The main thrust of our SIR-C research has been development and testing of coniferous forest radar backscatter models in order to better understand the relationship between forest biophysical characteristics and backscatter. The underlying rationale of this approach is as follows: By understanding in detail how backscatter is controlled by forest parameters such as tree height, branch density, and stem density, and how backscatter varies with changes in moisture conditions and forest phenology, it should be possible to develop model-based or regression-based inversion strategies to estimate forest parameters (particularly biomass) from SAR data. The limitations of SAR-biomass inversion are more clearly understood now than when the SIR-C investigations were planned in the late 1980's (eg., Dobson et al., 1995; Imhoff, 1995a). Yet, the above rationale remains relevant and continues to guide our approach in productive directions.

The Santa Barbara Microwave Canopy Backscatter Model (SBM) originated with Richards et al. (1987) and Sun and Simonett (1988a-b) who developed an early L-HH version of the composite forest backscatter model under the SIR-B program. As will be chronicled in a later section ("Evolution of the Santa Barbara Model"), the SBM has been completely reformulated in several stages during SIR-C (eg., Sun 1990, Sun et al. 1991, Wang 1992, Wang et al. 1993a). Model development continues to the present, together with model validation studies, sensitivity analyses, and related empirical studies. We had originally planned to focus our work on ponderosa pine and mixed coniferous forests near Mt. Shasta in northern California. However, when SIR-C flight plan was finalized, the Shasta site was not given thorough coverage. Therefore, we broadened the scope to include other forests, particularly loblolly pine forest at the Duke University "Supersite." At Duke, we have collaborated with Eric Kasischke and others who have made extensive forest measurements and generously shared their data. In this brief review, we trace the evolution of SBM and our other related SIR-C research from the early Shasta AirSAR period through the present studies of SIR-C data from Duke Forest. Some of the key results are summarized. Our current work and its implications are discussed.

Overview of Santa Barbara microwave canopy backscatter model (SBM)

In its current form, SBM is an integrated suite of models that allow some flexibility in modeling backscatter from different forest types. A continuous canopy version (CC) assumes forest canopy closure of >60%; a discontinuous canopy version (DC) is designed for more open canopies. The two versions can be used interchangeably, even within a single stand. The model is based on the radiative transfer equation. The CC version models canopy, tree trunks, and ground surface as scattering and attenuation layers. Four major components are included: canopy scattering, trunk-ground interactions, canopy-ground interaction, and direct-surface backscattering. The DC version is similar, but the canopy is modeled as a collection discrete ellipsoidal homogeneous scattering volumes (crowns) distributed randomly above the surface. Scattering and attenuation by the crowns are computed on the basis of the probability of a ray intersecting n crowns and of its crown path length.

Trunks are modeled as smooth dielectric cylinders, with some distribution of angle off-zenith. In addition to trunk height and diameter, and canopy depth, the DC model requires crown width. Branches are modeled as cylinders, and up to 3 size classes may be specified with independent dimensional and orientation distributions. Leaves and needles are modeled as disks and thin cylinders. In general, trunk and canopy components are specified by density, orientation, size, and dielectric constant. Inputs may be in the form of data histograms or functions. Several soil surface scattering models from the published literature have been implemented and can be used interchangeably as may be suitable for different surface conditions. Included are the small perturbation and geometric optics models and the recent models of Fung et al. (1992), Oh et al. (1992), and Oh et al. (1994). These typically require surface rms height, correlation length and volumetric soil moisture as inputs.

In use, the SBM is parameterized with stand-level and detailed tree measurements, estimated dielectric constants, soil parameters, and radar parameters (wavelength and incidence angle). Because the forest data is entered as distributions rather than fixed numbers, simulated backscatter for each pixel is stochastic. The mode is run repeatedly to build up a representative sample of pixels, from which backscatter statistics are calculated. Model outputs include pixel values and summary statistics for HH, VV, HV, and total backscatter, HH-VV phase difference, and the HH-VV correlation coefficient. A sequential summary of model developments, testing, and applications, referenced to our publications, is given below.

Study Sites and SAR data

Shasta study area

This site is located near the base of Mt. Shasta, California (4118' N, 12205' W) at about 1200 m elevation. The ground is moderately level. Soils are derived from volcanic material recently deposited in mud-flows. Most of the forest is second growth following logging in the late 1800's and early 1900's. There are also pockets of old growth and some recent plantations. Ponderosa pine (Pinus ponderosa) and white fir (Abies concolor) are the dominant species in the stands studied. Some stands are almost pure ponderosa, but most contain varying admixtures of fir. In most stands the canopy is either partially open or open with densities in the range of 200-800 ha. The largest trees measure 1.0-1.5 m in trunk diameter (dbh) and 40-50 m high. Litter deposits of 5-20 cm are typical. Understory vegetation is sparse and consists primarily of perennial grasses and small shrubs.

