Chapter 1: Introduction
Discussion of context of project: desire for better estimates and understanding
of land surface parameters from remote sensing data. Major example is that of
understanding scattering of radiation from land surfaces, and more specifically of land
surface albedo for climate modelling – lots of references indicating that for GCMs to
be able to isolate a weak trend such as that of an anthropomorphic 0.5 - 1o C rise in
global temperature above short term climatic variability far better estimates of surface
albedo are needed than are currently available. Given the spatial and temporal
resolution required this is only possible using EO data.
Advances in Earth Observation
In order to improve understanding of global circulation processes, concerted
effort to develop more accurate, long-term picture of global climate (e.g. NASA
MTPE). Missions/sensors developed to meet needs: examples of EOS AM-1 and PM1, in addition to ADEOS-I/II.
Land-surface processes – surface scattering, anisotropy and albedo
Recognised that most natural surface reflectances are anisotropic (early
attempts to remove directional effects from data with view/illum. angle variations). In
order to characterise surface albedo, need to be able to describe angular variation of
surface BRDF. Definition of BRDF, directional hemispherical reflectance,
bihemispherical reflectance, intrinsic properties of surface (unlike albedo). Realisation
that angular signature of surface represents another domain of information in addition
to spatial/spectral/temporal/polarisation, may have potential to yield information
regarding surface structure (structure controls anisotropy).
Surface scattering and vegetation
Lead on to particular example of vegetated surfaces – why do we need to
know about vegetation? Vegetation incredibly important part of biospheric transport
processes (CO2, biomass, absorption of moisture, energy) as well as economic
importance, hence great interest in being able to derive information regarding
biomass, fAPAR, LAI, cover fraction, and, more directly, canopy structure, from EO
data. This requires the construction of models of canopy reflectance in order to allow
interpretation of remotely sensed measurements of vegetated surfaces. Models must
have ability to describe angular variation – if sensors can be designed to explore
angular domain, possibility of retrieval of structural parameters.
Chapter 2: Review
Canopy reflectance modelling
Introduce idea of exploiting anisotropy of surface reflectance – define BRDF,
describe canopy reflectance models. Review of canopy reflectance modelling
approaches: physical (turbid medium/radiative transfer, geometric optic), empirical,
semi-empirical, numerical - Monte Carlo/radiosity. Advantages and disadvantages of
various methods, and state-of-the-art in each case, examples of use, operational
constraints.
Introduce BPMS
Introduce BPMS here – tool that will be used provide detailed analysis of
scattering within 3D description of measured canopy. Allows complete numerical
solution to scattering processes within canopy under arbitrary conditions re Monte
Carlo methods above.
Derivation of biophysical parameter information
Inversion of models. Theory – analytical/numerical. Introduce numerical
inversion techniques, advantages/disadvantages (only if going to use these methods
later e.g. in comparison of linear models with non-linear model inversion).
Linear kernel-driven models
Justification for this approach – only realistic method at the current time for
deriving BRDF/albedo estimates in near real-time and with required spatial coverage.
Describe in detail the semi-empirical approach – linear kernel-driven models. Indepth description of reasons behind this approach i.e. linear scaling (ignoring
adjacency effects); rapid inversion using matrices (leading to predictable error
estimates); pre-calculation of angular integrals of kernel parameter values, hence use
LUT approach to calculation of albedo; no assumption of homogeneity so flexible for
different surface types, and mixed pixel cases, so ideal for global moderate resolution
products; able to attach different physical meanings to weightings (e.g. areal
proportion).
Aims
Describe aims of thesis – try and provide a framework for detailed testing of
canopy reflectance modelling methods, specifically by investigating how linear
models describe observed canopy reflectances, but more generally using BPMS to
investigate type of information contained in BRDF model parameters. In case of
linear models, is the linear kernel-driven approach valid as a starting point? If so,
what information can be inverted from the models? Can inverted parameters be used
“as is”, or is structural information contained in some different form? Investigate
ways of formulating better kernel-driven models based on problems associated with
current limitations of kernels. Test such methods on real data (e.g. CASI, AVHRR,
SPOT etc.).
Chapter 3: Fieldwork data and validation
Description of data collected, in particular BPMS data, and estimates of LAI,
% cover etc. methods used, difficulties involved. Also, EO data: aerial photography,
CASI, ATM, SPOT, AVHRR, POLDER.
