FOREST STRUCTURE ESTIMATION USING SPACE BORNE POLARIMETRIC RADAR:

AN ALOS-PALSAR CASE STUDY

S. Cloude(1)

E. Chen (2), Z. Li (2), X. Tian (2), Y. Pang (2), S. Li (2)

E. Pottier(3) , L. Ferro-Famil(3), M. Neumann(3)

W. Hong(4), F. Cao(4), Y. P. Wang(4)

K. P. Papathanassiou(5)

(1) AEL Consultants, Cupar, KY15 5AA, Scotland, UK, e-mail: scloude@ieee.org

(2) Institute of Forest Resources Information Techniques, Chinese Academy of Forestry, Beijing, PR. China, e-mail: chenerx@caf.ac.cn

(3) SAPHIR Group, Institute of Electronics and Telecommunications of Rennes, University of Rennes 1, Rennes, France; email: eric.pottier@univ-rennes1.fr

(4) National Key Laboratory of Microwave Imaging Technology, Chinese Academy of Sciences, Beijing, PR. China, e-mail: whong@mail.ie.ac.cn

(5) Microwaves and Radar Institute, German Aerospace Centre (DLR), Wessling, Germany, e-mail: kostas.papathanassiou@dlr.de

ABSTRACT

Here we investigate the potential use of ALOS PALSAR L-band dual and quad polarization SAR data for improved forest structure mapping and parameter estimation. Several scenes of level 1.1 SLC data have been acquired under the DRAGON project in 2007 for the Tai-An test site of Shandong Province, P.R. China. The preliminary results for the first phase of this project are reported here. Firstly, the geo-coding method and its geo-location accuracy are evaluated with one geo-coded Landsat ETM as reference. Secondly, a high resolution DEM is used for implementing GTC processing. The benefits and limitations of GTC for forest mapping are evaluated with a SPOT5 image. Thirdly, various entropy-alpha polarimetric segmentation methods are evaluated for forest classification. Finally, interferometric coherence images are generated to investigate the possibility of studying polarimetric interferometry based forest structure information extraction methods.

1. INTRODUCTION

The need for reliable estimation of forest structure information over large areas is currently increasing because of the growing recognition of the potential role of forests in helping mitigate effects of climate change and global warming. Many studies have already been carried out using airborne SAR systems, as well as space systems such as SIR-C/X SAR, ERS and JERS-1. In particular, it has been observed that the backscattering coefficient at L band had some correlation with forest volume and biomass, with L-VH or HV better than HH or VV [1], although the cross polarization to co-

polarization ratio was also found to be useful for forest parameter estimation. Interferometric coherence has also been combined with these parameters to yield further improvements. For example, combining coherence and backscattering coefficient with simple scattering models such as the water-cloud model improves biomass estimation level and accuracy [2]. However, SAR signal saturation problems limit the biomass estimation level below 40-60 Tons/ha for L-band [3]. In an attempt to overcome this problem, references [4,5] first published a new method to extract forest height from repeat-pass polarimetric interferometeric SAR (POLInSAR) data. Forest biomass can then be estimated without saturation by using tree height and some known forest growth models [6]. Forest applications are the most important and successful field for POLinSAR techniques, especially at L-band. Importantly for PALSAR, even with dual-polarization interferometric observations, the POLInSAR model can still be applied for tree height inversion, albeit with reduced accuracy [5,6].

However, except for SIR-C/X-SAR, airborne SAR systems have been the only data source available for POLInSAR research until the launch of ALOS by JAXA in January 2006 [7]. ALOS is the first satellite to employ an L-band SAR sensor with both dual- and quad-polarimetric data acquisition modes. Hence it is very important to investigate and evaluate the capability and limitations of this space-borne sensor, so as to provide improved remote sensing tools and forest structure information products for management and environment protection at both local and regional scales.

