Content uploaded by Joel Timyan
Author content
All content in this area was uploaded by Joel Timyan on Jan 17, 2025
Content may be subject to copyright.
Decoding primary forest changes in Haiti and the Dominican Republic
using Landsat time series
Falu Hong
a,*
, S. Blair Hedges
b
, Zhiqiang Yang
c
, Ji Won Suh
a
, Shi Qiu
a
, Joel Timyan
d
,
Zhe Zhu
a
a
Department of Natural Resources and the Environment, University of Connecticut, Storrs, CT 06269, USA
b
Center for Biodiversity, Temple University, 1925 N 12th Street, Suite 502, Philadelphia, PA 19122, USA
c
US Forest Service, Rocky Mountain Research Station, Riverdale, UT, USA
d
Haiti National Trust, #20, rue Faubert, Suite 3, P´
etionville, Haiti
ARTICLE INFO
Edited by Marie Weiss
Keywords:
Primary forest
Land cover
Landsat
Disturbance
Resilience
Haiti
The Dominican Republic
ABSTRACT
Forest loss has greatly reduced habitats and threatened Earth’s biodiversity. Primary forest (PF) has an irre-
placeable role in supporting biodiversity compared with secondary forest (SF). Therefore, distinguishing PF and
SF using remote sensing observations is critical for evaluating the impact of forest loss on biodiversity. However,
continuous monitoring of PF loss through remote sensing time series observations remains largely unexplored,
particularly in developing tropical regions. In this study, we used the COLD algorithm (COntinuous monitoring of
Land Disturbance) and Landsat time series data to quantify PF loss on the island of Hispaniola, comprising Haiti
and the Dominican Republic, from 1996 to 2022. We considered the resilience of PF to different disturbance
agents and identied the primary drivers of PF loss in Hispaniola through a sample-based approach. Accuracy
assessment based on the stratied random sample shows that the overall accuracy of land cover classication is
80.5% (±5.2%) [95% condence interval]. The user’s, producer’s, and overall accuracies of PF loss detection are
68.8% (±9.3%), 73.6% (±38%), and 99.4% (±0.5%), respectively. Map-based analysis reveals a more pro-
nounced decline in PF coverage in Haiti (0.75% to 0.44% at 324 ha/year) compared to the Dominican Republic
(7.14% to 5.67% at 2,704 ha/year), with substantial PF loss occurring both inside and outside protected areas.
Furthermore, Haiti exhibits a higher degree of PF fragmentation, characterized by smaller and fewer PF patches,
than the Dominican Republic, posing signicant challenges for biodiversity conservation. The remaining PFs are
found on steeper slopes in both Haiti and the Dominican Republic, suggesting that atter, more accessible areas
are more vulnerable to PF loss. Fire, tree-cutting, and hurricanes were identied as the primary drivers of PF loss,
accounting for 65.7%, 20.9%, and 9.0% of the PF loss area in Hispaniola, respectively. These ndings underscore
the urgent need for conservation policies to protect remaining PF in Hispaniola, particularly in Haiti.
1. Introduction
Habitat loss, driven by human activities and climate change, poses a
major threat to global biodiversity (Alroy, 2017;Brooks et al., 2002).
Among forest ecosystems, primary forest (PF), also often referred to as
old-growth forest, stands out for its irreplaceable role in supporting
biodiversity, offering unique ecological niches and harboring a dispro-
portionately high number of endemic species compared to secondary
forest (SF) (Barlow et al., 2007;Gibson et al., 2011;Hedges et al., 2018;
Martin et al., 2004). While remote sensing observations have called
attention to the widespread forest degradation and deforestation
(Hansen et al., 2013), accurately assessing the impact of forest loss on
biodiversity necessitates differentiating between PF and SF. This
distinction, however, remains a persistent challenge, especially in the
highly biodiverse yet data-scarce tropical regions. Existing PF maps
often suffer from inconsistencies due to varying denitions and meth-
odologies. Some studies rely on strict spatial criteria (size and length),
rendering PF virtually absent on small islands, while others dene PF as
undisturbed forest within a Landsat time series, potentially including old
secondary forest (Turubanova et al., 2018; Vancutsem et al., 2021).
