Introduction and objective
The aim of the evaluation is Forest Monitoring using SIR-C/X-SAR data. It is intented to get information about tree species, canopy density, age-class structure and storm damaged areas. Of special interest are the following questions:
From what size on non-wooded areas in forests can
be recognized? Where is the limit in stocking degree among forest
and non-forest? Which biomass classes can be discriminated? Can
be detected differences in state of foliage between the two seasons?
Which further forest structures can be recognized? These objectives
were investigated in the two different testsites 'Oberpfaffenhofen'
and 'Bavarian Forest'.
Part A: Testsite Oberpfaffenhofen
Study Area
The forested surfaces of the supertestsite Oberpfaffenhofen
southwest of Munich/Germany are mainly located on moraines. The
investigated area contains an elevation difference of about 100
m maximum. The most important tree species are spruce (Picea abies),
beech (Fagus silvatica) and pine (Pinus silvestris). First of
all the area is -in addition to small monocultures- characterized
by a mixture in species, age-classes and management. The resulting
diversity of structure has been accelerated by storm damage and
bark beetle infestation during the last few years.
SAR Data
Overall 13 X-SAR and 6 SIR-C datasets from both missions were included in the presented work.
The used X-SAR data were available as GTC (Geocoded
Terrain Corrected)-products. The SIR-C data were
used as SLC (Single Look Complex)-products.
From the SIR-C data either dual-pol or quad-pol datasets were
derived depending on the recording modus. Quadpol data allow the
calculation of every polarization direction (linear, circular
and elliptical), but show a lower spatial resolution at the same
time. On the other hand only certain linear polarization configurations
are possible with dual-pol data, but they show higher spatial
resolution. The nominal resolution of the X-SAR GTC data is 25m
* 25m, the pixel spacing is 12.5m * 12.5m. In order to create
images with relative square formed pixels and to get a similar
spatial resolution like the already processed X-SAR data, the
SIR-C data were processed with 1 look in range and 3 looks in
azimuth [1].
Methods: Image to Image Registration and speckle
filtering
The X-SAR GTC-products of both missions could be combined directly to a multitemporal dataset. One subset of 15 km * 15 km and one of 7 km * 8 km within the Oberpfaffenhofen testsite has been choosen. The SIR-C data had to be adapted geometrically to the geocoded X-SAR data via image to image registration. This was performed with the aid of Ground Control Points and polynomial transformations. Here often the problem aroused finding enough qualified Ground Control Points in the corresponding datasets. As variations of the elevations were not taken into account within this method, the procedure owns clear inaccuracies and is therefore restricted to small areas. In this way multifrequent, multipolarized and multitemporal datasets have been created.
For application of a suitable speckle reduction different
speckle filters and different filter sizes have been tested. Finally
the Gamma Map filter with the size of 3x3 twice one after another
has been used [2].
Evaluation of the Aerial Photographs
Field observations and aerial photographs were used
for the aquisition of Ground Truth data. The analytical evaluation
of the aerial photographs has been a very important step during
the investigation of the information content. We were able to
use a stereoplotter for a detailed analysis of the aerial photographs.
For that purpose Ground Control Points have been measured with
the aid of a GPS (Globle Positioning System) -receiver. In this
way the aerial photos could be geocoded absolutely. Now different
measurements -for example stand heights or the size of deforested
or storm damaged areas- could be carried out. The results of this
processing step were digital thematic maps, which were linked
to the radar data via corresponding coordinates and used as training
and verification areas. Thus a method has been developed allowing
an objective selection and exact location of the objects of investigation.
In addition the classes of characters were marked independently
of the displayed color composits of the radar data.
Results: Visual Interpretation
X-SAR
This short wavelength provides poor information for
the object forest. Wooded and agricultured areas could not be
differentiated in a proper way. The use of multitemporal datasets
did not clearly improve the information content.
SIR-C combined with X-SAR
By combination of the three frequencies and the different
polarizations the content of information was significantly enhanced.
