CREATION OF LARGE AREA FOREST BIOMASS MAPS FOR NE USING ERS- 1/2 TANDEM COHERENCE

Oliver Cartus(1), Maurizio Santoro(2), Christiane Schmullius(1) , Pang Yong(3) , Li Zengyuan(4)

(1) Department of Earth Observation, Friedrich-Schiller University Jena, Grietgasse 6, 07745 Jena (Germany), Email: [email protected], [email protected] (2)Gamma Remote Sensing, Worbstrasse 225, 3073 Gümligen (Switzerland), Email: [email protected] (4) Center for Ecological Applications of Lidar, Colorado State University, Department of Forest Sciences, 131 Forestry Building, Fort Collins, CO 80523 (USA), Email: [email protected] (5) Forest Remote Sensing Lab, Institute of Forest Resource Information Technique, Chinese Academy of Forestry, Wanshoushan Hou, Haidian, Beijing, 1000091(China), Email: [email protected]

ABSTRACT Space Agency (ESA) and the Ministry of Science and Technology (MOST), P.R. China. After data selection ERS-1/2 tandem coherence is known to allow forest and processing the ERS tandem dataset available for stem volume mapping with reasonable accuracy. Large- forest mapping investigations consisted of 223 pairs. scale forest mapping, however, is hindered by the The dataset comprised data from all seasons and was variability of coherence with meteorological, acquired with a range of baselines between 0 and 400m. environmental and orbital acquisition conditions. The traditional way of stem volume retrieval is based on the In the project a simple empirical model was training of models, relating coherence to stem volume, used, which described the relationship between forest using forest inventory which is generally available for a stem volume and ERS tandem coherence primarily in few small test sites but not for large areas. In this paper terms of temporal decorrelation. While this can be a new approach is presented that allows the training of a considered fine for cases when a single data take is used semi-empirical model on a frame-by-frame basis using for the forest mapping, it is instead not sufficient for a the MODIS Vegetation Continuous Field product multi-seasonal and multi-baseline dataset, as the one without further need of ground data. A comparison of acquired over . This suggested the use the new approach with the traditional regression-based of a more general and robust model, which also and ground-data dependent model training procedure considers volume decorrelation effects, i.e. the and the application of the new approach to a multi- Interferometric Water Cloud Model [6]. Because of the seasonal and multi-baseline ERS-1/2 tandem coherence large area and the multi-temporal characteristic of the dataset covering Northeast China are presented. ERS dataset, coherence strongly varies with

meteorological and environmental conditions both in 1. Introduction space and in time. The model therefore needs to be During the ERS tandem mission a large dataset has been trained on a frame by frame basis, assuming that inner- generated for interferometric applications, in many frame variations of conditions are negligible. Typically ways being yet unexploited concerning forest model training exploits ground reference data to applications. Typically ERS-1/2 tandem data has been determine the model parameter unknown a priori. Since considered for methodology development over small ground-truth data was generally not available for most test sites, mostly in the boreal zone [1,2]. The SIBERIA of the coherence images, a training method needed to be project (SAR Imaging for Boreal Ecology and Radar found that does not depend on ground data. The Interferometry Applications) demonstrated for a 1 Mio SIBERIA model contained only one unknown that km2 large area in Central Siberia that large-scale forest could be determined by using the image histogram mapping is possible using ERS tandem coherence [3-5]. solely. This approach could however not be transferred Still the data used within this project was restricted to 1- to the multi-seasonal and multi-baseline dataset 2 data takes of images acquired during one season (fall covering Northeast China. We therefore analysed the 1997). To further advance knowledge in the field of use of the freely available MODIS Vegetation forest mapping with SAR interferometric data and Continuous Field product, providing global estimates of provide a further demonstration of the retrieval percent tree canopy cover at 500 m pixel size. In this capability of the ERS-1/2 tandem coherence, not only in paper a novel method will be presented, that aims at the boreal forests, the full archive of ERS-1/2 tandem data determination of the unknown parameters of the for Northeast China has been considered in the Forest Interferometric Water Cloud Model. Dragon Project. This is part of the DRAGON Cooperation Programme initiated by the European