Our project objectives are therefore to investigate advanced dual and quad-polarization POLInSAR data

_____________________________________________________ Proc. Dragon 1 Programme Final Results 2004–2007, Beijing, P.R. China 21– 25 April 2008 (ESA SP-655, April 2008)

processing techniques and forest structure parameter extraction methods with ALOS PALSAR data acquired in 46-day repeat-pass mode, and evaluate the accuracy of forest tree height, volume and biomass products using detailed ground truth. We aim to provide new SAR remote sensing based forest inventory techniques for forest management and the support of forest carbon estimation models with quantitative forest parameters. To investigate these new ideas we have employed ALOS data for a specific test site, Tai-An in P.R. China, which has extensive supporting data both in terms of in-situ measurements and also supporting EO data sets.

2. POLSAR PROCESSING METHODOLOGY

In order to counter image speckle effects, we begin by performing multi-look averaging of the SLC Quadpol data. Consequently we do not obtain a direct estimate of the pixel scattering matrix [S] itself but instead its 4x4 Hermitian coherency matrix <[T]> [8] (note that we employ all four channels of polarimetric data instead of the three demanded of backscatter reciprocity. This enables a more consistent treatment of noise and calibration errors in space borne systems as we shall demonstrate). For n samples we then obtain an estimate of the pixel coherency matrix as shown in equation 1 [8,9]

(1)

and this matrix is then expressed in an eigenvalue decomposition as shown in equation 2

(2)

For low frequency space-borne radars (such as the L-band PALSAR data used in this project), the 4th eigenvalue is associated with Faraday rotation and SNR effects. Faraday rotation arises from ionospheric propagation distortion of each pixel scattering matrix according to the following model [10]

(3)

where ψ is a one way propagation polarization rotation that depends on the integrated total electron concentration (TEC) along the radar path and its interaction with the local magnetic field vector. It can be directly calculated from the observed scattering matrix data using the difference between cross-polarized channels as shown in equation 4 [10]

(4)

In the ALOS-PALSAR data used in this project the mean rotation across the scene was found to be constant at –1.2 degrees. Note that for Quadpol data sets (the PLR mode of ALOS-PALSAR for example) this distortion can be removed on a pixel-by-pixel basis by using the average estimate of equation 4 to invert the matrices in equation 3. We note that such a correction is not possible in FBD or dual polarization mode (for ALOS-PALSAR this involves H transmit polarization and dual channel reception of H and V) where it represents a scene dependent error in the radiometric calibration, especially for the cross polarized HV channel, which becomes corrupted by leakage from the larger copolarized HH channel via the matrix products in equation 3.

Having removed any systematic Faraday rotation, the minimum eigenvalue then represents any residual noise in the data. In areas of high SNR however we have only three significant eigenvalues and corresponding eigenvectors available for analysis. These can be used for classification and physical parameter estimation using one of two approaches. In the first we consider only pure polarized scattering, when λ1 >> λ2,3. In this case we can employ up to five parameters per pixel as shown in equation 6, where we have used a secondary filter, a Cameron decomposition [8], to help isolate the symmetric scattering components of the pixel (these components are generally the most suitable for model based parameter estimation). Of particular importance is the so called alpha parameter, an angle representing a ratio of polarization scattering components and one that reflects differences in the boundary conditions on wave scattering at the pixel [9]. This makes it a robust parameter for classification and parameter estimation as it is based not on local data statistics but on the physics of wave scattering. We are then lead to the following two decomposition algorithms.

2.1 Polarized Decomposition

For coherent scattering (when λ1 is dominant) the following model is appropriate

(5)

where the backscatter intensity is given by λ1 and the normalized eigenvector e1 contains information on all possible polarimetric phase and amplitude ratios, including s and the additional scattering phase s as shown in equation 6. This approach is useful for identifying coherent point scatterers in the scene, most of which are usually associated with urban environments and man-made structures.

