Vegetation Structure Retrieval in Beech and Spruce Forests ... - Boku

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 5, NO. 1, FEBRUARY 2012

8

Vegetation Structure Retrieval in Beech and Spruce Forests Using Spectrodirectional Satellite Data Martin Schlerf and Clement Atzberger

Abstract—The structure of vegetation canopies largely controls the functioning of ecosystems. There is a substantial demand for spatial information on canopy structure. This paper examines the retrieval of an important forest structure property, leaf area index (LAI) from spectro-directional satellite observations (PROBA/CHRIS) using a forest reflectance model and a look-up table approach. Retrieved parameter estimates are compared to forest structure measured in 15 spruce stands (Picea abies L. Karst.) and 13 beech stands (Fagus sylvatica). For both species, off-nadir looking significantly reduced the normalized error (NRMSE) of forest LAI (spruce: NRMSE = 18.4%; beech: NRMSE = 26.1%) compared to near-nadir data (spruce: NRMSE = 32.6%; beech: NRMSE = 58.8%). At the same time acceptable R2-values were obtained. The best view angle for beech lies in forward direction due to foliar self shading in the canopy. With spruce, the forward direction is less favorable probably due to the very dark spruce leaves and dark shadows present in the canopy; instead the backward direction is more favorable as the canopy is brightly illuminated and shadows are minimal. Index Terms—Forestry, hyperspectral imaging, multi-angular remote sensing, vegetation mapping.

I. INTRODUCTION

T

HE structure of forests has an important effect on ecosystem functioning and carbon, water, and nutrients cycling [52]. For instance, the ability of plants to capture photosynthetically active radiation (PAR) is strongly affected by the amount, distribution and orientation of leaves in the canopy [19]. One of the most widely used descriptors of canopy structure is leaf area index (LAI), defined as the ratio of one-sided leaf area to ground area. LAI is an essential indicator of canopy structure, although it cannot alone fully describe canopy structure. The effective LAI measured with optical instruments like LAI-2000 or hemispherical photos aggregates elements of crown shape, canopy density, clumping, gaps and fractional vegetation cover into one single variable [39] and thus, may deviate from the true LAI. The increased need to understand local to global dynamics of ecosystems has created a substantial demand for spatial Manuscript received August 24, 2011; revised November 26, 2011 and December 30, 2011; accepted December 31, 2011. Date of publication February 10, 2012; date of current version February 29, 2012. M. Schlerf was with Trier University, Trier, Germany and Twente University, Enschede, The Netherlands, and is now with the Public Research Centre Gabriel Lippmann, L-4422 Belvaux, Luxembourg (corresponding author, e-mail: [email protected]). C. Atzberger was with the Joint Research Centre, Ispra, Italy, and Trier University, Trier, Germany, and is now with the University of Natural Resources and Life Sciences, Vienna, Austria (e-mail: [email protected]). Digital Object Identifier 10.1109/JSTARS.2012.2184268

information on ecosystem structure [52]. For instance, maps of vegetation structure are necessary to parameterise and constrain land surface models and biogeochemical models [1]. Networks such as the National Ecological Observatory Network (NEON, [31]) provide a growing amount of structure information, but cover only local scales. Terrestrial laser scanners allow measuring plant structure on the ground with high accuracy [17], but measurement of forest three-dimensional structure on the ground still bears inherent difficulties [5] and is time demanding. New LiDAR and RADAR remote sensing instruments have been proposed for spaceborne missions and will provide the capability to fill this gap [5], [60], but are not yet operational. Passive-optical remote sensing of vegetation structure typically relies on observations made from a single direction [30], [37], [50], [59]. Most land surfaces, due to their three-dimensional character, are strong anisotropic reflectors, that is, their reflectance changes with a change in the view angle [36]. The main factors influencing anisotropy of forests are at the leaf level the phase function of individual scatterers (e.g., leaves), at the canopy level the orientation (e.g., leaf angle distribution) and amount (e.g., leaf area index) of leaves, wood and litter, as well as leaf clumping and at the landscape level the size and spacing of the trees, as well as the shape of the crowns [1], [36]. A proper description of anisotropy is given by the bidirectional reflectance distribution function (BRDF) of the surface, which describes the distribution of the ratio of reflected radiance exiting from the surface along a specified direction to the irradiance incident on the surface from another specified direction [36], [46]. Spectro-directional (or multi-angle) remote sensing acquires surface radiance measurements—possibly in many narrow bands—at various observation and illumination angles [47], thus samples the surface BRDF. Repeated assessment of variations in vegetation structure over large areas requires directional data from space-borne platforms. Approaches based on airborne data [33] currently provide a very high spatial resolution of BRDF information but lack a sufficiently large areal coverage. Currently, the only spaceborne instrument that acquires spectro-directional data (multi-view angle and hyperspectral) simultaneously at high spatial resolution so that forest stands can be clearly defined in the imagery, is the Compact High Resolution Imaging Spectrometer (CHRIS) onboard PROBA (Project for Onboard Autonomy) platform [4]. Multi-angular observations potentially improve vegetation structure retrieval by explicitly viewing the effects of foliar self shading which is highly correlated with LAI and leaf angle distribution [1], [29]. Multi-angular remote sensing has recently been employed to detect structural attributes of vegetated