Stand data were collected in 1989, 1990, 1991, and 1992 for input to the canopy backscatter model. Stand data include stand density, tree trunk diameter at breast height (dbh), and species composition. Regression equations of tree height on dbh, tree crown depth on dbh, and tree crown width on dbh were derived. In 1989, trunk and branch dielectric constants were sampled at C-, L-, and P-bands using portable dielectric probes. Depth profiles of dielectric constant were made in pines and firs for a range of tree sizes. The pattern of dielectric constant with depth into the bole was different for pines and firs; pines showed a higher peak than firs at the cambium layer, but a lower dielectric constant in the heart. Measurements varied due to probe instability and problems with contact of probe tip and xylem tissues. The data, despite its limitations, permitted reasonable estimates of tree dielectric constant for model parameterization. It was not possible to make separate estimates for boles, branches and needles; for modeling purposes we have assumed them to be the same. Estimated pine tree dielectric constants for C-, L-, and P- bands are (22.5, -j7.5), (20.0, -j5.0), and (20.0, -j4.0). Limited soil dielectric measurements were also made. In 1992, by arrangement with a timber company, we made detailed crown measurements of 6 ponderosa pine trees ranging from 0.43 to 0.88 m dbh immediately after they were felled. Needle size and density, branch size and density, and branch angles were measured. The data were used to parameterize the SBM and to devise a strategy for partitioning branch size segments into classes. (see Sun 1990, Hess and Simonett 1990, Sun et al. 1991, Wang 1992, Wang et al. 1993a-b for further details on the Shasta site and data.)

Duke study area

The forest stands are in Duke Forest, Durham, North Carolina (3600' N, and 7900' W). Duke Forest includes approximately 3400 ha of forest stands, one quarter of which are pure stands of loblolly pine (Pinus taeda L.). These pine stands range in age from < 1 to about 100 years, and are typical of the successional chronosequence found in this region (Christensen and Peet, 1981). The forest is a long-standing experimental forest, well mapped and much studied.

Recent stand data and detailed tree data have been collected by Kasischke and co-workers (Kasischke 1992, Kasischke et al. 1994). The stand surveys were updated in 1994 to 1996 (Kasischke, pers. comm.). The stands used in SAR studies have stand densities varying from 120 to over 4000 trees/hectare. Mean dbh ranges from 5 to >50 cm, and height from <5 to >40 m. Canopy coverage is high; the canopy may be considered continuous. Tree dbh, and weights of bole, branches, and needles were measured for trees of different sizes. These data were used to develop allometric equations to estimate the biomass of tree trunks, branches and needles from the dbh measurements. Based on these estimates and measured stand dbh distributions and densities, stand biomass was estimated (Kasischke 1992). In addition, Kasischke (1992) estimated branch and needle characteristics including distribution of stem diameters, lengths, and branching angles.

During the SIR-C mission in April, 1994, Kasischke's group sampled volumetric soil moisture in young loblolly stands. A team from Santa Barbara sampled soil moisture and soil dielectric constant at L-band in loblolly stands ranging from 8 to 90 years old. We also measured litter depth, litter moisture content and soil surface roughness in 6 stands. Limited tree dielectric measurements were also made. During the October SIR-C mission, Kasischke's group revisited the stands sampled in April to resample soil moisture.

Kasischke has generously shared his data, enabling us to parameterize the SBM for the loblolly forest (Wang et al. 1994b, Wang et al. 1995, Day et al. 1996)

JPL AirSAR data

AirSAR data was acquired over the Shasta site in September 1989, March, 1990, and May, 1991. The 1989 conditions were very dry, and conditions were constant around the date of the overflights. Nine data takes were acquired at varying incidence angle. The data was calibrated with corner reflectors and active calibrators to an estimated accuracy of 1.0 dB. These data have been used extensively in our model development and validation (Sun 1990, Sun et al. 1991, Wang 1992, Wang et al.1993a-b, Day 1993, Saleta 1995, Wang and Dong 1997, Wang and Davis 1997). The data from March, 1990 are problematic. Snow cover was too deep to allow deployment of enough corner reflectors for accurate calibration. During the 24 hours preceding the flight, the weather warmed suddenly and the snow melted rapidly. These SAR data have not been used because of the complex, difficult to characterize environmental state, and the lack of calibration. The five data takes from May, 1991 were also acquired under rapidly changing conditions. Dry weather changed to rain and then snow immediately before the flight. By flight time the trees were coated with a layer of wet snow and the ground was snow-covered. Two days later, two additional AirSAR passes were made. The ground was wet, and trees were mostly snow-free. The scenes have been calibrated on the basis of five 8-foot trihedral corner reflectors deployed across the study area. These data were used by Saleta (1995).

A description of the AirSAR data acquired over the Duke Forest can be found in Kasischke (1992). The AirSAR data have been used in our model validation study for loblolly pine forest stands, and in the feasibility study of retrieving the forest biomass from the SAR data (Wang et al. 1995).

