About one hundred participated in the very full program of this three day meeting at the Joint Research Centre site at Ispra. Key presentations were oral, and all are summarised below, as are the plenary discussions at the end of each day. About seventy posters were presented with two-minute oral summaries and highlights of those are below. CSIRO was represented by Mike Caccetta of Exploration and Mining, Norm Campbell of Mathematical and Information Sciences, Darius Culvenor of Forestry and Forest Products, and Ian Grant of Atmospheric Research. Posters were presented by Darius Culvenor and Ian Grant.

Session 1: Estimating Land Surface BRFs from Space

MODIS and MISR, two of the five sensors aboard the TERRA space platform, were supposed to have been launched at the time of the meeting, but the earliest possible launch date is now said to be December 1999.

Alex Lyapustin of NASA Goddard examined the accuracy of atmospheric correction over non-Lambertian surfaces by 1D radiative transfer methods in comparison with 3D methods, as a function of spatial scale. For MODIS, the error from a 1D treatment is about 2% at the edge of the scan.

Crystal Schaaf of the MODIS team at Boston University described the at-launch MODIS BRDF and albedo products. If there are insufficient looks within a 16-day window to do a full BRDF inversion, a "magnitude inversion" is done, whereby an archetypal BRDF shape is assumed on the basis of a pre-existing land cover map and only a single overall reflectance factor is derived from the data. The BU team has applied the method to a period of AVHRR data to produce a global map of albedo which was almost complete at the time of the meeting; they apologized for Australia's absence from the global map! Their opinion was that an 8-day window was too short because of cloud, and a month was too long to capture vegetation changes.

John Martonchik of JPL described a method to separate the aerosol and land surface signatures in multidirectional MISR and Air MISR data. The method relies on the surface being inhomogeneous, and used empirical orthogonal functions and covariances.

Patrice Bicheron of CESBIO, France, described a dataset of over 400 POLDER-measured BRDFs, covering all biomes of the earth at two months (available at http://earth-sciences.cnes.fr:8060/POLDER/SCIEPROD/brdf.htm). He found that the Chen model (which is Roujean times a hotspot factor) fitted the BRDFs better than the Roujean, RPV, Engelson and some other models. The Kuusk model on which the POLDER 2 processing will be based takes a set of vegetation parameters as input. It will be operationally inverted with a neural network to retrieve these parameters.

Session 2: Modelling and Inversion of Surface BRF Models

Knyazikhin of Boston University discussed three-dimensional radiative transport in a vegetation canopy, with the object of inverting MODIS and MISR data to retrieve LAI and fAPAR. He noted that 1D and 2D transport theories were inadequate for the inversion.

Andres Kuusk discussed a new directional multispectral forest vegetation reflectance model developed by Kuusk and Nilson. Kuusk described how the transfer of radiation in forest canopies is extremely complex, and that we need new models to help determine the driving factors. These new models should incorporate existing forestry data and should produce optical images. The Kuusk-Nilson model represents crowns as combinations of ellipsoids, cylinders and cones. Single-order scattering in the model is computed by numerical integration. Understorey, background, trunks, bark and foliage are all valid reflecting surfaces within the model. The model uses the biochemical PROSPECT2 model (Jacquemoud et al.) for computing leaf angle and transmittance. The PROSPECT2 model works well with crops, however unrealistic biochemical parameters must currently be used in association with the Kuusk-Nilson model to give give reasonable values of forest reflectance (probably because of the added effects of crown structure and branch reflectance in forest environments). The Kuusk-Nilson model uses the 6S atmospheric model for calculating incoming fluxes. The Kuusk-Nilson model is used for simulating forest canopy reflectance versus wavelength (400-2500 nm) as a function of view zenith angle. Results presented by Kuusk showed a very sharp hotspot - a hotspot that Kuusk agreed was probably too sharp for natural scenes. The hotspot should be narrow for leaf-scale scattering elements, and become progressively broader as modelling occurs at the crown scale. Kuusk's conclusions were that the new model works well, but requires a large number of parameters, which may make inversion difficult. The model is available to anyone.

Bernard Pinty talked about RAMI - the Radiation Model Intercomparison he has been coordinating. The goal is to seek consistency between the models, there being no obvious "truth" to compare against. Homogeneous and heterogeneous models were compares separately. A dozen models in total were blindly compared in forward mode (the vegetation structure was specified and the models had to predict the BRDF). Several models dropped out of the exercise as code errors were discovered. Only one model was submitted to the inverse mode comparison, so there was no comparison there! Of all the heterogeneous models compared, the maximum deviation was only 5%. RAMI will continue.

