The performance of the classifications
was judged by comparing the obtained standwise estimates with the
corresponding standwise estimates based on field measurements
(totally 654 hectares). Correlation coefficients
weighted by stand area, difference (BIAS) and root mean square error (RMSE)
were calculated.
Twice mode-filtered ERS-1 images gave the best results in stand level
comparisons though none of the pre-processed images gave statistically
significant correlation coefficients for mean volume at stand level.
The mean of the estimated stand volume was 159.8 cubicm/ha, total BIAS
18.3 cubicm/ha and RMSE 104.6 cubicm/ha.
The estimates by volume classes (VCL) are presented in table 2.
The BIAS and RMSE are rather high in all relevant classes,
over 100 cubicm/ha.
<![CDATA[
TABLE 2. The estimated mean stem volume (V), BIAS and RMSE at the
forest stand level by volume classes, K-nn method.
VCL stand V BIAS RMSE
area
m3/ha ha m3/ha m3/ha m3/ha
0 - 49 0.1 45.0 -45.0 45.0
50 - 99 2.9 88.0 4.9 86.9
100 - 149 177.7 141.4 17.4 96.7
150 - 199 463.7 166.3 -31.2 107.7
200 - 299 10.7 206.8 -60.1 100.6
]]>
The stem volume in the stands of the Mäntsälä test area was estimated
by linear regression models fitted to data from NFI plots within a
distance of 30 km from the
test area. Using the means of backscattering coefficients in each segment,
the estimates were calculated for every pixel. The
estimates were compared with
known stem volumes in forest stands. Total area of stands was 620 hectares.
The correlation coefficient between observed and estimated
volumes were rather low.
Due to the noise of SAR images, the segment boundaries did not match the
real stand boundaries. Nevertheless, segmentation may facilitate the
detection of clear cuttings and other small-scale, but distinct, changes in
stem volume.
<![CDATA[
TABLE 3. Reliability characteristics of stem volume estimates based on
segmentation and regression analysis.
volume bias RMSE correlation
total 5.5 96 0.29
spruce 8.2 96 0.29
pine -2.6 37 0.08
deciduous -4.5 37 0.04
]]>
Results shown in Tables 2. and 3. can be improved
if the effect of moisture variation is taken into account.
Inversion method was applied for tree stem volume estimation
in two of the test sites.
The estimates were computed at forest stand level employing
a multi-temporal ERS-1 SAR data set. The 100 nearest sample plots
were applied for estimating the soil and canopy moistures.
Tree stem volume were estimated by means of formula 3.
Correlation coefficients between estimated and field measurement
based volume estimates are given in Table 4. For the two test sites,
they are 0.64 and 0.63. The rms errors are 90 - 98 cubicm/ha, respectively.
The correlation coefficient are calculated by weighting the stands
by the stand area. Known stand boundaries have been applied.
Table 4 shows some reliability characteristics for the two test
sites. The estimates of inversion method have also been justified
comparing the overall stem volume estimates of the two sites
with field measurement based ones and with mean volumes of
training sample plots (Table 4).
The results show that the inversion method significantly improves
the estimates based on the overall mean of training areas (NFI plots).
<![CDATA[
TABLE 4. ERS-1 SAR-based tree stem volume estimates.
Test site Number r RMS error SAR Field Mean
of stands (m^3/ha) vol. vol. volume of
est. est. NFI plots
Mäntsälä 198 0.64 90 161 154 144
Ohkola 91 0.63 98 126 126 131
]]>
Total volumes of stand level inventories were utilized as comparison
data in assesing the reliability of estimates at forest region level.
The field measurements of regional inventories have been carried
out in the years 1990-1993. The estimates are based on visual
assesment and only on few measurements.
Eight forest region were applied. The sizes vary from 559 to 6256
hectares, total area being 23 225 hectares.
Only K-nn based estimates have been tested so far.