______Proc. ‘Envisat Symposium 2007’, Montreux, Switzerland 23–27 April 2007 (ESA SP-636, July 2007)

In order to be able to develop and validate this method, be related to stand density directly. As suggested in [7], we considered an area for which forest inventory data RS will be used as a proxy for the structure of a forest in was available so that traditional model training could this study, i.e. high RS indicates a more homogeneous also be carried out. For this we considered data for the type of forest. For each stand, the polygon in the forest enterprises at Chunsky and Bolshe-Murtinsky in inventory data was edge-eroded, meaning that a 50 m Central Siberia, which were already used for the wide buffer zone along the stand boundaries was SIBERIA project. The method is then applied for forest removed from the rasterized inventory data in order to mapping of Northeast China, however, since most minimize the influence of mis-registrations between forests in Northeast China are located in mountainous satellite and ground data. For each stand the average areas, the influence of ‘topographic’ effects in the coherence resp. intensity was calculated. Stands with coherence data also need to be investigated. obvious errors, for example when coherence was high although the inventory data indicated high stem volume, Section 2 describes the test sites and the reference data were removed from the dataset. and Section 3 the ERS data available. The new VCF- based model training approach and the semi-empirical 3. SAR measurements model used are described in Section 4. Section 5 presents the results of the new model training procedure The multi-seasonal ERS-1/2 tandem dataset available compared to a regression-based training. Also in consisted of five and two image pairs for Bolshe- Section 5 the production of a forest map of Northeast Murtinsky and Chunsky respectively. The perpendicular baselines were between 65 and 313 m. The whole China is presented. dataset covering Northeast China consisted of 223 coherence images acquired in the period between winter 2. Test sites and ground data 1995 and summer 1998 with baselines up to 400 m. The methodological development was carried out using Interferometric processing consisted of co-registration the forest inventory data from a number of forest at sub-pixel level, multi-looking (1x5 for the Siberian compartments of two forest enterprises, Bolshe- dataset; 2x10 for the Chinese dataset), common-band Murtinsky (57° 5’ N, 92° 55’ E) and Chunsky (57° 45’ filtering and coherence estimation using adaptive N, 96° 43’ E) located in Central Siberia. The forest window sizes between 3x3 and 9x9 pixels. All images inventory data was last updated in 1998. The were finally geocoded using the SRTM-C DEM. During compartments are from now on referred to with an geo-coding, all images were resampled to a pixel size of index describing their geographical location within the 50x50 m. A detailed description of the processing of the territories (Bolshe NE, Bolshe-NW, Chunsky N and ERS-1/2 tandem coherence mosaic covering Northeast Chunsky E). The sizes of the compartments vary China was given in [8]. During geo-coding, slope, between 200 and 400 km 2. Bolshe NW is located west layover/shadow, pixel area normalization and local of the River where the topography is rather incidence maps were produced. gentle and soils are sandy or swampy. At Bolshe NE, located east of Yenisey river, a hilly relief can be found Table 1. Meteorological conditions at acquisition (T = with peat soils being dominant. The northern part of temperature, SD = snow depth, WS = wind speed), Chunsky N and the whole area of Chunsky E are located forest compartment and the perpendicular baseline. T in rather flat areas whereas the southern part of Chunsky and WS are averages for both acquisition dates. Only if N is characterized by steeper slopes. The forests in this larger differences occurred, values for both dates are part of the boreal zone are dominated by either mature given (index 1 & 2) (see [7]) stands of coniferous species like spruce, fir, larch, pine Acq. date Area B Weather conditions and cedar. When disturbance occurs, birch and aspen n T ≈-10° C, T ≈-23° C, 29.12.1995 171 1 2 and at Chunsky test sites also larch regenerate first and Chunsky N WS ≈6 m/s, WS ≈ 0 30.12.1995 m 1 2 are generally replaced after 60-100 years by evergreen m/s,SD: 18 cm coniferous species. The maximum stem volume in the T≈-20 °C, WS 1≈ 5-6 m/s, 3 01.01.1996 144 compartments is in the range of 400-470 m /ha [7]. Bolshe NE WS < 3 m/s, SD: 16 cm, 02.01.1996 m Snowfall The forest inventory data was provided in the form of 14.01.1996 Chunsky N 65 T1≈-18° C, T 2≈-23° C, digital forest stand boundary maps. A wide range of 15.01.1996 & E m WS < 2 m/s, SD: 27 cm parameters, like stem volume, height, dbh, age and 22.09.1997 260 T1≈16 °C, T 2≈19°C, Bolshe NE st relative stocking (abbr. RS), was given for each 23.09.1997 m WS< m/s, Rain on 21 polygon. Stem volume was given per hectare and 25.09.1997 Bolshe NE 233 T1≈20 °C, T 2≈13°C, accounted for all living species in a stand. RS is a 26.09.1997 & NW m WS < 2 m/s 27.10.1997 158 measure of stand density resp. basal area with respect to Bolshe NE T≈2 °C, WS < 1 m/s 28.10.1997 m an optimal stocked stand and was given from 0 to 100 28.05.1998 Bolshe NE 313 T1≈26 °C, T 2≈19°C, % in steps of 10%. It depends on site quality and cannot 29.05.1998 & NW m WS < 3 m/s