Forest and non-forest areas could be separated well within coherent
woodland. Between forest and build-up areas some intersections
appear. Especially the cross-pol combination of C and L band was
found to be a good combination for forest application. We noticed
as the most important feature the different spatial resolution
of quad-pol (mode x16) and dual-pol (mode x11) data. Dual-pol
data allow the recognition of non-wooded areas within forests
smaller than 1 ha (hectare), which cannot be detected by quadpol
data anymore. We found that the benefits of the higher spatial
resolution are much more important than the disadvantage of the
restricted polarimetric capacity. Nevertheless circular polarizations
with the duad-pol data were tested, but there could not be found
any advantage in comparison to the linear polarizations. So dual-pol
data were assessed to be more valuable and were used preferentially
in this heterogeneous investigation area.
Signature Analysis
As the used data were calibrated, it was possible to calculate the backscatter coefficient in dB values [3]. Now the mean dB values of many different training areas, subclasses and combined classes were compared. A good possibility for visualization are signature diagrams in form of dB-values as a function of frequency and polarization. It is important to take the variance of the means into consideration. It has been found, that if the means indicate for example in L-HV a possible separation of interesting features, the variance of the means was often significantly higher. For further detailed analysis scattergrams and histograms of single combinations were very helpful.
In our study we made several tests with different
training areas, subclasses and combined classes (similar to different
classification levels). We found out, that structural features
like altitude differences of trees and shrubs as well as vegetation
density or gaps within forested areas have more influence on the
backscatter coefficient than tree species and stand composition.
The results of the signature analysis are:
Classification: Maximum Likelihood Classification
for 6 landuse classes via transferred signatures
As an example of a performed Classification the result
of the multifrequent and multipolarized dataset of 07th October
is shown (fig.2). In this case, the transfer of signatures from
the other investigated subset has been tested. Therefore the areas
were classified using alternately the training areas and the signatures
of the other part for classification and the own training areas
for verification. For 6 landuse classes, only C and L-band were
included in the classification process. The result of this example
confirms the observations during visual interpretation and detailed
signature analysis. There exist certain signature intersections,
but overall, the investigated object 'forest' can be separated
quite well from the other landuse classes.
Testsite 'Oberpfaffenhofen' (subset 15 x 15 km)
Fig.1: 07 October 1994,
data take 110.00,
C-TP,L-TP,L-HV (RGB), speckle filtered
Fig.2: result of the classification of 6 landuse
classes with transferred signatures
Tab.1: Confusion matrix and legend for classification
of 6 landuse classes with transferred signatures (black area in
the nw-corner due to missing data, pixels not considered in the
confusion matrix)
verification areas results of classification
%
| no | class | 0 | 1 | 2 | 3 | 4 | 5 | 6 |
| 1 | build-up (white) | 17.7 | 17.3 | 29.5 | 29.5 | 4.2 | 0.0 | 1.8 |
| 2 | forest (green) | 0.4 | 2.2 | 81.4 | 14.9 | 0.2 | 0.0 | 0.9 |
| 3 | clearing (red) | 0.2 | 3.6 | 15.1 | 64.7 | 2.1 | 0.0 | 14.3 |
| 4 | agriculture (lightbrown) | 7.6 | 2.2 | 3.6 | 22.6 | 53.1 | 0.0 | 10.9 |
| 5 | water (blue) | 23.4 | 0.0 | 0.3 | 1.6 | 1.1 | 72.9 | 0.7 |
| 6 | wetland (darkbrown) | 0.5 | 0.0 | 0.0 | 8.7 | 6.3 | 0.0 | 84.5 |
black = class 0: not classified.
average accuracy = 62.34%
overall accuracy = 69.26%
The average accuracy is the average of the accuracies for each class, and the overall accuracy is the accuracy of each class weighted by the proportion of test samples for that class in the total training or testing set. Thus, the more accurate estimates of accuracy (i.e., those from larger test samples) are weighted more heavily in the overall accuracy.
Signature analysis and test classifications did not give a very clear tendency which further forest classes could be separated.