Localized mapping efforts frequently employ their own denitions based
on local expert knowledge, sometimes using terms like intact forest or
* Corresponding author.
E-mail address: faluhong@uconn.edu (F. Hong).
Contents lists available at ScienceDirect
Remote Sensing of Environment
journal homepage: www.elsevier.com/locate/rse
https://doi.org/10.1016/j.rse.2024.114590
Received 26 June 2024; Received in revised form 21 December 2024; Accepted 23 December 2024
Remote Sensing of Environment 318 (2025) 114590
Available online 8 January 2025
0034-4257/© 2025 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ).
old-growth forest interchangeably (Ar´
evalo et al., 2020; Decuyper et al.,
2022; Kormos et al., 2018; Mikol¯
aˇ
s et al., 2023; Turubanova et al., 2018;
Wang et al., 2019). These discrepancies hinder intercomparison and
create challenges for large-scale, standardized mapping.
The widely used qualitative denition of PF, provided by the Food
and Agriculture Organization (FAO), describes it as “naturally regen-
erated forest of native species where there are no clearly visible in-
dications of human activities and the ecological processes are not
signicantly disturbed”(FAO, 2015). Previous studies have used similar
PF denitions based on the FAO (Food and Agriculture Organization)
denition (Bernier et al., 2017;Kormos et al., 2018;Mikol´
aˇ
s et al., 2019;
Sabatini et al., 2021), including two key criteria. First, tree canopy cover
should be dense enough to support biodiversity. For example, recent
studies have used relatively high tree canopy thresholds for PF, with
thresholds ranging from 60% to 100% (Hedges et al., 2018;Rodrigues-
Eklund et al., 2021;Turubanova et al., 2018). Second, PF should exhibit
an ecosystem structure and function that remains minimally disturbed.
PFs include pristine PFs, which are free of disturbance, and recov-
ered PFs, which have experienced disturbances but the original
ecosystem structure and functions remain intact or quickly recover after
disturbances (FAO, 2015). Pristine PFs are often in remote regions that
are inaccessible to human activities or well protected. Recovered PFs
demonstrate their ability to maintain their pre-disturbance forest
structure and ecosystem functions after a disturbance, provided it does
not exceed the PF’s short-term resilience capacity (ITTO, 2002). For
instance, PFs may recover from minor and short events (Knorn et al.,
2013), such as ash droughts or weak hurricanes, within a few years.
However, recovery from human-induced or severe disturbances such as
low-intensity tree-cutting or severe wildres, landslides, and hurricanes
can take much longer, often over fty years, and may result in irre-
versible damage, transforming PFs into degraded PFs or other land cover
types, such as SFs (Heinrich et al., 2023;Poorter et al., 2016). Degraded
PFs denote that PFs have been severely impacted and cannot maintain
their original ecosystem structure and functions after disturbances. They
represent a transition stage and usually lead to replacement with SFs.
Consequently, degraded PFs are typically considered a subset of SFs. In
this study, we will refer to both degraded PFs and SFs collectively as SFs
hereafter.
While previous studies have identied key features of PFs, the cur-
rent qualitative denition by the FAO poses practical challenges in
identifying PFs (Bernier et al., 2017). Remote sensing provides a quan-
titative approach to determining tree canopy cover, and long-term time
series observations can be used to quantify disturbance agents and their
intensity, as well as to measure forest recovery through spectral signa-
tures (Hansen et al., 2013;Vancutsem et al., 2021;Wang et al., 2019;
White et al., 2017;Zhu et al., 2022). These insights are valuable for
distinguishing different forest conditions, including pristine PF, recov-
ered PF, and SF. Therefore, a workow based on remote sensing time
series has the potential to establish a standardized method for identi-
fying PF at large scales (Vancutsem et al., 2021;Wang et al., 2020).
