1. Noisiness of Retrieved Parameters. The anisotropic parameters retrieved on model inversion - and particularly the volume scattering parameter - can be rather unstable, owing to the sparse angular sampling provided by the AVHRRs. This means that modelling to certain sun-sensor geometries is not advisable; features may appear in modelled scenes which are related to sharp differences in noise inflation between adjacent locations rather than to features on the ground, even though the modelled values are reasonable estimates of bidirectional reflectance (at the specified geometry) when taken in isolation from neighbouring values.
Previous research into inverting linear semiempirical BRDF models with AVHRR data found the anisotropic parameters to be very noisy; however, this considered only data from a PM sensor (Leroy and Roujean, 1994). Flasse et al (1993), used data from the AVHRR on NOAA-11 (PM) to invert the nonlinear semiempirical RPV model (Rahman et al, 1993a,b), with good fits to observations but no quantitative assessment of the reliability of the (3) parameters. Privette et al (1996) used data from both NOAA-9 (PM) and -10 (AM) AVHRRs but used a turbid medium model (DISORD) and SOILSPECT with rather stringent parameterisation requirements (many adjustable parameters). Vives Ruiz de Lope and Lewis (1997) assessed the performance of the AMBRALS suite of linear semiempirical models over the HAPEX-Sahel supersites and obtained good fits to observations and reasonable temporal parameter trajectories; they used observations from NOAA-11 and -12 although the majority were from the PM sensor.
In the case of the data used for this study, the absence of observations from
just one AM overpass as a result of cloud and cloud-shadow screening leads to a
palpable increase in noise in the volume scattering parameter which is carried
through in modelling, although it only becomes an important source of error when
modelling to certain target geometries (interpolation or extrapolation). Note that
for this study data from only 4 NOAA-AM overpasses was available for model inversion
over the 17-day period (with data from 17 NOAA-PM overpasses making up the total of
21 scenes). Since receiving stations are capable of receiving HRPT data from
similar numbers of AM and PM overpasses, the real constraint is cloud contamination
rather than angular sampling per se. One of the main findings of this
research is that noise in model parameters is likely to be acceptable as long as all
available data from both AM and PM AVHRRs is used.
2. Retrieval of Negative Parameters. One problem which is
sometimes encountered in using simple linear semiempirical BRDF models is
that the parameters retrieved can be negative, which prevents any physical
interpretation. This happens because the best fit of observations to the
model is found when a kernel is inverted (i.e. used upside-down); this
is probably owing to canopy scattering processes which are not explicitly
accounted for in the models, the most important of which are multiple
scattering (where incoming radiation interacts with more than one
surface element via reflection or transmission from the first element
encountered) and specular effects (where radiation in a broad
region of the solar spectrum is reflected with little interaction with the
plant leaves or soil surface). For semiarid grassland applications both
these processes may play a role in shaping BRDF : grass leaves have a high
transmittance in the near-infrared wavlengths as well as high reflectance,
resulting in an increasing contribution in the forward-scattering
direction with increasing view zenith angle; they can also be shiny or
have shiny components such as the flower of the species Stipa grandis, resulting in increased
reflectance in the specular direction. However, since the rate and amount
of the increase in the forward-scattering direction appears to be related
to the density and physical structure of the vegetation, at least in
semiarid grasslands, negative parameters may still carry information on
the canopy which can be used in an implicit rather than explicit manner.
In addition, a kernel can be added to the basic three-kernel model to
account for multiple scattering effects; this has been shown to improve
model fits to observations in some validation experiments, although more
research is clearly required. For the semiarid grasslands investigated
here, the problem of negative parameter retrieval only afflicts linear
semiempirical models which incorporate the RossThick kernel (or a linear
scaling of it); those which include the RossThin kernel (e.g.
LiSparseMODIS-RossThin or Roujean-RossThin) do not demonstrate this
behaviour.
