Remote Sensing of Environment 229 (2019) 1–13

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Hurricane damage detection on four major Caribbean islands T

a,⁎ a a b,c
Kirsten M. de Beurs , Noel S. McThompson , Braden C. Owsley , Geoffrey M. Henebry
a
Department of Geography and Environmental Sustainability, University of Oklahoma, United States of America
b
Department of Geography, Environment, and Spatial Sciences, Michigan State University, United States of America
c
Center for Global Change and Earth Observations, Michigan State University, United States of America

A R T I C LE I N FO A B S T R A C T

Keywords: Tropical cyclones are natural events that transform into natural disasters as they approach and reach land. In
Hurricanes 2017 alone, tropical cyclones caused an estimated $215 billion in damage. While MODIS data are regularly used
Droughts in the analysis of hurricanes and typhoons, damage studies typically focus on just a few events without providing
MODIS a comprehensive overview and comparison across events. The MODIS record is now sufficiently long to enable
Disturbance
standardization in time, allowing us to extend previously developed disturbance methodology and to remove
Tasseled Cap
dependency on land cover datasets. We apply this new approach to detect the impact of both droughts and
hurricanes on the four largest Caribbean islands since 2001. We find that the percentage of disturbed land on the
four islands varies from approximately 0–50% between 2001 and 2017, with the highest percentages coinciding
with major droughts in Cuba, and Hurricane Maria in Puerto Rico. We demonstrate that (1) Hurricane Maria
resulted in significant disturbance across 50% of Puerto Rico (4549 km2), and (2) gradual recovery started about
2.5 months after the hurricane hit. While our approach focuses on the identification of damage arising from
hurricanes, it is also capable of identifying the damage from droughts. This approach ultimately enables a better
understanding of the combined effects of these two natural hazards across island landscapes.

1. Introduction excessive rainfall found during hurricanes have been linked to atmo-
spheric circulation patterns, such as the Atlantic Multidecadal Oscilla-
Tropical cyclones are natural events that transform into natural tion (AMO; Curtis, 2008, Elder et al., 2014, Fensterer et al., 2012,
disasters as they approach and reach land. In 2017 alone, tropical cy- Méndez and Magaña, 2010). The combined effect of these two natural
clones caused an estimated US$215B in damage (Faust and Bove, events can be highly destructive, and actual hurricane recovery may
2017). Some have argued that increases in tropical sea surface tem- take longer than expected following a major drought, such as was found
perature since the mid-1970s have increased the potential destruc- for lakes (Xuan and Chang, 2014). Teasing apart the combined effect of
tiveness of hurricanes as a result of longer storm lifetimes and greater even moderate hurricanes and moderate droughts is an important but
storm intensities (Emanuel, 2005). However, there is still great un- complicated endeavor requiring a better understanding of ecosystem
certainty with respect to the effects of climate change on hurricane recovery processes following disturbance (Beard et al., 2005; Ortegren
frequency and intensity (Lugo, 2000; Pielke Jr et al., 2005). For ex- and Maxwell, 2014) and a standardized dataset that allows for the
ample, some of the latest climate projections forecast an increase in the comparison of hurricane damage in a comprehensive manner.
intensity of tropical cyclones by 2–11%, while decreasing the frequency Since 2001, 15 hurricanes have hit the Caribbean islands of Cuba,
of the storms by 6–34% (Knutson et al., 2010). Hispaniola, Jamaica, and Puerto Rico, affecting the approximately 40
Climate models also consistently predict drying for many low to million people living on those islands (World Bank, 2018). Droughts
mid-latitude regions, with drier areas predicted to get drier and wetter have generally received less attention than other natural disasters in the
areas predicted to get wetter (Trenberth et al., 2014). In addition, as a Caribbean (Beard et al., 2005), but a new downscaled drought dataset
result of increased temperatures, the rate of drying is expected to in- demonstrated that between 2013 and 2016, the Caribbean faced its
crease, resulting in the establishment of droughts more quickly and most severe and widespread drought since the 1950s (Herrera and Ault,
with greater intensity (Trenberth et al., 2014). It is not uncommon for 2017). This drought was especially significant during the summer of
ecosystems exposed to drought to also experience hurricanes (Beard 2015, when in June drought conditions covered 95% of the four major
et al., 2005; Ortegren and Maxwell, 2014). Both droughts and the islands in the Caribbean, with 82% under at least severe drought

⁎
Corresponding author.
E-mail address: [email protected] (K.M. de Beurs).

https://doi.org/10.1016/j.rse.2019.04.028
Received 18 September 2018; Received in revised form 6 March 2019; Accepted 26 April 2019
Available online 06 May 2019
0034-4257/ © 2019 Elsevier Inc. All rights reserved.
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