The island of Hispaniola, comprising Haiti and the Dominican Re-
public, was once predominantly covered by PFs, but now faces signi-
cant pressure of biodiversity loss due to deforestation (´
Alvarez-Berríos
et al., 2013;Myers et al., 2000). Deforestation can cause forest frag-
mentation, which then facilitates additional deforestation through edge
effects and greater exposure to wind (Laurance et al., 1997). Haiti has
experienced a dramatic loss of its PFs, as detailed by Hedges et al.
(2018). The tree-cutting of forests during colonial times, shifting culti-
vation, poor conservation policy, frequent natural disasters, such as
earthquakes and hurricanes, and climate change have caused the decline
of PF (Cohen, 1984;Hedges et al., 2018). Although Haiti and the
Dominican Republic share similar natural environment backgrounds,
there is a substantial disparity in the PF distribution and change pattern
between Haiti and the Dominican Republic due to differences in the
present economic status, historical population structure, and land
management policy (Wilson et al., 2001). This PF disparity is also
manifested in the differences of PF fragmentation level, the effectiveness
of PF-protected areas, and the major drivers of PF loss. Continuous
mapping of Hispaniola’s PFs using a time series from remote sensing can
reveal PF disparities between Haiti and the Dominican Republic and
concurrently monitor their distinct trends of PF loss.
However, most previous studies on Hispaniola have largely focused
on overall forest cover change, often failing to distinguish between PF
and SF. This lack of discrimination has led to various, sometimes con-
tradictory ndings on whether the forest cover is increasing or
decreasing in Haiti (Churches et al., 2014;Pauleus and Aide, 2020;
Rodrigues-Eklund et al., 2021) and the Dominican Republic (John and
Yolanda, 2019;Sangermano et al., 2015a). Such discrepancies arise
from varying denitions of land cover categories and the mixing of PF
and SF (Aide et al., 2013;´
Alvarez-Berríos et al., 2013). Only two pre-
vious published studies have analyzed PF change in Haiti and one study
reported PF change in the Dominican Republic. Both of them reported
PF loss in the two countries. Cohen (1984) reported extensive PF loss in
Haiti and projected the complete loss of PF before 2000 using aerial
photogrammetry between 1956 and 1976. Hedges et al. (2018) used
Landsat imagery and time-series analysis to study the loss of PF, nding
less than 1% of PF in Haiti since 2000. The Global Forest Watch project
mapped the pan-tropical PF in 2001 based on the training knowledge
from Amazon, Indonesia, and the Democratic Republic of Congo
(Hansen et al., 2013;Turubanova et al., 2018). However, the subsequent
trend analyses may be affected by methodological changes. Specically,
the increased sensitivity of the change detection model after 2013 and
improved detection of forest loss from 2015 onwards could lead to in-
consistencies when comparing pre- and post-2013/2015 trends (Weisse
and Potapov, 2021). This raises potential concerns about the accuracy of
PF loss estimations derived from these datasets (Palahí et al., 2021).
Furthermore, those studies overlooked the resilience of PF, i.e., the
ability of PF to recover to the pre-disturbance level in a short time (e.g.,
one year) with unchanged forest composition, structure, and ecosystem
function (Seidl and Turner, 2022). They assumed PF that has undergone
disturbance will directly convert to non-PF, such as SF or other land
cover types. That could be true for PF impacted by severe disturbance,
but not for PF that recovers quickly after mild disturbances. Further-
more, a comprehensive assessment of the long-term PF disparities be-
tween Haiti and the Dominican Republic is still unexplored. For
instance, the disparities in PF conversion patterns (where PF changes to
different land cover categories) and fragmentation levels have not been
quantied. Additionally, the effectiveness of the designed protected
areas in conserving PF, as well as the major drivers causing PF loss, have
not been adequately studied. This has created difculties in under-
standing PF disparities between the two countries and in providing
support for biodiversity conservation policy in Haiti and the Dominican
Republic.