€

T[ ] =1n

kiki†

i=1

n

∑ k =12

Shh+SvvShh−SvvShv+SvhShv−Svh

€

T[ ] = λ1e1 + λ2e2 + λ3e3 + λ4e4 λ1 ≥ λ2 ≥ λ3 ≥ λ4 ≥ 0

€

Sψ[ ] =cosψ sinψ−sinψ cosψ

SHH SHVSVH SVV

cosψ sinψ−sinψ cosψ

€

tan4ψ =−2Re((SHV − SVH )(SHH + SVV )

*)SHH + SVV

2− SHV − SVH

2

€

TP[ ] = λ1e1

(6)

However in forestry applications the density of such points can be quite low and significant wave depolarization occurs. This observation can be used to help isolate different scattering environments in the scene by using the entropy/alpha/anisotropy decomposition, as shown in equation 7.

2.2 Entropy/Alpha/Anisotropy Decomposition

For depolarizing scatterers, the full eigenvalue spectrum of [T] must be considered. In this case the following model is appropriate [10,12], where H is called the entropy and A the scattering anisotropy.

- 7)

This approach can also be further combined with a Wishart unsupervised ML classifier approach to account for statistical fluctuations in the data, and when combined with class-merging techniques, be used as a robust template for analysis of polarimetric data sets [12,13,14,15]. We now turn to consider application of these ideas to ALOS-PALSAR data.

2. TEST SITE AND SUPPORTING EO DATA

Our test site is located in Tai-An district of Shandong Province, its geographic coordinate ranges from N35°59 to 36°5 in latitude and from E117°13 to 117°25 in longitude. The forest cover area of the test site includes Tai Mountain and Culai Mountain, whose forest coverage rate is above 80%. This site therefore not only poses the difficult task of identifying forest in the presence of strong topography (see Fig. 1) but also includes extensive urban development and other land use areas (see reference land use map in Fig. 2).

One regional remote sensing campaign has been carried out here from April to June of 2005, through which about

4.6 TB of earth observation data has been collected. The airborne sensor data acquired includes small footprint LIDAR, CCD and Hyper-spectral data (PHI). The space-borne sensor data includes ENVISAT ASAR-APP and APS (HH and HV) data, EO-1 Hyperion Hyper-spectral data, SPOT-5, Quick-bird and IKONOS. Ortho-rectified CCD images for the two mountains were produced separately. The DEM (Fig.1), forest component maps for Tai Mountain and Culai Mountain, and land use map (Fig.2) of 1:250 000 for Tai An district have been established using this data. Four scenes of L-band ALOS PALSAR level 1.1 data have then been acquired for the test site. The major imaging parameters for the four images are listed in Tab. 1.

Table 1. Major imaging parameters for the PALSAR data acquired for the test site

Imaging Date PolarizationAzim/rangeResolution

Inc. angle of image center

May 13, 2007 HH,HV,VH,VV 3.55/9.37m 23.8 degJune 21, 2007 HH, HV 3.18/9.37m 38.7 degJuly 20, 2007 HH, HV 3.19/9.37m 38.7 degSept 21, 2007 HH, HV 3.18/9.37m 38.7 deg

We note that we have access to only a single-pass in fully polarimetric Quadpol mode, but to several passes in the more restricted dual polarization FBD mode, two of which (June and September) are separated by 92 days, a multiple of the 46-day repeat time of the satellite and hence are suitable for investigation of polarimetric interferometry [4,5,6].

0 1020m

Figure 1: Reference DEM of Culai Mountain Test Site, P.R. China

€

TD[ ] = λ1e1 + λ2e2 + λ3e3

H = − Pii=1

3

∑ log3 Pi Pi =λiλ∑

A =λ2 − λ3λ2 + λ3

α = Pii=1

3

∑ α i

€

e1 =

cosαsinα cosβeiδ

sinα sinβeiχ

Cameron → e1s + +ens =

cosα s

sinα s cosβseiδ

sinα s sinβseiδ

+ ens

e1s =

1 0 00 cosβs −sinβs0 sinβs cosβs

.cosα s

sinα seiδ s

0

Figure 2: Reference Land use map for test area showing forested areas in green

Figure 3 :HH L-Band SAR image of test area

Fig. 3 shows the HH L-band image of the scene, with the river complex and urban features clearly associated with the reference map of Fig. 2. Note the significant layover in the mountainous regions and the high backscatter

levels from much of the urban area. Fig. 4 shows a corresponding polarimetric image of the scene. This is a Pauli decomposition image [8], with the RGB channels driven by the coherent sums, HH-VV, HV+VH and HH-VV. We see immediately a rich diversity of scattering behavior compared to the single channel image of Fig. 3. To proceed we now consider two important preprocessing stages, namely image geocoding of level 1.1 ALOS quadpol data and image classification results.