4. Coherence Modelling estimates of percent tree canopy cover at a pixel size of 500 m. Modelling of the relationship between coherence and stem volume was done using the semi-empirical Water In [8] a clear decrease of coherence with increasing Cloud Model IWCM [2,6,7]. The model describes the VCF tree cover was found. It was therefore analyzed if ERS-1/2 tandem coherence over forests γfor as a sum of the modes of the distributions of coherence and intensity a ground Γgr and a vegetation contribution Γveg . for low respectively high VCF tree canopy cover allow 0 0 0 σ gr the estimation of the unknown γ , γ , σ , σ Γ = γ e−βV (1) gr veg gr veg gr gr 0 parameters. An upper threshold of 10 % VCF tree cover σ for 0 was used for the estimation of γgr and σ gr resp. the 98- 0   − jωh −αh  σ veg −βV  α  e − e  percentile of VCF tree cover as lower threshold for the Γveg = γ veg 0 ()1− e   −αh  (2) 0 σ for  α − jω  1− e  estimation of γveg and σ veg . The 98-percentile, which was chosen to ensure that a sufficient number of The ground component (Eq. 1) accounts for the coherence and intensity pixels are being used for the temporal decorrelation of the ground, weighted by the 0 estimation, was 79 % for all Siberian test sites and was fraction of backscatter from ground σ gr with respect to 0 close to the maximum tree cover of boreal forests, the total backscatter σ for and the forest transmissivity 0 which hardly exceeds 80 %. Since γveg and σ veg expressed as a function of the stem volume, with β represent values for an ideally opaque canopy, a being an empirical parameter. The vegetation residual ground contribution needed to be compensated contribution (Eq. 2) accounts for the temporal for. For this we could assume that the forests that decorrelation of the canopy, weighted by the forest correspond to the highest VCF levels have a stem transmissivity and the fraction of backscatter from the 3 0 volume of ~400 m /ha. Using values between 300 or canopy σ veg with respect to the total backscatter. The 3 500 m /ha did in fact not affect the estimation of γveg terms in the large brackets account for volume 0 and σ veg. decorrelation and the effect of the InSAR geometry. These two factors are expressed as a function of the 5. Results two-way attenuation of the signal in the canopy α, the tree height h and the InSAR geometry coefficient ω, 5.1 Topographic effect where B represents the perpendicular baseline, R the n Before proceeding with model training and slant range and θ the local incidence angle (Eq. 3). The classification, the effect of uncompensated wavenumber backscatter from a forest is modeled according to [9] shift causing additional decorrelation had to be (Eq. 4). analysed. In [11] it was shown for hills covered with