To address these issues, this study has three major objectives: (1) to
develop a time-series-based approach to detect PF with consideration of
disturbance and resilience using dense Landsat time series; (2) to
continuously map the land cover in Haiti and the Dominican Republic
from 1996 to 2022 and quantify the PF conversion patterns; and (3) to
compare the PF conservation status between Haiti and Dominican Re-
public and identify the major drivers of PF loss in Hispaniola. Through
these objectives, this study aims to provide crucial data and insights to
advance PF conservation efforts in Hispaniola. By integrating novel
methodological approaches with an emphasis on PF resilience and the
drivers of change, our ndings contribute valuable knowledge to un-
derstanding and mitigating forest loss in this globally signicant biodi-
versity hotspot.
2. Study area and datasets
2.1. Study area
The island of Hispaniola, located in the Caribbean region, has a
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
2
tropical climate. It is the second largest island in the Caribbean with an
area of 76,192 km
2
, after the island of Cuba (Fig. 1). Hispaniola is
divided into two countries: Haiti on the western side covering an area of
27,750 km
2
and the Dominican Republic on the eastern side covering an
area of 48,442 km
2
. The topography of Hispaniola varies greatly, with
elevations in Haiti ranging from −28 to 2678 m (mean ±standard de-
viation, 408 ±373 m) and in the Dominican Republic from −58.0 to
3097 m (mean ±standard deviation, 400 ±487 m).
Hispaniola has a rich biodiversity with a high level of endemism in
many groups, including orchids, arthropods, amphibians, and reptiles
(Fern´
andez, 2007;Hedges et al., 2018;Sangermano et al., 2015b).
However, the biodiversity faces signicant human pressure due to the
island’s high population density, with over 11 million people in each
country (Haiti and the Dominican Republic) in 2021 (UN, 2022). This
demographic pressure poses a threat to local biodiversity and PF
conservation.
Historical factors and economic disparities between Haiti and the
Dominican Republic contribute to distinct pressure on PF conservation
in the two countries (Marzelius and Droste, 2022). Haiti is regarded as
the poorest country in the Western Hemisphere with a GDP per capita of
around $1,800 in 2022. Most of the economic activity in Haiti is
centered around agriculture. In contrast, the Dominican Republic has
the second largest economy in the Caribbean region with a GDP per
capita of around $10,000 in 2022. Tourism has replaced agriculture to
become the largest contributor to the economy (Sangermano et al.,
2015b).
2.2. Remote sensing and auxiliary dataset
We used all available Landsat Collection 2 Level 2 data to monitor
land disturbances in Hispaniola. The Landsat data, spanning from
January 1, 1984, to April 1, 2023, was downloaded from https://eart
hexplorer.usgs.gov/ (access date: Apr 2023). However, due to the
sparse data spanning from 1992 to 1995, we constrained our mapping
period from 1996 to 2022. The dataset, comprising 5,780 images, covers
eleven Landsat Worldwide Reference System (WRS)-2 scenes. We
created the Landsat Analysis Ready Data for Hispaniola to facilitate
analysis by projecting the original Landsat images to the Albers Equal
Area coordinate system and clipping to predened tiles.
Additionally, we compiled the 30-m resolution Shuttle Radar
Topography Mission (SRTM) elevation data from https://dwtkns.com/
srtm30m/ (access date: Apr 2023). This dataset was used to calculate
the slope and aspect for each pixel, aiding in land cover classication
and analysis of topographic characteristics in PFs. We also included 30-
m Height Above the Nearest Drainage (HAND) data, which normalizes
topography based on local relative heights within the nearest drainage
network. This dataset provided insights into PF changes in relation to
the topographic characteristics of the local environment (Donchyts
et al., 2016;Heinrich et al., 2023;Nobre et al., 2011).
Furthermore, we obtained the protected area boundaries of Haiti and
the Dominican Republic from the World Database on Protected Areas
(WDPA) (UNEP-WCMC and IUCN, 2023). Our objective was to analyze
the spatial distribution of PFs inside and outside protected areas. After
close examination, we retained the major protected areas that contain
the PF for further analysis (Fig. 1).
