HR: 1340h
AN: B43B-0275 [Abstracts]
TI: A User's Version View of a Robustified, Bayesian Weighted Least-Squares Recursive Algorithm for
Interpolating AVHRR-NDVI Data: Applications to an Animated Visualization of the Phenology of a
Semi-Arid Study Area
AU: * Hermance, J F
EM: John_Hermance@brown.edu
AF: Geological Sciences Department, Box 1846, Brown University, Providence, RI 02912
United States
AU: Jacob, R W
EM: Robert_Jacob@brown.edu
AF: Geological Sciences Department, Box 1846, Brown University, Providence, RI 02912
United States
AU: Bradley, B A
EM: Bethany_Bradley@brown.edu
AF: Geological Sciences Department, Box 1846, Brown University, Providence, RI 02912
United States
AU: Mustard, J F
EM: John_Mustard@brown.edu
AF: Geological Sciences Department, Box 1846, Brown University, Providence, RI 02912
United States
AB:
In studying vegetation patterns remotely, the objective is to draw inferences on the development of specific or general land
surface phenology (LSP) as a function of space and time by determining the behavior of a parameter (in our case NDVI), when
the parameter estimate may be biased by noise, data dropouts and obfuscations from atmospheric and other effects. We describe
the underpinning concepts of a procedure for a robust interpolation of NDVI data that does not have the limitations of other
mathematical approaches which require orthonormal basis functions (e.g. Fourier analysis). In this approach, data need not
be uniformly sampled in time, nor do we expect noise to be Gaussian-distributed. Our approach is intuitive and
straightforward, and is applied here to the refined modeling of LSP using 7 years of weekly and biweekly AVHRR NDVI data for
a 150 x 150 km study area in central Nevada. This site is a microcosm of a broad range of vegetation classes, from irrigated
agriculture with annual NDVIvalues of up to 0.7 to playas and alkali salt flats with annual NDVI values of only 0.07.
Our procedure involves a form of parameter estimation employing Bayesian statistics. In utilitarian terms, the latter
procedure is a method of statistical analysis (in our case, robustified, weighted least-squares recursive curve-fitting) that
incorporates a variety of prior knowledge when forming current estimates of a particular process or parameter. In addition
to the standard Bayesian approach, we account for outliers due to data dropouts or obfuscations because of clouds and snow
cover. An initial "starting model" for the average annual cycle and long term (7 year) trend is determined by jointly fitting
a common set of complex annual harmonics and a low order polynomial to an entire multi-year time series in one step. This is
not a formal Fourier series in the conventional sense, but rather a set of 4 cosine and 4 sine coefficients with fundamental
periods of 12, 6, 3 and 1.5 months. Instabilities during large time gaps in the data are suppressed by introducing an
expectation of minimum roughness on the fitted time series.
Our next significant computational step involves a constrained least squares fit to the observed NDVI data. Residuals between
the observed NDVI value and the predicted starting model are computed, and the inverse of these residuals provide the
weights for a weighted least squares analysis whereby a set of annual eighth-order splines are fit to the 7 years of NDVI
data. Although a series of independent 8-th order annual functionals over a period of 7 years is intrinsically unstable when
there are significant data gaps, the splined versions for this specific application are quite stable due to explicit
continuity conditions on the values and derivatives of the functionals across contiguous years, as well as a priori
constraints on the predicted values vis-a-vis the assumed initial model.
Our procedure allows us to robustly interpolate original unequally-spaced NDVI data with a new time series having the
most-appropriate, user-defined time base. We apply this approach to the temporal behavior of vegetation in our 150 x 150 km
study area. Such a small area, being so rich in vegetation diversity, is particularly useful to view in map form and by
animated annual and multi-year time sequences, since the interrelation between phenology, topography and specific usage
patterns becomes clear.
DE: 0430 Computational methods and data processing
DE: 0434 Data sets
DE: 0439 Ecosystems, structure and dynamics (4815)
DE: 0466 Modeling
DE: 0480 Remote sensing
SC: Biogeosciences [B]
MN: Fall Meeting 2005