Within a PI project for ecology, two study sites in Germany have been investigated for vegetation mapping and forest characterization: the mainly forested area of the Harz Mountains in Northern Germany and the region of Oberpfaffenhofen supersite (Southern Germany), including the area around the lake "Ammersee".
The forests of the Harz Mountains are in parts severely affected by forest damages. Especially in the high altitudes above 800 m (connected with rough climatic conditions), the dominant spruce forest can be characterized by stand destruction over the last 1 ½ decades. This is a result of complex impact by forest decline and storm damages. Stand destruction, large clearings, and regeneration areas can be observed around the Brocken massive (the highest peak in the Harz, 1142 m) and on the quarzitic ranges of the "Ackerbruchberg" Southwest of it. The Central Harz around the Brocken has been declared a National Park in 1993.
Complimentary to various studies and observation programs by optical remote sensing in the Western Harz (Förster 1989, Kenneweg et al. 1992 and 1996), multifrequency and multipolarization SAR observations have been performed to investigate forest information content of SAR data concerning:
forest types (deciduous / coniferous / mixed), and
forest states (age classes, canopy density and regeneration classes).
After preparatory evaluations based on JPL's AIRSAR system within the MAC Europe ´91 Campaign (Keil et al., 1995), multipolarimetric data of both SIR-C / X-SAR missions have been investigated. The special aspect was to study the influence of topography on SAR data in this mountaineous area and to test methods to compensate slope effects on radar backscatter.
The datatakes over Oberpfaffenhofen supersite cover a region around the lake Ammersee South-west of Munich up to the city of Munich, with Oberpfaffenhofen in the scene center. Various SIR-C / X-SAR datasets of different modes and looking angles are available. Landcover is characterized by a mixture of settlements, agriculture, and forest areas with dominating spruce. Large parts of the forest areas are affected by storm damages. Main destructions resulted from orcans in 1990, when pure spruce stands were affected dominantly; in some areas complete stands were broken or blown down. Thus, objectives for landcover and vegetation classification by multifrequency SAR had to be orientated towards regeneration states of these areas and to heterogeneous forest environments.
The study area of Northern and Central Harz was covered by the SIR-C / X-SAR shuttle on April 13 and Oct. 4, both in full polarimetric mode, look angle 31°. Geocoded, terrain corrected AIRSAR scenes of July 1991 were used as reference data. Infrared airphotos of autumn 1991 and about 200 reference sites were available through a joint research project with the Technical University of Berlin (Winter et al., 1994),They also show a quite inhomogeneous vegetation distribution, especially at high altitudes. In order to integrate reference data and SIR-C / X-SAR data, it was necessary to obtain geocoded, terrain corrected L-, C- and X-band data products (GTC). The GTC products have been available since Sept. 1996, as so-called "combi-products", i.e. co-registrated multiband and multiseasonal products. The generation of these products was enabled through additional funding by DARA.
The Oberpfaffenhofen testsite data include full polarimetric datatakes of April 11 and Oct. 2, a dual polarimetric datatake (April 16) and three interferometric datatakes of Oct. 8 to Oct. 10. SIR-C data were available as MGD (multilook ground range detected) and SLC (single look complex) products, X-SAR data as MGD and GTC products, infrared airphotos from 1992 were used as reference.
Investigation of SIR-C / X-SAR data was mainly performed on standard polarisations, i.e. HH-, VV- and HV-Polarisation for linear polarisation of C- and L-band. Also circular polarisation states (LL-Pol., lefthand, RR-Pol., righthand) were used. To compare backscatter behaviour between different acquisitions (e.g. April vs October) calibrated products were applied, where backscatter values are expressed into dB. In order to test a Maximum Likelihood approach for supervised classification, the data was filtered with a MAP-filter of window size 3 by 3 or / and 5 by 5. For texture classification approaches, unfiltered data was used.
