«int. j. remote sensing, 1997, vol. 18, no. 6, 1319± 1332 Decomposition of polarimetric synthetic aperture radar backscatter from upland and ¯ ooded ...»
int. j. remote sensing, 1997, vol. 18, no. 6, 1319± 1332
Decomposition of polarimetric synthetic aperture radar backscatter
from upland and ¯ ooded forests
Department of Geography, East Carolina University, Greenville, NC 27858,
F. W. DAVIS
Institute for Computational Earth System Science, University of California,
Santa Barbara, CA 93106, U.S.A.
(Received 22 August 1995; in ® nal form 24 June 1996)
Abstract. The goal of this research was to decompose polarimetric Synthetic Aperture Radar (SAR) imagery of upland and ¯ ooded forests into three backscatter types: single re¯ ection, double re¯ ection, and cross-polarized backscatter.
We used a decomposition method that exploits the covariance matrix of backscatter terms. First we applied this method to SAR imagery of dihedral and trihedral corner re¯ ectors positioned on a smooth, dry lake bed, and veri® ed that it accurately isolated the di erent backscatter types. We then applied the method to decompose multi-frequenc y Jet Propulsion Laboratory ( JPL) airborne SAR (AIRSAR) backscatter from upland and ¯ ooded forests to explain scattering components in SAR imagery from forested surfaces. For upland ponderosa pine forest in California, as SAR wavelength increased from C-band to P-band, scattering with an odd number of re¯ ections decreased and scattering with an even number of re¯ ections increased. There was no obvious trend with wavelength for cross-polarized scattering. For a bald cypress-tupelo ¯ oodplain forest in Georgia, scattering with an odd number of re¯ ections dominated at C-band. Scattering power with an even number of re¯ ections from the ¯ ooded forest was strong at L -band and strongest at P-band. Cross-polarized scattering may not be a major component of total backscatter at all three wavelengths. Various forest structural classes and land cover types were readily distinguishable in the imagery derived by the decomposition method. More importantly, the decomposition method provided a means of unraveling complex interactions between radar signals and vegetated surfaces in terms of scattering mechanisms from targets. The decomposed scattering components were additions to the traditional HH and V V backscatter. One cautionary note: the method was not well suited to targets with low backscatter and a low signal-to-noise ratio.
1. Introduction Recently Cloude ( 1992 ) and van Zyl ( 1994) showed that polarimetric SAR backscatter can be decomposed into three scattering types: a scattering power with an o
matrix into constituent re¯ ection types should simplify the backscatter data and render them in a form where they can be more readily related to surface features.
Cloude (1992 ) and van Zyl ( 1994) laid the theoretical foundation for decomposing a polarimetric SAR image. Here we make an addition to the model put forward by van Zyl ( 1994 ). The addition is the development of criteria to identify the scattering mechanisms of the decomposed components. We then implement the revised model, and decompose SAR imagery of upland and ¯ ooded forested environments. First we do an empirical test of the target decomposition method by applying it to SAR imagery of dihedral and trihedral corner re¯ ectors positioned on a smooth, dry lake bed. Based on the good results obtained there, we proceed with analyses of the JPL AIRSAR imagery of upland ponderosa pine forests in California and ¯ oodplain bald cypress-tupelo forest in Georgia.
2. M ethod of radar target decomposition
2.1. Decomposition of covariance matrix for azimuthly symmetric and reciprocal targets Using measurements of a monostatic polarimetric SAR system (e.g., AIRSAR), one can de® ne the covariance matrix ( U ) for azimuthly symmetric and reciprocal
Detailed expressions for the eigenvalues and the elements of the eigenvectors can be found in van Zyl ( 1994 ). It is easy to show that Ki (i=1, 2, and 3) are unit vectors, and are orthogonal to each other.
Decomposed SAR scattering components in forests
2.2. Identi® cation of scattering mechanisms of decomposed components To identify the scattering mechanisms of the decomposed components from U,
3. Study areas and AIR SAR data To test the decomposition method we analysed two AIRSAR images of Goldstone Lake, California, a dry lake bed 60 km north of Barstow. The study site was selected by scientists from the JPL to test SAR calibration methods, and they deployed a number of calibration devices including trihedral and dihedral corner re¯ ectors. The 1322 Y. Wang and F. W. Davis corner re¯ ectors were arranged on the lake bed, which is a level surface of low re¯ ectance at C-, L -, and P-bands ( Freeman et al. 1990 ). SAR images were acquired on 23 May 1988 and 26 July 1989. We applied the decomposition method to both images with similar results, and so present only the May 1988 imagery here.
The second study site is located at the base of Mt. Shasta, California. Since 1989 this site has been the focus of several SAR studies (e.g., Sun et al. 1991, Wang et al.