These results are caused on the one hand by the forested
surface itself, by its very heterogeneous and small sized pattern
in forest stands and on the other hand by the missing of a 'combi'-product,
that means by the inaccuracies of the combination of C and L band
to the geocoded X band using the image to image registration method.
As C and L band data were available as GTC product in December
1996 for the second investigation area ;Bavarian Forest', the
work in the testsite 'Oberpfaffenhofen' has been interrupted temporary.
Part B: Testsite Nationalpark 'Bavarian Forest'
Study Area
The National Park 'Bavarian Forest' is located in
north east Bavaria, bordered by the young National Park of Sumava
(Czech Republic) and covers an area of about 13 000 ha. The region
is part of the Central European mountain ranges (700 - 1400 m
above sea level) and is built by metamorphic and magmatic rocks.
The interior Bavarian Forest -a larger area surrounding the National
Park- builds together with the adjacent Bohemian Forest (Sumava)
the largest coherent woodland area in Middle Europe. The Nationalpark
'Bavarian Forest' is a densely forested region which is dominated
by beech (fagus silvatica) mixed with fir (abies alba) and spruce
(picea abies) in the lower areas and by spruce (picea abies) in
the higher regions. As main parts of the Nationalpark has not
been managed with regard to forestry since more than 25 years,
natural regeneration with remaining dead wood could take place.
In consequence the economically used forests with typical age
classes are lacking. This system of natural forest growth in connection
with storm damages and bark beetle infestation within spruce stands
in the last few years causes a huge variety of structure in the
investigated area. Another aspect in comparison to the Oberpfaffenhofen
testsite is the stronger influence of topograpy to the radar backscatter.
SAR Data as 'Combi-product'
Based on our existing work with SIR-C/X-SAR data
in the Supertestsite Oberpfaffenhofen we noticed, that only geocoded
products of SIR-C/X-SAR data provide a starting point for successful
evaluations, especially in hilly terrain. This kind of data was
available first time at the end of 1996, when we were received
four GTC (Geocoded Terrain Corrected)-products
from the two missions in April and October 1994. These so called
'combi-products' containing the three frequencies and the appendant
local incidence angle mask deliver the coordinates in the local
cartesian coordinate system for each pixel. Until now four data
takes were included in the work.
Methods: Visual Interpretation
In addition to the results of the testsite Oberpfaffenhofen
the visual interpretation of different RGB images of these
products shows a very strong influence of the topography in this
mountainous region. Again the dual-pol data promise better results
for this application caused by the better spatial resolution.
Furthermore the quality of the combi-product, the performed corregistration
of the three bands can be assessed. It was found out, that there
are existing differences in the quality of various data takes
as well as in the different sections within one dataset. On average
the accuracy in fitting the L and C band to the X band is approximately
1 to 2 pixels. As in the Oberpfaffenhofen testsite the analytical
evaluation of aerial photographs has been a very important step
during the evaluation.
Illumination correction
In addition to the investigated object the radar
backscatter is effected by the local slope of the relief. Different
inclinations of surfaces relating to the look direction of the
SAR sensor cause strongly modified backscatter coefficients. Without
correction of this topographic effect, major mistakes during classification
process of different surface types are to be expected. The aim
of different basic approaches for the correction of this effect
is to provide means of backscatter coefficient largely independent
from slope for statistical analysis and further use. The value
, which represents the backscatter of flat
terrain is deemed to be a qualified aim.
The existing models to reach this aim can be structured in the three categories of geometrical, statistical and semi-empirical models.
The used model in our work was developed by BEAUDOIN
et al. [4] and belongs to the semi-empirical category:
The local incidence angle (loc)
is given by the geocoded layover, shadow and incidence angle mask
„GIM". Factor n has to be determined empirically. This
factor stands for the slope of a linear regression between 
and 
. In our case values between 0.71 to
0.96 were ascertained taking the landuse class forest (dominated
by conifer) into account.
Visual comparisons of the resulting output file (fig.
4) and the uncorrected original file (fig.3) show very clear the
effect of removal or at least of reduction of the influence of
the topography. The plastical impression, the illumination differences,
generally the tracing of the relief has been removed. The red
polygon in this figures marks the border of the National Park.
