Structure-based forest biomass from fusion of radar and hyperspectral observations Robert N. Treuhaft Jet Propulsion Laboratory, California Institute of Technology, USA Gregory P. Asner Department of Global Ecology, Carnegie Institution of Washington, Stanford University, USA Beverly E. Law College of Forestry, Oregon State University, USA Received 30 December 2002 revised 21 February 2003 accepted 5 March 2003 published 8 May 2003. [1] Forest biomass was estimated from the remotely sensed profiles of leaf area density. Biomass estimated from forest structure profiles may be more accurate than that determined from microwave power or optical radiance measurements. Multialtitude, airborne, C-band, radar interferometry produced relative density profiles, which were normalized by leaf area indices from airborne hyperspectral optical imagery, yielding the forest canopy leaf area density for 11 structurally diverse stands in Central Oregon. Fits of field biomass measurements to model functions of remotely sensed leaf area density produced agreement between the field and remotely sensed biomasses at the level of 25 tons/ha, or 16% of the average stand biomass. The errors in the field and remote sensing observations indicated that this level of agreement was significant with greater than 99.5% confidence. These results suggest that further demonstrations may lead to a set of model functions that enable global, structure-based biomass remote sensing. INDEX TERMS: 1615 Global Change: Biogeochemical processes (4805) 1640 Global Change: Remote sensing 1694 Global Change: Instruments and techniques. Citation: Treuhaft, R. N., G. P. Asner, and B. E. Law, Structure- based forest biomass from fusion of radar and hyperspectral observations, Geophys. Res. Lett., 30(9), 1472, doi:10.1029/ 2002GL016857, 2003. 1. Remote Sensing of Biomass Via Forest Structure [2] Measurements of forest biomass are central to under- standing carbon sequestration and change resulting from natural and anthropogenic processes. Balancing the terres- trial carbon budget, as well as detecting changes in land-use patterns, calls for global remote sensing of forest biomass. Global monitoring, in turn, requires model functions or algorithms, F, which transform a set of remote sensing observations into biomass, schematically represented as: Biomassrs �� F��Remote Sensing Observations a b c . . .�� ��1�� The model function determining remotely sensed biomass, Biomassrs, depends on parameters a, b, c. . .. A long-range goal necessary for global biomass remote sensing is to determine model functions, including associated parameters, either a priori or from remote sensing data, for all regions of interest around the globe. However, current demonstrations determine F using field measurements, estimating a, b, c. . . by minimizing the difference, in a least squares sense [Hamilton, 1964], between Biomassrs and field biomass measurements, Biomassf, for a set of ������training������ stands. [3] Until remote sensing of 3-dimensional forest structure became possible via radar interferometry [Treuhaft et al., 1996 Treuhaft and Siqueira, 2000 Reigber and Moreira, 2000] and lidar [Lefsky et al., 1999 Drake et al., 2002], relationships like (1) often derived remotely sensed biomass from microwave power [Moghaddam et al., 1994 Dobson et al., 1995 Luckman et al., 1997 Paloscia et al., 1999] or optical radiance [Sader et al., 1989 Wu and Strahler, 1994 Steininger, 2000]. Accuracies of these approaches, which are generally worse than the desired 10���20% (10���40 tons/ hectare (ha) for most stands), in part reflect the radar and optical ������saturation������ phenomenon. When saturation occurs, radar power and optical reflectance generate biomass esti- mates with large errors, particularly at large biomass values. Additionally, for a given power or reflectance, biomass depends strongly on the vertical distribution of vegetation, as can be demonstrated from simple scattering and empirical considerations [Imhoff, 1995 Treuhaft et al., 1996 Stei- ninger, 2000]. Biomass estimates can erroneously span values corresponding to the range of unmeasured vertical structures. Structure measurements enable a new generation of model functions, which describe the physical apportion- ing of vegetation volume. They may therefore be simpler, accurate, robust, and may lead to the long-range goal of global biomass remote sensing. 2. Remotely Sensed Leaf Area Density [4] In this demonstration, multialtitude, airborne C-band radar interferometry generated relative density profiles, which were constrained to be Gaussian. Hyperspectral optical observations yielded the one-sided leaf area index (LAI), which normalized the radar relative profiles to produce the leaf area density (LAD), as detailed in Treuhaft et al., 2002. Our version of F first generates the biophysical vertical profile of LAD by performing a model-and-param- eter-estimation process on the remote sensing observations, and from LAD the biomass is then estimated, as shown GEOPHYSICAL RESEARCH LETTERS, VOL. 30, NO. 9, 1472, doi:10.1029/2002GL016857, 2003 Copyright 2003 by the American Geophysical Union. 0094-8276/03/2002GL016857$05.00 25 -- 1