1993 ). Topography is level and the soil is a coarse sandy loam with veneer of pine needles and du of 1± 5 cm depth. Land cover is a mosaic of ponderosa pine (Pinus ponderosa ) and ponderosa pine-white ® r (Abies concolor) stands ranging from recent clear-cuts to dense plantations and mid-seral forests with 600± 800 trees haÕ to mature, open pine woodlands with large, scattered trees and densities of less than 30 trees haÕ. For this study we examined ® ve stands that span forest structures from open pine woodland to mid- and late-seral forests of medium to dense stocking (see table 1 for a summary of ® ve ponderosa pine stands). We analysed eight images of the Shasta site that were acquired over the course of the day on 6 September 1989.
The data were processed and calibrated by JPL, as described by Sun et al. ( 1991) and Wang et al. ( 1993 ). Because similar results were obtained from the eight data takes, we con® ne this discussion to one data take at C-, L -, and P-bands.
The ¯ ooded forest site is located in the Altamaha River ¯ oodplain in coastal Georgia ( Hess and Melack, 1994 ). The ¯ oodplain supports emergent freshwater marshes and swamp forest. The ¯ ooded forest is dominated by tupelo gum (Nyssa aquatica ), black gum (N. bi¯ ora), and bald cypress ( Taxodiu m distichum). The understory consists of Fraxinus spp. and Acer rubrum saplings and a variety of shrubs (Myrica cerifera, Lyonia lucida, Clethra alnifolia.) We analysed a single AIRSAR scene acquired on 28 March 1990 during high water conditions when much of the bottomland forest was ¯ ooded to depths of 0´5± 2 m.
4.1. Decomposed backscatter f rom trihedral and dihedral corner re¯ ectors The trihedral corner re¯ ectors deployed on the Goldstone Lake bed were used by JPL in calibrating the imagery and thus do not provide an independent test of the decomposition method. These re¯ ectors were oriented horizontally ( parallel to the soil surface) and pointed towards the AIRSAR, and thus should have produced high SP-O backscatter, and low SP-E and SP-C backscatter at C-band. As predicted, they are bright, conspicuous features in the decomposed image of SP-O backscatter and are not apparent in images of SP-E and SP-C backscatter (® gure 1 ). These results do not validate the decomposition method but do con® rm that our implementation of the decomposition method is performing as expected.
Table 1. Stand characteristics of ® ve ponderosa pine stands, Mt.
Figure 1. Scattering powers decomposed from the JPL AIRSAR C-band data, Goldstone, CA (23 May 1988).
(a) SP-O, (b) SP-E, and (c) SP-C.
The three dihedral corner re¯ ectors that appear in ® gure 1 (b) were not used to calibrate the imagery and thus provide a stronger test of the decomposition method.
Because they were oriented parallel to the ground surface and pointed towards the sensor, one would predict from theory that they should have produced a high SP-E return but contributed little SP-O or SP-C backscatter. The decomposed C-band imagery is certainly consistent with theory (® gure 1 (b)), as are the results at L - and P-bands ( table 2 ). Typically, 70± 90 per cent of total backscatter from each of the dihedral corner re¯ ectors was modelled as SP-E backscatter (table 2).
If a dihedral corner re¯ ector is oriented 45ß o a horizontal ground surface and aimed at the SAR, the re¯ ector should in theory return only cross-polarized backscatter. Three such corner re¯ ectors at the Goldstone site are conspicuous in ® gure 1 (c), again providing another strong evidence that the decomposition method e ectively separates the three scattering types. The re¯ ectors are shown as white spots, pointed by an arrow, in the decomposed C-band cross-polarized scattering image (® gure 1 (c)). This result is another veri® cation of the decomposition method.
Table 3 quanti® es three scattering types, averaged from a 3 by 3 window near the re¯ ectors, of these dihedral corner re¯ ectors; cross-polarized scattering is dominant at C-, L -, and P-bands.
1324 Y. Wang and F. W. Davis Table 2. Percentage of scattering powers of SP-O, SP-E, and SP-C normalized by total power from three dihedral corner re¯ ectors (Dcr) oriented 0ß ( horizontally) on ground surface. AIRSAR data, Goldstone, CA (23 May 1988).
Table 3. Percentage of scattering powers of SP-O, SP-E, and SP-C normalized by total power from three dihedral corner re¯ ectors (Dcr) oriented 45ß from ground surface.
AIRSAR data, Goldstone, CA ( 23 May 1988 ).
4.2. Decomposed backscatter f rom ponderosa pine forest Figure 2 is a C-band total power ( black and white grey scaled) image covering our Mt. Shasta study area. The white shows high backscatter, grey intermediate backscatter, and black low backscatter. The scattered dark areas are open ® elds and clear-cut areas, and the grey areas are ponderosa pine forest. The colour composite of scattering powers decomposed from the SAR data at C-, L -, and P-band is shown in ® gure 3. The colours are coded as red standing for SP-O, green SP-E, and blue SP-C.
Open ® elds and clear-cut areas are bright red showing strong SP-O return. The forested areas are also dominated by red color, showing strong SP-O return. There
Figure 3. Colour composite of scattering powers decomposed from multi-frequenc y AIRSAR data, Mt.
Shasta, CA (6 September 1989 ). SP-O ( Red ), SP-E (Green), and SP-C (Blue).