4πBn o o −βV o −βV grassland that slopes steeper than ~10° and tilted ω = (3) σ for = σ gr e + σ veg ( 1 − e ) (4) λR sin θ towards the sensor were somewhat stronger decorrelated than flat areas or areas on the backside of the hills. The tree height is generally replaced by an allometric Areas affected by additional decorrelation needed to be equation of the form: h=(a*V) b, with a and b being masked out in order to avoid misclassifications. Fig. 1 regression parameters. Parameters found for shows how strong the dependence of residual spatial Scandinavian boreal forests [6] were also found to be decorrelation upon the length of the baseline is. Fig. 1 valid for Siberian [7] and Northeast-Chinese forests and shows the average coherence of areas with low VCF were therefore used in this study (a=2.44, b=0.46). The tree cover (<2%) for 4 slopes versus aspect angle. 0 0 IWCM includes five unknowns: γgr , γveg , σ gr , σ veg and β 22-23 Feb. 96, B = 40 m 28-29 Mar. 96, B = 101 m 18-19 Jan. 96, B = 148 m n n n that should be determined via model training. 1 1 1 Rigorously, α should be considered an unknown as well. Recent results, however, showed consistent modeling 0° 0.5 0.5 0.5 results using a value of 1 dB/m in case of frozen and 2 5° dB/m in case of unfrozen conditions [7]. Coherence Coherence Coherence 10° 15° 0 0 0 0 100 200 300 0 100 200 300 0 100 200 300 Traditionally model training is based on the availability Aspect angle [°] Aspect angle [°] Aspect angle [°] of forest inventory data. In case of large datasets with Figure 1: Coherence of non-forested areas for different high variability of coherence with respect to slopes and aspect angles meteorological and environmental acquisition Uncompensated spatial decorrelation gains importance conditions and lack of area-wide ground-truth data, the at even lower slopes with increasing length of the modelling is severely limited. To overcome these baseline. For baselines longer than 100 m even slopes problems a novel training approach has been developed, tilted away from the sensor were affected. We also utilizing the MODIS Vegetation Continuous Field found that the spatial estimation of coherence, in this (VCF) product [10]. The VCF product provides global case up to 9x9 pixels, affected the coherence located near steep slopes. As a consequence, spatial was lower for the image acquired in December 1995 decorrelation propagates into flat areas. Almost over Chunsky N compared to the image acquired in complete decorrelation could be observed for long January 1996 over the Chunsky territory. This should baseline images in areas characterized by steep slopes have been related to the longer baseline and stronger located close to each other. winds.