Validation
Describe simulation process in detail. Attempts at validation of fieldwork data
i.e. BPMS reflectance simulated at same viewing and solar illumination conditions in
an effort to allow comparison with radiometry data (difficulty of weather conditions
in obtaining at least some reasonable radiometric measurements, small FOV, and
variable illumination conditions). Simulate reflectance using same FOV to assess
impact that has on measured reflectance (could also simulate with 15o FOV to see
how much of the directional signal would be lost in this case)). Comparison of
measured LAI and % cover values with BPMS-derived values. LAD from LAI-2000,
and gap-probability.
Sensitivity analysis of BPMS data i.e. try and quantify the effects of
measurement errors on directional reflectance, LAI and % cover. Suspect most
sensitivity due to errors in tiller azimuth/zenith angle and leaf angle distribution
measurements, causing incorrect amounts of sunlit/shaded canopy components in
simulations.
Chapter 4: Investigation of linear kernel-driven approach
Linear assumption
Investigation of linear models ability to model canopy reflectance as a linear
combination of a volumetric and geometric-optic component. Justification from
theory? Does this assumption hold? If not, then serious reservations about the whole
approach.
Simulate BPMS reflectance using same assumptions as in formulation of
linear models (refl=trans, single scattering, Lambertian soil). Extract volumetric and
geometric-optic components of reflectance (GO=rohsoil*prop. sunlit soil, so vol =
rohcanopy-GO - this is done independent of wavelength). Compare two components
directly with linear model volumetric and GO kernel parameter values. Regress
results of reflectance simulations from range of solar zenith and row azimuth angles
against linear model values.
Important factors are: do BPMS-modelled components show behaviour we
would expect of the two separate components? Do row azimuth and solar zenith angle
variations make much difference? Do the relationships change as we would expect as
the canopy develops i.e. does the volumetric parameter start small in comparison with
the GO parameter, and increase as the canopy develops, and vice-versa? Examine
possibility that relationship between BPMS-derived and linear model is not
necessarily linear (results indicate this in some cases) – possibility of e.g. exponential
or x2 relationship between BPMS-derived and linear model parameters. Is there a
theoretical justification for such a hypothesis? In the more extreme cases (i.e. across
rows, and at higher solar zen. angles, in sparse canopy) GO parameters appear to be
less linearly related than vol.
Linear models 2
Seperability of canopy reflectance components i.e. do the model parameters
act independently of each other in describing volumetric and GO components of
canopy reflectance? Compare results of generating forward modelled canopy
reflectance using estimates of volumetric and GO components derived previously
with those generated using linear models. If the components are perfectly separable
they should be the same – is this the case? How do they differ? Is there a possibility
that the component of canopy reflectance not modelled by one parameter is modelled
by the other? In this case, the volumetric parameter describes an element of the GO
scattering and vice versa. Leads on to looking at simulating reflectance considering
only GO scattering, and only volumetric scattering independently.
[Look at reciprocity – what is the justification for how the kernels have been
adapted to be reciprocal (no problem with why it is done, but how it was done)? Is
there justification for other methods of providing reciprocity?]
Chapter 5: Isolating GO and volumetric scattering components of canopy
reflectance, assumptions made in formulations of kernels
Look at assumptions behind the formulation of the two components of canopy
reflectance in linear modelling approach.
GO component
Examine how realistic it is that a vegetation canopy can be described as GO
scattering primitives, even one containing well-defined vegetation ‘envelopes’ such as
a forest canopy of discrete crowns. Is this appropriate for describing the GO
component of reflectance from other types of canopy? Look at approach taken in
Roujean and Li kernels.
Start by looking at simplest cases i.e. blocks on a plane as in Roujean kernel.
Look at scattering behaviour of solid blocks arranged in rows (mimic the row nature
of the barley canopy for example). Will have very different scattering properties
depending on viewing and illumination azimuth angle (along or across rows), due to
changes in shadowing proportions (e.g. if rows are infinite in length, there will be no
shadowing looking along the rows in the principle plane, whereas in the xpp,
reflectance will be dominated by shadowing, depending on size of blocks and row
spacing).
Abstraction to solid objects
Abstraction of GO component of canopy reflectance to arrangement of
ellipsoids on a surface. Considering individual ellipsoids: what dimensions should an
ellipsoid have to generate the same proportions of sunlit/shadowed soil/crown as for a
real plant? This relationship directly controls the contribution of GO scattering to
canopy reflectance. Can a relationship between dimensions and shadowing be found
that is general (give or take the natural variation of the plant sizes and shapes) and
holds at all viewing and illumination angles? If not, how much does it vary, and how
(un)realistic is it to model a plant as an ellipsoid of fixed dimensions, given variations
of viewing and illumination? Compare with cases using b/r values of 0.75 and 2.5,
and h/b ratios of 1 and 2 suggested by Wanner et al. in formulation of Li kernels
Examine distribution of objects on plane i.e. random as compared with regular
distribution. Consider the cases of mutual/no shadowing and examine differences in
scattering behaviour compared with ‘real’ reflectance. In what case are the two
approximations more appropriate (dense/sparse) and do these conform to those cases
specified in model formulation? Young barley (sugar beet?) is a reasonable
approximation to a sparse canopy, and mature barley/wheat would appear to be
reasonable approximation to a dense canopy. Do simulations using ellipsoids yield
reflectance ditributions in any way related to that of real canopy?