Poster highlights for sessions 1 and 2

Chen of the Canada Center for Remote Sensing showed that the hotspot / coldspot ratio was sensitive to the vegetation type, where the coldspot is the BRDF minimum in the forward scatter direction.

Culvenor and Coops from CSIRO presented a geometric-optical ray-tracing model developed specifically to replicate canopy architecture representative of eucalypts. The model, called SORTE (Simulation Of Remote Tree Environment), allows crowns to be modelled at either the whole-crown, branch or leaf scale. This approach allows crown architecture to be specifically investigated for its relative contribution to canopy BRDF. The model may also be run using a topographically variable background based on an imported DTM. A case study was presented to demonstrate the relative influence of internal crown structure on canopy BRDF as a function of canopy cover. The model produces real images and has also been used for validating automated image interpretation algorithms.

Gao and the MODIS team from Boston University presented two posters. One described a new kernel, the Li-Transit kernel, which smoothly changes between Li-Sparse and Li-Dense to give better extrapolation to high view and sun zenith angles. The other described constraining the inversion of BRDF models by a priori knowledge when angular sampling is sparse.

Lewis of University College London tried a kernel approach to parameterizing reflectance spectra. Studying existing sets of spectra for a range of surface types, he found that seven spectral kernels were enough to describe the spectrum over 400-2400 nm.

According to Craig Trotter of Landcare Research, New Zealand, topographic corrections have developed little in the last twenty years. He described a NZ experiment where reflectance measurements are made of small trees arrayed on a tiltable platform - artificial topography.

Marc Leroy argued theoretically that the BRDF is not strictly reciprocal, essentially because for a scattering layer (e.g. canopy) of non-zero thickness, a sensor views different scattering volumes when looking at the same target from different directions.

Discussion of sessions 1 and 2

The discussion focussed on the evaluation of models, particularly on RAMI. The question was raised of comparing the models with thorough field measurements such as BOREAS. Pinty's reply, which was echoed by Verstraete and others, was that errors in field measurements are too large for this to be useful. It would probably happen eventually, but the first step should be model intercomparisons and simple experiments in labs where conditions can be tightly controlled. It was suggested that figures of merit for model performance should be based on the angular coverage of real sensors rather than treating all angles equally. Anne Nolin of the US National Snow and Ice Data Centre wanted to see snow tests in RAMI, including snow-covered vegetation; snow is unusual in that it scatters predominantly forward, compared with backscatter for soil and vegetation. The question was also asked: should more models be developed to specifically investigate more spectral ranges? One suggestion is that the geometric and spectral parts of the problem can be treated separately. In particular, if the vegetation components are modelled as lambertian surfaces, then there is a generic geometric solution that is independent of wavelength.

Session 3: Current and Future Airborne Space Missions

The first three papers presented proposed but unconfirmed missions.

Wiscombe of NASA Goddard described LEONARDO, a novel idea to fly a formation of 6-10 small satellites which would simultaneously view a target from different directions. The fleet would be reconfigurable and have some intrinsic intelligence to reduce the ground segment requirements. The mission's primary goal would be to reduce errors in measurements of atmospheric radiative forcing due to sparse angular and diurnal sampling. Whereas current efforts aim to measure a global and static radiation budget, this mission aims to enable the measurement of radiation forcing which is local and dynamic, such as that associated with hurricanes and fires.

Gerstl of Los Alamos described the Triana satellite which, if it survives current consideration by Congress, is planned to image the Earth continuously from the L1 point between the Earth and the sun from July 2001. The L1 point is the gravity-neutral point between the earth and the sun (1.6 million kms away) from which the whole earth can be viewed over a solid angle of 0.39 degrees. Triana has five UV bands and five VIS/NIR bands, all with 8 km resolution.

Over six months it will sample the hotspot (that part of the BRDF in the antisolar direction) at all points of the globe, over directions 2-15 degrees from the antisolar direction. Previous modelling and airborne measurements show that the hotspot width and amplitude indicate leaf size and shape, so that there is a mapping from hotspot parameters to landcover ecology. The evolution of leaf shape could reveal vegetation stress.