1939-1404/$31.00 © 2012 IEEE

SCHLERF AND ATZBERGER: VEGETATION STRUCTURE RETRIEVAL IN BEECH AND SPRUCE FORESTS USING SPECTRODIRECTIONAL SATELLITE DATA

surfaces including LAI [20], [41], [18], tree cover [26], woody shrub cover [12] [11] and forest background portion [7]. Compared to nadir bands, multi-angular data often produced more accurate estimates [23]. This can be partly explained by the fact that the disturbing contribution from the understory decreases with more oblique viewing angles [44]. Experimental data and radiative transfer theory reveals that canopy reflectance is most variable in the principal plane compared to other azimuthal planes [21]. This should lead to more efficient inversions. However, it is in general not clear which zenith view angles are most useful for canopy structure retrieval [42], [32] and if species-specific differences exist. In the past, most methods to retrieve maps of vegetation structure relied on empirical approaches and were based on either nadir view data only or on angle-normalized vegetation indices (e.g., [26], see also [54]). Recently, a number of papers were published that exploit multi-angle data explicitly [13], [35], [40]. The use of statistical techniques makes empirical approaches region- and species-specific and requires extensive field sampling efforts to provide the necessary field data. Semi-empirical BRDF models such as the Rahman–Pinty–Verstraete model [43] and the Ross–Li model (e.g., [38] are successfully used to retrieve BRDF and albedo from MODIS (Moderate Resolution Imaging Spectroradiometer) and MISR (Multi-angle Imaging Spectro-Radiometer) data. However, similar to the purely statistical techniques, field data are necessary for relating the retrieved model parameters to canopy structural variables [56]. More sophisticated model inversion methods employing physically based radiative transfer models are potentially general, reduce the amount of field measurements, but are not yet fully mature [36]. One general problem seems to be lack of coherence between real and modelled BRDF measurements [20]. Physically based BRDF models can be divided into turbid-medium radiative-transfer models [36], geometric-optical models [10], Monte Carlo ray-tracing models [16], and radiosity models [22]. Hybrid models such as Invertible Forest Reflectance Model (INFORM, [2], [49]) combine aspects of geometric-optical models with turbid-medium radiative transfer models. As a result of this combination, models such as INFORM are capable of assessing not only vegetation canopy structural properties but also leaf biochemical properties from spectro-directional signatures. Some preliminary assessment showed that INFORM simulations compare well with competing models and also with spectro-directional measurements when correctly parameterized in forward mode [51]. For efficient model inversion, look-up tables (LUT) and neural network methods are recommended techniques to handle even complex models without any simplification [32]. Both approaches are efficient for regional and global applications. LUT overcome limitations of iterative optimization algorithms as they permit a global search in the parameter space. Compared to neural nets, LUT are less sensitive to measurement/model errors although computationally more demanding. This paper examines the retrieval of forest LAI from spectrodirectional satellite observations (PROBA/CHRIS) using the

9

INFORM radiative transfer model and a LUT approach. We examine two managed vegetation types that are widespread in Europe: beech forest (Fagus sylvatica) and spruce forest (Picea abies L. Karst.). Retrieved parameters are compared to ground measurements in 28 forest plots. Field measurements were acquired concurrently to the satellite overpass and results are assessed in terms of precision and accuracy. The general research question consists in assessing the potential of multi-angular satellite observations for forest LAI mapping in spruce and beech stands. Specific research questions address the following issues: 1) Does off-nadir looking significantly improve forest LAI retrieval compared to nadir looking? 2) Of the five PROBA/CHRIS view angles, what is the best single view angle to retrieve forest LAI of spruce, respectively beech stands? 3) What is the retrieval accuracy of a single band multi-angle dataset compared to a multi-band single view dataset if both data sets contain the same number of observations? 4) What is the retrieval accuracy of a full multi-angle multiband dataset (i.e., a dataset consisting of five wavebands observed at five view angles)? II. MATERIALS AND METHODS A. Study Site , , Fig. 1) is part of the The Idarwald ( Hunsrück mountain chain (Germany) with remnants of peneplains forming gentle slopes and quartzite ridges up to heights of 800 m above mean sea level. Climatically the area is dominated by oceanic influences with primarily westerly winds; a relatively high mean annual precipitation of about 800–1000 mm and a moderate mean annual air temperature of 6.5 C. Weathering products of Devonian quartzite are almost pure sands while the shale weathers to sandy loam. Loess sedimentation and solifluction during the Pleistocene had a major influence on the soil formation. The predominant soil types are rankers, podzols, and dystric cambisols. The soils range from well to poorly drained, are generally low in nutrient status and are partially acidified. The dominant forest species are Norway spruce (Picea abies L. Karst.), European beech (Fagus sylvatica), Sessile oak (Quercus petraea) and Douglas fir (Pseudotsuga menziesii). They cover 53%, 18%, 11%, and 5% of the forested area, respectively. The predominant natural forest community in the area is Luzulo-Fagetum typicum, but afforestation starting at the end of the 18th century lead to a wide spreading of Norway spruce [48]. B. PROBA/CHRIS Image Data PROBA/CHRIS is a hyperspectral sensor that provides up to five angular looks via tilting and nodding of the satellite. PROBA/CHRIS images were acquired over Idarwald on 5 September 2005 in mode 1 (411–1004 nm, 62 bands) with a spectral resolution of 5–12 nm at five nominal fly-by view zenith angles ( , ,0 , , and ). Nominal ground sampling resolution is 34 m covering an area of 13 km 13 km. The images were acquired at 44 solar zenith

10


Fig. 2. Observation geometry of PROBA/CHRIS images acquired at the study site.

under a single view angle and five angular datasets of a single waveband observed at five different angles and one full spectro-angular dataset (Table I). C. Field Data Collection

Fig. 1. Location of forest plots in the study area Idarwald.