Volume scattering component
Examine the assumptions made in the volumetric (RossThick and RossThin)
kernels i.e. azimuthally invariant, horizontally homogenous semi-infinite assembly of
randomly oriented infinitesimal scatterers, approx. to solution of rad. trans. for single
scattering. Is it realistic to model volume scattering component of canopy refl. in this
way, and how far from this ideal is the scattering behaviour of a real canopy? How
much of canopy structural information contained in mulitple scattered radation?
Simulate reflectance from layer of randomly orientated scatterers (disks for
e.g.) – examine: single scattering approx., phase function of medium, impact of size
of scatterers, orientation distribution, and density function on scattering properties.
Abstraction to volumetric objects
Abstraction of scattering behaviour to volumetric objects i.e. model GO and
volumetric scattering components of reflectance represented in kernel-driven models
by using volumetric ellipsoids. Compare with simulated canopy reflectances – same
questions as in case of solid objects: what dimensions do ellipsoids require to exhibit
same scattering behaviour as real plants? What phase function, leaf area density, and
leaf angle distributions are appropriate? Are these comparable to those exhibited by
the real plants? Look at distribution of objects, as before. At what point does either
type of scattering become dominant, and do the linear models handle this
appropriately - is there always a component of both in scattering from a real canopy,
or can they be isolated? Possible to develop a method of deciding how inverted model
parameters can be interpreted.
Operationally, AMBRALS has been reduced to Iso + RossThick + LiSparse –
can this combination possibly hope to describe wide range of surfaces likely to be
encountered in real data? Even if it does, can any meaningful biophysical information
be inverted using this model configuration? There is a worry that in reducing the
operational model choice to one kernel comibnation, the flexibility of choice that
using RMSE of inversion to decide the combination is lost. There may be cases where
if only correction for BRDF effects are required (rather than model parameter
information) then a purely empirical model may produce better results than the
RossThick LiSparse combination.
Chapter 6: Comparison of scattering behaviour treated by non-linear model e.g.
Kuusk model
Examine the way in which a non-linear canopy reflectance model treats
scattering from vegetation, and compare with linear model treatment. For e.g. look at
Kuusk model – assumptions made are mostly same as in Ross kernels (azimuthally
random distribution of leaf normals, horizontally homogenous and infinitely
extended). Kuusk model considers both single and multiple scattering, as opposed to
just single in linear models. Also has modification to Ross’s treatment of gap
probability for hot-spot direction. The probab. of light ray exiting a gap in the canopy
is modified by a correlation function to account for greater probability of exiting in
direction of viewer i.e. hot spot direction. Mutiple scattering treated as analytic
approximation to solution of radiative transfer in homogenous medium.
Determine what proportion of directional signal from canopy (containing info.
on canopy structure) is contained in single and multiple scattered components of
reflectance respectively. Is mutiple scattered component large enough compared with
single scattered component to invalidate single scattering approximation of linear
models invalid? Is hot-spot treatment of Kuusk model capable of describing hot-spot
features seen in simulated reflectances in more detail than simple shadowing
treatment of Li kernels (or no hot-spot treatmant at all in Ross kernels)?
Chapter 7: Application of linear models to EO data
Look at application of linear models to EO data (aerial, SPOT, AVHRR, poss.
POLDER). Data at various scales would allow testing of the assumption that scaling
properties of linear models permits reprsentation of inhomogenous surfaces at range
of scales (ignoring adjacency effects). Limited by the small size of the field area, but
could extend this to AVHRR (or could degrade aerial data to AVHRR scale in order
to simulate lower resolution data). Invert linear models against reflectance data attempt to obtain biophysical parameter information based on knowledge of the
operation of linear models. cf with field data of these areas (limited by lack of
directional sampling in aerial data – provided (hopefully) by overlap of ground tracks
- AVHRR data prob. better source of variation in viewing and illumination but
problem of resolustion and size of field site). Do inverted parameters contain i)
information expcted from model formulation ? ii) information as predicted by
findings from previous experiments? How far are previous results borne out by
inversion against real data?