2.3. Training labels
We dened eight land cover types for Hispaniola, which include
developed, primary wet forest, primary dry forest, secondary forest,
shrub/grass, wetland, water, and other. Table 1 and Fig. A1 showed the
descriptions of each land cover type and examples of high-resolution
images, respectively. Through combining Google Earth high-resolution
images, Landsat time-series and local expert knowledge, we collected
high-quality training data with stable conditions. In total, we digitized
550 polygons, encompassing 342,013 pixels at 30-m resolution, be-
tween 2003 and 2022 (Table A1). The “Count of selected training
sample”column means the number of pixels used to train the random
forest classier (see Section 3.1.2 for the details).
3. Methodology
The workow for generating land cover (including PF) in Haiti and
the Dominican Republic consists of three main components (Fig. 2). The
rst component is to prepare training data, which includes using the
COLD (COtinuous monitoring of Land Disturbance) algorithm to detect
land disturbance and extract temporal trajectory features (Zhu et al.,
2020). The second component is to generate the land cover map with a
random forest model, followed by post-processing steps. The third
component is to evaluate the map’s accuracy using the collected refer-
ence sample.
3.1. Land disturbance detection and cover classication
3.1.1. Land disturbance detection using all available Landsat data
We used the COLD algorithm to detect land disturbance and classify
Fig. 1. Study area of Hispaniola. Inset (b) shows the location of Hispaniola. The background denotes elevation. The interior polygons denote the major protected area
boundaries containing the primary forest in Haiti and the Dominican Republic, respectively.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
3
land cover with all available Landsat images from 1996 to 2022 (Zhu
and Woodcock, 2014;Zhu et al., 2020). The COLD algorithm uses the
combination of linear and harmonic models (Eq. (1)) to t clear Landsat
observations for each band, then dynamically compares the consequent
clear observations with the model predictions. If the differences between
a certain number of consecutive anomaly observations and model
predictions exceed the change probability, land disturbances are
conrmed, and the corresponding temporal segment is created. When all
the clear observations are processed, the extracted temporal trajectory
parameters are used for land cover classication. Considering the rela-
tively limited clear observations in Hispaniola and the fact that PF is
sensitive to disturbance, we chose the sensitive parameter settings for
the COLD algorithm by adjusting the number of consecutive anomaly
observations from six (default) to four and change probability from 0.99
(default) to 0.95 (chi-squared distribution) (Zhu et al., 2020). More
discussion is provided in Section 5.2.
ρ
=a0+c1x+a1cos2
π
Tx+b1sin2
π
Tx+a2cos2
π
Tx
+b2sin2
π
Tx+a3cos2
π
Tx+b3sin2
π
Tx(1)
where
ρ
is the surface reectance from model prediction. Tis the number
of days per year (set to 365.25) and xis the Julian data. Parameter a
0
is
the overall surface reectance of the temporal segment. Parameter c
1
represents the inter-annual change (slope) for the temporal segment.
Parameters a
1
,b
1
,a
2
,b
2
,a
3
, and b
3
are the coefcients for the harmonic
functions.
3.1.2. Land cover classication based on COLD outputs
After running the COLD algorithm, we trained the random forest
(RF) machine learning model to classify the land cover of each temporal
segment. The predictor variables include the spectral-related parameters
derived from COLD-tted temporal segments and topographic charac-
teristics (elevation, slope, and aspect) derived from the digital elevation
model (Zhu and Woodcock, 2014). The RF model was selected since it
has proven successful on large-scale classication tasks in previous
research (Brown et al., 2020;Friedl et al., 2022).
The training sample distribution signicantly affects the perfor-
mance of the RF model (Belgiu and Dr˘
agut¸, 2016;Zhu et al., 2016).
Training sample with proportional distribution following the actual land
cover map achieve better accuracy than training sample with equal
distribution (Zhu et al., 2016). However, the proportion of each land
cover in the training sample remains unknown without the wall-to-wall
land cover map. To make the training sample proportion close to the
actual distribution and achieve better accuracy, we rst used the equal-
Table 1
Standards when collecting training labels.
Land cover Description
Developed Land containing constructed materials with impervious surface
percentages greater than 10%. Examples include roads, buildings,
parking lots, and low-density residential.