conditions and 58% under extreme drought. There are several global Table 1
and a few regional drought datasets available for regions affected reg- Country population and GDP from the World Bank (2018).
ularly by hurricanes, such as the Caribbean (Herrera and Ault, 2017). Island Country Population Total GDP Per capita GDP
However, while remote sensing methods have been developed for the (US$B) (US$)
detection of hurricane damage over the past few years, especially fo-
Cuba Cuba 11,475,982 80.656 7602
cused on the detection of hurricane damage on forests (Negrón-Juárez
Hispaniola Haiti 10,847,334 8.023 766
et al., 2014), there are few long time series that allow for the stan- Dominican Republic 10,649,791 71.584 7052
dardized comparison of the effect of hurricanes. Puerto Rico Puerto Rico 3,411,307 103.135 30,833
The dense temporal resolution of MODIS data (daily, with 8-day Jamaica Jamaica 2,871,934 14.187 4879
composites) and the relatively long time series available (2000–current)
make MODIS data ideally suited for the study of hurricane damage.
MODIS data are indeed regularly used in the analysis of hurricane and Fig. 1 provides an overview map of the four islands based on the
typhoon damage (Chambers et al., 2007; Long et al., 2016; Parker et al., MODIS Land Cover data (MCD12Q1) with the Plant Functional Type
2017; Rogan et al., 2011; Rossi et al., 2013); however, studies have (PFT) classification scheme from 2001. The most dominant land cover
typically focused on just one hurricane or, at best, a few and do not type was evergreen broadleaf forest for Jamaica (61%) and Puerto Rico
provide a comprehensive overview—an exception being Potter (2014). (42%); whereas, it was broadleaf cropland for Cuba (46%) and grass-
In addition, many of the MODIS analysis use simple vegetation indices land for Hispaniola (28%).
such as the NDVI and the EVI, with few extending to indices such as the In the Atlantic Basin and Eastern Pacific, tropical cyclones with
Tasseled Cap Water Index and the NDII (Jin and Sader, 2005; Wang and sustained wind speeds of > 118 kph and center pressure of 980 mbar
Xu, 2010). In contrast, Potter (2014) performed a more comprehensive are considered to be hurricanes. The Saffir-Simpson Hurricane Intensity
analysis and presented a global assessment of damage to coastal eco- Scale is used to further classify storms from Category 1 to Category 5.
system vegetation from tropical storms between 2006 and 2012 based Each island has been hit by at least one hurricane since 2001, with Cuba
on quarterly differences in vegetation index imagery. The Quarterly enduring 11 hurricanes (Table 2).
Indicator of Cover Change identified regions where at least a 40%
change in green vegetation cover occurred, demonstrating that ex-
tensive tropical forest damage could be detected following extreme 3. Data
storm events (Potter, 2014). These data were not extended past 2012,
and the relatively coarse spatial resolution used (5600 m) make them 3.1. MODIS NBAR BRDF product – MCD43A4
less than ideal for the study of the Caribbean islands.
We have previously shown that a MODIS-derived disturbance index We used the MODIS Nadir BRDF-Adjusted Reflectance Distribution
based on standardized tasseled cap brightness, greenness, and wetness Product (MCD43A4) version 6, which simulates surface reflectance
data as derived by Healey et al. (2005) can successfully detect forest values and is delivered daily, based on 16 days of observations (Schaaf
clearance (de Beurs et al., 2016; Tran et al., 2016). We also demon- et al., 2002), with 463 m spatial resolution. Both Terra and Aqua data
strated that the disturbance index can be adapted for use in grasslands are used in the generation of the product to increase the probability for
(de Beurs et al., 2016). We have applied disturbance index analysis to a quality input to the models. We downloaded the data at eight-day
the central United States to demonstrate that it is quite effective in intervals between 2001 and 2017 for four tiles: H10V06, H10V07,
identifying damage from tornados and can track the subsequent re- H11V06 and H11V07. For each of 782 time steps, we calculated the
covery phase (Kingfield and de Beurs, 2017). MODIS Tasseled Cap brightness, greenness, and wetness values (Lobser
A fundamental limitation of the disturbance index approach is that and Cohen, 2007) by multiplying each band coefficient by the corre-
it is relative: it relies on a comparison of potentially damaged pixels sponding band value and then summing these values for each index
against a normal distribution of typical “undamaged” pixels. For ex- listed in Table 3.
ample, in the case of forest disturbance, the approach requires a large After calculating these Tasseled Cap indices, we stacked the data in
sample of normal, non-disturbed, forest pixels to serve as the baseline temporal order. Missing data were filled using a median filter centered
for comparison. As a result, the approach is highly dependent on a high- on five sequential composites.
quality land cover map, and potentially information about aspect and
slope (de Beurs et al., 2016). Yet, the quality of land cover datasets
varies widely across the globe and low accuracies are prevalent, espe- 3.2. Hurricane paths
cially in areas with heterogeneous land cover (Gómez et al., 2016;
Herold et al., 2008) The Atlantic Hurricane Database (Atlantic HURDAT2, 2018;
The MODIS record is now sufficiently long to enable standardization Landsea and Franklin, 2013) was downloaded from the National Hur-
in time. Thus, here, we are extending the disturbance index approach ricane Center (NHC), a division of the National Oceanographic and
by removing the dependency on a land cover dataset by using the Atmospheric Administration (NOAA). Atlantic HURDAT2 provides the
length of the MODIS time series. We apply this new approach to detect most accurate possible information on the track of historical hurricanes
the impacts of hurricanes since 2001 on the four largest Caribbean is- and tropical cyclones in the Atlantic Basin. HURDAT 2 provides six-
lands. By comparing the results against other standardized datasets, hourly information on the location, maximum winds, and central
specifically developed for the detection of droughts, we demonstrate pressure of all known tropical and subtropical cyclones from 1851 to
that it is also capable of identifying the impact of drought on the ve- 2016. Data for individual storms were also downloaded for Hurricanes
getated land surface. Irma and Maria. These two storms fell outside of the Atlantic HURDAT2
database, but within the timeframe of the collected MODIS data. We
2. Study region only evaluated the data for cyclones occurring since January 1, 2001.
We generated point files for the cyclone records, which we then inter-
We investigate the impact of hurricanes and droughts on the four sected with the extent of the four major Caribbean islands to identify
largest islands in the Caribbean Sea: Cuba, Hispaniola (hosting Haiti when (dates) and where (locations) the hurricane centers passed di-
and the Dominican Republic), Puerto Rico, and Jamaica. The total rectly over land. These extracted point data for each event were con-
population across these four islands was nearly 40 million in 2016 verted to a line representing the track of the storm across the island.
(Table 1).

2
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Fig. 1. Study region overview figure based on MODIS Land Cover data from 2001.

3.3. Drought data Table 3

MODIS Tasseled Cap Coefficients, from Lobser and Cohen (2007).
The Caribbean has been affected by several severe droughts during Band Brightness Greenness Wetness
our study period. To better understand the spatio-temporal landscape
variability resulting from drought, we downloaded a high-resolution Red 0.4395 −0.4064 0.1147
drought dataset statistically downscaled to 4 km spatial and monthly NIR 1 0.5945 0.5129 0.2489
Blue 0.2460 −0.2744 0.2408
temporal resolution (Herrera and Ault, 2017). This dataset consists of Green 0.3918 −0.2893 0.3132
self-calibrating Palmer Drought Severity Index data. For this study, we NIR 2 0.3506 0.4882 −0.3122
used all monthly observations between 2001 and 2016, which is the last SWIR 1 0.2136 −0.0036 −0.6416
year available in this dataset. SWIR 2 0.2678 −0.4169 −0.5087

3.4. Validation data

Hispaniola and Hurricane Maria over Puerto Rico were available from
the Open Data Program. For validation of Hurricane Matthew, we
To validate the disturbance index results, we downloaded high-re-
downloaded images from September 13, 2013 and October 15, 2016;
solution, optical satellite data from the Open Data Program provided by
for Hurricane Maria, we selected images from August 21, 2017 and
Digital Globe (https://www.digitalglobe.com/opendata). The Open
October 16, 2017. These images were not entirely cloud-free, with
Data Program provides high resolution, commercial satellite images for
cloud cover between 10% and 30%, but they provided the best op-
use in natural disaster damage identification. Of the 15 hurricanes that
portunity for validation at high spatial resolution.
directly hit the islands, only data for Hurricane Matthew over

