HR: 0800h
AN: B41A-0093    [Abstracts]
TI: A Robust, Bayesian Approach to the Analysis of Vegetative Phenology Using Satellite Vegetative Indices in the Presence of Outliers, Noisy Data and Data Gaps
AU: * Jacob, R W
EM: Robert\_Jacob\@brown.edu
AF: Dept. of Geological Sciences, Brown University Box 1846, Providence, RI 02912 United States
AU: Bradley, B
EM: Bethany\_Bradley\@brown.edu
AF: Dept. of Geological Sciences, Brown University Box 1846, Providence, RI 02912 United States
AU: Hermance, J F
EM: John\_Hermance\@brown.edu
AF: Dept. of Geological Sciences, Brown University Box 1846, Providence, RI 02912 United States
AU: Mustard, J
EM: John\_Mustard\@brown.edu
AF: Dept. of Geological Sciences, Brown University Box 1846, Providence, RI 02912 United States
AB: Vegetated cover is affected over time by such factors as long and short term climate changes, inter-annual climate variability, changes in the hydrologic cycle, and anthropogenic land cover change. Although remote sensing has revolutionized the way vegetation changes are observed, the full potential is often hampered by missing data due to instrumentation, weather conditions (clouds) or ground cover (snow). These breaks in the time series make it difficult, if not impossible, to employ a number of classical time series methods - such as standard Fourier analysis - to characterize the phenology of vegetation for individual years and between years. We have constructed a recursive least-square (LS) algorithm, drawing on concepts from Bayesian statistics, that generates a robust polynomial interpolation of a noisy, biased, discontinuous data set. The heart of our procedure employs an asymmetric weighted LS technique to jointly (simultaneously) fit time series data on a pixel-by-pixel basis to two classes of polynomials. One polynomial accounts for long-term change and inter-annual fluctuations. The second class of polynomials consists of a set of annual 8-th order spline functions constrained to observe certain continuity constraints between years. Estimated data points that are substituted for missing data are down-weighted by an adjustable measure to influence the high-order intra-annual curve-fit. Thus the curve-fit model relies mainly on observed data, when they are available, but is stabilized by the estimated data. Preliminary testing of our algorithm demonstrated its stability was vulnerable when dealing with data sets having substantial gaps over extended periods of time. Our solution was to pre-condition these data gaps by introducing estimated data based on an a priori estimate of the average annual cycle plus long term, low-order variations for each pixel. In addition, a minimum roughness criterion was invoked for the average annual cycle during times when the expectation of data gaps was highest. Currently, we represent the mean annual cycle by the first four annual harmonics (T = 12, 6, 4 and 3 months). To evaluate the algorithm, we applied it to satellite vegetation data from a semi-arid ecosystem in the Great Basin. The data consisted of 1 km$^{2}$ pixels of weekly Normalized Difference Vegetation Index (NDVI) from the Advanced Very High Resolution Radiometer (AVHRR) between 1990-2001. Long-term sensor drift - typical of NDVI data - is accounted for by removing a jointly estimated average trend plus annual fluctuation based on non-vegetated regional salt flat data. Therefore, the resultant data set consists of "differential NDVI data" (DNDVI) - the difference between observed NDVI values and those average NDVI values from regional non-vegetated areas. Conditioning the DNDVI "signal" with this algorithm makes it possible to clearly visualize patterns in the time series of each pixel, as well as to continuously simulate temporal changes over a gridded area. One can accordingly identify the overall vegetation phenology, specifically the timing of green-up, peak greenness, senescence, and duration of greenness, among other characteristics; thus setting the stage for identifying and discriminating among, principle vegetative classes.
DE: 1640 Remote sensing
DE: 0400 Biogeosciences
DE: 0910 Data processing
DE: 0933 Remote sensing
SC: Biogeosciences [B]
MN: 2004 AGU Fall Meeting