Fig.3: 17 April 1994, data take 126.0, X-VV,C-HV,L-HH (RGB)
Fig.4: 17 April 1994, data take 126.0, X-VV,C-HV,L-HH
(RGB), illumination corrected
As the GIM mask contains as well information of layover
and shadow areas, these parts have been excluded for further use.
Due to the old age of the investigation area the terrain is softly
morphed. Therefore only small sections of the image had to be
erased.
Signature analysis
As in the Oberpfaffenhofen testsite, the signature
analysis is based on the evaluation of aerial photographs. On
one side the results were very similar to Oberpfaffenhofen, especially
looking for a raw landuse separation of 5 classes. On the other
side a clear improvement was realized within coherent wooded areas.
There exists a clear trend for the separability of deciduous and
coniferous forests. In the following a few examples of signature
analysis are given.
Fig.5: signature diagramm Fig.6: diagram of the
variance of the means
Fig.7: histogram C-HV Fig.8: histogram L-HH
In addition to the means of the backscatter coefficient
(fig.5) it is important to take into account the variance of the
means (fig.6). The two examples of histograms (fig.7 and fig.8)
show on the one hand a good separability of clearings including
glades from forest classes, on the other hand a big overlapping
within the forest classes. Nevertheless there exists a clear tendency
for a distinction of forests dominated by conifers and forests
dominated by deciduous. The classes 'conifer' and 'deciduous'
meaning pure stands could not be separated from there mixed variant
'mixed conifer', wich means clearly dominated by conifer but mixed
with deciduous up to about 30 %, the corresponding situation is
valid for the class 'mixed deciduous'. The class mixed forest
means a balanced proportion of deciduous and conifer trees. The
signature analysis contains numerous other combinations, but in
general the problem of to small training samples and to big signature
intersections was pointed out. Even though the signature analysis
is not final judged, classifications were performed.
Classification
As a first step a Maximum Likelihood Classification
for 5 landuse classes was performed. For that purpose X, C and
L-Band together were found to give the best results. The area
which was evaluated with aerial photos was used for training areas
and also as reference ground truth map covering approximately
30 to 35 % of the National Park area.
Tab.2: Confusion matrix of the ML-Classification
for 5 landuse classes:
training areas <==> digital ground truth
map (derived by Evaluation of Aerial Photographs) in %
| no | class | 1 | 2 | 3 | 4 | 5 |
| 1 | water | 90.8 | 5.7 | 0.0 | 3.5 | 0.0 |
| 2 | clearings, glades | 6.0 | 77.8 | 9.4 | 3.4 | 3.4 |
| 3 | forest | 0.2 | 1.5 | 93.9 | 0.4 | 4.0 |
| 4 | agriculture | 2.2 | 7.4 | 1.1 | 88.5 | 0.8 |
| 5 | build-up areas | 0.7 | 5.5 | 51.5 | 1.4 | 40.9 |
average accuracy = 78.37%
overall accuracy = 92.77%
These parts of the image, which were classified as forest and
which were lying within the border of the National Park were preserved
for further use, the other parts have been erased.
As a second step the forested areas were classified into three
classes, mixed forest, deciduous with conifer (pure deciduous
and mixed deciduous) and conifer with deciduous (pure deciduous
and mixed deciduous). This classification was performed in two
different ways.
Maximum Likelihood Classification
The first strategy was to use the combination of X-VV, C-HV and L-HH for Maximum Likelihood Classification. The confusion matrix below (tab.3) shows the result for the dataset of 17th April 94 (data take 126.02).
Tab.3: Confusion matrix of the ML-Classification for 3 forest
classes:
training areas <==> digital ground truth map (derived
by Evaluation of Aerial Photographs) in %
| no | class | 1 | 2 | 3 |
| 1 | mixed forest | 17.3 | 34.8 | 47.9 |
| 2 | deciduous with conifer | 12.5 | 60.4 | 27.1 |
| 3 | conifer with deciduous | 12.6 | 12.0 | 75.4 |
average accuracy = 51.04%
overall accuracy = 66.76%
Classification via threshold values for 3 forest classes
The second strategy was to use the difference between C-HV (fig.7)
and L-HH (fig.8) for an easy threshold value method.