Table 2
Named hurricanes with maximum windspeeds in kilometers/h (miles/h). Some hurricanes lasted multiple days.
Hurricane Category Max Wind Speed Max Wind Speed Date Island
(km/h) (miles/h)

Michelle 3 185 115 11/04/01 Cuba

Isidore 1 121 75 09/20/02 Cuba
Lili 1 145 90 10/01/02 Cuba
Charley 2 169 105 08/13/04 Cuba
Dennis 3 193 120 07/08/05 Cuba
Gustav 3 193 120 08/31/08 Cuba
Ike 3 185 115 09/08/08 Cuba
Paloma 1 145 90 11/09/08 Cuba
Sandy 1 121 75 10/24/12 Jamaica
Sandy 2 161 100 10/25/12 Cuba
Matthew 3 209 130 10/04/16 Hispaniola
Matthew 3 185 115 10/05/16 Cuba
Irma 5 266 165 09/09/17 Cuba
Maria 4 249 155 09/20/17 Puerto Rico
Maria 3 201 125 09/22/17 Hispaniola

3
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

4. Methods sample sizes (Steinskog et al., 2007). Here we applied the Shapiro-Wilk
test to test for normality. The Shapiro-Wilk test is based on the corre-
4.1. MODIS disturbance index with temporal standardization lation between our observations and corresponding normal scores, and
it can be used for very small sample sizes. In addition, for small sample
Following methodology described previously in other papers, we sizes, this test provides better power than the Kolmogorov-Smirnov test
standardized the TC brightness, greenness, and wetness before com- (Steinskog et al., 2007). We carried out the test for normality for each of
bining these into a Disturbance Index (DI) (de Beurs et al., 2016; Healey the 46 composites, evaluating the anniversary composites over the 17-
et al., 2005; Tran et al., 2016). The normalization was carried out as year period.
follows:
4.3. Validation
Brightnessn (t ) = (Brightness (t ) − μBrightness )/σBrightness (1)

Greennessn (t ) = (Greenness (t ) − μGreenness )/σGre enness (2) Before and after images provided by the Open Data Program for
Hurricane Matthew were overlaid to establish the overlapping area. We
Wetnessn (t ) = (Wetness (t ) − μWetness )/σWetness (3) randomly selected 25 independent validation regions consisting of four
MODIS pixels each. Each validation region was visually inspected to
where t is a composite at time t and μ and σ are the mean and the determine the dominant land cover type and to determine the presence
standard deviation, respectively, of the particular index for that com- of identifiable hurricane damage. For the samples that visually pre-
posite over the years. In previous papers, the mean and standard de- sented hurricane damage, we identified the disturbance index values
viation calculations were carried out spatially, where each pixel was based on the MODIS data both before the hurricane hit and three weeks
compared to a distribution of values extracted from pixels with the after the hurricane went through. The same process was completed
same land cover. The disadvantage of this approach is that it is highly using the Hurricane Maria images over Puerto Rico. We randomly se-
dependent on high quality land cover classification. In addition, aspect lected 46 independent validation regions consisting of four MODIS
and slope can have a strong influence on pixels even within the same pixels each for Hurricane Maria. We were able to select more validation
land cover (de Beurs et al., 2016). The MODIS time series now extends points in Puerto Rico, because the overall area with hurricane damage
17 full years (2001–2017). As a result, there are a sufficient number of was larger with a greater diversity in land cover types. Note that in this
observations to enable the development of normal distributions of va- validation, we focus specifically on the ability of the disturbance data to
lues in time. Thus, in Eqs. (1) to (3), the mean and the standard de- quantify disturbance. As a result, the validation focuses on areas with
viations were calculated based on the anniversary composites from hurricane damage and we have not carried out a specific validation for
every year. In other words, the first pixel in the first composite in the areas without damage. Nevertheless, we do confirm that the dis-
first year is compared to the mean and the standard deviation of the turbance index is close to zero (the expected value for no damage)
first pixel for all first composites of each year. Thus, January 1, 2001 is before the hurricanes hit.
compared with the mean and the standard deviation of all January 1
composites between 2001 and 2017. 5. Results
When natural vegetation is damaged as the result of a hurricane or
tornado (Kingfield and de Beurs, 2017), greenness and wetness typi- 5.1. MODIS DI and standardization results
cally decline as a result of vegetation loss while brightness increases,
because damaged areas typically absorb less solar radiation than areas Fig. 2 provides the mean brightness, greenness, and wetness images
with healthy vegetation. As a result, we can calculate the disturbance for Cuba. With brightness values displayed in red, greenness values
index as follows: displayed in green, and wetness values displayed in blue, the map of the
DI (t ) = Brightnessn (t ) − (Greennessn (t ) + Wetnessn (t )) (4) mean values of the Tasseled Cap Indices closely resembles the MODIS
derived land cover data. Urban areas have higher values in the
Since normalized distributions typically have a mean of 0 and a brightness index, but lower values in wetness and greenness. These
standard deviation of 1, we expect high DI values when there is a lot of areas show up in orange and red colors. Forests present higher values in
disturbance. For example, if a pixel brightness value is two standard greenness, moderate values in wetness, and lower values in brightness.
deviations brighter than average, i.e., its normalized value is 2, and the As a result, forest areas show up in darker green colors. Water has
greenness and wetness are two standard deviations less than average, virtually no brightness or greenness and, thus, appears blue. Croplands
i.e., their normalized values are −2, then the resulting DI value is 6 and grasslands appear in mixtures from brown to yellows.
(=2 − (−2 + −2)). We applied the Shapiro-Wilk test to determine the number of
When normalized distributions are added or subtracted, their mean composites that do not follow a normal distribution in brightness,
and standard deviations are added (they are never subtracted). In de greenness, and wetness data separately and found that 99.9% of all
Beurs et al. (2016), we demonstrated that if we set the threshold of pixels exhibited normality for > 90% of all composites. As a result, we
disturbance at 3, we have a 15.9% probability of identifying a pixel as concluded that our standardization method based on the normal dis-
disturbed when, in fact, it is not (i.e., a false positive). Here we will tribution generates valid observations.
identify disturbed areas both based on a threshold of 2 (probability of
false positive equals 25.2%) and a threshold of 3. Note that the prob- 5.2. Disturbance detection: droughts
ability of a false positive is 9.1% for a disturbance value of 4, and 4.8%
for a disturbance value of 5. The percentage of disturbed land on the four islands varies from
approximately 0–50%, with the highest percentages coinciding with
4.2. Testing for normality major droughts (Fig. 3) and immediately after hurricane Maria in
Puerto Rico (Fig. 6).
To compare composites properly, we need to assess whether the It is apparent that longer periods of drought result in greater overall
sampling distribution of pixels follows a normal distribution. Several disturbance and that rainfall quickly alleviates drought effects, e.g., in
tests have been developed to test for normality. For all those tests, the Cuba there was a major drought through 2004 and early 2005, ending
null hypothesis is the “sample distribution is normal”; thus, a significant later that year. The percentage of disturbed pixels increased almost
result indicates that the sampling distribution may deviate from nor- immediately at the beginning of the drought and remained high
mality. One drawback of normality tests is their low power for small throughout the period. By 2005, 45% of Cuba (49,446 km2) was