In the study site of Harz Mountains, backscatter values were strongly influenced by the topographic relief. A local incidence mask modeled by the digital elevation model and the shuttle path geometry was available for the GTC products. This mask was used to investigate the influence of slope or local incidence angle. In addition, an illumination correction was performed using the approach of Beaudoin et al (1996). In this approach, backscatter values are transferred to a reference incidence angle, corresponding to the incidence angle on a flat surface, by following formula:
The factor n depends on frequency and polarisation state and can be derived by linear regression functions for sigma-0 in dB. For the looking angle of 31° of the Harz passes, values between 0.95 and 1.4 were determined for n.
In preparation for landcover classification, e.g. for band selection and training set selection, backscatter signatures of various polarisation states have been studied for selected reference sites. Two classification approaches were used: Conventional Maximum Likelihood classification, performed on prefiltered SIR-C / X-SAR band combinations, and the EBIS texture classifier ("Evidence Based Interpretation of Satellite data") by Lohmann (1994).
EBIS is a window based classifier which takes the speckle and textural information of radar data in a pixel environment into account. One option is to evaluate local histograms in the window environment for their class contribution, based on multinomial distributions and criteria of evidence, according to Dempster -Shafer decision rules (Lohmann, 1991). In another option, textures can be classified based on several co-occurence feature vectors (horizontal, vertical, diagonal co-occurence of direct neighbours) modelled by multinomial density functions.
Additionally, within a joint research program of SAI and DLR-DFD, combined amplitude and coherence data from three interferometric datatakes of Oberpfaffenhofen were classified. For this approach, a multi-layer feed-forward neural network classifier was used at JRC, trained by a backpropagation algorithm.
Main issue of the interferometric investigation was to test additional information content by integrating coherence data and multitemporal amplitude data. Studies of interferometric datasets were performed on a subarea of about 20 km by 13 km around Oberpfaffenhofen. L- and C-band pairs from 8th and 9th Oct. and an X-band pair from 9th and 10th Oct. were successfully used to extract coherence. A description of the method applied for coherence estimation can be found in (Schwäbisch, 1995) and (Schwäbisch,Geudtner, 1995).
Fig. 1: Backscatter coefficients of reference stands of
spruce, canopy density loose and sparse, in dependence of local
incidence angle (before illumination correction and prefiltering
Illumination correction using the approach of Beaudoin (1996)
was found to create a sufficient base for classification. In Fig.
1 and Fig. 2 backscatter values for L-TP are shown
for 15 reference sites as function of local incidence angle, before
and after illumination correction, prefiltering. In the illumination
corrected datasets, layover and shadow areas were masked out,
as well as areas of extreme illumination ( (loc) > 55, (loc)
< 20). Remaining effects of illumination, especially in areas
of extreme illumination, can be reduced by an illumination stratification
and a separate classification in more illuminated and more shadowed
areas, an approach also used for the AIRSAR data. The illumination
corrected October data set is shown in in Fig. 3, with bands C-HV,L-HV
and X-VV in R,G,B.
Fig. 2: Backscatter coefficients of reference spruce stands of Fig. 1 after illumination correction
Spruce is the dominant tree species in the main study area of
the central Harz around the Brocken massive, between Bad Harzburg
and Braunlage. By signature analysis and classification tests,
the canopy density of spruce stands was found to have a larger
influence on backscatter than stand age, with respect to natural
age classes from thickets (about 20 years old) and pole timber
to mature timber. Several stages of regeneration were investigated.
Differentiation of deciduous, coniferous and mixed stands was
not a main task in this area. Larger parts of the deciduous and
mixed stands occuring on stronger sloped areas of Northern Harz
edge had to be masked out because of layover or high illumination
effects.
Fig. 3: SIR-C/X-SAR Image (October 4) of Central Harz Mountains after Illumination correction,
Multiband Composite of C-HV-pol., L-HV-pol., X-VV-pol. in R,G,B.
Layover and Shadowed Areas are masked out.