To avoid that slope-dependent decorrelation affects Since VCF-based training does not allow the modeling and classification, a masking strategy had to determination of the forest transmissivity parameter β, be considered. The masking of all slopes above 10° and generalized values needed to be defined, describing the widening the mask for half of the maximum side length two main scenarios occurring: (i) stable frozen imaging of the coherence estimation window was considered as a conditions and (ii) early saturation of coherence with suitable masking strategy. In case of images acquired respect to stem volume as a consequence of unstable with short baselines (up to ~100-150 m) only masking environmental conditions. Regression-based training of slopes tilted towards the sensor was necessary. delivered a narrow range of values for β between 0.005 and 0.007 for the first case. A value of 0.006 was 5.2 VCF-based model training therefore used for the VCF-based training for all images acquired under stable frozen conditions. For all other To validate the VCF-based model training we compared images β equal to 0.012 resulted in reasonable modelled the results to model parameters estimates obtained using coherence. It was considered a reasonable compromise the traditional training method based on ground since it well described the steep decrease of coherence reference data. For each test site, the polygons were at Bolshe NW, cf. Fig. 2, and the less defined trend therefore divided into a training and a test set by sorting found for all images acquired over Bolshe NE (when the data for increasing stem volume and assigning every doing regression with a fixed value for β). At Bolshe other stand to the test and training set respectively. Only NE the considerable spread of measurements did not stands larger than 3 ha and with a RS >50% were used suggest a certain value for β since a relatively wide for modelling in case of the images acquired in winter range of values could be used to somehow describe the [7]. Since stands with high RS were predominantly at trend. the mature stage and the sensitivity of coherence reduces to a range up to ~100 m 3/ha if rain, freeze/thaw, Chunsky N 29-30 Dec. 95 Chunsky E 14-15 Jan. 96 1 1 etc. occurs, a lower threshold of 30% had to be used for Rel. RMSE (VCF) = 38.0 Rel. RMSE (VCF) = 60.4 Rel. RMSE (LSQ) = 38.1 Rel. RMSE (LSQ) = 48.3 the images acquired under changing environmental conditions. The use of higher thresholds caused the 0.5 0.5 model training to generate unrealistic estimates since too few measurements were available below the Coherence Coherence saturation level. 0 0 0 100 200 300 400 0 100 200 300 400 Stem volume [m 3/ha] Stem volume [m 3/ha] Fig. 2 shows four examples of the model lines obtained Bolshe NE 22-23 Sep. 97 Bolshe NW 25-26 Sep. 97 from both training procedures with respect to reference 1 1 data. The top row includes examples of images acquired under stable frozen. The bottom row includes examples of images acquired under unstable conditions, showing. 0.5 0.5 early saturation of coherence with respect to stem Coherence Coherence volume. Coherence clumped mostly at a low level for 3 0 0 volumes above 100 m /ha. Differences in ground 0 100 200 300 400 0 100 200 300 400 coherence could be found when regarding the coherence Stem volume [m 3/ha] Stem volume [m 3/ha] of sparse forest, i.e. high influence of the ground Figure 2: Regression- and VCF-based model training coherence, at Bolshe NE and NW. The coherence of results. The dashed lines show the model line obtained sparse forest was generally higher at Bolshe NW. In [7] from regression, the solid line shows the result of the it was concluded that this observation should be related VCF-based training. For the two images acquired under to the sandy soils that can be found at Bolshe NW frozen conditions, the stem volume retrieval accuracy in whereas Bolshe NE is characterized by peat soils. It was form of the relative RMSE [%] is shown as well assumed that differences in soil-moisture variation of (LSQ=least-squares regression, VCF=VCF-based both soil types after rainfall caused the differences in training). ground coherence.

In case of the winter images, no or a much later saturation could be observed with a level of ground coherence around 0.8. The coherence of dense forest