SIR-C/X-SAR data

The SIR-C/X-SAR missions in 1994 acquired 10 data takes of Duke Forest at C- and L-bands. Four were from the April flight (SRL-1) and six from the October flight (SRL-2). X-band data were also acquired. We have received calibrated, single-look slant-range data from JPL for all passes. Incidence angles are 19-32 for April and 23-54 for October. Calibration uncertainty is discussed in Freeman et al. (1995). Nominal resolution varies from 6.2-8.2 m in azimuth and 11.4-24.4 m in range. All data were acquired at about dawn or within two hours before. During the April mission there was intermittent rain at Duke Forest, and the soil varied from moist to saturated. During the October mission conditions were dry, since no rain had fallen for two weeks, and only 2.5 cm had fallen in the month preceding the flight. (Light rain fell immediately before the final data take.) Forest conditions changed rapidly over the 10 days of the April mission, visibly transforming the forest from brown to green with the leafing out of deciduous species. October conditions were stable, as the flight took place before the autumn leaf-fall.

Evolution of the Santa Barbara Model

The following is a chronological outline of the development, validation, and applications of SBM, highlighting some of the more important results.

Sun (1990); Sun, Simonett, and Strahler (1991)

• Extended the L-HH backscatter model to C-, L-, and P-bands and reformulated it as polarimetric. The model is composite, comprising 4 scattering components: canopy volume, ground surface, trunk-ground, and crown-ground.
• Developed the discontinuous canopy model based on geometric-optical canopy modeling (Li and Strahler, 1988).
• Analyzed sensitivity of backscatter components to many parameters (soil dielectric constant, incidence angle, branch density, stem density).
• Validated the model using Shasta AirSAR data.

The model made reasonable estimates of ⁰ at L- and P-bands, but overestimated C-band. L- and P-band backscatter were dominated by the trunk-ground term for the Shasta forests, even for dry soil conditions. ⁰ in all bands increased asymptotically with increasing tree height and stand density. L-HV increased almost linearly with branch density (i.e., saturation did not occur). L- and P-band dynamic ranges were less than 3 dB over the wide range of tree sizes and stand densities modeled.

Wang (1992); Wang, Day, and Sun (1993a)

• Refined the discontinuous canopy backscatter model with explicit conditional probability formulation of scattering and attenuation pathways.
• Significantly improved model implementation and documentation.
• Added computation of phase difference and correlation coefficient.
• Validated the model at L-band (AirSAR) for very sparse stands at Shasta.

Surface, canopy, and trunk-ground backscatter terms all contributed significantly to modeled co-polarized backscatter (the proportions depending in incidence angle), while the canopy term dominated modeled cross-polarized backscatter. Co-polarized backscatter decreased with increasing ₀. SAR-model agreement was generally good at L-HH and L-VV, but the model slightly underestimated HH at shallow incidence. At L-HV, ⁰ was almost constant with ₀. The model underestimated L-HV. This may be attributable to use of the first order small perturbation model (which predicts no cross-polarized backscatter) for ground surface modeling. The lack of a cross-polarized surface component may have a large effect on total HV backscatter because the test stands are very sparse and the ground surface exposed.

Wang and Imhoff (1992)

• The L-HH continuous canopy model was adapted for mangrove forests Bangladesh.
• Modeled backscatter and SIR-B data were compared under flooded and non-flooded surface conditions.

Flooding increased the L-HH backscatter by 1.6-3.8 dB depending on forest stand and data take. Simulations indicate that canopy volume scattering strongly dominates backscatter for non-flooded conditions, and that the trunk-ground term is co-dominant for flooded conditions. Since trunk-ground backscatter under flooded conditions is most strongly enhanced at ₀<35, small incidence angles should be best suited for mangrove flood boundary delineation.

Wang (1992)

• Ran extensive sensitivity analyses based on detailed tree measurements made at the Shasta site (branch diameter, branch length, branch density, branch angle, needle density, crown width, crown depth, dbh, stand density).
• Partitioned branches into multiple size classes.

L-band model sensitivity analysis showed pine needles and small branches were weak scatterers and weak absorbers. Even for dry soil conditions and average tree water status, the trunk-ground term dominated when crown volume was small. As tree crown volume increased, the importance of the trunk-ground term diminished with increases in crown attenuation and crown volume scattering. The effects of changing crown volume were more pronounced than those resulting from large changes in branch size and shape, and branch and needle densities.

Wang, Davis, and Melack (1993b)

• Modeled backscatter, correlation coefficient, and phase difference at C-, L-, and P-bands over a range of incidence angles for 3 discontinuous stands at Shasta.
• Compared model predictions to AirSAR data and interpreted the observed backscatter, phase difference, and correlation coefficient based on modeled scattering components.