4
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Fig. 2. Three colour composite of the mean Brightness, Wetness, and Greenness Tasseled Cap Indices for the island of Cuba. Lower left inset showing forests on the
eastern end of the island. Lower right inset showing Havana and surrounding area.

Fig. 3. Orange and blue provide the Palmer Drought Severity Index for Cuba (upper left), Hispaniola (upper right), Puerto Rico (lower left), and Jamaica (lower
right). The dotted line provides the percentage (divided by 10 for visualization) of the island that is classified as disturbed (DI > 3). Large drought periods are linked
with strong disturbances, for example the 2004/2005 drought in Cuba disturbed 45% of the island. The vertical lines indicate the timing of the hurricanes; in the case
of two hurricanes within one month only one vertical line is drawn. (For interpretation of the references to colour in this figure legend, the reader is referred to the
web version of this article.)

5
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Fig. 4. Maximum drought months (upper pair) on Jamaica (left) and Puerto Rico (right) with the corresponding disturbance index data (lower pair) for that month.
Severe droughts have negative PDSI values and disturbed land has values over 2.

Michelle
(11/04-
05/2001)

Isidore
(09/20/2002)
& Lili
(10/01/2002)

Charley
(08/13/2004)

Charley
occurred
during a
widespread
severe
drought. Cf.
Figure 3.
Dennis
(07/08/2005)

Fig. 5. Major hurricanes affecting Cuba and the observed disturbance data.

6
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Gustav
(08/31/2008)
& Ike
(09/08/2008)

Paloma
(11/09/2008)

Sandy
(10/25/2012)

Mahew
(10/05/2016)

Irma
(09/09/2017)

Disturbance Miles/hour

Fig. 5. (continued)

7
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Mahew
(10/04/2016)

Hurricane Puerto Rico

Maria
(09/20/2017)

Fig. 6. Major hurricanes affecting Hispaniola and Puerto Rico, and the observed disturbance data. The legend for the hurricane speeds and the disturbance data can
be found in Fig. 5.

Fig. 7. Before hurricane Michelle (top), Cuba al-

Michelle
ready displayed a relatively large amount of dis-
(11/04- turbance. This disturbance increased substantially
05/2001) immediately after hurricane Michele made landfall
on Cuba (middle). About two months after the hur-
ricane made landfall, the drought has been alle-
viated somewhat and some of the immediate dis-
turbance has recovered (bottom). The legend for the
hurricane speed and the disturbance data can be
found in Fig. 5.

significantly disturbed (DI > 3). The percentage of disturbed area de- is greater variation in the DI data than in the PDSI data resulting from
clined immediately when drought conditions eased. Similar responses the difference in spatial resolution: approximately 4 km for the PDSI
are evident in Hispaniola after the 2003 drought, as well as in Puerto versus 500 m for the DI (Fig. 4). Note that the PDSI data provides a
Rico (e.g., 2015) and Jamaica (2004–2005 and 2013–2016), although measurement of climatological drought severity, while the DI provides
the percentage of disturbed land on these islands rarely exceeded 25%. a measurement of the damage to the vegetation. Not every drought
When we compare the PDSI with the DI spatially, we observe that there event will directly result in disturbance apparent in the landscape. For

8
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

landscape, with greater windspeeds causing significantly more damage

(Figs. 5, 6). We also find that while the effect of the hurricanes can be
devastating, the total area affected is relatively localized, except for
hurricane Maria, which affected all areas of Puerto Rico. The damage
from Michelle and Charley on Cuba appears larger and more wide-
spread than expected, as a result of concurrent drought (Fig. 3). Hur-
ricane Michelle struck Cuba in the middle of a short but severe drought
(Fig. 7). The widespread drought effect is clearly visible on Cuba's
landscape right before the hurricane hit, but the disturbance index re-
veals increased disturbance in the area where Michelle hit. The drought
is clearly alleviated following the rains delivered by Michelle (Fig. 7,
bottom).
Fig. 8 also shows that the rainfall brought by hurricane Matthew
alleviated drought conditions on Hispaniola and Cuba.

