poorly modelled or poorly measured (Darvishzadeh
et al., 2008).
Figure 2. The average absolute error between best-fit and the
measured HyMap reflectance.
The relation between the measured and estimated grass canopy
chlorophyll content based on the smallest RMSE criterion is
demonstrated in Figure 3.
Measured canopy chlorophyll content (g m" 2 )
Figure 3. Estimated versus measured canopy chlorophyll using
the PROSAIL model and the minimum RMSE criterion in the
LUT search.
We also evaluated the retrieval accuracy if multiple solutions
are used. Table 3 compares the “multiple solutions” with the
“best-fit” LUT solutions. This demonstrates how different
solutions affect the accuracy of the estimated variables.
No. of
Stat.
LCC
ccc
Solu.
Param
(UR cm' 2
)
(g m' 2 )
Best
spectra
/
R 2
RMSE
nRMS
R 2
RMSE
0.24
nRMS
0.12
0.35
3.8
0.17
0.84
First 10
Median
0.36
3.8
0.17
0.84
0.24
0.12
Mean
0.36
3.7
0.17
0.85
0.23
0.11
First
Median
0.39
3.1
0.14
0.81
0.25
0.12
100
Mean
0.40
3.1
0.14
0.82
0.24
0.11
Table 3. R\ RMSE and normalized RMSE between measured
and estimated leaf and canopy chlorophyll content from
PROSAIL inversion.
An appropriate band selection is known to improve radiative
transfer model inversion and prevents bias in the estimation of
the variables of interest (Schieri and Atzberger, 2006).
Therefore, to account for band selection the inversion of the
model was also tested with wavelengths that had an AAE
smaller than 0.02. We considered bands with an AAE greater or
equal to 0.02 as wavelengths with high errors (Figure 2). These
bands were systematically excluded in the inversion process,
and each time the AAE between the measured and best-fit
reflectance spectra was re-calculated until all remaining
wavelengths had an AAE smaller than 0.02. The elimination of
wavelengths stopped after 19 iterations. The remaining
wavebands (n=107) are called subset II and was used in the
inversion procedure.
The assignment of the spectral subset II in the estimation of
grass chlorophyll was again evaluated on the basis of the R 2 and
the normalized RMSE between the measured and estimated
variables. The results showed that, after removing the
wavelengths with high AAE (AAE>0.02), the relationships
between measured and estimated leaf and canopy chlorophyll
content were considerably improved (Table 4).
Spectral
Stat.
LCC
CCC
sampling set
Param
(pg cm' 2
)
(gm' 2
Using all
R 2
RMSE
nRMS
R 2
RMSE
nRMS
bands
(n=126)
Best fit
0.35
3.8
0.17
0.84
0.24
0.12
median
0.36
3.8
0.17
0.84
0.24
0.12
mean
0.36
3.7
0.17
0.85
0.23
0.11
Best fit
0.37
3.7
0.17
0.84
0.25
0.12
Subset II
median
0.38
3.4
0.15
0.87
0.23
0.11
(n=107)
mean
0.39
3.2
0.14
0.87
0.22
0.10
Table 4. R 2 , RMSE and normalized RMSE between measured
and estimated leaf and canopy chlorophyll content from
PROSAIL inversion using subset II.
Overall, the estimation accuracies between measured and
estimated leaf and canopy chlorophyll content improved using
the spectral subset (Table 4). This reflects the danger with