Model predictions of ⁰ agreed with observed backscatter within calibration uncertainty for most bands and stands. Observed backscatter was nearly constant with incidence angle for all wavelengths and polarizations except for a decrease at VV polarization at shallow ₀ (above 40). For relatively dense stands of ponderosa pine (but not for a very sparse stand) the trunk-ground term dominated modeled ⁰ at HH and was important at VV at L- and P-bands. Crown volume scattering dominated at C-band. Double-bounce trunk-ground backscatter was detected in P-band SAR data by the mode of the HH-VV phase difference approaching 180. However, at L-band the canopy and double-bounce terms were comparable at VV, and as a result the mode of the phase difference was close to 0. The ratio of HH/VV backscatter is also an indicator of double-bounce backscatter because double-bounce scattering is greater at HH than VV. This effect was demonstrated for both model and SAR results.

Wang, Day, Davis, and Melack (1993d)

• Modeled Alaska boreal forest under frozen and thawed conditions.
• Compared model with AirSAR at L-band.

Modeled backscatter was significantly less under frozen (low dielectric constant) conditions than thawed conditions. This supports the prevailing interpretation of the observed decrease in SAR backscatter under frozen conditions. The simulations underestimated backscatter for frozen conditions, as compared to AirSAR. The frozen/thawed ⁰ ratio was thereby overestimated. There is evidence indicating inadequate ground surface modeling is one cause of the discrepancy.

Wang, Kasischke, Melack, Davis, and Christensen (1994)

• Applied the continuous canopy model to loblolly pine forest stands.
• Analyzed effects of soil moisture on regressions to retrieve biomass from ERS-1 SAR backscatter.
• Simulated effects of soil moisture and surface roughness on C-VV backscatter components for young pine stands (biomass 0.4-5.3 kgm^-2).

Under wet soil conditions, backscatter was not significantly correlated to biomass; for dry soil, moderate correlation was observed. Even for dry soil, the biomass signal at C-VV saturated at very low biomass levels (<1.5 kg m^-2). At such low biomass levels, modeled backscatter is strongly influenced by surface roughness and moisture; increases in either factor lead to increases in the surface backscatter component which may equal or exceed the canopy component. Saturation at low biomass levels and sensitivity to surface conditions limit the usefulness of ERS-1 SAR for measuring forest biomass.

Wang, Davis, Melack, Kasischke, and Christensen (1995)

• Compared AirSAR and modeled backscatter at C-, L-, and P- bands for loblolly pine stands in the range of 0-15 kg m^-2
• Modeled sensitivity of backscatter to biomass.

Backscatter at all polarizations at C-band, and at VV for L- and P-bands, was insensitive to biomass differences. L-HH, L-HV, P-HH and P-HV increased more than 5 dB as biomass increased from 0 to 15 kgm^-2. Most of the increases were for biomass <5 kgm^-2. L-HV and P-HV appeared to saturate at about 5 and 14 kgm^-2 respectively. Combinations of L-HH + P-HH or L-HV + P-HV might be useful for biomass retrieval, though the HH combination would be more affected by surface conditions.

Wang, Hess, Filoso, and Melack (1995b)

• Applied the continuous canopy model to Amazon forests to investigate potential for SAR detection of flooding beneath closed canopy.
• Modeled sensitivity of C-, L-, and P-band backscatter to soil moisture, flooding, surface roughness and leaf area index.

The ratio of ⁰(flooded)/⁰(non-flooded) is higher at HH than at VV polarization and at steep incidence angles than at shallow ones. Therefore, flooding should be most detectable at HH and steep incidence. As the surface soil moisture beneath the nonflooded forest is increased from 10% to 50% volumetric moisture, the flooded/nonflooded ⁰ ratio decreases; the decreases are small at C- and L-band but large at P-band. When the leaf size is comparable to or larger than the wavelength of C-band, the leaf area index (LAI) has a large effect on the simulated C-band (not L-band or P-band) backscatter from both flooded and nonflooded forests.

Wang and Dong (1997)

• Estimated stand density and mean tree dbh from Shasta AirSAR data using a neural network trained by the discontinuous canopy backscatter model.

A backscatter model-based neural network (NN) was used to retrieve two important stand parameters, density and dbh, from SAR data. In tests on two ponderosa pine stands, relative errors of estimated mean dbh ranged from 4.3-15.6% for the two stands and seven SAR wavelength combinations tested. Relative errors of estimated stand density ranged from 7.1-23%. Unexpectedly, the NN's employing multiple SAR frequencies in many cases produced larger errors than single-frequency NN's; for each stand and inverted parameter the best estimate was made by a single-frequency NN. The method is a promising inversion option that deserves further study. Success with this method depends entirely on having available a well specified backscatter model to train the NN.

Wang, Paris, and Davis (1996, in revision)

• Developed a semi-empirical multiple scattering model and integrated it into both the continuous and discontinuous versions of SBM.
• Modeled C- and L-bands for Shasta and Duke Forest stands.
• Compared modeled loblolly pine backscatter with SIR-C data.