5.4. Disturbance validation

Hurricane damage detection is not necessarily straightforward, even

when high resolution images are available. Fig. 9 reveals a damaged
area in Puerto Rico after Hurricane Maria.
As mentioned earlier, cloud cover can make it difficult to find un-
obscured high-resolution imagery to map hurricane damage in the
Caribbean. We selected before and after images for Hurricane Matthew
in Hispaniola and Hurricane Maria in Puerto Rico to validate the dis-
turbance data (Fig. 10). For Hurricane Matthew, we identified that 21
of the 25 samples visually revealed hurricane damage, two samples
were too cloudy to identify any damage, and we were unable to identify
damage on the remaining two sites (Fig. 10). We identified the dis-
turbance index values both before the hurricane hit and three weeks
after the hurricane event based on the MODIS data for the 21 samples
for which we confirmed hurricane damage (Fig. 10). Before the hurri-
cane, the disturbance index was not significantly different from zero,
which is our expected value for no damage. After the hurricane the
disturbance index was significantly higher, with an average DI value of
5.92, and 20 out of 21 samples revealed disturbance values well above
zero.
We were able to identify a greater number of samples for Maria,
which allowed us to separate the effect of hurricanes on developed
lands from the impact on vegetated land (Fig. 10, Table 4). We found
that the disturbance index was not significantly different from 0 before
Maria hit Puerto Rico. After Maria hit the average disturbance index
Fig. 8. Hurricane Matthew causing disturbance on Hispaniola and Cuba, but
increased to 2.40 for developed land and 3.37 for vegetated land. Thus,
relieving drought. The legends for the PDSI and the disturbance data can be
the disturbance index was capable of measuring disturbance on both
found in Fig. 4.
land covers, but the increase in the disturbance index was higher over
vegetated land compared to developed land and, while the developed
example, irrigation or urban landscapes might not experience any dis- areas revealed an increase in DI, the ultimate DI values was often not
turbance from the lack of precipitation because soil moisture is main- above two after the hurricane (Table 3).
tained through human intervention. In Puerto Rico, the severe drought It is important to emphasize here that it is much easier to identify
of August 2015 did not affect the landscape evenly, with no significant that a sample displayed damage, than to identify the absence of da-
disturbance in large parts of the central forested highlands (Fig. 4). A mage. As a result, we focused our validation efforts on the detection of
similar pattern is visible in Jamaica, where the drought that started in damage. Nevertheless, we note that before the hurricane hit the average
the beginning of 2015 and lasted through the end of the observation disturbance values were zero (the expected value for areas with no
period (December 2016) is evident almost everywhere across the island, damage).
while the main area of disturbance is visible only in the central portion
of the island. 6. Discussion

5.3. Disturbance detection: hurricanes 6.1. Remotely sensed hurricane damage detection

The disturbance index can clearly reveal the impact of hurricanes. A review of the literature results in several papers which rely, at
We have mapped out all hurricanes that hit Cuba between 2001 and least partly, on the use of vegetation indices, such as the Normalized
2016 (Fig. 5) as well as the impact of Matthew on Hispaniola and Maria Difference Vegetation Index (NDVI) or the Enhanced Vegetation Index
on Puerto Rico (Fig. 6). (EVI) for the detection of hurricane damage. For example, before and
Isidor and Lili in 2002 were combined into one map, as were Gustav after MODIS Enhanced Vegetation Index (EVI) imagery were used to
and Ike in 2008, because the hurricanes occurred within one month of detect damage of Hurricane Felix on the broadleaf and pine forests of
each other, making the damage from their crossing paths difficult to Nicaragua (Rossi et al., 2013). They reported greater damage in the
separate. The impacts from each hurricane are clearly visible on the broadleaf forests, with > 75% of the trees blown down in some areas,

9
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Fig. 9. Before and after damage detection for a mixed urban/vegetated area Puerto Rico. The MODIS based disturbance index had a value of 1.43, which increased to
2.98 after the storm. Green vegetation is visible on both images. However, the after image shows the disappearance of several trees in the urban area, as well as a
browning of the forested region in the upper left. Note that the disturbance value before the storm was relatively high (1.43) as a result of the mixed nature of this
pixel and a drought. Nevertheless, the disturbance index increased substantially as a result of the storm. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

than in the pine forest. They concluded that the spatial resolution of the the electromagnetic spectrum. For example, a MODIS instantaneous
250 m MODIS data was sufficient for damage assessments after hurri- Global Disturbance Index using EVI and Land Surface Temperature
canes in (sub-)tropical forests. Two other studies investigating hurri- (LST) was developed to detect disturbances, such as hurricanes, across
cane landfall damage along the northern Gulf of Mexico and south- North America (Mildrexler et al., 2009). An evaluation of the effects of
eastern Yucatan in Mexico also used differences in EVI to measure Hurricanes Katrina, Rita, and Wilma revealed that this disturbance
hurricane damage (Rogan et al., 2011; Wang and D'sa, 2009) with si- index was able to pick up the expected effect of high severity dis-
milar results. While it is possible to use NDVI or EVI to investigate turbance near the shoreline and moderate severity disturbance inland,
hurricane damage (Hu and Smith, 2018), we have demonstrated pre- although the authors acknowledged that more research was necessary
viously that the defoliation and surface disturbance detected as a result to evaluate the moderate severity disturbance detections.
of tornadoes disappears more quickly when analyzed with NDVI data Wang and Xu (2010) compared four different algorithms and sev-
than with disturbance index data (Kingfield and de Beurs, 2017). We eral different spectral indices to assess the damage of Hurricane Katrina
have found a similar effect in this paper for the hurricane damage on on forests. They found that the impact of the selection of classification
the Caribbean islands; for example, Hu and Smith (2018) indicated that method was less important than the effect of the selection of the
vegetation returned to near-normal about 1.5 months after Hurricane spectral index. They also found that the Tasseled Cap Wetness index
Maria affected Puerto Rico and Dominica in 2017. Recovery from dis- outperformed five other tested indices and that a post-classification
turbance appeared much slower when analyzed using our disturbance comparison proved the most accurate methodology (Wang and Xu,
index approach. After Hurricane Maria hit Puerto Rico, we found that 2010). (See Wang et al. (2010) for an overview of other post-storm
50% of Puerto Rico (4549 km2) showed significant disturbance (DI > damage assessment studies, predominantly using vegetation indices.)
3). We found signs of recovery in the disturbance data starting about
8–12 weeks after hurricane Maria hit; however, by the end of 2017, the
disturbance index was still > 3 across 31% of the island (2822 km2). 6.2. Hurricane and drought interaction
We believe that the quick recovery in NDVI data is often the result
of increased rainfall generating a leaf pulse in hurricane-affected ve- Vegetation gains due to additional rainfall are regularly observed in
getation (Parker et al., 2017). Although the remotely sensed data might the aftermath of hurricanes. For example, in dryland forests, where an
reveal a return to normal levels of NDVI relatively quickly, the actual increase in the NDVI was visible within a few months after Hurricanes
vegetation canopy can take much longer to recover to pre-disturbance Jova and Patricia affected Mexico (Parker et al., 2017). We have de-
conditions. For example, tall mangrove stands might take > 20 years to monstrated that the MODIS disturbance index standardized in time is
recover fully; whereas, shrublands can recover much more quickly effective at identifying disturbance resulting from both hurricanes and
(Imbert, 2018). In Taiwan, it took two years for litterfall to return to droughts. An example of drought and hurricane interaction can be
pre-typhoon levels after a major event in 1994, and annual peak leaf found after Hurricane Matthew (Fig. 8). It was the strongest storm to
area index only returned to pre-event levels after ten years (Lin et al., impact Haiti (western Hispaniola) since the early 1960s, resulting in
2017). major damage and as many as 546 deaths. Hurricane Matthew hit both
Instead of one of the more common vegetation indices, such as Cuba and Hispaniola as a category 4 hurricane, depositing between
NDVI or EVI, a few studies have investigated changes in the leaf water 400 mm and 500 mm of rain in these countries and effectively alle-
content after hurricanes (Jin and Sader, 2005; Wang et al., 2010), for viating a drought in parts of these islands (Fig. 8). This impact is not
example using the Normalized Difference Infrared Index (NDII) that unlike the effect of Atlantic tropical cyclones on drought over the
combines the near infrared (NIR) and shortwave infrared (SWIR) bands. Eastern United States (Kam et al., 2013; Maxwell et al., 2013). We have
Wang et al. (2010) indicated that NDII was more sensitive to vegetation also demonstrated that not every climatological drought condition re-
changes after hurricanes, especially in areas dominated by forests, sults in an observable disturbance to the landscape, with drought effects
which results in loss of NDVI sensitivity due to higher levels of leaf area being often slow to appear in forested areas (e.g., Fig. 4).
index (Vina et al., 2004). Others have also expanded to other parts of