angle and 167 solar azimuth angle. Actual view zenith angles differed from the nominal values and are assumed constant across the spatial extent of the image (Fig. 2). The five images were atmospherically corrected to obtain surface bidirectional reflectance factor (BRF) estimates using AtcPro3 software [24], [25]. AtcPro3, like other atmospheric modelling codes enables the processing of data from tilted sensors by accounting for varying path length through the atmosphere. As no data on atmospheric parameters during the satellite overpass were available standard atmosphere conditions were assumed. The images were subsequently geocoded to the local reference system using sets of more than 20 ground control points per image. Reference coordinates stemmed from digital topographic map sheets of scale 1:25.000. The images were resampled to 30 m pixels using nearest-neighbour interpolation resulting in an accuracy of 1–3 pixels with the lowest accuracies for the large view-angle images. The geometric accuracy allowed to easily locate the selected forest stands in the image. For each forest stand four pure pixels were extracted next to the GPS-measured plot location from each (directional) image and the mean spectral BRF was calculated. Before calculating the average signatures, the data were screened for possibly occurring outliers. For the analysis all five view angle images were used. From each view angle image five spectral wavebands out of 62 available bands at 550 nm, 675 nm, 740 nm, 805 nm, and 970 nm were selected to represent green, red, red edge and NIR domains. By reducing the number of spectral bands to five, full comparability with the five angular observations was ensured when comparing different datasets. In total, eleven datasets were compared: five spectral wavebands observed

After notification of the successful and cloud-free image acquisition of the study area on September 5, 2005, a field campaign was launched and finalized within two weeks. Fifteen stands of Norway spruce (Picea abies L. Karst.) and 13 stands of European Beech (Fagus sylvatica) were selected (Fig. 1). Only forest stands of minimum size of about 150 m 150 m were selected to minimize border effects. All forest stands were managed (one-age) plantations with relatively homogenous canopy structure. Ages of the selected stands ranged from 20 to 250 years. Stand age information was available from forest inventory data provided by the state forest authorities. In each forest stand, a plot of 30 30 m was established and its central position determined using a hand-held GPS device to an estimated accuracy of approximately 5–10 m. Within the plots, Leaf Area Index (LAI; m2 m-2) was measured by a Li-Cor LAI-2000 (LICOR Inc., Lincoln, NE, USA) device with 270 degree view restrictor in flat terrain; 10 below canopy measurements were taken at regularly spaced points in a plot, and one single above canopy measurement was taken in nearby open fields. The LAI-2000 was only operated under overcast sky conditions during 10–16 h daytime and attention was paid to ensure stable illumination conditions between the below canopy and above canopy measurements. Despite the non-random distribution of leaves, no corrections for shoot level clumping and stand level clumping were applied [9]. Likewise, contributions of woody surfaces were neglected as the influence of woody components on the measurements is likely to be quite small. It was assumed that the underestimation of LAI due to clumping effects was somehow compensated by the overestimation of LAI through woody structures. Tree density (TD; ha-1) was derived by counting the total number of trees in the plot. Crown diameter (CD; m) was measured in two directions as an average of five trees per plot. Canopy height (CH; m) was calculated as the average from the height of five representative trees in a plot determined from angular measurements. Canopy closure (CC; in %), the amount


11

Fig. 3. Averaged spectral radiances observed over spruce forest (A) and beech forest (B) for five observation geometries illustrating the anisotropy of forests.

SUMMARY

OF

TABLE I ELEVEN ANALYZED DATASETS

of the sky obscured by the canopy from a certain point on the ground, was visually estimated in steps of 5%. Coverage of understory vegetation (CU; in %), the amount of forest floor covered by understory green plants, was derived visually in steps of 5% using the traditional grid quadrat technique. For all metrics, we followed standard guidelines used in German forest inventories.

Sampled forest plots cover a wide span of stand ages ranging from young to old forest stands (Table II). This ensured a relatively large variation in forest structural attributes. D. Model Parameterisation and Parameter Retrieval For retrieval of forest structural parameters the INFORM model was used. INFORM [2], [49] is a hybrid model


12

TABLE II SUMMARY STATISTICS OF COLLECTED FIELD DATA

, , ,

,

,

combining FLIM, SAILH, and PROSPECT or LIBERTY submodels. It accounts for 1-dimensional turbid-medium radiative-transfer within the crowns and 3-dimensional effects such as crown created shadows, the hotspot and the clumping of leaves in crowns. The FLIM submodel within INFORM [45] explicitly considers the canopy geometrical structure. While theoretically accurate for very sparse canopies without significant mutual shadowing among tree crowns [10], it proved to be useful in medium-dense and dense managed conifer stands [49], [51], [58] with low variation in tree height and thus limited overlapping of crowns. The combined SAILH [53], [34] and PROSPECT [27], [28] or LIBERTY [15] submodels implement a full multiple scattering scheme that allows to simulate vegetation optical properties (i.e., reflectance, absorption and transmittance) from 400 to 2500 nm, rather than a few discrete bands, as pure geometric-optical models typically do [10]. Successful parameter retrieval requires complete and comprehensive model parameterisation. In particular, parameters that will not be retrieved through the inversion procedure should be fixed at reasonable values. The parameterization of INFORM is summarized in Table III for the two tree species. Parameters which will be retrieved through inversion are indicated together with a plausible range of values (and the number of regular sampling steps). The lower and upper bounds minimize the ill-posedness of the inverse problem [3]. The species-specific leaf parameters were kept constant across age classes. The leaf parameters were manually tuned to fit the modelled forest reflectance to the measured spectra of three spruce and three beech stands. For this purpose, one young, one intermediate and one old stand were manually selected for the two species. In the same way, the mean leaf inclination angle was tuned. A look-up table inversion method was selected for model inversion because it is a conceptually simple technique that can be