The model assumes that multiple scattering increases absolute backscatter of all 4 polarization combinations equally. Therefore, addition of the multiple scattering model resulted in a proportionally greater increase in cross-polarized than co-polarized backscatter. Accordingly, cross-polarized backscatter expressed in dB increased substantially and co-polarized backscatter only slightly with the addition of the model. The enhancement was greatest at short wavelengths (C-band) where canopy scattering dominates. Accuracy of model predictions of SIR-C backscatter at C-HV was significantly improved by addition of the multiple scattering model.

Day, Wang, and Davis (1996, in revision)

• Modeled sensitivity of backscatter to soil moisture, surface roughness and correlation length for three size/density classes of loblolly pine forest at C- and L-bands, 3 polarizations, over a range of incidence angles.
• Investigated which SAR bands and incidence angles are sensitive and which are insensitive to the expected range of variability of surface conditions under existing canopies.
• Compared model predictions with SIR-C data for wet and dry surface conditions.

Within the roughness-moisture parameter region modeled, L-band was sensitive primarily to soil moisture. C-band was sensitive to both roughness and moisture. Of the wavelength-polarization combinations considered, L-HH was the most sensitive to surface conditions, followed by L-VV and C-HH. C-VV had intermediate sensitivity for the forest stand class with lowest biomass and was insensitive for the other classes. C-HV and L-HV were insensitive except for the steepest incidence angle (₀ = 20) and lowest biomass stand class tested, where they were slightly sensitive. A comparison of SIR-C backscatter under moist conditions (April, 1994) and dry conditions (October, 1994) failed to confirm model predictions. We believe this may be the result of unknown (and unmodeled) phenologically-caused changes in canopy scattering and extinction, compounded by calibration uncertainties.

Other Related Studies

Day (1993), Day (1997, in preparation)

• Developed a qualitative "range-bin" model to visualize the interaction of SAR pixel geometry with forest gaps.
• Compared backscatter and phase difference of gap-containing and gap-free coniferous forest using Shasta AirSAR data.
• Investigated effects of gaps on the statistical distribution of backscatter.

Response to gaps was greatest at C-VV at shallow incidence, registering as a drop in backscatter at the low end of the ⁰ distribution. However, the L-VV gap signal was more robust than C-VV due to the lower variance at L-band. An offset of gap location 1-3 pixels toward the near-range was observed at C-band and shallow incidence; this may be a canopy "layover" effect.

Saleta (1995)

• Applied discriminant analysis and a decision tree classifier to SAR data to discriminate and classify coniferous forest stands according to the canopy cover and crown size classes used by the U.S. Forest Service.
• Analyzed backscatter, phase difference, band ratios, and derived indices from Shasta AirSAR data acquired under different surface conditions.

Misclassification rates were high (53-61%), suggesting SAR may be inappropriate for discriminating the U.S.F.S. forest management classes in this forest. By aggregating the stands into 3 coarse timber volume classes, misclassification rates were reduced to 23-28%.

Wang and Davis (1997, in press)

• Verified a radar target decomposition method (based on Cloude, 1992; van Zyl, 1994) using AirSAR calibration target data.
• Tested the method on AirSAR data from Shasta forest and Georgia floodplain forest.

As wavelength increased from C-band to P-band, scattering with an odd number of reflections decreased and even number of reflections increased. The method was found useful for unscrambling forest scattering mechanisms and for discriminating forest structural types and cover classes.

Current Research and its Implications

Our current research focuses on three questions:

I Can we explain the seasonal changes in SIR-C backscatter from the Duke loblolly forest?

We have preprocessed sub-images of the Duke study area from all the SIR-C scenes into HH, HV, and VV backscatter images. Backscatter for many loblolly stands has been extracted using a digital stand map in ArcInfo (provided by L.L Bourgeau-Chavez at ERIM). We have made careful comparisons of backscatter from 15 loblolly stands from two April-October pairs of same-incidence data takes. We compared the observed seasonal ⁰ changes to model predictions based on differences between April and October soil moistures and pine needle densities. We have not yet been able to satisfactorily explain observed increases in ⁰ (particularly L-VV and L-HV) from the wet April to the dry October conditions (Day et al., 1996 in revision). Possible explanations include changes in forest phenologic state or canopy water distribution, and calibration error.

The task of analyzing the full SAR data set is in progress. The SIR-C data offers a good opportunity for comparing effects of radar incidence angle, seasonal change, and soil moisture change, but the effects must be unscrambled. A recurring problem is the variability of the loblolly stand backscatter. The variability is attributable in part to the small size of the Duke stands and in part to stand inhomogeneity characteristic of loblolly stands. To improve estimates of backscatter and its variance, more sample stands are needed. Thus, one current task is to improve accuracy of SIR-C registration to the Duke Forest stand maps so that a larger number of stands can be accurately and efficiently sampled and so that marginally small stands can be utilized. We also plan to develop an accurate empirical model of the relation of incidence angle and ⁰, relying mainly on the October data (which was acquired under constant environmental conditions). There is an obvious increase in backscatter with decreasing ₀ that can be roughly corrected by dividing ⁰ by cos₀ (i.e., using ), but the relation appears to vary with wavelength, polarization, and possibly stand maturity in addition to incidence angle. An improved incidence angle model should allow us to adjust backscatter data at different incidence angles so that comparisons among data takes can be freely made, and soil moisture and seasonal trends revealed. Additional sensitivity analyses are being planned to shed more light on the Duke SIR-C results. In particular, the effects of varying branch and needle dielectric constant will be studied and compared to observed seasonal backscatter changes. We anticipate that these studies will improve our understanding of the causes of stand-to-stand, day-to-day, and season-to-season backscatter variability of Duke Forest, and lead to two or more additional journal articles.