10
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

opportunities, and it is difficult to plan large field campaigns around

the potential for hurricanes. As a result, it is very difficult to tie the
disturbance index values to ground damage assessments.
2) Hurricane damage often occurs during periods and in regions with
very dense cloud cover. As a result, it is difficult to find high re-
solution, or even moderate resolution cloud-free before and after
satellite images as we experienced in this study. As an example, we
have downloaded all Landsat images (L1T) from 1984 until now
during the hurricane season (July–December) for two Landsat WRS-
2 Path/Row scenes (P16/R44 and P11/R46) in the East and West of
Cuba. Over the entire time period we found a total of 301 images in
Western Cuba, and 361 images in Eastern Cuba. For each image we
used the provided quality assessment data to identify if a pixel
contains cloud cover. Fig. 11 provides a cloud climatology based on
these Landsat footprints over Cuba showing that almost half
(48.31%) of P11//R46 was cloud-free < 50% of the time and 70%
of P16/R44 was cloud-free < 50% of the time. In addition, coastal
areas, where the hurricane damage is often the most severe, show
even higher percentages of clouds, some regions missing as much as
80% of the data. This cloud climatology also explains why it is so
difficult to carry out standardized “before-and-after” comparisons to
identify hurricane damage, e.g., the probability of finding a cloud-
free before-and-after pixel pair can be calculated by multiplying the
values in Fig. 11. Lastly, expanding the potential “before” and
“after” time periods for comparison, might aid in the probability of
finding cloud-free data, but as described above, hurricane damage
can recover rather quickly. Thus, the potential for finding a stan-
dardized hurricane response will be greatly diminished.
For comparison purposes, we also investigated New Orleans (P22/
R39) and Oklahoma City (P28/R35). We found significantly more
images for New Orleans (496), but determined that 50% of the area
was cloud-free < 50% of the time, similar to Cuba. In contrast,
Oklahoma City, which is more inland, showed that the 50% of the
pixels were cloud-free at least 80% of the time.

3) While we note that droughts are visible in the disturbance index

Fig. 10. (top): Boxplot shows the distribution of the mean DI for all validation data, we did not provide an explicit validation of the effect of
points showing visually identified hurricane damage one week before droughts (other than Fig. 3). We also did not attempt to use the
Hurricane Matthew and three weeks after the hurricane. (N = 21). (bottom): disturbance index data to identify whether a disturbance is the re-
Box plot showing the distribution of the mean DI of validation points on Puerto sult of drought, or the result of hurricane damage. Since we had
Rico when sorted between developed (urban and suburban) and vegetated good hurricane track data available (Hurdat 2), we assumed that
(forest, scrub, and other natural vegetation). Not that the y-axes are different disturbances detected immediately after a hurricane are due to that
for the two sets to better visualize the distributions. event. Vice versa, since we had good drought data available (Herrera
and Ault, 2017), we assumed that disturbance identified to coincide
Table 4 with observed droughts are due to that hazard. Instead, we proposed
Number of samples that are visibly damaged and their corresponding DI class this standardized dataset approach as a way to identify the damage
according to MODIS. caused by these two natural hazards in a standardized manner that
Developed Vegetated Total enables comparison of the effects of hurricanes over time and for
different ecosystems.
MODIS DI < 2 14 4 18
MODIS DI > 2 7 21 28
6.4. Other available data sources
Total 21 25 46

Hurricanes often cause significant flooding in addition to major

6.3. Validation of hurricane damage wind damage. It is not possible to use the MODIS disturbance data to
map the flooding as it is ongoing, because the optical MODIS sensor is
While the MODIS-derived disturbance index can be used to track unable to penetrate the significant cloud cover that accompanies hur-
damage and recovery from both hurricanes and droughts on vegetated ricanes. A recent paper demonstrates the use of Cyclone Global
and developed lands, we have found that it is more sensitive to dis- Navigation Satellite Systems (CYGNSS) data, which has a spatial re-
turbances on vegetated landscapes. This sensitivity is a result of the solution of a few kilometers and a temporal resolution of just a few days
incorporation of the Tasseled Cap greenness index, which measures (Chew et al., 2018). Combining these types of data with the MODIS
declines in greenness on the landscape and the exposure of soils after disturbance index might generate an opportunity to better understand
disturbances. It is also important to point out that large-scale validation where flooding occurs and how long the damage remains visible on the
of the effect of hurricanes is complicated in several ways: landscape. Visual evaluations reveal very similar damage agreements
between the CYGNSS data and the MODIS disturbance data reported
1) Hurricane damage significantly hampers field validation here for Hurricane Irma.
The Visible Infrared Imaging Radiometer Suite (VIIRS) flying on the

11
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

Fig. 11. Cloud climatology for Landsat WRS-2 P16/R44 and P11/R46 over Cuba.