Menenti of University Louis Pasteur described the Land Surfaces Processes and Interactions Mission (LSPIM), a candidate mission of ESA. It will observe the BRDF in the shortwave, and be the first mission to measure the temperature directional distribution. Resolution is 50 m and bands will be dynamically selected from 142 in the range 0.4-2.35 microns . A field segment is an integral part of the mission - 100 sites have been chosen globally.

Cahoon of NASA Langley described a helicopter-based BRDF measurement system, which typically measures a single 30-m footprint at view zenith angles up to 75 degrees and from flight lines at five azimuths. His measurements of the central ARM site disturbingly showed variations of +/- 10%. A system on a fixed-wing aircraft is being developed to measure surface variability over the large area (10-20 km) of the CERES footprint. Field data may be made available in the future online at http://www-svg.larc.nasa.gov/~ceres/data/index.html.

Minnis of NASA Langley described investigations of the angular dependence of outgoing thermal radiation, the understanding of which is useful for the retrieval of skin temperature and the outgoing thermal radiation flux. Data from an airborne scanner or from the two GOES satellites showed that the apparent temperature could vary by up to 4 degrees with view direction. The temperature difference appeared to vary linearly with the BRDF anisotropy, with a slope that depended on the type and altitude variability of the terrain. Night observations show that emittance variations are not the main factor.

Session 4: Demonstration of Laboratory and Field Measurements

Thursday afternoon was devoted to an inspection of two JRC facilities, the European Goniometer Laboratory (EGO) and the European Microwave Signature Laboratory (EMSL), and three instruments for the field measurement of BRDF brought by Workshop participants. We were each given detailed brochures on the EGO and EMSL. The three field instruments are now briefly described.

Demircan of the German Aerospace Center demonstrated his BRDF field imager. This uses a CCD-line camera which captures a 1D image of the scene from nadir to horizon and which is rotated about the vertical axis to build up the hemisphere. The instrument's support allows it to be flipped 180 degrees to measure the sky.

Another design's key feature was that it supported a spectrometer 10 m above the ground from the top of a van so that it could be moved quickly to sample several areas in a short time. The configuration resembled Fred Prata's car-mounted system but with the advantage of much greater height. The influence of the car on the measurement had not been assessed.

The third supported a spectrometer on a semi-circular vertical arc, the two ends of which ran along a circular track on the ground. The sensor was about two metres from the target surface, and the motion in two directions and data acquisition were computer-controlled. Incrementing the sensor through an entire view zenith was reasonably slow, and a faster system would be preferable to minimise variation in sun azimuth and zenith angle during the sample period.

Poster highlights for sessions 3 and 4

Demircan is compiling a library of BRDFs to be published in 2000 as a CD with book. So far he has 440 BRDFs of 14 surface types.

Jones of the University of Dundee explored the retrieval of canopy structure from the variation with angle of the frequency distribution of radiance, by analysing field images of canopies for the fraction seen of soils and leaves, and of sunlit and shaded components.

Gerhard Meister of the University of Hamburg described BRDF studies, including EGO measurements, for components of urban landscapes, to develop the automated processing of digital multispectral imagery for cadastral and urban planning applications. A significant component of this work involved characterising the BRDF of roof tiles in the European Goniometer Laboratory a the JRC.

Nandy of the University of Arizona described a digital imaging system for the field measurement of BRDF in vicarious calibration campaigns. It consisted of a fish-eye lens, filters and a thermoelectrically cooled, astronomical grade CCD. Resolution was 0.2 degrees and exposures were a few seconds per filter. A great deal of effort had to go into the geometric and radiometric characterization of the system; camera polarization dominated the error. I heard Nandy comment that characterization of a field of view approaching 180 degrees was so difficult that pointing a camera with a narrower field to several directions might be more practical.

Discussion for sessions 3 and 4

The question was posed as to what further BRDF measurements are needed. David Diner emphasised that the requirements (spectral, temporal, angular, etc.) should be driven by the science questions or applications. Ralph Kahn of the MISR team noted that radiation budget requirements are more stringent than most applications. Yves Govaerts of EUMETSAT questioned the requirement for any surface BRDF measurements, other than for vicarious calibration.
Nick Strugnell of Boston University commented it was difficult to find measurements for some of BU's 25 BRDF archetypes. Alan Strahler of BU suggested a need for detailed hotspot measurements to retrieve vegetation structure, to separate LAI from fAPAR, for instance, for the estimation of carbon fluxes. The narrow and wide hotspots from fine and coarse structure will overlap, and we need to learn how to separate them. Field hotspot measurements are obstructed by the shadow of most instruments in use now; either small instruments or aircraft measurements are required.