II How does backscatter uncertainty caused by stand structural variability, calibration error, and forest phenologic change affect biomass estimates?

One of the best-learned lessons to come from our work with AirSAR and SIR-C data is that backscatter varies among stands that have apparently similar biophysical statistics, and varies in the same stand for different data acquisitions under environmental conditions that appear similar. Dobson et al.(1995) and Imhoff (1995a) have shown that accurate biomass retrieval by regression can only be accomplished within uniform forest structural types. In many forests, including Shasta and Duke forests, structure varies more or less continuously across the landscape due to gradients and discontinuities in site history and diverse site characteristics at many spatial scales. In regional biomass surveys it will not be feasible to develop separate biomass regressions for each structurally variant forest patch. (Model inversion approaches, including neural networks, face an equivalent limitation.) Inevitably, backscatter varies across the forest, and algorithms to retrieve biophysical parameters must also vary. In our experience, backscatter for similar forest stands usually varies at least 1-2 dB (often more), while backscatter across the full range of stand maturities and biomass (excluding the lowest biomass stands of seedlings and saplings) may be only 2-3 dB (eg., Harrell et al., 1997; Day et al., 1996, in revision).

Temporal backscatter variability is also large in comparison to the biomass "signal". SIR-C calibration uncertainty is estimated at 2.2 dB at C-Band and 1.3 dB at L-band (Freeman et al., 1995). To this must be added a stand imaging uncertainty term to account for radar fading and sampling variation in the non-uniform forest medium. (Imaging uncertainty is minimized in large, uniform stands.) These errors are difficult to separate from actual changes in forest scattering properties arising from changes in canopy water distribution, soil moisture content, and forest phenologic state, whether known or unknown. For example, in the Duke SIR-C data we observe an average increase in ⁰ (L-HV) of 1.7 dB from April to October. It is difficult to determine what fraction of the change reflects calibration error (if any) and what fraction is from unknown canopy changes. (We rule out soil moisture change as a probable cause on the basis of surface modeling.) In the absence of long-term canopy dielectric constant and scatterometry studies far more extensive than any attempted to date, it may not be possible to to understand changes in forest backscatter with the accuracy required for practical biomass retrieval.

Because spatial and temporal uncertainties in backscatter both appear large in comparison to the dynamic range of the forest biomass "signal", it is important to access the effect of backscatter error on biomass estimates. We plan to analyze how ⁰ uncertainty translates into biomass uncertainty for several published biomass retrieval algorithms.

III Given the current knowledge about forest backscatter, what are the realistic options for SAR biomass sensing?

The uncertainty issues outlined above lead us to wonder if more productive alternative strategies can be devised to measure forest biomass or monitor its change with SAR. Synoptic surveys will not be feasible if they require large homogeneous forest stands and highly structure-specific biomass regressions or exacting canopy modeling. With current approaches these requirements seem necessary in order to overcome biomass saturation at 5-10 kgm^-2, a level that limits biomass survey to a small percentage of the world's forests (see Imhoff, 1995b). We plan to compare possible alternative SAR forest monitoring strategies, at least conceptually, and to demonstrate them with examples if possible.

Publications

Sun, G., D. S. Simonett, and A. H. Strahler, 1989, A radar backscattering model for discontinuous forest canopies, Proceedings of IGARSS'89, vol. 4, pp. 2832-2835.

Sun, G., 1990, Radar Backscattering Modeling of Coniferous Forest Canopies, Ph.D. dissertation, The University of California at Santa Barbara.

Sun, G. and D. S. Simonett, 1990, Polarimetric radar backscatter modeling for discontinuous canopies, Proceedings of IGARSS'90, vol. 1, pp. 483-486.

Hess, L. and D. S. Simonett, 1990, Dielectric properties of two western coniferous tree species, Proceedings of IGARSS'90, vol. 1, pp. 503.

Sun, G., D. S. Simonett, and A. H. Strahler, 1991, A radar backscattering model for discontinuous coniferous forests, IEEE Trans. on Geosci. and Remote Sens., vol. 29, pp. 639-650.

Day, J. and F. W. Davis, 1992, SAR backscatter from coniferous forest gaps, Summaries of the Third Annual JPL Airborne Geoscience Workshop, JPL Publication 92-14, vol. 3, pp. 12-14.