Suomi National Polar-orbiting Partnership (NPP) satellite since 2011 is analysis of land surface dynamics in North and South America using mul-
the operational replacement for the MODIS sensors that are currently in tiple sensor datastreams. The authors thank Prof. Toby Ault of Cornell
service well beyond their six-year life span. To transfer the disturbance University for the use of his dataset. The MCD43A4 product was re-
index methodology presented here to the VIIRS data, it will be crucial trieved from the data pool of the NASA EOSDIS Land Processes
to understand how the VIIRS data with a much shorter temporal record Distributed Active Archive Center (LP DAAC), USGS/Earth Resources
can be standardized against the MODIS observed time series. The VIIRS Observation and Science (EROS) Center, Sioux Falls, South Dakota. The
spectral bands are slightly different than the MODIS bands. Tasseled authors thank the helpful input of two anonymous reviewers that
Cap coefficients will need to be newly developed for VIIRS, and we are helped improve the clarity of the presentation.
not aware that any such coefficients have yet been developed.
It is important to point out that our disturbance dataset is unable to References
estimate monetary damages: economic impacts on developed land are
usually much greater than the impacts on vegetated areas. While the Atlantic Hurdat2, 2018. Atlantic HURDAT2. pp Page. https://www.nhc.noaa.gov/data/
spectral bands available on VIIRS are comparable to the MODIS bands, hurdat/hurdat2-format-atlantic.pdf.
Beard, K.H., Vogt, K.A., Vogt, D.J., et al., 2005. Structural and functional responses of a
VIIRS also has a Day/Night Band, which is particularly sensitive to very subtropical forest to 10 years of hurricanes and droughts. Ecol. Monogr. 75, 345–361.
low levels of light and allows for the tracking of nighttime lights. These Cao, C., Shao, X., Uprety, S., 2013. Detecting light outages after severe storms using the S-
data have been used to understand the effect of hurricanes on power NPP/VIIRS day/night band radiances. IEEE Geosci. Remote Sens. Lett. 10,
1582–1586.
outages that can be directly linked to the socio-economic impact of Chambers, J.Q., Fisher, J.I., Zeng, H., Chapman, E.L., Baker, D.B., Hurtt, G.C., 2007.
these storms (Cao et al., 2013; Miller et al., 2018). Another potential Hurricane Katrina's carbon footprint on US Gulf Coast forests. Science 318, 1107.
opportunity for the use of the VIIRS data would be to link our dis- Chew, C., Reager, J.T., Small, E., 2018. CYGNSS data map flood inundation during the
2017 Atlantic hurricane season. Scientific Reports (Nature Publisher Group) 8, 1–8.
turbance dataset with data from the VIIRS day/night band (Román Curtis, S., 2008. The Atlantic multidecadal oscillation and extreme daily precipitation
et al., 2018), which might prove useful for the understanding of damage over the US and Mexico during the hurricane season. Clim. Dyn. 30, 343–351.
on both vegetated and developed land surfaces (Wang et al., 2018). de Beurs, K.M., Owsley, B.C., Julian, J.P., 2016. Disturbance analyses of forests and
grasslands with MODIS and Landsat in New Zealand. Int. J. Appl. Earth Obs. Geoinf.
45, 42–54.
7. Conclusions Elder, R.C., Balling Jr., R.C., Cerveny, R.S., Krahenbuhl, D., 2014. Regional variability in
drought as a function of the Atlantic Multidecadal Oscillation. Caribb. J. Sci. 48,
31–43.
To better study the combined effects from droughts and hurricanes,
Emanuel, K., 2005. Increasing destructiveness of tropical cyclones over the past 30 years.
it is important to have a stable disturbance dataset that (1) is not Nature 436, 686.
strongly influenced by seasonal variation and (2) can be used to track Faust, E., Bove, M., 2017. In: Re, M. (Ed.), The Hurricane Seasons 2017: A Cluster of
Extreme Storms, pp Page. https://www.munichre.com/topics-online/en/2017/12/
recovery from hurricanes in a standardized way. Based on this dataset,
hurricane-season-2017#experts (Munich RE).
it will be possible to evaluate whether the size of the damage area from Fensterer, C., Scholz, D., Hoffmann, D., Spötl, C., Pajón, J.M., Mangini, A., 2012. Cuban
hurricanes can be predicted based on the hurricane strength, whether stalagmite suggests relationship between Caribbean precipitation and the Atlantic
there is a seasonal difference in damage severity and recovery time, or Multidecadal Oscillation during the past 1.3 ka. The Holocene 22, 1405–1412.
Gómez, C., White, J.C., Wulder, M.A., 2016. Optical remotely sensed time series data for
whether certain ecosystems recover more rapidly than others. land cover classification: a review. ISPRS J. Photogramm. Remote Sens. 116, 55–72.
Furthermore, it will be possible to start teasing apart the effects of Healey, S.P., Cohen, W.B., Zhiqiang, Y., Krankina, O.N., 2005. Comparison of Tasseled
droughts and hurricanes. For example, do hurricanes cause damage in Cap-based Landsat data structures for use in forest disturbance detection. Remote
Sens. Environ. 97, 301–310.
larger areas if preceded by a drought? Is the damage more severe after Herold, M., Mayaux, P., Woodcock, C., Baccini, A., Schmullius, C., 2008. Some challenges
drought (e.g., higher disturbance index values) or does it take longer for in global land cover mapping: an assessment of agreement and accuracy in existing 1
the ecosystem to recover? The disturbance index approach developed in km datasets. Remote Sens. Environ. 112, 2538–2556.
Herrera, D., Ault, T., 2017. Insights from a new high-resolution drought Atlas for the
this paper can be used in such a manner to not only report the damage Caribbean spanning 1950–2016. J. Clim. 30, 7801–7825.
from one or a few hurricanes, but to report the damage of multiple Hu, T., Smith, R.B., 2018. The impact of hurricane Maria on the vegetation of Dominica
hurricanes in a standardized fashion. The dataset is also capable of and Puerto Rico using multispectral remote sensing. Remote Sens. 10, 827.
Imbert, D., 2018. Hurricane disturbance and forest dynamics in east Caribbean man-
identifying the damage from both droughts and hurricanes, which can
groves. Ecosphere 9, e02231.
be used to better understand how the combined effects of these two Jin, S., Sader, S.A., 2005. Comparison of time series tasseled cap wetness and the nor-
natural hazards impact landscapes. malized difference moisture index in detecting forest disturbances. Remote Sens.
Environ. 94, 364–372.
Kam, J., Sheffield, J., Yuan, X., Wood, E.F., 2013. The influence of Atlantic tropical cy-
Acknowledgements clones on drought over the eastern United States (1980–2007). J. Clim. 26,
3067–3086.
Kingfield, D.M., de Beurs, K.M., 2017. Landsat identification of tornado damage by land
This research was supported in part by the NASA Science of Terra &
cover and an evaluation of damage recovery in forests. J. Appl. Meteorol. Climatol.
Aqua project NNX14AJ32G entitled Change in our MIDST: Detection and

12
K.M. de Beurs, et al. Remote Sensing of Environment 229 (2019) 1–13

56, 965–987. tropical storms. Remote Sensing Letters 5, 315–322.