3. DATA PROCESSING AND ANALYSIS

3.1 Geo-coding of PALSAR data

The 1.1 level PALSAR products provided by JAXA are in slant-range radar co-ordinates and must be geocoded by the user. With this in mind, we developed our own Range Doppler (RD) geo-location model based geocoded elevation corrected (GEC) data processing method using JAXA supplied metadata and studied its corresponding accuracy as part of our project. After generating the GEC images of all the data listed in Tab. 1 with the same kind of GEC method, the geo-location performance was validated in two ways.

Figure 4 : Pauli Coded Polarimetric Image (red=hh-vv, green=hv+vh,blue=hh+vv) of test site area

Firstly, each pair of the GEC images was overlapped routinely to check the fitness of image features. It has been found that the three dual-polarization images (Tab.

1) of different dates can be stacked together with a relative positive bias around 1.0 pixel. However, the location bias between image of MAY (the quad-polarization data) and the other three is currently around 203m in East-West direction and 64m in North-South direction. The second way to check accuracy is to compare the GEC image with another sensor image already in map projection (geo-coded). One scene of Landsat ETM+ image extracted from USGS EROS GLOVIS server was used as reference to validate the performance of the GEC method. The Landsat ETM+ image was in UTM map projection with WGS 84 ellipsoid and datum. The PALSAR images were also geo-coded to the same map projection. When using the SEPT image (Tab.1, acquired in Sep. 2007) upon ETM+ image, we found out that the location bias between them is about 200m in East-West direction and 70m in North-South direction.

The absolute geo-location accuracy of the PALSAR level 1.1 products with the developed GEC method cannot be validated because of the lack of ground control points of high accuracy. However we realize the need in the longer term to be able to generate geo-coded terrain correction (GTC) images in order to integrate level 1.1 PALSAR data acquired from different dates, in different descending/ascending orbit with each other or with optical data. Although the absolute geo-location performance of the current geo-location method applied seems unsatisfactory, it is sufficient for driving topography (DEM) based SAR image simulation procedures and supporting further development of programs to generate GTC products.

The error sources for geocoding are not yet clear, but in the near future JAXA are planning improved support for geocoding of their 1.1 products, so enabling easier use of their polarimetric data sets in multi-temporal processing formats. However our results can still be usefully applied for checking image quality and for first stage analysis of Quadpol single image applications and of multi-temporal PALSAR dual-polarization images.

For example, Fig. 4 shows the GTC image of the MAY quad-polarization data in the Pauli-basis representation, taking HH-VV as red, HV+VH as green and HH+VV as blue. The DEM used and the GTC image are all of pixel size 10m*10m. The image coverage shown is only of Culai Mountain. However we can see the increased volume scattering (HV or green) over the mountain areas. We note again however the strong topography effects. To what extent the foreshortening and layover effects the polarimetric segmentation and whether it is possible to correct its effect through only radiometric terrain correction (RTC), as proposed in [16] needs further study. We now turn to consider polarimetric segmentation and classification of these data sets.

3.2 Quad-polarization SAR data classification

The entropy/alpha (H/α) classification [9] was applied to

the multi-look (5 looks in azimuth, 1 look in range) MAY quad-polarization data (Coherency matrix) in original slant range geometric frame.

Figure 5: Quad polarization Pauli composite image following GTC processing

The land terrain type image generated was then geo-coded using the GEC method, and Fig.6 shows the resulted segmentation result. Also shown is the color-coding palette used in the entropy-alpha plane and the distribution of pixels in the scene. Comparing Figs. 6 and 2, it can be found that the dark green class in Fig. 6 corresponds to forest cover in Fig. 2 (with light green color) very well. Water bodies are of yellow color in Fig. 2, most of them can be detected as blue color in Fig. 6, but apparently there are also many other small blue areas which are not water.