Fig.9: Classification via threshold values of 17th April 94 (data
take 126.02), chv-lhh
Tab.3: Confusion matrix of classification result and the digital ground truth map
(derived by Evaluation of Aerial Photographs) in %
| no | class | 1 | 2 | 3 |
| 1 | mixed forest (red) | 34.6 | 33.1 | 32.3 |
| 2 | deciduous with conifer (yellow) | 27.3 | 54.4 | 18.3 |
| 3 | conifer with deciduous (green) | 27.4 | 12.3 | 60.3 |
average accuracy = 49.77%
overall accuracy = 56.59%
CONCLUSIONS
Multipolarimetric multifrequency SIR-C/X-SAR data proved to be a valuable tool for forest monitoring. In the centre of attention is the observation of increase or decrease of woodland. In the presented study areas the best spatial resolution is more important than the full polarimetric resolution. The availability of Combi-products enables to work in mountainous regions and to get the full advantage of the three frequencies. Furthermore the condition for combination with additional remote sensing data or additional digital data derived by a GIS is given. The reduction of topographical effect improves the information content. The work is not yet finished.
Future works and transfer of the result
The started work has to be continued. Particularly in the region
of the Bavarian Forest it is expected to get more interesting
and better results. This can be reached by the combination of
the radar imagery with optical data and with additional data like
a DGM for further detailed analysis. Up to now, the phase information
of the data is not used. If this additional source of information
of the data can be used, the classification results probably can
be improved. There is a new project planned to use the SIR-C/X-SAR
data for the assesment of forest structure in a larger region,
including the area of the present National Park Bavarian Forest,
its planned extension area in northwest and the National Park
Sumava (Czech part). Of special interest is the detection of areas
with dead wood and the differentiation within young stands and
regeneration areas.The administration of the National Park is
very interested in this work.
The authors thank the staff of DFD of the DLR, especially our
PI Manfred Keil for support. The evaluation of the arial photos
was supported by Dr. Peter Zeilhofer, Institute for Landuse Planning
and Nature Conservation, many thanks to him. The work was funded
by the German Space Agency (DARA), contract number 50 QS 900049.
[1] Chapman, B.(1994): SIR-C Data Compression Software User Guide. Jet Propulsion Laboratory Internal Note JPL D-11427, Revision 2.0, June 1994
[2] Lee, J.S.,Jurkevich, I., Dewaele, P., Wambacq, P., Oosterlinck, A.(1994): Speckle Filtering of Synthetic Apertur Radar Images: A Review. Remote Sensing Reviews, Vol.8, 1994, pp.313-340
[3] Zink, M.(1995): Results of X-SAR Calibration, International Geoscience and Remote Sensing Symposium (IGARSS'95), July 10 - 14, 1995, Florence, Italy, Vol.1, pp. 590-592
[4] Beaudoin, A, Castel, T., Deshayes, M., Stachs, N., Stussi, N., Le Toan, T.(1996): Forest Biomass retrieval over hilly terrain from spaceborne SAR data , Proceedings Symposium Toulouse, France, 10-13 October 1995, pp.131-140
[5] Werle, D. (1989): RADAR REMOTE SENSING FOR APPLICATION IN FORESTRY, December 1989, CCRS, Ottawa
[6] Coenradie, B. (1992): Waldklassifizierung und Totholzkartierung
im Nationalpark Bayerischer Wald unter Verwendung von Landsat-TM-Daten
und digitalen Zusatzdaten, Diplomarbeit Universität Bonn,
DLR-FB 92-10
[7] Keil, M., Coenradie, B., Rall, H., Sima, M. (1992): The National Parks of Bavarian Forest and Sumava - A Landsat TM perspective on the structure and dynamics of the largest woodland of Central Europe.-European „International Space Year" Conference 1992, Munich, Germany, 30 March - 4 April 1992, ESA-ISY 1, Vol.II, pp.751-756