Fig. 4. Comparison of simulation outputs (solid lines) with PROBA/CHRIS data (dotted lines) for young, intermediate and old beech and spruce stands in the spectral domain at near-nadir. Stand-ID’s link to stand attributes listed in Table III.

easily implemented. In previous studies, it often yielded good retrieval performances (e.g., [8]). In our study, the LUT was based on a regular sampling of parameters ranges (LAIs, TD, CD, and CH; see Table III) resulting in 24,570 parameter combinations. For these parameter combinations, the corresponding canopy reflectance spectra were simulated in five view angles. To retrieve the INFORM parameters corresponding to a given CHRIS spectrum the summed squared residuals between measured and simulated spectra were calculated. Previous research found the median of a number of the closest matching points within the LUT improved the estimates compared to the single best value [57]. Following this suggestion, we used the median of first 100 best matches [14]. Thus, after sorting the residuals, the 100 best simulated spectra were identified along with the corresponding canopy parameters values. From the 100 parameter values the median was calculated to receive the final parameter estimate. Estimated parameters were compared to field measurements per forest species using all forest plots. The coefficient of determination (R2) was used to assess the precision of the parameter retrieval, while the root mean squared error (RMSE) was used to assess the retrieval accuracy. Normalized RMSE values (NRMSE) were calculated from predicted and measured LAI in percent of the field-measured average LAI values. III. RESULTS A. Observed and Modeled Spectro-Directional Forest Signatures Spectral and directional simulations obtained in forward mode are shown in Fig. 4 and Fig. 5. For the simulations, stand specific measurements for CD, CH, SD and LAIs were used as input to INFORM. The remaining parameters were fixed according to Table IV. Overall, reasonable agreement between


13

TABLE III RANGE OF INPUT PARAMETERS USED TO GENERATE THE LOOK UP TABLE USING INFORM FOR SIMULATION OF PROBA/CHRIS REFLECTANCE

modelled and measured reflectance values were obtained in the visible. In particular, the influence of stand age and forest species is well captured. However, INFORM underestimated NIR reflectances for beech stands and the young spruce stand. Relatively good fits were achieved in the directional mode (Fig. 5) with the exception of the pronounced bowl shape in the red waveband seen in the PROBA/CHRIS data. INFORM does not follow this shape simulating a continuous decrease in reflectance.

B. Vegetation Structure Retrieval The results of the LAI estimates are presented in Table V for each of the 11 datasets. In general, spectral datasets (five spectral wavebands observed at a single angle) show higher accuracies than directional datasets (single waveband observed at five different view angles). Spectral off-nadir data produced considerably lower estimation errors (spruce: NRMSE = 18.4%; beech: NRMSE = 26.1%) compared to spectral near-nadir data (spruce: NRMSE = 32.6%; beech: NRMSE = 58.8%) and also showed acceptable values. For both forest types, the near-nadir observations (dataset 3) produced larger precisions but also lower accuracies (larger RMSE values). Other view angles (for

Fig. 5. Comparison of simulation outputs (stars) with PROBA/CHRIS data (circles) for young, intermediate and old spruce and beech stands in the directional domain (B) at 780 nm (black) and red reflectance at 670 nm (grey). Stand-ID’s link to stand attributes listed in Table III.

14


TABLE IV STAND CHARACTERISTICS AND MODEL INPUTS USED FOR THE FORWARD SIMULATIONS SHOWN IN FIGS. 4 AND 5

(1) Field Measurement; (2) Field Measurement Used as Model Input; (3) Model Input (Not Measured) TABLE V RESULTS OF THE RETRIEVAL OF LEAF AREA INDEX USING DIFFERENT DATASETS (SEE TABLE I)

Near-nadir results in italic, the best off-nadir results per species in bold.

instance, dataset 2 for beech forest) also resulted in larger precisions with much lower accuracies. The results of the LAI estimates are presented in Table V for each of the 11 datasets. In general, spectral datasets (five spectral wavebands observed at a single angle) show higher accuracies than directional datasets (single waveband observed at five different view angles). Spectral off-nadir data produced considerably lower estimation errors (spruce: NRMSE = 18.4%; beech: NRMSE = 26.1%) compared to spectral near-nadir data (spruce: NRMSE = 32.6%; beech: NRMSE = 58.8%) and also showed acceptable values. For both forest types, the near-nadir observations (dataset 3) produced larger precisions (R2) but also lower accuracies (larger RMSE values). Other view angles (for

instance, dataset 2 for beech forest) also resulted in larger precisions with much lower accuracies. According to the present analysis, the best view direction available from PROBA/CHRIS to estimate spruce forest LAI is the backward scattering direction with a view angle of 44 degrees (dataset 1). In contrast, beech forest LAI is best retrieved through the forward scattering direction with a view zenith angle of 44 degrees (dataset 5). The estimated LAI for these two datasets are plotted in Fig. 6 against the corresponding field measurements. Interestingly, the full (5 5 observations) dataset (dataset 11) yields only unsatisfactory results. The positive contribution of some of the view angles is apparently counterbalanced by several uninformative or corrupted observations. The latter can be demonstrated (not shown) when choosing small subsets (here: 10 bands) within the 25 features for LUT inversion. The best performing 10 band dataset for spruce LAI yields R2 of 0.60 (RMSE: 0.78) and for beech R2 of 0.85 (RMSE: 0.6). During the inversion process, parameters other than LAI were also retrieved. Table VI reports by age class the averaged estimated CC and TD together with LAI for the two tree species. Both CC and TD show plausible trends with age class and reflect known species related differences. In particular, we observe a decreasing canopy closure and tree density with increasing stand age. IV. DISCUSSION Overall, INFORM simulated successfully the spectral and directional reflectance curves of six forest stands. For these stands reasonable matches with the PROBA/CHRIS curves were obtained. However, INFORM produced less angular