Alex Lyapustin and Don Deering suggested that a protocol for BRDF measurements is required, so that all relevant parameters are measured. Alex emphasised the need for the atmospheric correction of surface BRDF measurements (diffuse skylight and surface-atmosphere coupling).

After listening to much of the discussion, Juhan Ross declared, "We measure too much". He thought it was easier to measure than to think, and felt many had become lazy. He believed a lot more planning should go into the measurements, and that we should give some thought to what we can't measure.

Ross also suggested that the BRDF community should collaborate with the microwave community, perhaps holding joint conferences or measurement campaigns. Nick Strugnell suggested that lidar and BRDF studies be combined.

Session 5: Benefits of Multiangular Measurements

David Diner of the MISR team outlined the characteristics and objectives of the expected-to-have-been-launched-but-not-quite-yet MISR (4 bands, 9 view angles along track). Diner stated that MISR had three main interests, all of them climatic: clouds, aerosol, and surface-atmosphere interactions. I didn't appreciate MISR's uniqueness until Michel Verstraete emphasised it at the Les Diablerets Workshop the following week: all of its view angles are acquired within seven minutes, compared to several days for MODIS; and its 275 m resolution greatly enhances homogeneity within its field of view compared to POLDER. However only the red band data (672 nm) is supplied at 275 m resolution. The resolution is averaged to 1.1 km for the other bands (446, 558 and 866 nm).

Marshak of NASA Goddard has developed a novel method to "exploit the sharp spectral contrast in vegetation surface reflectance across 700 nm wavelength to retrieve cloud properties from ground-based radiance measurements" of the sky. The vegetated surface is a source of photons, which has a distinctive spectral signature, illuminating the underside of clouds. The radiative transfer has been studied theoretically. CIMEL sunphotometer observations under clear, overcast and broken cloud conditions were made, and the ranking of intensities at 440, 670, 870 and 1020 nm depended on the sky conditions. A Normalized Difference Cloud Index,

NDCI = (I870 - I670) / (I870 + I670),

was negative for clear skies, positive for cloudy skies, and well correlated with cloud optical depth inferred from the ARM microwave radiometer. Satellite estimates of surface albedo, uniformity, and 870/670 contrast would help.

The description by Ralph Kahn of the MISR team of the MISR aerosol retrieval method centred on extensive sensitivity tests. MISR is expected to distinguish three size groups, two compositions (dirty/clean), spherical and non-spherical particles. The inability to distinguish carbonaceous from sulphate aerosols is the main deficiency. It is expected that MISR will identify aerosol airmasses, while ground measurements will have to supply the detailed microphysical properties.

Session 6: Multiangular Applications over Land Surfaces

Bernard Pinty of the JRC described the algorithm for a new albedo product from Meteosat. The method takes all observations during one day (every 30 minutes, sunrise to sunset) to exploit the range of sun directions as angular samples of the BRDF. The RPV model, which treats the BRDF as amplitude times shape, was fitted to the observations for each pixel and outliers (usually clouds or their shadows) were iteratively rejected until a cost function was acceptably small. Pinty demonstrated the product by showing that summer-winter albedo difference maps of Africa identified burnt areas.

Ranga Myneni of Boston University discussed the generic inversion of multiangular multispectral data for the retrieval of LAI and fAPAR. Notably, he explicitly treated the uncertainty caused by the spread of LAI, etc. within each biome.

Poster highlights for sessions 5 and 6

I didn't see this poster but it's in the book of abstracts and worth noting. Acharya et al. described MODTRAN4, an upgrade of MODTRAN which accounts for the dependence of multiple scattering on the azimuth of the line of sight, and also accommodates parameterized surface BRDFs.

Gobron of the JRC described the steps needed to create vegetation indices optimized to particular applications and sensors.

Grant of CSIRO presented results from an exploration of the POLDER surface BRDF product for Australia by comparison with an existing map of vegetation structural types. Maps of the parameters of the Roujean model reveal many regions of distinct landcover type clearly, but there are often large variations within a nominally uniform landcover type. In June these variations occur almost entirely at large spatial scales, but in December strong structure down to the pixel (7 km) scale is widespread. In either month the spread of parameters within a class is generally much greater than the differences between classes.

Ni of Raytheon was absent but her poster investigated the information obtainable from the spatial variance of the directional radiance, the BRVF.