Wang, Y., 1992, Radar backscatter canopy modeling and applications in forested environments, Ph.D. dissertation, The University of California at Santa Barbara.

Wang, Y., F. W. Davis, and J. M. Melack, 1992, Modeled response of L-band radar backscatter from conifer woodland to changes in tree canopy volume, Proceedings of IGARSS'92, vol. 1, pp. 776-778.

Wang, Y., F. W. Davis, and J. M. Melack, 1992b, Comparison of modeled backscatter with SAR data at P-band, Summaries of the Third Annual JPL Airborne Geoscience Workshop, JPL Publication 92-14, vol. 3, pp. 9-11.

Day, J. L., 1993, Imaging coniferous forest gaps with synthetic aperture radar, M.A. thesis, The University of California at Santa Barbara.

Saleta, J., F. W. Davis, and J. L. Day, 1993, Stand discrimination in a western coniferous forest using multifrequency polarimetric SAR data, Proceedings of PIERS'93, pp. 863.

Wang, Y. and M. L. Imhoff, 1993, Simulated and observed L-HH radar backscatter from tropical mangrove forests, Int. J. Remote Sens.,vol. 14, no. 15, pp. 2819-2828.

Wang, Y., J. L. Day, and G. Sun, 1993a, Santa Barbara microwave backscattering model for woodlands, Int. J. Remote Sens., vol. 14, no. 8, pp. 1477-1493.

Wang, Y., F. W. Davis, J. M. Melack, 1993b, Simulated and observed backscatter at P-, L-, and C- bands from ponderosa pine stands, IEEE Trans. on Geosci. and Remote Sens., vol. 31, no. 4, pp. 871-879.

Wang, Y., F. W. Davis, J. M. Melack, E. S. Kasischke, and N. L. Christensen, Jr., 1993c, Relating P-band AirSAR backscatter to forest stand parameters, Summaries of the Fourth Annual JPL Airborne Geosci. Workshop, JPL Publication 93-26, vol. 3, pp. 81-84.

Wang, Y., J. L. Day, F. W. Davis, and J. M. Melack, 1993d, Modeling L-band radar backscatter of Alaskan boreal forest, IEEE Trans. on Geosci. and Remote Sens., vol. 31, no. 6, pp. 1146-1154.

Urs, W., F. Holecz, Y. Wang, and G. Kattenborn, 1994, Theoretical sensitivity of ERS-1 SAR backscatter over forest, Proceedings of IGARSS'94, pp. 2477-2479.

Wang, Y., E. S. Kasischke, J. M. Melack, F. W. Davis, and N. L. Christensen, Jr., 1994, The effects of changes in loblolly pine biomass and soil moisture on ERS-1 SAR backscatter, Remote Sens. of Environment, vol. 49, no. 1, pp. 25-31.

Wang, Y., F. W. Davis, and E. S. Kasischke, 1994b, Effects of variation in soil moisture on ERS-1 SAR backscatter, Proceedings of IGARSS'94, pp. 1475-1477.

Wang, Y., L. L. Hess, and S. Filoso, and J. M. Melack, 1994c, Canopy penetration studies: modeled radar backscatter from amazon floodplain forests at C-, L-, and P-band, Proceedings of IGARSS'94, pp. 1060-1062.

Holecz, F., U. Wegmuller, E. Rignot, and Y. Wang, 1995, Observed and predicted radar backscatter from forested areas with terrain variations, Proceedings of IGARSS'95, pp. 613-615.

Saleta, J. L., 1995, Stand Discrimination in a Western Coniferous Forest Using Airsar Data, M.A. thesis, The University of California at Santa Barbara.

Wang, Y., F. W. Davis, J. M. Melack, E. S. Kasischke, and N. L. Christensen, Jr., 1995, The effects of changes in forest biomass on radar backscatter from tree canopies, Int. J. Remote Sens., vol. 16, no. 3, pp. 503-513.

Wang, Y., L. L. Hess, S. Filoso, and J. M. Melack, 1995b, Understanding the radar backscatter from flooded and nonflooded Amazonian forests: results from canopy backscatter modeling, Remote Sens. of Environment, vol. 54, no. 3, pp. 324-332.

Wang, Y., and F. W.. Davis, 1996, Radar backscatter components from ponderosa pine forests, Proceedings of IGARSS'96, pp. 1077-1079.

Wang, Y., J. L. Day, and F. W. Davis, 1996, Sensitivity of modeled C-band backscatter from loblolly pine forests to surface soil roughness and moisture, Proceedings of IGARSS'96, pp. 1083-1085.

Wang, Y., and D. Dong, 1997, Retrieving forest stand parameters from SAR backscatter data using a neural network trained by a canopy backscatter model, Int. J. Remote Sens., vol. 18, pp. 981-989.

Wang, Y., and F. W. Davis, 1997, Decomposition of polarimetric synthetic aperture radar backscatter from upland and flooded forests, Int. J. Remote Sens. (in press).