Fig. 6. Predicted LAI using the best directional dataset against field measurements for two forest types. TABLE VI RETRIEVED PARAMETERS PER AGE CLASS

The individual retrievals were averaged within each age class. n indicates the number of available observations per age class.

reflectance variation compared to PROBA/CHRIS curves. A similar pattern was also observed in a study by [44] where the forest reflectance model FRT produced a smaller angular reflectance effect compared to PROBA/CHRIS data. The results presented in this paper are also in line with [7] who found that in forward scattering directions coniferous forests have lower reflectance values when compared to deciduous forests. This effect is also present in backward scattering direction, but less pronounced. The tendency of INFORM to overestimate beech NIR reflectance may be explained by the PROSPECT submodel parameterisation. In a local sensitivity analysis a large influence of leaf and needle parameters (such as leaf structure parameter and dry matter content) on forest canopy reflectance has been shown [35]. These leaf and needle parameters may vary with stand age and thus, an age class specific parameterisation could reduce the differences between simulated and measured spectra.

15

The smallest retrieved relative RMSEs in LAI estimates were 26.1% (beech) and 18.4% (spruce). The reduction in relative RMSE by using spectral off-nadir data compared to spectral nadir data amounts to 32.7% (beech) and 14.2% (spruce). Previous research confirms the advantage of off-nadir compared to nadir data. At a comparable spatial scale using multispectral data (Landsat 5 TM and Beijing-1 small satellite) acquired at four different days and the INFORM model, [58] achieved an improvement of inversion accuracy of forest LAI with multi-angle image data by 30% compared to the average accuracy of the inverted LAI with the single angle data. In a study on estimating corn LAI through inversion of a canopy reflectance model on PROBA/CHRIS data, [55] reported an absolute RMSE of 0.41 using five view angles and 62 spectral bands compared to absolute RMSE value of 1.42 when using only one view angle and only four wavebands. In our study, the spectral datasets (1–5) achieved larger accuracies and precision in LAI retrieval than the directional datasets (6–10). This is in line with the argumentation of [6] who found that better results are obtained using several spectral bands and one viewing angle, instead of several viewing angles at one spectral band. There has been an argument at what view angle canopy structure could be retrieved best. Two contrary suggestions are given: [1] stated that “effects of foliar self shading in forward scattering direction are highly correlated with LAI and leaf angle distribution.” Conversely, it was stated by [36], [42] that “the hot spot region (backward scattering direction) is the most information rich BRDF sub region.” From these interpretations, the results of the present study suggest that foliar self shading effects help to estimate beech LAI from satellite imagery. Conversely, spruce LAI is best retrieved from very bright images near the hot spot region. A possible explanation could be that spruce canopies with their relatively rough canopy surface already contain quite large amounts of shadow with over-enhancement of shadows when looking against the sun; instead the backward direction is more favorable as the canopy is brightly illuminated and shadows are minimal. Looking against the sun, however, towards rather smooth beech canopies with little shadow present at nadir views would produce the amount of self shading required for an accurate LAI retrieval. With spruce, also the forward direction is less favorable because spruce leaves are already very dark (at least in the red wavelengths) so that changes in shaded area with varying viewing directions are small, providing less information. In the same logic, the relatively poor performance of the full (5 5 observations) dataset can be explained. The positive contribution of the mentioned view angles is apparently counterbalanced by several poorly performing observations. For example, several 10 band subsets (not shown) showed higher retrieval accuracies compared to the full feature set suggesting that some of the 25 PROBA/CHRIS observations are either uninformative or corrupted, respectively, not well simulated by INFORM. V. CONCLUSIONS This study shows that the spectrodirectional approach offers more accurate information about forest structural properties than multi-spectral signals observed at nadir. The study thus

