IEEE Geoscience and Remote Sensing Magazine - September 2019 - 117

signed-rank test [38] was used to evaluate the difference
between the two GPval and GPtrain pairs. We obtained P
values greater than 0.1, indicating that, for all scenarios,
the results from validation data did not differ significantly from the test data, suggesting that we have not faced
overfitting problems in our experiment. All variations of
the proposed approach overcame all baselines, with computed P values lower than 0.01.
For almost all dissimilarity-measure classifiers used in
isolation, increasing the noneucalyptus sample size led to
worse classification accuracies. This phenomenon was already expected, because increasing sample-size values produce a group of noneucalyptus points with more diversity
(e.g., with pixels representing other vegetation species, pastures, urban construction, or water from rivers), potentially
hindering the classification performance. Finally, GPraw
and GPdec were also tested through the paired Wilcoxon
signed-rank test (P > 0.1), indicating that no statistical difference was observed.

The h parameter is the minimal segment size between potentially detected breaks in the trend model given as fraction relative to the sample size (i.e., the minimum number
of observations in each segment divided by the total length

AMAZON BASIN
In Figure 8, we show the spatial distribution of the number of extreme drought events in the Amazon Basin (i.e.,
the number of times the SPI value was below -1.5 for the
entire analyzed period). Extreme events were concentrated
in the central, northwestern, and southwestern regions. In
the southwestern region, we highlight a point (previously
described in the "Data" section under "Amazon Basin")
that faced the greatest number of extreme events-116 in
the period from 1984 to 2018.
In Figure 13, we present the spatial distribution of the SPI
for July 1992, April 1998, and 2010 May as well as the SPI
values for the highlighted point. Due to a limitation of our
collected Landsat data (available from 1984 onward), we did
not analyze the effects of the 1982/1983 drought. We also
did not consider the 2015/2016 drought in our analysis due
to the lack of time needed to fully investigate how the vegetation responded (and is still responding) to this drought.
In general, we observe that the 1991/1992 and 1997/1998
droughts extend over a broader area than the 2010 drought.
For the 2010 drought, we notice two main drought outbreaks-one in the northwest and another in the southwest. By analyzing the SPI values of the highlighted point,
we conclude that it was heavily exposed to the 2010 drought
(SPI = -4.07), followed by the 1997/1998 drought (SPI =
-1.72), and was least affected by the 1991/1992 drought
(SPI = -0.29).
In Figure 14, we show the EVI2 time series for the highlighted point presented in Figure 8. The time-series data
range from 1983 May to November 2018. Green, orange,
and purple points in the time series indicate data captured
by Landsat mission 5, 7, and 8, respectively. The plot also
presents geographical coordinates of the selected point:
latitude (-8.6260°) and longitude (-69.3750°).
In Figure 15, we present the outputs of the BFAST algorithm considering two distinct values for the h parameter.

FIGURE 11. The MODIS NDVI time series for the selected regions A
(red) and B (green).

SEPTEMBER 2019

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

NDVI

0.8
0.6
0.4
0.2
Eucalyptus

2005

2010
Time

2015

0.9
0.7
0.5
0.3

0
−0.02

0.9
0.8
0.7
0.6
0.5
0.4
0.3

0.2
0
−0.3

2000

Noneucalyptus

2005

2010
Time
(a)

2015

0.75
0.65
0.55
0.45

0.05

2000

−0.05

0.75
0.65
0.55
0.45
et

0.10
0
−0.10

2000

2005

2010
Time
(b)

2015

FIGURE 12. The decomposition and structural-break-detection

outputs of NDVI time series for regions (a) A and (b) B using the
BFAST algorithm.
117

IEEE Geoscience and Remote Sensing Magazine - September 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2019