The estimated and measured mean volumes by tree species are given
in Table 5. The deviations of estimates from the total means
are not very high. It means that ERS-1 SAR -images as such are not
good information source in estimating forest characteristics.
However, moisture estimation and inversion method may
improve the results. It has not yet been tested at region level.
<![CDATA[
TABLE 5. Region level comparisons of tree stem volume estimates
Region Area Vol Vol Pine Pine Spruce Spruce Non-con. Non-con
(ha) m3/ha est. est. est. est.
hapu 1704 143 151 51 31 73 97 18 21
hirv 2694 128 148 24 29 89 95 14 21
kyta 3120 138 146 33 30 89 94 15 20
lauk 2092 136 146 34 30 84 91 16 22
mapo 6256 160 150 22 26 119 98 17 20
ohko 3721 158 147 31 29 106 99 20 21
oitt 3079 149 152 29 29 105 100 14 20
torp 559 123 148 17 28 87 96 19 22
Total 23225 147 148 29 29 100 97 16 21
]]>
ERS-1 SAR images from the dates 24th of July, 28th
of August, 2nd of October and 6th of November were applied
for analyzing the co-effect of soil type and moisture variation
on backscattering coefficients.
SAR images were compared pixelwise with the soil map.
SAR images were geocoded using the terrain elevation model. Mean values of
the backscattering coefficient for different soil-classes are presented
in Table 5.
<![CDATA[
TABLE 5.Mean values of backscattering coefficient for the soil classes
Date Open Moraine Moraine Eskers Clay Peat Gravel,
rocks ridges and sand,
silt fine sand
24.7.93 -9.87 -7.79 -8.99 -9.27 -10.00 -7.72 -8.80
28.8.93 -9.02 -10.26 -7.85 -8.75 -9.02 -9.79 -7.83
2.10.93 -9.07 -9.29 -10.26 -7.66 -8.49 -8.83 -10.17
11.6.93 -7.68 -8.64 -8.99 -10.01 -7.72 -8.75 -8.92
]]>
High values of fine sand and clay classes
are mainly caused by
high moisture content of those soil types.
There was a strong rain during the SAR
data acquisition on the 24th July.
In order to justify the reliability of the inversion method
is assessing soil and needle moisture,
weather condition measurements have been carried out
on some of NFI sample plots.
The comparisons have been conducted using ERS-1 SAR data
corresponding to the NFI sample plots from the Porvoo test
area. The semi-empirical backscattering model has been fitted
into the sample plotwise SAR results using two free parameters
which correspond to (1) volumetric soil moisture and (2)
volumetric forest canopy moisture.
The
highest positive correlations are obtained in cases where a long
non-rainy period (over one week) has occurred before the SAR image
acquisition Hence, the soil moisture
has evidently had low values as compared with e.g. with the results for
when a substantial rain occurred a day before the
image acquisition. The higher the rain rate has
been, the higher is the level of backscatter from ground (since
the soil moisture is evidently higher). The variation of the
backscattering coefficient appears to be higher for unforested
areas (V= 0 m3/ha) than for heavily forested areas. This
indicates that the changes in forest canopy moisture are smaller
than those at forest floor. The model fittings suggest the
average soil moisture to vary from 9.5 % (6 November) to 15.5 %
(13 October).
The analysis of winter time images in Sodankyl\"a test area show
that the response to forest stem volume is at its
highest level with the presence of wet snow-cover
This phenomenon
occurs probably either (1) due to the high signal attenuation
caused by the wet snow or/and (2) due to the low level of
backscatter from the (smooth) air-snow boundary.
Contact information
- Finnish Forest Research Institute METLA
- Unioninkatu 40A, FIN-00170 Helsinki, Finland
- tel. +358-0-857 051, fax +358-0-625 308
- Internet: Erkki.Tomppo@metla.fi, Petri.Mikkela@metla.fi
[METLA] [National Forest Inventory]
[ERS1-SAR Project]
PM, August 30, 1995.