16


contributes to open fundamental questions on the science drivers and applications of multi-angle remote sensing raised by the International Forum on BRDF [36]. We conclude that off-nadir compared to nadir looking significantly improved retrieval of forest leaf area index. Coniferous and broadleaf trees have different “optimal” angles for LAI retrieval. The best view angle for beech lies in forward direction due to foliar self shading in the canopy. With spruce, the forward direction is less favorable due to dark shadows present in the canopy; instead the backward direction is more favorable as the canopy is brightly illuminated and shadows are minimal. More research is warranted to identify the “optimum” spectro-directional feature set for different forest species. If possible, this assessment should use an additional radiative transfer model to ensure (i) that the findings are not solely related to possible shortcomings of INFORM and (ii) to figure out why INFORM only partially reproduced the observed CHRIS data. ACKNOWLEDGMENT The authors would like to thank Joachim Hill, Henning Buddenbaum, and Johannes Stoffels (Trier University, Remote Sensing Department) for assistance with the field data collection and atmosphere correction of PROBA/CHRIS imagery. The authors further thank Bianca Hoersch (European Space Agency), Peter Fletcher (Remote Sensing Applications Consultants) as well as Mike Cutter and Lisa Johns (Sira Technology Ltd.) for acquiring and providing the PROBA/CHRIS images. They thank the three anonymous reviewers for their requests and suggestions that substantially improved the manuscript. REFERENCES [1] G. P. Asner, “Contributions of multi-view angle remote sensing to landsurface and biogeochemical research,” Remote Sens. Rev., vol. 18, pp. 137–162, 2000. [2] C. Atzberger, “Development of an invertible forest reflectance model: The INFOR-Model,” in A Decade of Trans-European Remote Sensing Cooperation. Proceedings of the 20th EARSeL Symposium Dresden, Germany, 14-16. June 2000, M. Buchroithner, Ed., 2000, pp. 39–44. [3] F. Baret and S. Buis, “Estimating canopy characteristics from remote sensing observations: Review of methods and associated problems,” in Advances in Land Remote Sensing: System, Modeling, Inversion and Application, S. Liang, Ed. Berlin, Germany: Springer, 2008, pp. 173–201. [4] M. J. Barnsley, J. J. Settle, M. A. Cutter, D. R. Lobb, and F. Teston, “The PROBA/CHRIS mission: A low-cost smallsat for hyperspectral multiangle observations of the earth surface and atmosphere,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 7, pp. 1512–1520, 2004. [5] K. M. Bergen, S. J. Goetz, R. O. Dubayah, G. M. Henebry, C. T. Hunsaker, and M. L. Imhoff et al., “Remote sensing of vegetation 3-D structure for biodiversity and habitat: Review and implications for lidar and radar spaceborne missions,” J. Geophys. Res.–Biogeosci., vol. 114, p. G00E06, 2009. [6] P. Bicheron, M. Leroy, O. Hautecoeur, and F. M. Breon, “Enhanced discrimination of boreal forest covers with directional reflectances from the airborne polarization and directionality of Earth reflectances (POLDER) instrument,” J. Geophys. Res.–Atmospheres, vol. 102, no. D24), pp. 29517–29528, 1997. [7] F. Canisius and J. M. Chen, “Retrieving forest background reflectance in a boreal region from Multi-angle Imaging SpectroRadiometer (MISR) data,” Remote Sens. Environ., vol. 107, no. 1–2, pp. 312–321, 2007.

[8] R. Casa and H. G. Jones, “LAI retrieval from multiangular image classification and inversion of a ray tracing model,” Remote Sens. Environ., vol. 98, no. 4, pp. 414–428, 2005. [9] J. M. Chen, “Optically-based methods for measuring seasonal variation in leaf area index of boreal conifer forest,” Agricult. For. Meteorol., vol. 80, pp. 135–163, 1996. [10] J. M. Chen, X. Li, T. Nilson, and A. Strahler, “Recent advances in geometrical optical modelling and its applications,” Remote Sens. Rev., vol. 18, pp. 227–262, 2000. [11] J. M. Chen, C. H. Menges, and S. G. Leblanc, “Global mapping of foliage clumping index using multi-angular satellite data,” Remote Sens. Environ., vol. 97, no. 4, pp. 447–457, 2005. [12] M. Chopping, L. H. Su, A. Laliberte, A. Rango, D. P. C. Peters, and N. Kollikkathara, “Mapping shrub abundance in desert grasslands using geometric-optical modeling and multi-angle remote sensing with CHRIS/Proba,” Remote Sens. Environ., vol. 104, no. 1, pp. 62–73, 2006. [13] M. Chopping, C. B. Schaaf, F. Zhao, Z. Wang, A. W. Nolin, and G. G. Moisen et al., “Forest structure and aboveground biomass in the southwestern United States from MODIS and MISR,” Remote Sens. Environ., vol. 115, no. 11, pp. 2943–2953, 2011. [14] R. Darvishzadeh, A. Skidmore, M. Schlerf, and C. Atzberger, “Inversion of a radiative transfer model for estimating vegetation LAI and chlorophyll in a heterogeneous grassland,” Remote Sens. Environ., vol. 112, pp. 2592–2604, 2008. [15] T. P. Dawson, P. J. Curran, and S. E. Plummer, “LIBERTY—Modeling the effects of leaf biochemical concentration on reflectance spectra,” Remote Sens. Environ., vol. 65, pp. 50–60, 1998. [16] M. Disney, P. Lewis, and P. Saich, “3D modelling of forest canopy structure for remote sensing simulations in the optical and microwave domains,” Remote Sens. Environ., vol. 100, no. 1, pp. 114–132, 2006. [17] J. U. H. Eitel, L. A. Vierling, and D. S. Long, “Simultaneous measurements of plant structure and chlorophyll content in broadleaf saplings with a terrestrial laser scanner,” Remote Sens. Environ., vol. 114, no. 10, pp. 2229–2237, 2010, [Article]. [18] Z. Fu, J. Wang, J. L. Song, H. M. Zhou, Y. Pang, and B. S. Chen, “Estimation of forest canopy leaf area index using MODIS, MISR, and LiDAR observations,” J. Appl. Remote Sens., vol. 5, p. 053530, 2004, doi:10.1117/1.3594171. [19] J. L. Funk and M. T. Lerdau, “Photosynthesis in forest canopies,” in Forest Canopies, M. D. Lowman, Ed. Burlington, VT: Elsevier, 2004. [20] F. Gascon, J. P. Gastellu-Etchegorry, and M. Leroy, “Using multi-directional high-resolution imagery from POLDER sensor to retrieve leaf area index,” Int. J. Remote Sens., vol. 28, no. 1–2, pp. 167–181, 2007. [21] N. S. Goel, “Models of vegetation canopy reflectance and their use in estimation of biophysical parameters from reflectance data,” Remote Sens. Rev., vol. 4, pp. 1–212, 1998. [22] N. S. Goel and R. L. Thompson, “A snapshot of canopy reflectance models and a universal model for the radiation regime,” Remote Sens. Rev., vol. 18, pp. 197–225, 2000. [23] J. Heiskanen, “Tree cover and height estimation in the Fennoscandian tundra-taiga transition zone using multiangular MISR data,” Remote Sens. Environ., vol. 103, no. 1, pp. 97–114, 2006. [24] J. Hill, P. Hostert, and C. Atzberger, “Calibration, atmospheric correction and validation of DAIS-7915 hyperspectral data,” in 1st Earsel Workshop on Imaging Spectroscopy, M. Schaepman, D. Schlapfer, and K. Itten, Eds., 1998, pp. 137–146. [25] J. Hill and B. Sturm, “Radiometric correction of multitemporal thematic mapper data for use in agricultural land-cover classification and vegetation monitoring,” Int. J. Remote Sens., vol. 12, no. 7, pp. 1471–1491, 1991. [26] A. R. Huete, G. Hua, J. Qi, A. Chehbouni, and W. J. D. Vanleeuwen, “Normalization of multidirectional red and NIR reflectances with the SAVI,” Remote Sens. Environ., vol. 41, no. 2–3, pp. 143–154, 1992. [27] F. Jacquemoud and F. Baret, “PROSPECT: A model of leaf optical properties spectra,” Remote Sens. Environ., vol. 34, pp. 75–91, 1990. [28] S. Jacquemoud, S. L. Ustin, and J. Verdebout, “Estimating leaf biochemistry using the PROSPECT leaf optical properties model,” Remote Sens. Environ., vol. 56, pp. 194–202, 1996. [29] H. G. Jones and R. A. Vaughan, Remote Sensing of Vegetation: Principles, Techniques, and Applications. Oxford, U.K.: Oxford Univ. Press, 2010.