Knutson, T.R., Mcbride, J.L., Chan, J., et al., 2010. Tropical cyclones and climate change. Rogan, J., Schneider, L., Christman, Z., Millones, M., Lawrence, D., Schmook, B., 2011.
Nat. Geosci. 3, 157. Hurricane disturbance mapping using MODIS EVI data in the southeastern Yucatán,
Landsea, C.W., Franklin, J.L., 2013. Atlantic hurricane database uncertainty and pre- Mexico. Remote Sensing Letters 2, 259–267.
sentation of a new database format. Mon. Weather Rev. 141, 3576–3592. Román, M.O., Wang, Z., Sun, Q., Kalb, V., Miller, S.D., Molthan, A., Schultz, L., Bell, J.,
Lin, K.-C., Hamburg, S.P., Wang, L., Duh, C.-T., Huang, C.-M., Chang, C.-T., Lin, T.-C., Stokes, E.C., Pandey, B., Seto, K.C., 2018. NASA's Black Marble nighttime lights
2017. Impacts of increasing typhoons on the structure and function of a subtropical product suite. Remote Sens. Environ. 210, 113–143.
forest: reflections of a changing climate. Sci. Rep. 7, 4911. Rossi, E., Rogan, J., Schneider, L., 2013. Mapping forest damage in northern Nicaragua
Lobser, S., Cohen, W., 2007. MODIS tasselled cap: land cover characteristics expressed after hurricane Felix (2007) using MODIS enhanced vegetation index data. GIScience
through transformed MODIS data. Int. J. Remote Sens. 28, 5079–5101. & remote sensing 50, 385–399.
Long, J., Giri, C., Primavera, J., Trivedi, M., 2016. Damage and recovery assessment of Schaaf, C.B., Gao, F., Strahler, A.H., et al., 2002. First operational BRDF, albedo nadir
the Philippines' mangroves following Super Typhoon Haiyan. Mar. Pollut. Bull. 109, reflectance products from MODIS. Remote Sens. Environ. 83, 135–148.
734–743. Steinskog, D.J., Tjøstheim, D.B., Kvamstø, N.G., 2007. A cautionary note on the use of the
Lugo, A.E., 2000. Effects and outcomes of Caribbean hurricanes in a climate change Kolmogorov–Smirnov test for normality. Mon. Weather Rev. 135, 1151–1157.
scenario. Sci. Total Environ. 262, 243–251. Tran, T.V., de Beurs, K.M., Julian, J.P., 2016. Monitoring forest disturbances in Southeast
Maxwell, J.T., Ortegren, J.T., Knapp, P.A., Soulé, P.T., 2013. Tropical cyclones and Oklahoma using Landsat and MODIS images. Int. J. Appl. Earth Obs. Geoinf. 44,
drought amelioration in the Gulf and southeastern coastal United States. J. Clim. 26, 42–52.
8440–8452. Trenberth, K.E., Dai, A., Van Der Schrier, G., Jones, P.D., Barichivich, J., Briffa, K.R.,
Méndez, M., Magaña, V., 2010. Regional aspects of prolonged meteorological droughts Sheffield, J., 2014. Global warming and changes in drought. Nat. Clim. Chang. 4, 17.
over Mexico and Central America. J. Clim. 23, 1175–1188. Vina, A., Henebry, G.M., Gitelson, A.A., 2004. Satellite monitoring of vegetation dy-
Mildrexler, D.J., Zhao, M., Running, S.W., 2009. Testing a MODIS global disturbance namics: sensitivity enhancement by the wide dynamic range vegetation index.
index across North America. Remote Sens. Environ. 113, 2103–2117. Geophys. Res. Lett. 31.
Miller, S.D., Straka Iii, W.C., Yue, J., et al., 2018. The dark side of hurricane Matthew: Wang, F., D'sa, E.J., 2009. Potential of MODIS EVI in identifying hurricane disturbance to
unique perspectives from the VIIRS day/night band. Bull. Am. Meteorol. Soc. 99, coastal vegetation in the northern Gulf of Mexico. Remote Sens. 2, 1–18.
2561–2574. Wang, F., Xu, Y.J., 2010. Comparison of remote sensing change detection techniques for
Negrón-Juárez, R., Baker, D.B., Chambers, J.Q., Hurtt, G.C., Goosem, S., 2014. Multi-scale assessing hurricane damage to forests. Environ. Monit. Assess. 162, 311–326.
sensitivity of Landsat and MODIS to forest disturbance associated with tropical cy- Wang, W., Qu, J.J., Hao, X., Liu, Y., Stanturf, J.A., 2010. Post-hurricane forest damage
clones. Remote Sens. Environ. 140, 679–689. assessment using satellite remote sensing. Agric. For. Meteorol. 150, 122–132.
Ortegren, J.T., Maxwell, J.T., 2014. Spatiotemporal patterns of drought/tropical cyclone Wang, Z., Román, M.O., Sun, Q., Molthan, A.L., Schultz, L.A., Kalb, V.L., 2018.
co-occurrence in the Southeastern USA: linkages to North Atlantic climate variability. Monitoring disaster-related power outages using NASA black marble nighttime light
Geogr. Compass 8, 540–559. product. Int. Arch. Photogramm. Remote. Sens. Spat. Inf. Sci. 1853–1856.
Parker, G., Martínez-Yrízar, A., Álvarez-Yépiz, J.C., Maass, M., Araiza, S., 2017. Effects of World Bank, 2018. World Development Indicators. pp Page. https://data.worldbank.
hurricane disturbance on a tropical dry forest canopy in western Mexico. For. Ecol. org/indicator/The World Bank.
Manag. 426, 39–52. Xuan, Z., Chang, N.-B., 2014. Modeling the climate-induced changes of lake ecosystem
Pielke Jr., R.A., Landsea, C., Mayfield, M., Layer, J., Pasch, R., 2005. Hurricanes and structure under the cascade impacts of hurricanes and droughts. Ecol. Model. 288,
global warming. Bull. Am. Meteorol. Soc. 86, 1571–1576. 79–93.
Potter, C., 2014. Global assessment of damage to coastal ecosystem vegetation from

13