Breakpoint analysis with the BFAST algorithm applied to global vegetation indexLaura Holtzman1, Kirsten M. de Beurs1

1Geography and Environmental Sustainability, College of Atmospheric & Geographic Sciences, University of Oklahoma

lauraholtzman@ou.edu, kdebeurs@ou.edu

OverviewDetecting abrupt changes in time series of remotely sensed data is animportant approach to monitoring land use and land cover change. Timeseries change detection can be used to analyze several types of data includingtemperature, carbon emissions and NDVI.NDVI is Normalized Difference Vegetation Index, derived from Near infraredand red bands of satellite data and measures health of vegetation. NDVIranges from -1 to 1, where values closer to 1 indicates healthier and denservegetation.

In this study, we will use BFAST on the global GIMMS3G NDVI product to testthe sensitivity of the BFAST algorithm in determining break changes.

Objective

What is BFAST?

Figure 1: Median NDVI values for the adjusted GIMMS3G data from January 1982 to December 2011

Study RegionsFigure 1 (below): Land cover images of the contiguous United States and Australia. Images courtesy of the National Land Cover Dataset and Australian Government.

Case 1

Figure 2 (below): landcover of Case Study 1.Yellow box highlightspixel location.

DataGIMMS3G• This project used global NDVI values from the Global Inventory Modeling

and Mapping Studies (GIMMS).• GIMMS scaled values are:

• NDVI = 0.0 – 10000• Water = -10000• Null = -5000

• Temporal Resolution = 15 day composites (2 observations a month).• Spatial Resolution = Resampled to a 1 degree lat/lon resolution.• Dates Observed = Adjusted to January 1982 – December 2011.

Future Work

References:

ConclusionsBFAST is able to detect abrupt changes in locations where events that affectNDVI occur. Issues that arise with using BFAST:

• breakpoints exist that cannot be explained• magnitude results are difficult to interpret• amount of breaks inputted can not be individually adjusted for trend and

seasonal componentCase Study 1: A breakpoint that occurs in 2007 (Fig. 8) indicating an abruptdecrease in NDVI is not apparent in the drought monitor graph (Fig. 6).Case Study 2: The greatest magnitude change occurred in 2010, which hasgreater decrease in NDVI values than the Yellowstone National Park Wildfire in1988.Case Study 3: A break in seasonality occurs, however there is no option toadjust the number of breaks in the seasonal component separate from thetrend.It is possible that there was a different driver affecting NDVI values for theissues above.

• Apply BFAST to entire global time series and compare breakpoints using other change detection methods.

• Derive an algorithm that analyzes breakpoints that are not driven byseasonality.

Case StudiesCase 2: Yellowstone National Park, Wyoming

44°59'59.94"N, 110°59'59.94"W

Case 3: Southern Australia30°0'0.06"S, 141°0'0.06”E

Figure 6 (above): U.S Drought Monitor Statistics Graph for Beckham County, OKfrom January 2000 to December 2011. Courtesy of the National DroughtMitigation Center.

Case 1: ResultsComparing the BFAST output for breakpoints (Fig. 7) to thedrought monitor (Fig. 6), BFAST correctly identified droughts in2006 and 2011, specifically recognizing that 2011 had a significantmagnitude decline in NDVI values. However, a breakpointoccurred in September 2007, which cannot be explained by Fig. 6.All the breakpoints that occur have a negative slope, indicatingthat the each step change has decreasing values in NDVI.

Case 2: ResultsThe original time series (𝑌𝑡 in Fig. 11) depicts a strong seasonality,where NDVI values are lowest during the winter and the highestduring the summer. Four breaks occur in July 1988, June 1994,June 2007 and November 2010. Fig. 9 shows Yellowstone NationalPark was severely in drought. BFAST detects the most significantdrop in NDVI values occurs in 2010, while the drought monitor(Fig. 10) shows minimal drought in 2010.

Case 3: ResultsBFAST calculated an abrupt increase in NDVI values in February2010. In January 2010, Fig. 13 shows that rainfall was belowaverage in the area of interest and received 400% its averagerainfall in February 2010 (Fig. 14). The seasonal time series (𝑆𝑡 inFig. 15) displays a seasonal shift after the breakpoint occurs.Originally, the seasonality varied only slightly. After the seasonalbreak, the seasonal time series shows a strong seasonality.