Figure 6 : Entropy-Alpha segmentation results, the legend for 8 terrain types, and segment occurrence

histogram

The city of Tai An is located in the center of Fig. 6, just under Tai Mountain. Apparently the city cannot be detected effectively by the segmentation method, only a few pixels of purple color in the entropy alpha diagram can be safely thought of as urban, most of the pixels covered by the city were classified as light green color,

which was confused with low vegetation (crop field and shrubs). Our explanation for this is considered in section 4.1. Most of the forests are located in the mountainous region. It seems the simple segmentation can identify most of the forest covering area effectively through the simple comparison of the entropy-alpha segmentation result with the land use map.

After applying the GTC processing to the entropy-alpha segmentation result, we get the topography corrected segmentation result in map coordinates (UTM) as shown in Fig. 7.

Figure.7: GTC image of the entropy-alpha segmentation result (a) and the layover, shadow region map (b).

Fig. 7 covers most part of the Culai Mountain region, where there is large area of forest distributed continuously. However, Fig.7-(a) shows only some parts of the forest cover area as forest (dark green color), a large fraction of the coverage is missed. If we compare Fig.7-(a) with (b), the layover (red) and shadow (blue) map, it is easy to find out that the lost forests are located in these layover and shadow regions. The topography caused layover and shadowing changes the polarimetric signal, the coherency matrices in these region seem to have lost their original land cover type polarimetric signature. One approach investigated was to mask these regions before or after entropy-alpha segmentation. However we also investigated this behavior further by considering more advanced segmentation and physical decomposition analyses as follows.

4. ADVANCED SEGMENTATION AND IMAGE INTERPRETATION METHODS

We saw in the previous section that the basic entropy/alpha scheme can be used to separate forest from non-

forest but has two main limitations, namely the confusion of other land-use classes and sensitivity to topography effects. To try and overcome the first of these we employed more advanced segmentation techniques based on the inclusion of anisotropy (see equation 7) and image amplitude (or eigenvalue) information. The technique we employed has two major innovations [14]. The first is to employ an adaptive number of clusters instead of the fixed number employed in the H/α method. The second is to employ the span (sum of eigenvalues) amplitude information in the initial classification. The amplitude dynamic range is then segmented initially into 3 classes and for each a classical H/alpha/A segmentation is employed into 16 classes, merging then into a 48-class final image product. We then employ a “cluster merging criterion” based on the Wishart test statistic and a threshold based on a desired probability of false alarm or PFA to reduce this set. The final merged set for the image then comprises 13 classes in a segmented image as shown in Fig. 8.

Figure 8: Results of new Span-Entropy-Alpha-Anisotropy unsupervised segmentation into 13 classes

Here we see much better separation of different land use classes (especially in the city and agricultural areas) but still some sensitivity to topography effects, with amplitude modulations in the mountainous area leading to different segments for the same land use type. To try and combat these, we return to a basic interpretation of the scattering mechanisms present in the coherency matrix.

4.1 Polarimetric Image Interpretation

One important advantage of radar polarimetry over other multidimensional signal processing techniques is the ability to identify structure in the coherency matrix with physical scattering mechanisms in the scene. The ratio of eigenvalues, through entropy H and anisotropy A may be used to identify the most significant scattering contributions as illustrated in Fig. 9. High and low entropy respectively correspond to random and quasi-deterministic scattering. Global scattering with intermediate entropy values are associated to two scattering mechanisms with equal importance or one dominant scattering mechanism perturbed by secondary terms, according to the anisotropy value.

Figure 9: Selection of scattering mechanisms in the entropy/anisotropy plane

Specific identification procedures may then be applied to each of the following three cases discriminated in the H-A plane

- One dominant mechanism : single and double bounce

scattering are then separated by and .

- Two significant mechanisms: a distributed matrix, , is constructed from the first two elements of the eigenvector expansion. The nature of the scattering is determined by comparing its first two Huynen

generators, and [8,9,13,15]

- Three significant mechanisms: the random polarimetric scattering is associated to volume diffusion.

