text
stringlengths
0
820
attacks) and anthropogenic (e.g. deforestation, urbanization, farming)
disturbances ( Jin and Sader, 2005 ).The ability of any system to detect change depends on its capacity
to account for variability at one scale (e.g. seasonal variations), while
identifying change at another (e.g. multi-year trends). As such, change
in ecosystems can be divided into three classes: (1) seasonal change ,
driven by annual temperature and rainfall interactions impacting plant
phenology or proportional cover of land cover types with different
plant phenology; (2) gradual change such as interannual climate
variability (e.g. trends in mean annual rainfall) or gradual change in
land management or land degradation; and (3) abrupt change , caused
by disturbances such as deforestation, urbanization, floods, and fires.
Although the value of remotely sensed long term data sets for
change detection has been firmly established ( de Beurs and Henebry,
2005 ), only a limited number of time series change detection methods
have been developed. Two major challenges stand out. First, methods
must allow for the detection of changes within complete long term
data sets while accounting for seasonal variation. Estimating change
from remotely sensed data is not straightforward, since time series
contain a combination of seasonal, gradual and abrupt changes, in
addition to noise that originates from remnant geometric errors,
atmospheric scatter and cloud effects ( Roy et al., 2002 ). ThoroughRemote Sensing of Environment 114 (2010) 106 –115
⁎Corresponding author. Tel.: +61 395452265; fax: +61 395452448.
E-mail address: [email protected] (J. Verbesselt).
0034-4257/$ –see front matter. Crown Copyright © 2009 Published by Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2009.08.014
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
reviews of existing change detection methods by Coppin et al. (2004)
andLu et al. (2004) have shown, however, that most methods focus on
short image time series (only 2 –5 images). The risk of confounding
variability with change is high with infrequent images, since
disturbances can occur in between image acquisitions ( de Beurs and
Henebry, 2005 ). Several approaches have been proposed for analyzing
image time series, such as Principal Component Analysis (PCA) ( Crist
and Cicone, 1984 ), wavelet decomposition ( Anyamba and Eastman,
1996 ), Fourier analysis ( Azzali and Menenti, 2000 ) and Change Vector
Analysis (CVA) ( Lambin and Strahler, 1994 ). These time series analysis
approaches discriminate noise from the signal by its temporal
characteristics but involve some type of transformation designed to
isolate dominant components of the variation across years of imagery
through the multi-temporal spectral space. The challenge of these
methods is the labeling of the change components, because each
analysis depends entirely on the speci fic image series analyzed.
Compared to PCA, Fourier analysis, and wavelet decomposition, CVA
allows the interpretation of change processes, but can still only be
performed between two periods of time (e.g. between years or
growing seasons) ( Lambin and Strahler, 1994 ), which makes the
analysis dependent on the selection of these periods. Furthermore,
change in time series is often masked by seasonality driven by yearly
temperature and rainfall variation. Existing change detection techni-
ques minimize seasonal variation by focussing on speci fic periods
within a year (e.g. growing season) ( Coppin et al., 2004 ), temporally
summarizing time series data ( Bontemps et al., 2008; Fensholt et al.,
2009 ) or normalizing re flectance values per land cover type ( Healey
et al., 2005 ) instead of explicitly accounting for seasonality.
Second, change detection techniques need to be independent of
speci fic thresholds or change trajectories. Change detection methods
that require determination of thresholds often produce misleading
results due to different spectral and phenological characteristics of
land cover types ( Lu et al., 2004 ). The determination of thresholds
adds signi ficant cost to efforts to expand change detection to large
areas. Trajectory based change detection has been proposed to move
towards a threshold independent change detection by characterizing
change by its temporal signature ( Hayes and Cohen, 2007; Kennedy
et al., 2007 ). This approach requires the de finition of the change
trajectory speci fic for the type of change to be detected and spectral
data to be analyzed (e.g. short-wave infrared or near-infrared basedindices). Furthermore, the method will only function if the observed
spectral trajectory matches one of the hypothesized trajectories.
Trajectory based change detection can be interpreted as a supervised
change detection method while there is a need for an unsupervised,
more generic, change detection approach independent of the data
type and change trajectory.
The purpose of this research is to develop a generic change detection
approach for time series, involving the detection and characterization of
Breaks For Additive Seasonal and Trend (BFAST). BFAST integrates the
iterative decomposition of time series into trend, seasonal and noise
components with methods for detecting changes, without the need to
select a reference period, set a threshold, or de fine a change trajectory.
The main objectives are:
(1) The detection of multiple abrupt changes in the seasonal and
trend components of the time series; and
(2) The characterization of gradual and abrupt ecosystem change
by deriving the time, magnitude, and direction of change
within the trend component of the time series.
We assessed BFAST for a large range of ecosystems by simulating
Normalized Difference Vegetation Index (NDVI) time series with
varying amounts of seasonal variation and noise, and by adding
abrupt changes with different magnitudes. We applied the approach
on MODIS 16-day image composites (hereafter called 16-day time
series) to detect major changes in a forested area in south eastern
Australia. The approach is not speci fic to a particular data type andcould be applied to detect and characterize changes within other
remotely sensed image time series (e.g. Landsat) or be integrated
within monitoring frameworks and used as an alarm system to
provide information on when and where changes occur.
2. Iterative change detection
We propose a method that integrates the iterative decomposition
of time series into trend, seasonal and noise components with
methods for detecting and characterizing changes (i.e. breakpoints)
within time series. Standard time series decomposition methods