Thus, following forest classes were defined besides the nonforest classes of agriculture, settlements/quarries, water bodies, and moors (similar to the classes of AIRSAR investigation, Keil et al., 1994):
As high altitudes above 800 m had been still snow covered in April,
the October scene was used for classification.
Signature analysis showed the importance of L-HV polarisation and other L-band data for differentiation of canopy density classes, together with C-HV polarisation also for regrowth stages. In Fig. 4 and 5 backscatter coefficients are shown for typical training areas of these classes, for prefiltered data (MAP filtered 3 by 3, than 5 by 5) of the October dataset. X-band data was found to be very useful for separation of coniferous and deciduous/mixed stands.
Classification was performed based on MAP filtered data (MAP 3 by 3, than 5 by 5), using a Maximum Likelihood approach. Because of heterogeneous, small scaled distribution of many of the envisaged 10 classes, EBIS texture approach was not yet applied, following tests in Oberpfaffenhofen testsite. For the ML classification result shown in Fig. 6, band combination L-HH, L-HV,C-HV and X-VV was used. In this Fig., three main zones of canopy density for spruce are evident, with loose spruce stands dominating in the lower parts (in NW and SE), sparse stands at higher altitudes, and large portions of open stands, as well as clearings/cultures, in the Brocken area (around UTM 612 000 E, 5 740 000 N) and on the ridge of Ackerbruchberg (around UTM 603 000 E, 5 737 000 N).
Fig. 4: Polarimetric Signatures of October data in dB:
spruce loose, x
spruce sparse, + spruce open
Fig. 5: Polarimetric Signatures of October data in dB
clearing open culture closed
culture/young thicket
A first accuracy assessment for this classification result was done based on 67 reference sites (35 independent control areas, 32 training areas integrated in classification). In Table 1, percentages of accuracies are marked for the assignment of ten classes, added by the assignment to the first competitive class and a note on misclassification: For a number of competitive classes, transition stages have to be taken into account. Thus, the class closed cultures / young thickets shows a small accuracy, with 20% misclassified as class spruce dense/loose, with older thickets and pole timber included. Especially unsatisfying is the low accuracy for the class settlements/quarries (high overlap with spruce open), an inhomogeneous class of rural environment in several subclasses, where double bounce scattering plays a similar role as for the open stands with high trunk - ground interaction. For improvement of that separability, integration of phase information, e.g. estimate of coherence of L-HH and L-VV (Soyris,1996), could be very useful. Also, for separation of the open spruce and regeneration class closed cultures/young thickets, phase information is supposed to be helpful. Separation of agriculture and moors was not a main issue, it can possibly be improved by adding more C-band data.
Fig. 6: Classification result in the central Harz region around Brocken.
| low Backscatter both | ||||
| high double Bounce (Phase Inf. needed) | ||||
| low Backscatter | ||||
| Transition Class | ||||
| Transition Class | ||||
| Transition Class | ||||
| Transition Class | ||||
| Transition Class | ||||
| Backscatter similar (Phase Inf. needed) | ||||
| deciduous parts in Spruce Cultures | ||||
Tab. 1: Estimated accuracies in % for 10 landcover classes; for comparison misclassification is added for the first competitive class
Classification of related polarimetric datasets (April 11, Oct. 2) were performed, especially under the aspect of comparing seasonal data. But signature analysis demonstrated that both dates did not have the same preconditions concerning wheather influences: the April dataset was strongly influenced by the humid wheather conditions of that period, some backscatter coefficients showed more diurnal variations in April (e.g. in X-band) than saisonal variations. Differentiation of deciduous and coniferous forest was more difficult in the October dataset, but altogether this record showed better conditions.
The following main classes have been investigated: settlements, pasture land, acre land, water bodies, wetlands, deciduous and mixed forest, coniferous forest, clearings/cultures on storm damaged areas.