Table 2. Producers’ Accuracy, overall accuracy and In order to get an impression of the possible kappa coefficient for the classification based on the classification accuracies, the images covering the VCF training procedure utilizing the class division of Siberian test sites were classified according to the the SIBERIA map. The second value represents the SIBERIA legend into 4 stem volume classes (‘0- accuracy when doing classification with regressed 20’,’20-50’,’50-80’,’>80 m 3/ha’) using simple model parameters thresholds. Tab. 2 shows the producers accuracy for each class, the overall accuracy and the kappa coefficient Test site & 0- 20- 50- >80 Over kappa image 20 50 80 [m 3/ -all obtained from pixel-based accuracy assessment. Only ha] Acc. the accuracies of the 4 examples shown in Fig. 2 are [%] given since they were representative. Chunsky N 78.6 38.8 12.4 93.9 81.1 0.69 29-30 Dec.95 80.4 26.4 8.0 97.2 82.1 0.68 The highest overall accuracy of 81 % and a kappa of Chunsky E 65.6 39.8 29.2 87.4 70.5 0.54 0.69 were achieved for the image from 29-30 December 14-15 Jan.96 73.3 26.9 31.4 84.9 72.5 0.54 covering the Chunsky N test site closely followed by the Bolshe NE 8.1 14.1 51.1 92.8 37.0 0.22 image acquired on 25-26 September 1997. In both cases 22-23 Sep.97 72.3 28.7 30.0 84.6 69.0 0.52 the classes ‘0-20’ and ‘>80 m 3/ha’ showed the highest Bolshe NW 81.0 67.8 34.1 78.1 75.6 0.62 accuracies whereas the intermediate classes showed 25-26 Sep.97 82.0 67.8 30.9 76.8 75.0 0.62 lower accuracies. The overall accuracy, however, was mainly influenced from the extreme classes since they When comparing regression- and VCF-based training were the dominant classes. For the coherence image results, slight differences occurred, depending on the from 22-23 September 1997, the accuracy of the lowest environmental conditions. For the images acquired class dropped to 8 %, being a consequence of the high under unstable conditions, the estimation of γveg resulted value for γgr obtained from the VCF-based model in reasonable outcomes. As can be seen in Fig. 2 for the training. Also in case of the scene from 14-15 January coherence from 22-23 September 1997 at Bolshe NE, 1996, the overestimation of γgr resulted in a lower the estimated γgr was considerably higher than indicated accuracy for the lowest class and, since this class by the coherence measurements within the test site. At represents a large part of the in situ data, a lower overall Bolshe NW, however, almost identical values for γgr accuracy. The low accuracies of the intermediate classes were estimated from both training approaches. This was questioned the delineation of these two classes. Still the likely caused by the differences in ground coherence at main confusion did not occur in between these classes the sandy and peat soil areas. The largest fraction of low but with the extreme classes. Therefore, joining the VCF tree cover consisted of open areas west of Yenisey intermediate classes would only have improved the river, i.e. for sandy soils. Thus VCF-based training accuracy slightly. When comparing the VCF-based delivered the (higher) ground coherence of these areas. classification results with the result of a classification using the regressed model parameters, comparable For the winter images acquired over the Chunsky test results could be found. Thus the low accuracies of the sites, the VCF-based training delivered a γveg ~0.15 intermediate classes can be considered a systematic lower than the regressed value in case of the image from problem rather than a problem of the training method. 29-30 December 1995. This finding was at least to some extent the consequence of a lower coherence in some areas of the image with high correspondent VCF tree 5.3 Production of the forest map cover. This might have been caused by different wind The production of the forest map of Northeast China speeds. This difference, however, accounted only for comprised following steps: ~0.04 of the difference in γ . Another reason for the veg 1. Topography: (i) masking of all slopes above 10° difference might be found in uncertainties in the (long baseline) resp. all slopes above 10° tilted regression-based training since the distribution of stem towards the sensor (short baseline < ~150m), (ii) volume shows a gap in the range of 50-100 m 3/ha at the masking of layover/shadow areas, (iii) extension of Chunsky N test site. In the second example of a winter the masked areas by 4 pixels in order to account for image, the VCF-based training delivers a model line the spatial estimation of coherence. following the same trend like the regression line. Still a 2. Land Cover: A track wise analysis of ground certain offset was found. Also in this case spatial coherence and intensity, i.e. VCF<10%, was carried variation of environmental conditions could explain the out. The influence of an exclusion of urban and results. Still an analysis of this assumption is difficult, agricultural areas as well as the Inner Mongolian since a demarcation of areas of consistent coherence arid grasslands was tested using the MODIS Land behaviour is generally not possible. Similar Cover product [12]. With the aid of the land cover observations were made for the VCF-based estimation map, variations of coherence and intensity not of σ0 and σ0 but are not discussed here. gr veg related to forest underground, for example