(a) C-band (top), (b) L -band (middle), and (c) P-band ( bottom).
1326 Y. Wang and F. W. Davis are some green areas (SP-E) and blue areas (SP-C) scattered within the forested areas, but these two types of scatterings are small. These results show that at C-band the SP-O dominates in the ponderosa pine forest (® gure 3 (a)).
At L -band, the whole image may be still dominated by the SP-O or red colour (® gure 3 (b)). However, in the forested areas there is more green at L -band than at C-band. The increase of the green colour or SP-E may be explained as a result of the increase of trunk-ground interactions at L -band. As the wavelength increases to P-band (® gure 3 (c)), there are even more greens within the forests than at L -band;
more trunk-ground interactions within the forests at P-band than at L -band. These results are consistent with those from the analysis of HH± V V phase di erences of the SAR backscatter and modeled backscatter from the ponderosa pine forest ( Wang et al. 1993).
To quantify above analysis, we have selected ® ve ponderosa pine stands where ground data were collected (see table 1 for a summary of the stand data). As the basal area of a stand increases, the total power of the backscatter at C-, L -, and Pbands increases (® gure 4 (a)). The total power at L -band and P-band may be similar, and may be saturated once the basal area is over 54 cm mÕ (table 1).
At C- and L -bands the SP-O increases as the basal area increases. The SP-O at P-band is smaller than that at C-band or L -band for stands with basal areas over 40 cm mÕ (® gure 4 (b)). This may be explained as less canopy scattering at P-band than at C- and L -bands.
As the wavelength increases from C-band to P-band, there is more SP-E in the stands with large basal areas (® gure 4 (c)). This may be a result of more trunk-ground interactions at a long wavelength than at a short wavelength. Also, at P-band, as the stands changes from Sp2 (with a basal area of 13´4 cm mÕ ) to St11 (with a
basal area of 54´3 cm mÕ ), the increase of the SP-E is about 9 dB (® gure 4 (c))
whereas the increase of the total power of backscatter is around 5 dB (® gure 4 (a)).
This enlarged dynamic range of the SAR data may provide a better potential in applications such as the retrieving of the biophysical parameters of forests because the decomposed data may be more sensitive to the change of the parameters.
Because the total cross-polarized scattering derived from the decomposition method is simply twice of H V or VH backscatter (equations ( 1) and ( 6)), no new results are anticipated. We show the plot of the cross-polarized scattering only for the completion of the presentation (® gure 4 (d )).
As noticed in ® gure 4, the scattering power varies not only among stands, but among wavelengths. These variations make it di cult to generalize our ® ndings. We have normalized the powers of the three scattering types by the total power of backscatter at C-, L -, and P-bands, respectively. In general, for the ponderosa pine forest, as SAR wavelength increases from C-band to P-band, ( 1) the normalized SP-O decreases (® gure 5 (a)), ( 2) the normalized SP-E increases (® gure 5 (b)), and ( 3 ) there may be no obvious trend for the normalized SP-C (® gure 5 (c)).
4.3. Decomposed scattering powers f rom ¯ oodplain forest Figure 6, an L -band AIRSAR total power image with a black and white grey scale, shows the Altamaha River ¯ oodplain study area. A black straight-line on the right is a highway. Black curved features are rivers. Grey and light grey areas are ¯ ooded forests. Decomposed powers of three scattering types at C-, L -, and P-bands are shown in ® gure 7 as a colour composite. The colours in ® gure 7 are coded as red for SP-O, green for SP-E, and blue for SP-C.
Decomposed SAR scattering components in forests
Figure 4. Total power (dB) and powers (dB) of three scattering types at C-band (c), L -band (l ), and P-band ( p) from ® ve ponderosa pine stands, Mt.
Red and magenta (red and blue mixed ) dominates decomposed image at C-band (® gure 7 (a)). Some green areas are scattered over the image; these areas are ¯ ooded forests. The curved river channels are noticeable. Parts of river channels are bright red showing strong SP-O. However, there are some green scattered on the river channels showing the existence of the SP-E. A possible explanation is that the backscatter from the channels is small (dark in ® gure 6 ), and may be close to the noise level of the SAR data. The noise a ects the decomposition.
There is more green at L -band (® gure 7 (b)) than at C-band (® gure 7 (a)). Most of the green areas are ¯ ooded forests. Thus, there is a strong SP-E in the ¯ ooded forest. As wavelength increases from C-band to L -band, attenuation from tree canopy becomes less; more microwave energy can penetrate the canopy and hit the ground water surface. Because the ¯ ooded forests are dense and trees in the forests are tall and large ( Hess and Melack 1994 ), and the water surface is a perfect re¯ ection medium, the interactions of double-bounce trunk-ground or the SP-E can be strong.
1328 Y. Wang and F. W. Davis
Figure 5. Percentage of scattering powers of three scattering types normalized by total power at C-band (c), L -band (l ), and P-band ( p) from ® ve ponderosa pine stands, Mt.
Figure 6. L -band total power of AIRSAR image, Altamaha River, GA (28 March 1990 ).