Because of heterogeneous vegetation distribution in the wooded areas (concerning mixture of deciduous and coniferous as well as clearings/cultures of storm damaged areas), texture classification was done with a special focus: the EBIS texture approach was used for forest / nonforest separation, in order to improve separation of deciduous forest and rural settlement areas. These two subclasses showed higher overlaps in backscatter distributions in all standard polarisations, both with higher double bounce scattering and multiple scattering effects, because of garden land within rural settlements, but differences in texture. Thus, EBIS texture helped to improve separability. For detailed classification, a Maximum Likelihood approach on filtered data was used.
Overall classification accuracy for October data resulted in 67% for the above mentioned 8 landuse classes, and 63% for April data (Siegmund, 1996). A main problem was still separation of rural settlements and deciduous forest on the one hand, and separation of clearings/cultures against agriculture and on the other hand. On these issues, application of phase information seems to be necessary, using complex L-HH and L-VV information and an estimate of coherence between L-HH and L-VV (Soyris,1996) or using interferometric phase information.
Multiband coherence was found to have additional potential for landcover classification. Parallel to an LCX amplitude composite (in RGB), an LCX coherence composite is shown in Fig. 7 and Fig. 8. Here coherence levels in forested areas appear mainly in red, with much higher values in L-band than in C- and X-band. Settlement areas are connected with high coherence levels in all three bands.
Comparison of multitemporal amplitude classification and amplitude / coherence classification was done for seven main classes, comparable with the classes in the polarimetric investigation, but pasture and acre land were combined to agriculture. As the images were resampled to 20 m by 20 m pixel size, no other prefiltering was carried out. The amount of improvement in the Neural Network classification, by integrating LCX coherence, can be found in Tab. 2, showing producer´s accuracy :for classification of LCX amplitude alone and LCX amplitude and coherence in combination.
Fig. 7:SIR-C/X-SAR Multiband Image of Oberpfaffenhofen site, 10-Oct.-1994,
Amplitudes of L-TP, C-TP, X-VV in R,G,B.
Fig. 8: Multiband Coherence Image, Interferometric Data Pairs between Okt 8 and Okt. 10,
Coherence Layers of L-band, C-band and X-band in R,G,B>
| LCX amplitude | LCX amplitude and coherence | |
| Settlements | ||
| Agriculture | ||
| Decoiduous Forest | ||
| Coniferous Forest | ||
| Clearings / Cultures | ||
| Lakes | ||
| Wetlands |
Tab. 2: Producer's accuracy in % for Neural Network classification of LCX amplitude and combined LCX amplitude/coherence information
The main improvement is found for the class of (mainly rural) settlements, here classification accuracy has increased from 47% to 75%, especially against deciduous forest and agriculture. Accuracy for clearings/cultures has improved, too, from 53% to 73%. Separation of deciduous and coniferous forest has increased. Thus, integration of phase information is shown to be valuable, especially for separation of rural settlements.
In the Harz testsite, a forest and vegetation mapping has been performed in mountaineous terrain, where topographic relief situation had to be taken into account for radar backscatter. The approach of illumination correction by Beaudoin (1996), developed especially for forest targets, proved to be successful. The quite steep look angle of 31° resulted in larger areas of layover and some smaller shadowed areas to be masked.
For vegetation classification, combined L- and C-band data proved successful in mapping regenerating areas and young stands, e. g. at higher altitudes. For detailed differentiation of regeneration stages like clearings, open cultures, closed cultures, and thickets, more up-to-date ground truth data should be available. Canopy density information was more relevant in SAR data, especially in L-band, than natural age classes. Three canopy closure classes could be separated. Integration of X-band data was essential for differentiation of deciduous and coniferous stands.
In the Oberpfaffenhofen testsite, forest areas were characterized by extended storm damages of 1991, having left quite heterogeneous stand structures. In classification of main landcover types, separation of rural settlements and deciduous forest types formed a main problem. Separation was improved by integrating multiband coherence, an approach which is of course limited to suitable interferometric datasets. Therefore, in further investigations, integration of polarimetric phase information is going to be included, especially L-HH and L-VV phase differences. Phase information is assumed to give valuable information on forest classes (Soyris et al., 1996), e.g. for older stands of low canopy density in comparison to younger, more closed stands of the same biomass level.