decorrelation because of irrigation, harvest, etc., were identified. The information obtained could be The first results of the accuracy assessment confirm that used to stabilize the model training in distinct winter data are more suited for forest mapping, even transition zones like the one between the forests of though mapping of sparse v. dense forest seems to be Lesser Hinggan Mountains and the agricultural areas more realistic than stem volume mapping if high of the or the transition zone accuracy needs to be obtained. We have however to between Greater Hinggan Mountains and the Inner consider that the strength of an ERS-1/2 coherence Mongolian arid grasslands, which were based retrieval of forest biomass lies in the multi- characterized by very low backscatter. temporal combination, which is known to significantly 3. VCF-based training of the IWCM for each frame: improve classification accuracies. This option is The forest transmissivity parameter β was set to unfortunately possible only at local scale both in Siberia 0.006 for all images acquired under stable frozen and Northeast China. conditions and to 0.012 for the rest. Potential differences in forest transmissivity between Siberian 7. Acknowledgments and Northeast Chinese forests could not be analysed. 4. Calculation of coherence thresholds for 20, 50 and Ground and satellite data were available from the EC- 80 m3/ha using the IWCM and application of the funded SIBERIA (ENV4-CT98-0743), SIBERIA-II thresholds to the coherence images. (EVG1-CT-2001-00048) and Forest Dragon (C1 5. The water class was generated using the SRTM F.2583) projects. For SAR processing the Gamma Remote Sensing software was used. Weather data were Water Bodies Mask. provided by DWD. The forest map of Northeast China is shown in Fig. 3. The SRTM Water Body Mask was downloaded from: Validation has not been done yet but is foreseen in http://e0srp01u.ecs.nasa.gov/srtm/version2/SWBD. The cooperation with the Chinese project partners. VCF product was downloaded from: http://edcdaac.usgs.gov/modis/dataproducts.asp. using The MODIS Land Cover product was downloaded from http://duckwater.bu.edu/lc/mod12q1.html

8. References [1] Pulliainen, J. et al. (2003). Feasibility of multi- temporal interferometric SAR data for stand-level estimation of boreal forest stem volume. Remote Sensing of Environment , 85, 397-409. [2] Santoro, M. (2002). Stem Volume Retrieval in Boreal Forests with ERS-1/2 Interferometry. Remote Sensing of Environment , 81, 19-35. [3] Schmullius, C. et al. (2001). SIBERIA - SAR Imaging for Boreal Ecology and Radar Interferometry Applications. Final Report, EC- Center for Earth Observation, Project Reports, Contract No. ENV4- CT97-0743-SIBERIA. [4] Wagner, W. et al. (2003). Large-scale mapping of boreal forest in SIBERIA using ERS tandem coherence and JERS backscatter data. Remote Sensing of Environment , 85, 125-144. Figure 3: Forest map of Northeast China produced [5] Balzter, H. et al. (2002). Accuracy assessment of a from 223 ERS-1/2 tandem coherence images large-scale forest cover map of central Siberia from synthetic aperture radar. Canadian Journal of Remote 6. Conclusions Sensing , 28, 719-737. In this paper a novel classification approach has been [6] Askne, J. (1997). C-band repeat-pass interferometric IEEE Transactions on presented, which aims at the fully-automated SAR observations of the forest. Geoscience and Remote Sensing classification of forest stem volume using ERS-1/2 , 35, 25-35. [7] Santoro, M. et al. (2007). Properties of ERS-1/2 tandem coherence. The method developed at test sites in Siberian forests was transferred to map forest biomass coherence in the Siberian boreal forest and implications Remote Sensing of of Northeast China. The new classification approach for stem volume retrieval. Environment represents a fast and easy to apply method to map , 106 (2), 154-172. forests.

[8] Santoro, M. et al. (2005). On the generation of a forest biomass map for Northeast China: SAR Interferometric processing and development of classification algorithm. In Proc. of Fringe 2005 Workshop , ESA-ESRIN, Frascati. [9] Pulliainen, J. et al. (1994). Backscattering properties of boreal forests at the C- and X-bands. IEEE Transactions on Geoscience and Remote Sensing , 32, 1041-1050. [10] Hansen, M. C. et al. (2002). Towards an operational MODIS continuous field of percent tree cover algorithm: Examples using AVHRR and MODIS data. Remote Sensing of Environment , 83, 303-319. [11] Castel, T. et al. (2000). ERS INSAR data for remote sensing hilly forested areas. Remote Sensing of Environment , 73, 73-86. [12] Strahler, A. et al. (1999): MODIS Land Cover Product: Algorithm Theoretical Basis Document, Version 5.0. Boston University, Boston, MA, USA, pp72.