schematically in Figure 1. The goal of this paper is to gauge whether the accuracy and feasibility of using a remotely sensed biophysical vegetation profile to estimate biomass warrant pursuing this approach. [5] The radar interferometric and hyperspectral optical data were taken over the Metolious River Basin in Central Oregon to produce remotely sensed LAD. LAI, tree dimen- sions, and biomass were measured in the field for 11 1-ha sites [Law et al., 2001a, 2001b] to generate field estimates of LAD. These sites are diverse in species and successional stage, many containing predominantly ponderosa pine, others dominated by grand fir and larch. Some stands are a mix of mature and large old trees, where the mature trees exist at higher density, and are the first successful cohort following exclusion of fire for the past 100 years. Other stands are spatially uniform in age, and some contain sparse vegetation. These stands span a biomass range from 24��� 295 tons/ha, with a mean of 161 tons/ha. [6] The NASA Airborne Synthetic Aperture Radar (AIR- SAR) [Zebker et al., 1992] and the Airborne Visible and Infrared Imaging Spectrometer (AVIRIS) [Green et al., 1998] flew over a 20 10 km area in the spring and summer months of 1998, 1999, and 2000. 3. From Leaf Area Density to Biomass [7] The general form of F relating the remotely sensed biomass to LAD in this paper is: Biomassrs �� F Z LAD z��f �� z��dz �� a b : ��2�� In (2), F involves a set of functions, f(z), which weight LAD(z) in a set of integrals (denoted by {}) over the vertical coordinate z. Nonlinear least-squares approaches estimate two parameters a, b [Hamilton, 1964]. The observations and parameter-generated quantities are related by Biomassf 1 Biomassf 2 . . . BiomassfN z0i f g sig f LAIig f 0 B B B B B B B B B B B B B B B B B B @ 1 C C C C C C C C C C C C C C C C C C A �� Biomassrs1 ^01 z ^1 s L^ AI 1 a b Biomassrs2 ^02 z ^2 s L^ AI 2 a b . . . BiomassrsN ^0N z ^N s L^ AI N a b ^0i z f g ^ig s f L^ AI i 0 B B B B B B B B B B B B B B B B B B @ 1 C C C C C C C C C C C C C C C C C C A �� ef 1 ef 2 . . . ef N ez0i o esi f g eLAIi f g 0 B B B B B B B B B Bn B B B B B B B B @ 1 C C C C C C C C C C C C C C C C C C A ��3�� where field biomass for stand i is Biomassfi, and efi is the field biomass measurement error. Below the field biomasses in the observation vector on the left are the remote sensing observations z0i, the location of the Gaussian peak of LAD, si, the Gaussian standard deviation, and LAIi, the remotely sensed leaf area index, as explained in Treuhaft et al., 2002. Parameters in the vector to the right of the equal sign, corresponding to each of the remote sensing structural observations and indicated with hats, are estimates of the actual remote sensing observations, given their errors and the fit suggested by (3). These parameters are used to calculate the model Gaussian LADs and biomasses, explicitly shown in (4a). Allowing estimates of these remote sensing parameters to shift away from the observations, constrained by the remote sensing observation errors ez0i, esi, eLAIi in the error vector, is a way of accounting for those errors, and is an important component of assessing the performance of variants of F [Bierman, 1977]. Minimizing the difference between the observations to the left of the equal sign and the quantities in the vector to the right, in a least-squares sense, yields all the parameter estimates. Values of LAI observations in (3) ranged from 0.3 to 5.7, and the LAI observation errors from hyperspectral data were about 5��� 10%. Radar observations of Gaussian peak locations ranged from 0 to 24 m, while Gaussian standard deviations were between 0.4 and 20 m. Errors on the radar structure parameters generating LAD were typically in the 1���2 meter range. [8] Demonstrating the feasibility of the approach in (2) entails finding a model function F which, when the param- eters in (3) are estimated, brings the reduced c2 (c2 per degree of freedom) close to unity [Hamilton, 1964] and produces an RMS difference between the field and remotely sensed biomasses that is in the 10���40 tons/ha range. Four trial LAD model functions, (4a) through (4d), were used to estimate the parameters in (3). The resulting RMS differ- ences between field and remote sensing biomasses and reduced c2s are in Table 1 for each trial F. The first three model functions, (4a) through (4c), were simple functions of LAD involving its total integral, LAI, its standard deviation, szi, and its mean, zi. The fourth variant of F, chosen on the basis of inspection of the remote sensing LADs and the field biomasses, depends on all of the LAD-related quantities in (4a), (4b), and (4c). The 4 structure-based trial functions, plus one additional involving C-band radar power for the ith stand, pi, (for completeness), are Figure 1. Schematic diagram of estimating biomass from radar and hyperspectral data via leaf area density. Leaf area density is produced by performing a model-and-parameter- estimation process on the remote sensing observations. The dashed box contains F of (1). 25 - 2 TREUHAFT ET AL.: STRUCTURE-BASED FOREST BIOMASS