A cluster based estimation of the canonical scattering mechanisms then prevents excessive sensitivity of the classification process to the hard-decision limits with respect to the parameters H, A and α [15]. For example, we can classify the scattering mechanisms based on simple physical classes as shown in Fig.10. Here we show a scattering type segmentation of the test site data. In green we see those regions where depolarizing volume scattering is present, and see good agreement with the forest coverage map of Fig. 2.

Figure 10 :Identification of scattering behaviour from coherency matrix eigenvectors SR = surface reflection, DR = dihedral returns, VD = volume diffuse scattering

In red we show the dihedral scattering components based on the assumption that they are characterized by eigenvectors with α > 45o. We see only a very small segment in this class, despite the presence of extensive urban areas. This we traced to the space-borne geometry of ALOS. The α > 45o criterion was originally developed for airborne sensors that operate at larger angles of incidence (typically around 40 degrees [9]). The ALOS PALSAR polarimetric mode is however restricted to operation at 21.5o [7]. For this steep angle the Brewster angle effect in dihedral scattering can cause the alpha parameter to fall below 45o. For example, Fig. 11 shows the predicted variation of alpha for dihedral scattering over the range of angles of incidence for ALOS-PALSAR Quadpol mode [11]. The x–axis shows variation of the first surface dielectric constant (as defined in the diagram in figure 11) and the y-axis the range of alpha angles (note that because of the steep angle there is only a weak dependence of alpha on variations in the dielectric constant of the second surface reflection). Note that as the dielectric constant increases so alpha increases (giving some potential to estimate dielectric constant from alpha, as discussed in [11]) but for dry materials alpha can be as low as 30 degrees, below the normal threshold used in polarimetric classification techniques. This is supported qualitatively by reference to Fig. 12. Here we show a section of the PALSAR image around the city of Tai An.

Figure 11: Model prediction of the variation of alpha with dielectric constant of the first reflection in dihedral

scattering

To further emphasize the urban components, we employ the polarized decomposition of equations 5 and 6 and image the alpha parameter of the dominant eigenvector (αs in equation 6) and its corresponding amplitude (λ1). To visualize only the polarized parts, we employ the entropy as a saturation variable so that depolarizing areas appear in black and white (the forested mountain at the top of the image for example). We see that the bright scattering elements in the city are mostly green, (alpha around 45 degrees as shown in the lower color template diagram), with a small set of orange/red points (corresponding to high alpha). According to Fig. 11 these green areas can be interpreted as dihedral scattering from non-metallic structures. Without this knowledge, the green regions could be misclassified as dipole (vegetation) scattering in the original entropy/alpha method. Hence for proper interpretation of ALOS-PALSAR Quadpol data a revised segmentation of the entropy/alpha plane is required.

The above image interpretation ideas can be used to explain some of the misclassification features in the Quadpol imagery. However there remains the issue of topography effects in forest classification. To try and counter this, we have developed an algorithm based on the eigenvector approach designed to minimize topography effects in forest studies. The technique employs the use of three polarimetric channels, the first is the alpha parameter for the dominant (not the mean) eigenvector. This helps characterize urban and agricultural areas when used as a color channel in the imagery (we use the same color template as shown in the lower part of Fig. 12). The second is the entropy, again used to control the saturation of this color so that forested areas remain black and white, while other areas, surface and non-forested areas, as well as urban features remain colored. Finally we try and reduce the amplitude modulations due to topography variations by considering only the ‘diffuse’ backscatter component, given formally by λ2+λ3 i.e., the sum of the minor eigenvalues. The idea

is that at L-band, slopes towards the radar will enhance backscatter mainly through the polarized direct surface component and hence by subtracting this from the radar image intensity we reduce such sensitivity.

Figure 12: Polarized Entropy/Alpha/λ1 HSV image of Tai An city using the color coding and saturation levels

shown in lower entropy/alpha plane

Fig. 13 shows an application of this new approach to our test data. Here we see an HSV composite image of these 3 channels. Note the following features:

1) All polarized returns appear black. This tends to make it easier to identify bare surface, open water courses etc. in the imagery. It also provides improved image contrast between forested and non-forested regions.

