126 
signal and since the radar equation is applied for 
each plane individually the model supplies an objec 
tive way, independent of sensor measurement geometry 
to calculate these contributions. In fact the coef 
ficients are directly related to y values for 
each layer (section 5.2). 
Since the n-level model equation contains n un 
known coefficients at leastn samples of the return 
signal are needed to solve these coefficients from 
the model. The calculation of the coefficients 
together with their confidence intervals is most 
effectively done by lineair regression analysis 
techniques. 
Figure 7. (a) Simulated radar return signals origi 
nating from 4 equeally distanced identical planes 
in C-band at 16.5 degrees incidence angle. The height 
of fleight is 323 meter above level 0, 323-9 meter 
above level 1, etc. 
(b) Measured returns from a grass field. 
The largest signal peak closely matches the simula 
ted return of level 0. The signal part exceeding a 
range distance of 360 meter is an artifact intro 
duced by the receiver. 
(c) Measured returns from a poplar field. 
Two peaks are visible. The first one matches level 
2, the second one level 0. Again the signal part 
exceeding a range distance of 360 meter is an 
artifact. 
In some cases it was found that the significance of 
the estimated values was low. This can be explained 
from measurement geometry and the model. When returns 
from the scatter planes are highly correlated as is 
the case in figure 6, it is difficult to separate 
the measured signal into individual returns follo 
wing the model prediction. A strong correlation in 
the individual returns results in a strong correla 
tion between the estimates of the coefficients A^. 
The correlation matrix for the estimates for this 
case is shown in table 4(a). The estimates are nega 
tively correlated in pairs. This means that e.g. 
an over-estimation of Ao is easily compensated by 
an under-estimation of Ai (r=-0.78), which in turn 
can be compensated by an over-estimation of Az 
(r=-0.80), etc.. 
When using a 3-level model with 12.5 meter spacing 
the correlation matrix for the estimates, in table 
4(b), shows lower correlations and correspondingly 
higher significances for the estimates were found. 
Table 4 Correlation matrices of estimates for 
L-band, HH-polarization, 14.5 degrees incidence 
and 250 meter flying height. 
(a) 4-level 9 meter spacing model 
CORRELATIONS OF ESTIMATES 
0 1 .00 
1 - 0.78 1.00 
2 0.52 -0.80 1.00 
3 - 0.31 0.51 -0.78 1.00 
(b) 3-level 12.5 meter spacing model 
CORRELATIONS OF ESTIMATES 
0 1 .00 
1 -0.52 1.00 
2 0.23 -0.51 1.00 
The correlation matrix of the estimates of A^ (or the 
covariance matrix of the model returns) together 
with the number of samples and speckle level can 
serve as a measure to indicate the minimum level 
spacing allowed when a certain accuracy is speci 
fied. Since the correlation of the returns as well 
as the number of relevant samples is influenced by 
flight geometry, the degree of separability of the 
radar return in contributions of individual forest 
layers is an element of experiment design. 
Figure 8. The backscattering might originate from 
(A) the ground, (B) the canopy surface, (C) the 
leaves, twigs and branches of the tree crown's 
volume or (D) from trunk-ground reflections. 
5. SOURCES OF SCATTERING IN FORESTS AT L- AND 
C-BAND 
Though numerous radar images of forests have been 
acquired during the past 20 years one still spe 
culates on the question which elements of the 
forest volume contribute to the backscatter sig 
nal (figure 8) and how this depends on sensor 
parameters. In order to be able to indicate poten 
tial applications of radar remote sensing in fores 
try one desires to know the relation between the 
backscatter signal and forest parameters. Scatter 
models for vegetation are being developed, with some 
success, but the forest structure is far more com 
plicated to model than the structure of most other 
vegetation types. If one could indicate experimen 
tally which parts of the forest dominate the back 
scatter signal this would simplify the model-making 
effort. 
The DUTSCAT, when operated from relatively low alti 
tudes, can provide such information through inversion 
of the multi-level model introduced in section 4. In 
this section results will be given for several poplar