[30] M. Kalacska, G. A. Sanchez-Azofeifa, B. Rivard, T. Caelli, H. P. White, and J. C. Calvo-Alvarado, “Ecological fingerprinting of ecosystem succession: Estimating secondary tropical dry forest structure and diversity using imaging spectroscopy,” Remote Sens. Environ., vol. 108, pp. 82–96, 2007. [31] T. U. Kampe, B. R. Johnson, M. Kuester, and M. Keller, “NEON: The first continental-scale ecological observatory with airborne remote sensing of vegetation canopy biochemistry and structure,” J. Appl. Remote Sens., vol. 4, 2010, doi 10.1117/1.3361375. [32] D. S. Kimes, Y. Knyazikhin, J. L. Privette, A. A. Abuelgasim, and F. Gao, “Inversion methods for physically-based models,” Remote Sens. Rev., vol. 18, pp. 381–439, 2000. [33] T. Koukal and C. Atzberger, “Potential of digital aerial frame camera derived BRF information for forest classification,” IEEE JSTARS, vol. 5, no. 1, pp. 30–43, Feb. 2012. [34] A. Kuusk, “The hot spot effect of a uniform vegetative cover,” Soviet J. Remote Sens., vol. 3, no. 4, pp. 645–658, 1985. [35] V. C. E. Laurent, W. Verhoef, J. Clevers, and M. E. Schaepman, “Estimating forest variables from top-of-atmosphere radiance satellite measurements using coupled radiative transfer models,” Remote Sens. Environ., vol. 115, no. 4, pp. 1043–1052, 2011. [36] S. Liang, A. Strahler, M. Bernsley, C. Borel, S. Gerstl, and F. Prata et al., “Multiangle remote sensing: Past and future,” Remote Sens. Rev., vol. 18, pp. 83–102, 2000. [37] D. S. Lu, M. Batistella, and E. Moran, “Satellite estimation of aboveground biomass and impacts of forest stand structure,” Photogramm. Eng. Remote Sens., vol. 71, no. 8, pp. 967–974, 2005. [38] W. Lucht, C. B. Schaaf, and A. H. Strahler, “An algorithm for the retrieval of albedo from space using semiempirical BRDF models,” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 2, pp. 977–998, 2000. [39] S. V. Ollinger, “Sources of variability in canopy reflectance and the convergent properties of plants,” New Phytol., vol. 189, pp. 375–394, 2010. [40] J. Pisek, J. M. Chen, R. Lacaze, O. Sonnentag, and K. Alikas, “Expanding global mapping of the foliage clumping index with multi-angular POLDER three measurements: Evaluation and topographic compensation,” ISPRS J. Photogramm. Remote Sens., vol. 65, no. 4, pp. 341–346, 2010. [41] A. Pocewicz, L. A. Vierling, L. B. Lentile, and R. Smith, “View angle effects on relationships between MISR vegetation indices and leaf area index in a recently burned ponderosa pine forest,” Remote Sens. Environ., vol. 107, pp. 322–333, 2007. [42] W. Qin and S. Liang, “Plane-parallel canopy radiation transfer modeling: Recent advances and future directions,” Remote Sens. Rev., vol. 18, pp. 281–306, 2000. [43] H. Rahman, B. Pinty, and M. M. Verstraete, “Coupled surface-atmosphere reflectance (CSAR) model: Semiempirical surface model usable with NOAA advanced very high resolution radiometer data,” J. Geophys. Res., vol. 98, no. D11, pp. 20791–20801, 1993. [44] M. Rautiainen, M. Lang, M. Mottus, A. Kuusk, T. Nilson, and J. Kuusk et al., “Multi-angular reflectance properties of a hemiboreal forest: An analysis using CHRIS PROBA data,” Remote Sens. Environ., vol. 112, pp. 2627–2642, 2008. [45] A. Rosema, W. Verhoef, H. Noorbergen, and J. J. Borgesius, “A new forest light interaction model in support of forest monitoring,” Remote Sens. Environ., vol. 42, pp. 23–41, 1992. [46] G. Schaepman-Strub, M. E. Schaepman, T. H. Painter, S. Dangel, and J. V. Martonchik, “Reflectance quantities in optical remote sensing-definitions and case studies,” Remote Sens. Environ., vol. 103, pp. 27–42, 2006. [47] M. E. Schaepman, “Spectrodirectional remote sensing: From pixels to processes,” Int. J. Appl. Earth Observ. Geoinf., vol. 9, pp. 204–223, 2007. [48] M. Schlerf, “Determination of Structural and Chemical Forest Attributes Using Hyperspectral Remote Sensing Data—Case Studies in Norway Spruce Forests” Ph.D. dissertation, Trier University, Trier, Germany, 2006 [Online]. Available: http://ubt.opus.hbz-nrw.de/volltexte/2006/369/ [49] M. Schlerf and C. Atzberger, “Inversion of a forest reflectance model to estimate structural canopy variables from hyperspectral remote sensing data,” Remote Sens. Environ., vol. 100, pp. 281–294, 2006. [50] M. Schlerf, C. Atzberger, and J. Hill, “Remote sensing of forest biophysical variables using HyMap imaging spectrometer data,” Remote Sens. Environ., vol. 100, pp. 177–194, 2005.