Two projects for tasks of vegetation mapping in Germany have been
defined and prepared, based on geocoded, terrain corrected SIR-C/X-SAR
„combi-products", proposed for funding by German Agency
for Space Affairs (DARA): In a subarea of Ore Mountains, SIR-C/X-SAR
data are to be used especially for mapping of regenerating forest
areas, in order to get additional information on the success of
forest regeneration program. In the context of planned enlargement
of Bavarian Forest National Park, the National Park administration
is very interested to get information by combined radar and optical
data on dead wood areas and the differentiation within young stands
and regeneration areas.
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.
Förster B. (1989), Untersuchung der Verwendbarkeit
von Satellitendaten (Thematic Mapper) zur Kartierung von Waldschäden,
Diss. TU Berlin, DFVLR-FB 89-06, Oberpfaffenhofen.
Keil, M., D. Poll, Winter, R. (1995): Investigation
of Polarimetric AIRSAR Data in a Forest Area of the Harz Mountains",
MAC EUROPE 91 Final Results Workshop Proceedings, 4-6 October
1994, Lengries, Germany, p. 171 - 178, ESA WPP-88, Jan. 95
Keil, M., D. Scales, R. Winter (1995): Investigation
of Forest Areas in Germany and Brazil using SAR Data of the SIR-C
/ X-SAR and other SAR Missions, Proceedings IGARSS' 95 Conference,
10 - 14 July 1995, Vol. II, p. 997 - 999, Firence, Italy.
Kenneweg H., Schardt M., Sagischewski H.(1992): Observation
of forest damages in the Harz Mountains by Means of Satellite
Remote Sensing, European "International Space Year"
Conference 1992, Munich, Germany, 30 March - 4 April 1992,
ESA - ISY 1, Vol II, pp. 805-810.
Kenneweg H., Schardt M., Sagischewski H. (1996):
Beobachtung von Waldschäden in Gesamtharz mit Methoden der
Fernerkundung, Forschungsbericht, im Selbstverlag, Techn Univ.Berlin
.
Lohmann, G.: An Evidential Reasoning Approach to the Classification
of Satellite Images. DLR-FB 91-29, Oberpfaffenhofen (1991).
Lohmann, G.: "Co-occurence based Analysis and Synthesis of
Textures", Intern. Conf. on Pattern Recognition, Oct. 1994,
Jerusalem, Israel.
Schwäbisch, M. (1995): Die SAR-Interferometrie zur Erzeugung
digitaler Geländemodelle, DLR Technical Report: DLR-FB 95-25,
1995.
Schwäbisch, M., D. Geudtne (1995): Improvement of Phase and
Coherence Map Quality Using Azimuth Prefiltering: Examples of
ERS-1 and X-SAR, Proc. IGARSS '95, Florence, pp. 205 - 207, 1995.
Siegmund, R. (1996): Untersuchungen zur Differenzierung von Waldbeständen
in Radardaten der SIR-C/X-SAR-Missionen - Informationsgehalt multipolarimetrischer,
mehrfrequenter SAR-Daten zweier Jahreszeiten am Beispiel des Ammersee-Gebietes,
Thesis for Diploma, University of Regensburg / DLR Oberpfaffenhofen.
Soyris, J.-C., Th. Le Toan, N. Floury, L. Thomasson, Ch. Ch. Hsu,
J. A. Kong (1996): Use of Polarisation Synthesis for Deforestaion
Studies based on SIR-C/X-SAR Data Analysis, Proc. IGARSS'96, Lincoln,
Nebraska,pp. 836-838.
Winter, R., M. Keil, D. Poll (1994): Forest Information Extraction
by Polarimetric AIRSAR Data in the Harz Mountains. PIERS '94 Conference,
11-15 July 1994, Noordwijk.