Richaume of CESBIO, France, described the neural network inversion to be applied to POLDER 2 data to retrieve vegetation parameters. The POLDER 1 data will have been processed through this scheme by mid-2000, and may serve as inputs for Australian biospheric models.

Stefan Sandmeier of NASA Goddard was absent but his poster showed that vegetation structural information and increased accuracy of classification could be obtained from the spectral variability of BRDF effects. This is mostly due to the multiple scattering effect which is dependent on the absorption characteristics of the canopy. The poster stated that "in erectophile canopies, such as coniferous forests, the spectral variability is very pronounced due to the large canopy gap fraction. In planophile canopies, such as closed deciduous forests or bush vegetation on swamps, the spectral BRDF variability is rather limited."

Nick Strugnell of Boston University described his approach to treating BRDFs as an amplitude times an archetypal shape for each landcover class. Nick emphasised that the vegetation canopy hotspot occurs for two main reasons: because objects hide their own shadow, but also because of a volumetric effect, whereby in a field of dispersed scatterers, a multiply-scattered photon is statistically most likely to exit the volume along the same path of ingress.

Discussion for sessions 5 and 6

David Diner requested that the community strive for precision in nomenclature, particularly in the proceedings of the Workshop. He also made a plea for precision in specifying the accuracy requirements of science questions. Diner and Ralph Kahn noted that the global albedo data needed to achieve 1% accuracy in the radiation budget is a clear and well specified requirement for multiangular information. They raised biomass estimation via LAI and fAPAR as another, less clear cut, application.

Yves Govaerts of EUMETSAT stated that albedo only improved weather forecasting in the presence of snow, otherwise a fixed albedo was sufficient; cloud had more influence on the surface radiation budget.

Marc Leroy described a mission he has proposed to begin ~2005: the Terrestrial Ecosystem Monitoring System (TEMS). This would effectively be three POLDERs on one platform: VIS/NIR (1 km resolution), SWIR, thermal IR (6 km). It would aim to contribute to vegetation function models for carbon budgeting, landcover change monitoring, and surface-atmosphere boundary conditions for meteorological forecasts. David Diner and Ralph Kahn asked what science demands were driving the instrument specifications, to which Leroy replied that the instrument specifications were limited to what was technologically feasible.

Wolfgang Lucht commented that while both the remote sensing and ecological modelling communities talked about using earth observation data in carbon budget models, neither seemed committed and little had happened. He suggested that the MODIS and MISR teams should actively push the ecologists to use their products. David Diner observed that this might be viewed poorly by funding authorities if the ecologists don't ask for the products.

Don Deering thought the BRDF community had done a poor job of demonstrating and selling BRDF applications to funding agencies and resource managers. David Diner warned that if the BRDF community remains closed, without applications, then it may cease to exist. Someone commented earlier that it is hard to convince ESA that clouds and surfaces are non-Lambertian. Flasse suggested that remote sensing case studies be presented at non-remote sensing meetings. This echoes an observation made at a recent CAR workshop on external earnings: applications have arisen from presenting research at meetings of peripherally-related communities. Flasse also suggested that overview articles be written for non-remote sensing journals, which parallels Fred Prata's call at the International Forum on BRDF for an EOS article on BRDF.

Michel Verstraete commented that BRDFers and, say, foresters knew nothing about each others fields, and moreover that they shouldn't need to. Earth observation should supply only geophysical parameters to the various experts on ecological systems; only the remote sensing scientist needs to worry about instrument characteristics, atmospheric correction, BRDF, etc. In a conversation in Les Diablerets he expanded on this, suggesting the need for a third group of experts: those specialising in particular geophysical parameters (say land surface temperature or snow) who could sensibly integrate datasets which nominally described the same parameter but which had different spatial resolutions, temporal coverages, errors, etc.

Verstraete went on to observe that BRDF did not have a constituency, a community that depended on us, which resulted in users continuing to use simplistic products such as NDVI.

David Diner suggested that case studies be done to demonstrate what is lost if multiangular data are ignored, and suggested that BRDF workers should keep the users in mind.

Conclusion

The Workshop closed with an exchange on the publishing of proceedings and the timing of the next workshop. The most likely vehicle will be a special issue of Remote Sensing Reviews, edited by Michel Verstraete and Bernard Pinty, with publication expected in August 2000. It was generally agreed that meeting in one year's time would be too soon after the likely commencement of MODIS and MISR operation, and so the next Workshop would be held in 2002.