Methods
MODIS Observations. Our study encompasses 16 MODIS tiles (h10v08 and
h13v11), which is a land area of 16.75 million km2 spanning 10° N to 30° S in
latitude and 80° W to 42° W in longitude. The area is characterized by
seasonal tropical savannah and seasonal deciduous forest in the north and
south and tropical evergreen forest in the equatorial region. Tropical
regions frequently experience high cloud cover and high variability in atmospheric
aerosols because of seasonal biomass burning (59). These atmospheric
conditions are challenging to assess with traditional algorithms that
rely on spectral and thermal reflectance thresholds alone (30). MAIAC is
a new generation cloud screening and atmospheric correction algorithm
that uses an adaptive time-series analysis and processes groups of pixels to
derive atmospheric aerosol and surface reflectance without typical empirical
assumptions. MAIAC implements a sliding window approach, storing up to
16 d of MODIS observations gridded to 1-km resolution. These data are used
to derive a spectral regression coefficient relating surface reflectance in the
blue (0.466 μm) and short-wave IR (2.13 μm) for aerosol retrievals (60) and
obtain parameters of surface bidirectional reflectance distribution function
(61) in the MODIS reflective bands (31). Using time-series analysis over 4–16 d,
MAIAC is able to separate stable surface background with its characteristic
spatial pattern from generally random and changing fields of clouds (62).
MAIAC also features a dynamic land–water–snow classification and change
detection algorithm to account for rapid and seasonal changes in surface
reflectance of the time-series data (31). Currently, MAIAC is undergoing
operational code conversion and testing, and it is expected to become an
operational MODIS algorithm in 2014. For this work, we used MODIS C6
Level 1B (calibrated and geometrically corrected) data, which removed
major sensor calibration degradation effects present in earlier collections.
Detailed descriptions of MAIAC and quality testing (31, 60) as well as an
assessment of errors and uncertainties over the tropical regions (30) are
provided elsewhere.
Precipitation. Monthly rainfall was obtained from the TRMM at 0.25° spatial
resolution between 2000 and 2012 (3B43, version 7) (13). TRMM provides
monthly estimates of precipitation in millimeter hour−1. The 3B43 dataset
combines microwave images with the IR-based Geostationary Operational
Environmental Satellite Precipitation Index and a network of gauge data
(63). Although previous research (59) has shown good agreement with independent
rain gauge data in the Brazilian Amazon, rainfall is somewhat
underestimated during very wet months (>300 mm mo−1). TRMM data were
resampled to match MODIS data at 1-km spatial resolution using a bilinear
interpolation technique. Estimates of monthly precipitation (P) were normalized
by subtracting the mean annual precipitation of 2000 and 2002 as
baseline from monthly data (58). Net changes were then assessed by summation
of the normalized changes to assess gains or losses over time,
where n is the total number of monthly observations. Although we recognize
that the magnitude of trends will depend on the selection of the
baseline, the goal of this analysis was to investigate sensitivity of vegetation
to changes in precipitation rather than establish long-term drying or wetting
trends. These years were selected, because they both represent non-El Niño
events and therefore, fairly normal conditions. We have also tested the
single year (2000) and the whole-record average baseline approach and
found little effect on spatial domains of detected change.
Changes in TWSC. Changes in TWSC were obtained from the Global Data
Assimilation System, a land surface modeling system that integrates data
from advanced Earth-observing satellite- and ground-based observations to
support and improve hydrometeorological predictions (64). Monthly TWSC
estimates were obtained at 1° spatial resolution between 2000 and 2012
from the Earth Sciences Data and Information Services Center from NASA
(grace.jpl.nasa.gov). Net changes in TWSC were derived analogous to
changes in precipitation.
Time-Series Analysis. Time-series analysis of NDVI was used to derive trends in
vegetation over time (37). Several methods exist for assessing changes in
NDVI time series, including those based on simple thresholds or more comprehensive
time-series models. Here, time-series fitting is based on the
TIMESAT approach (37). Originally developed for the NDVI series from the
Advanced Very High Resolution Imaging Spectroradiometer, the technique
has been adapted for MODIS (65). The Savitzky–Golay filter method was
used to fit splines to a temporally moving window of satellite observations.
The use of this adaptive filtering method allows fitting of time series of
observations without assuming data stationary. To take into account that
most noise in NDVI is negatively biased, the algorithm was fitted to the
upper envelope of the observed NDVI (37). Net changes in NDVI were
obtained analogs to changes in precipitation (Eq. 1). MODIS data before late
February of 2000 were substituted with corresponding monthly averages
from 2001 to 2002 to obtain a full set of observations for the year 2000.
ACKNOWLEDGMENTS. We thank Dr. Richard H. Waring (Oregon State
University) for helpful discussions and comments and Dr. Lars Eklundh (Lund
University) for help with the time-series algorithm. We also thank the NASA
Center for Climate Simulation for computational support and access to their
high-performance cluster. This work was supported by the Science of Terra
and Aqua Program of NASA (A.I.L. and Y.W.).