100.00%

80.00%

60.00%

40.00%

20.00%

0.00%

Figures 9 and 10 (above): Palmer Drought Severity Index for 1988 and StatisticsGraph for Park County, WY from January 2000 to December 2011. Both courtesyof the National Drought Mitigation Center.

• Breaks For Additive Seasonal Trend is an additive decomposition linearmodel that decomposes time series accordingly: 𝑌𝑡 = 𝑇𝑡 + 𝑆𝑡 + 𝑒𝑡 .• 𝑌𝑡 is observed data at time t, 𝑇𝑡 is the trend component, 𝑆𝑡 is the

seasonal component, 𝑒𝑡 is the remainder component or residualcomponent (output is shown on Fig. 7, 11, and 15).

• Determines magnitude and date of significant breakpoints in time series• Ordinary least squares residuals-based moving sum test is applied to the

trend and seasonal component to see if breakpoints occur.• The BFAST algorithm allows the user to set a minimal time between

breakpoints and number of breaks to identify.• Analyses were carried out using R statistical software: http://bfast.r-forge.r-

project.org/ (Verbesselt et al. 2010)

Figure 4 (below): landcover of Case Study 3.Yellow box highlightspixel location.

2001 2003 2005 2007 2009 2011

100.00%

80.00%

60.00%

40.00%

20.00%

0.00%

Figure 8 (above): Negative step change at each breakpoint.

3000

4000

5000

6000

2001 2003 2005 2007 2009 2011

Verbesselt, J., Hyndman, R., Newnham, G., & Culvenor, D. (2010). Detectingtrend and seasonal changes in satellite image time series. Remote sensing ofEnvironment, 114(1), 106-115.

1985 1990 1995 2000 2005 2010

3000

2000

4000

5000

1985 1990 1995 2000 2005 2010 1000

2000

3000

4000

To recover NDVI values ranging -1 to 1 use NDVI = float(raw/10000)

10000

0

Figure 7 (above): BFAST output. The minimum time between breakpointsinputted is one year and the number of breaks inputted is three.

1985 1990 1995 2000 2005 2010

10

00

40

00

𝑌 𝑡𝑒 𝑡

𝑇 𝑡𝑆 𝑡

15

00

3

00

0

-40

0

0

80

0-5

00

0

1

00

0

Case 1: Western Oklahoma and North Texas34°59'59.94"N, 99°59'59.94"W

2000 2005 2010

Figures 13 and 14 (above): Rainfall percentage for Australia in January 2010 andFebruary 2010. Both courtesy of the Bureau of Meteorology for the Australiangovernment.

Figure 11 (above): BFAST output. The minimum time between breakpointsinputted is one year and the number of breaks inputted is four.

Figure 15 (above): BFAST output. The minimum time between breakpointsinputted is one year and the number of breaks inputted is one.

Figure 12 (above): Negative step change at each breakpoint. Figure 16 (above): Positive step change at the breakpoint.

Below: Deseasonalized time series (black line) with trends (green lines) and breakpoints (dashed green lines) where a purple line signifies the slope of values during the breakpoint and a red line indicates the breakpoint with greatest magnitude of change

Figure 5: Median NDVI values for the adjusted GIMMS3G data from January 1982 to December2011

1985 1990 1995 2000 2005 2010

-20

00

0

10

00

20

00

7

00

0𝑌 𝑡

𝑒 𝑡𝑇 𝑡

𝑆 𝑡3

00

0

6

00

0

-20

00

0

30

00

2000 2002 2004 2006 2008 2010 2012

𝑌 𝑡𝑒 𝑡

𝑇 𝑡𝑆 𝑡

30

00

70

00

40

00

60

00

-15

00

0

1

00

0-5

00

10

00

Case 2 Case 3

Figure 3 (below): landcover of Case Study 2.Yellow box highlightspixel location.

400 %300200150125100806040200

400 %300200150125100806040200

Percentage of Mean Percentage of Mean