2) The urban areas and vegetated agricultural regions provide some diffuse scattering components but appear green in color due to their combination of low to moderate entropy and high alpha. This makes it easier to separate them from forested areas.

3) Non-forested mountain terrain is characterized by blue linear features, indicating polarized returns from slopes pointing towards the radar. In shadow they appear black due to loss of return

θ = 25 degrees

θ = 23 degrees

signal. Only for forested terrain is the image bright, and in black and white, allowing for easier image segmentation based on physical properties.

These ideas can be combined for improved identification of the physical structure of Quadpol segmentation algorithms and will form the basis for further improvements and refinements of the techniques, especially for application to space-borne geometries.

Figure 13: Entropy/Alpha/λ2+λ3 HSV PALSAR Diffuse Scattering image of test site

5. POLInSAR USING DUALPOL DATA SETS

Finally we consider some initial results from polarimetric interferometry for our test site. POLInSAR has the potential to provide important additional forest structure parameters, particularly forest height information, and as such has provided an important focus for our DRAGON activities [17]. However, due to limited space-borne data availability, our activities have been confined mainly to airborne and SIR-C data analyses. Further, due to the restricted use of the fully polarimetric (but experimental) mode of ALOS-PALSAR, we were unable to obtain repeat pass fully Quadpol data for our site and currently only have available one baseline pair in FBD or dualpol HH and HV mode with 92 days separation. Image to image co-registration was therefore carried out between JUNE and SEPT images with sub-pixel accuracy. Fig.14

shows the coherence images for the two polarization channels from the InSAR pair of JUNE and SEPT. We note that the coherence image contains complementary information to backscatter, although the temporal effects over forested areas make the coherence of this data set too low for quantitative parameter estimation. The coherence information could however be integrated with intensity images of HH and HV polarization diversity from both dates for improved classification. Such techniques will form the focus for future studies.

6. CONCLUSIONS A major forest test site for advanced radar studies has been established in two mountainous forest regions in China, where high resolution optical images, land use maps and DEMs are available for classification validation. One scene of quad-polarization and three of dual-polarization L-band PALSAR data have then been acquired under the DRAGON project. Custom GEC and GTC processing algorithms have been applied to PALSAR level 1.1 products and their geo-location performance has been validated. Although some problems were found, the performance is good enough for generating initial GTC products based on SAR simulation using a DEM.

Various Entropy-Alpha segmentation techniques were then validated with a land-use map of the test site as ground truth, and it has been found that the entropy/alpha approach can be used to identify forest from other land cover types in general, but there are problems in mountainous regions where layover and shadowing occur. We have subsequently developed more advanced segmentation and image interpretation techniques and used them to highlight several important features of space-borne polarimetric radar data analysis. In particular we have found that a new segmentation of the entropy/alpha diagram is required for space-borne geometries.

Although one pair of dual-polarization data can be registered precisely for interferometric coherence analysis, the coherence image quality is poor, the mean coherence is well below normal values for quantitative InSAR applications (due to temporal decorrelation). How to integrate this kind of coherence information with multi-temporal intensity images for the development of efficient classification and forest structure parameter methods needs to be investigated further. More images will be acquired and ground truth data of forest structure parameters, such as forest canopy height, volume density and above ground biomass will be collected in year 2008.

7. ACKNOWLEDGEMENTS

We would like to acknowledge the support of JAXA for providing the ALOS-PALSAR data used in this project. Thanks also to the MOST-ESA DRAGON program for their support in establishing this collaboration.

(a)

(b)Fig. 14 L-Band Interferometric coherence images for

HH-HH (a) and HV-HV (b) polarizations

REFERENCES

[1] K Jon Ranson , Guoqing Sun. “Mapping Biomass of a Northern Forest Using Multi-frequency SAR Data”. IEEE Trans. Geosci. Remote Sensing, 1994, 32: 388~396

[2] C E Schmullius et al., “SIBERIA – SAR Imaging for Boreal Ecology and Radar Interferometry Applications”, EC-Center for Earth Observation, Project Reports, Contract No. ENV4-CT97-0743-SIBERIA, Final Report, 2001.