Manuscripts in Review/Revision

Wang, Y., J. F. Paris, and F. W. Davis, 1996, Inclusion of a simple multiscattering model into a microwave canopy backscatter model, Remote Sens. of Environment.

Day, J. L., Y. Wang, and F. W. Davis, 1996, Sensitivity of C- and L-band SAR backscatter to soil surface roughness and soil moisture content in loblolly pine forest, Int. J. Remote Sens.

Other References Cited

Christensen, Jr., N. L. and R. K. Peet, 1981, Secondary forest succession on the North Carolina Piedmont, in Forest Succession: Concepts and Applications (D. C. West, H. H. Shugart, and D. B. Botkin, Ed), Springer-Verlag, New York, New York, pp. 230-245.

Cloude, S. R., 1992, Uniqueness of target decomposition theorems in radar polarimetry, in Direct and Inverse Methods in Radar Polarimetry, Part I, (ed. by W. M. Boerner et al.), Kluwer Academic Publishers, Dordrecht, The Netherlands, pp. 267-296.

Dobson, M.C., F.T. Ulaby, L.E. Pierce, T.L. Sharik, K.M. Bergen, J. Kellndorfer, J.R. Kendra, E. Li, Y.C. Lin, A. Nashashibi, K. Sarabandi, and P. Siqueira, 1995, Estimation of forest biophysical characteristics in northern Michigan with SIR-C/X-SAR, IEEE Trans. on Geosci. and Remote Sens., vol. 33, no. 4, pp. 877-895.

Freeman, A., M. Alves, B. Chapman, J. Cruz, Y. Kim, S. Shaffer, J. Sun, E. Turner, and K. Sarabandi, 1995, SIR-C data quality and calibration results, IEEE Trans. on Geosci. and Remote Sens., vol. 33, no. 4, pp. 848-857.

Fung, A. K., Z. Li, and K. S. Chen, 1992, Backscattering from a randomly rough dielectric surface, IEEE Trans. on Geosci. and Remote Sens., vol. 30, no. 2, pp. 356-369.

Harrell, P.A., E.S. Kasischke, L.L. Bourgeau-Chavez, E.M. Haney, and N.L. Christensen, Jr., 1997, Evaluation of approaches to estimate aboveground biomass in southern pine forests using SIR-C data, Remote Sens. of Environment (in press).

Imhoff, M. L., 1995a, A theoretical analysis of the effect of forest structure on synthetic aperture radar backscatter and the remote sensing of biomass, IEEE Trans. on Geosci. and Remote Sens., vol. 33, no. 2, pp. 341-352.

Imhoff, M. L., 1995b, Radar backscatter and biomass saturation: Ramifications for global biomass inventory, IEEE Trans. on Geosci. and Remote Sens.,vol. 33, no. 2, pp. 511-518.

Kasischke, E. S., 1992, Monitoring changes in aboveground biomass in loblolly pine forests using multi-channel synthetic aperture radar data, Ph.D. dissertation, The University of Michigan, Ann Arbor.

Kasischke, E. S., N. L. Christensen, Jr., and E. Haney, 1994, Modeling of geometric properties of loblolly pine tree and stand characteristics for use in radar backscatter models, IEEE Trans. on Geosci. and Remote Sens., vol. 32, no. 4, pp. 800-822.

Li, X. and A. H. Strahler, 1988, Modeling the gap probability of a discontinuous vegetation canopy, IEEE Trans. on Geosci. and Remote Sens., vol. 26, no. 2, pp. 161-170.

Oh, Y., K. Sarabandi, and F. T. Ulaby, 1992, An empirical model and an inversion technique for radar scattering from bare soil surface, IEEE Trans. on Geosci. and Remote Sens., vol. 30, no. 2, pp. 370-381.

Oh, Y., K. Sarabandi, and F. T. Ulaby, 1994, An inversion algorithm for retrieving soil moisture and surface roughness from polarimetric radar observation, Proceedings of IGARSS'94, vol. 3, pp. 1582-1584.

Richards, J. A., G. Sun, and D.S. Simonett, 1987, L-band radar backscatter modeling of forest stands, IEEE Trans. on Geosci. and Remote Sens., vol. 25, no. 4, pp. 487-498.

Sun, G. and D.S. Simonett, 1988a, A composite L-band radar backscattering model for coniferous forest stands, Photogrammetric Engineering and Remote Sens., vol. 54, no. 8, pp. 1195-1201.

Sun, G. and D.S. Simonett, 1988b, Simulation of L-band and HH radar backscattering for coniferous forest stands: a comparison with SIR-B data, Int. J. Remote Sens., vol. 9, no. 5, pp. 907-925.

van Zyl, J. J., 1994, Application of Cloude's target decomposition theorem to polarimetric imaging radar data, Proceedings of the Third Spaceborne Imaging Radar Symposium, held at the Jet Propulsion Laboratory, Pasadena, California, Jan. 18-21, 1994, pp. 207-216.