17

[51] M. Schlerf, W. Verhoef, H. Buddenbaum, J. Hill, C. Atzberger, and A. K. Skidmore, “Comparing three canopy reflectance models with hyperspectral multi-angular satellite data,” in Proceedings of the 10th International Symposium on Physical Measurements and Spectral Signatures in Remote Sensing (ISPMSRS’07), Davos, Zwitserland, 12–14 March, 2007, M. E. Schaepman, S. Liang, N. E. Groot, and M. Kneubühler, Eds., Davos, 2007, vol. XXXVI, pp. 404–407, ISPRS. [52] H. H. Shugart, S. Saatchi, and F. G. Hall, “Importance of structure and its measurement in quantifying function of forest ecosystems,” J. Geophys. Res.-Biogeosci., vol. 115, 2010. [53] W. Verhoef, “Light scattering by leaf layers with application to canopy reflectance modeling: The SAIL model,” Remote Sens. Environ., vol. 16, pp. 125–141, 1984. [54] J. Verrelst, M. E. Schaepman, B. Koetz, and M. Kneubuhler, “Angular sensitivity analysis of vegetation indices derived from CHRIS/PROBA data,” Remote Sens. Environ., vol. 112, pp. 2341–2353, 2008. [55] F. Vuolo, L. Dini, and G. D’Urso, “Retrieval of leaf area index from CHRIS/PROBA data: An analysis of the directional and spectral information content,” Int. J. Remote Sens., vol. 29, no. 17–18, pp. 5063–5072, 2008. [56] J.-L. Widlowski, B. Pinty, N. Gobron, M. M. Verstraete, D. J. Diner, and A. B. Davis, “Canopy structure parameters derived from multiangular remote sensing data for terrestrial carbon studies,” Climatic Change, vol. 67, pp. 403–415, 2004. [57] M. Weiss, F. Baret, R. B. Myneni, A. Pragnere, and Y. Knyazikhin, “Investigation of a model inversion technique to estimate canopy biophysical variables from spectral and directional reflectance data,” Agronomie, vol. 20, no. 1, pp. 3–22, 2000. [58] G. Yang, C. Zhao, Q. Liu, W. Huang, and J. Wang, “Inversion of a radiative transfer model for estimating forest LAI from multisource and multiangular optical remote sensing data,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 3, pp. 988–1000, 2011. [59] Y. Zeng, M. E. Schaepman, B. F. Wu, J. Clevers, and A. K. Bregt, “Quantitative forest canopy structure assessment using an inverted geometric-optical model and up-scaling,” Int. J. Remote Sens., vol. 30, no. 6, pp. 1385–1406, 2009. [60] F. G. Hall, K. Bergen, J. B. Blair, R. Dubayah, R. Houghton, G. Hurtt, J. Kellndorfer, M. Lefsky, J. Ranson, S. Saatchi, H. H. Shugart, and D. Wickland, “Characterizing 3D vegetation structure from space: Mission requirements,” Remote Sens. Environ., vol. 115, no. 11, pp. 2753–2775, Nov. 2011.

Martin Schlerf received the M.Sc. degree in geography from Justus-Liebig University, Giessen, Germany, in 1999, and the Ph.D. degree in remote sensing from the University of Trier, Trier, Germany, in 2005. From 2006 until 2011, he held a position as Assistant Professor in Earth Observation of Natural Resources at University of Twente—ITC, The Netherlands. Since November 2011, he has been working as Project Leader in Remote Sensing at the Public Research Centre Gabriel Lippmann (CRP-GL), Luxemburg. Clement Atzberger received the Ph.D. degree on crop growth modeling and remote sensing data assimilation from Trier University, Germany, in 1997. He is currently full Professor and Head of the Institute for Surveying, Remote Sensing and Land Information (IVFL) at the University of Natural Resources and Life Sciences (BOKU), Vienna, Austria. He worked for two years as Assistant Professor at ITC, Enschede, The Netherlands, and further two years in private industry (GeoSys SA, Toulouse, France). Between 2007 and 2010, he was with the Joint Research Centre (JRC) of the European Commission, Ispra, Italy.