[3] M L Imhoff, “Radar backscatter and biomass saturation – ramifications for global biomass inventory”, IEEE Trans. Geosci. Remote Sensing, 1995, 33(2): 511-518

[4] S R Cloude, K P Papathanassiou, “Polarimetric SAR Interferometry”, IEEE Trans. Geosci. Remote Sensing, 1998, 36 (5): 1551-1565

[5] K. P. Papathanassiou and S. Cloude, 2001, “Single-baseline polarimetric SAR interferometry”, IEEE Trans. GRS. 39 (6), 2352-2363, November 2001

[6] K P Papathanassiou, S R Cloude, A Liseno, T Mette, and H Pretzsch, “Forest Height Estimation by means of Polarimetric SAR Interferometry: Actual Status and Perspectives”, Proceedings of 2nd ESA POLInSAR Workshop, Frascati, Italy, January 2005.

[7] A Rosenqvist, M Shimada, N Ito, M Watanabe “ ALOS PALSAR: A Pathfinder Mission for Global-Scale Monitoring of the Environment”, IEEE Trans. GRS 45 (11), pp 3307-3316, November 2007

[8] S. R. Cloude, E. Pottier "A Review of Target Decomposition Theorems in Radar Polarimetry ", IEEE Trans. GRS, vol. 34, n°2, pp 498-518, September 1995.

[9] S. R. Cloude, E. Pottier, "An Entropy Based Classification Scheme for Land Applications of Polarimetric SAR ", IEEE Trans. GRS, vol. 35, n°1, pp 68-78, January 1997.

[10] S H Bickel, R H T Bates, “Effects of Magneto-Ionic Propagation on the Scattering Matrix”, Proc. IEEE, vol. 53 (8), pp 1089-1091, August 1965

[11] S. R. Cloude, Z. S. Zhou, B. Bates, “Temporal Change Analysis of Polarimetric ALOS-PALSAR Data”, Proceedings of 5th International ESA Symposium on Retrieval of Bio-and Geophysical Parameters from SAR Data for Land Applications, Bari, 25-28 September, 2007

[12] L. Ferro-Famil, E. Pottier, J. S. Lee, "Unsupervised classification of multifrequency and fully polarimetric SAR images based on the H/A/Alpha-Wishart classifier", IEEE Trans. GRS, vol. 39, n°11, pp 2332-2342, November 2001.

[13] J.S. Lee, M.R. Grunes, T.L. Ainsworth, L. Du, D.L. Schuler and S.R. Cloude "Unsupervised Classification of Polarimetric SAR Images by Applying Target Decomposition and Complex Wishart Distribution ", IEEE Trans. GRS, vol. 37, no.5, pp2249-2258, Sept. 1999.

[14] E. Pottier, C. Fang , Y. Wang , and H. Wen,” Validation of Forest Area and Forest Gap Mapping Derived from Polsar Data Analysis”, Proceedings of 2007 ESA-MOST DRAGON Symposium, Aix-en-Provence, France, June 2007

[15] L. Ferro-Famil, E. Pottier and J. S. Lee, "Unsupervised classification of natural scenes from polarimetric interferometric SAR data", in Frontiers of Remote Sensing Information Processing, C. H. Chen, Singapore : World Scientific Publishing, 2003, pp. 105-137.

[16] D Small, M Jehle, E Meier, and D Nuesch. “Radiometric Terrain Correction Incorporating Local Antenna Gain”. In Proc. of EUSAR 2004 - 5th European Conference on Synthetic Aperture Radar, Ulm, Germany, 929-932, May 25-27 2004.

[17] M. Neumann, L. Ferro-Famil, and A. Reigber. Multibaseline Polarimetric SAR Interferometry Coherence Optimization. IEEE Geosci. Remote Sensing Lett. , 5(1), January 2008. in press.