Primary wet
forest
A large patch of closed forest with nearly 100% tree cover.
Primary wet forests are usually above 600 m in Hispaniola. The
time series of forest change was considered when collecting the
training sample of the primary wet forest. Primary wet forests
should have abundant forest cover and have not undergone
signicant forest cover loss during the past four decades. The
pristine and recovered primary wet forests are both collected.
Primary dry
forest
A large patch of closed forest with nearly 100% tree cover.
Primary dry forests are usually below 600 m in Hispaniola. The
time series of forest change was considered when collecting
primary dry forest as we did for primary wet forest.
Secondary
forest
Areas with nonprimary forest tree cover greater than 10%, which
include but are not limited to (1) degraded primary forests after
signicant natural disturbances, such as hurricanes and severe
drought; (2) open forests after selective logging; (3) regrowth
forests after clearance, such as regenerated forests from
abandoned agricultural land; (4) tree plantations.
Shrub/Grass Areas covered with naturally generated grassland, herbs, and
bushes. The vegetation cover is greater than 10%, but usually, it is
hard to identify any forest stands from high-resolution images.
Water Areas with open water, generally with a water percentage greater
than 50%. Examples include rivers, lakes, reservoirs, bays, and
oceans.
Wetland Areas where water saturation is the determining factor in soil
characteristics and vegetation types. The soil or substrate is
periodically saturated with or covered with water. A typical
example in Hispaniola is mangrove forests.
Other Areas not belonging to the land cover types dened above. It
mainly consists of barren, croplands, and transitional types.
Examples include cultivated crops, pasture for cattle, sand, rocks,
bare soil, abandoned agricultural land with unnatural grassland,
and deforested areas with shallow vegetation.
Fig. 2. Flowchart of land cover (including primary forest) map generation in Haiti and the Dominican Republic. COLD denotes the COtinuous monitoring of Land
Disturbance algorithm (Zhu et al., 2020).
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
4
distribution training sample to generate the initial land cover classi-
cation to approximate the actual land cover proportion, and then
adjusted the training sample proportion based on the proportion
calculated from the initial results of land cover classication and the
optimized strategy from Zhu et al. (2016). Around 25,000 sample points
were randomly extracted from each digitized polygon to make sure each
polygon contributed to the land cover classication. The number of
training sample for each polygon was based on the polygon area. The
nal counts of selected training sample used for the RF training are
provided in Table A1.
3.1.3. Primary forest post-processing with the consideration of resilience
and patch size
RF classication results contained errors that were not consistent
with the real-world evolution of PF. For instance, PF which fully
recovered to the pre-disturbance condition in a short time (recovered
PF) was misclassied as SF. The degraded PF, which should have been
classied as SF after disturbance, was still classied as PF. Additionally,
regrowth SF was misclassied as PF. However, even in an ideal situation
where seeds and stems of the original forest species are present, SF
usually takes over 50 years to recover to PF (Poorter et al., 2016), which
could not be observed during our mapping period (1996–2022). We also
observed classication errors in the scattered PF pixels. Those scattered
PF pixels, which are not large enough to support the biodiversity that PF
normally supports, were misclassied as PF. To correct the classication
errors, we applied three post-processing steps.
First, we separated the recovered PF and SF, which originate from
degraded PF, by considering the resilience—specically, whether the PF
recovered shortly (e.g., one year) after the disturbance. We evaluated
the condition of post-disturbance PF and calculated the corresponding
vegetation index (VI) change before and after the disturbance. The VI
change was calculated by subtracting the maximum vegetation index
one year after the disturbance from the maximum vegetation index one
year before the disturbance. We linked the short-term VI change to the
PF recovery within a short time (one year) to indicate the resilience of PF
in various disturbance agents.
To determine the optimal VI and threshold for assessing whether the
post-disturbance PF had recovered, we randomly selected 120 poten-
tially disturbed PF pixels from the initial land cover classication map in
Section 3.1.2. We scrutinized the Landsat time series and interpreted the
disturbance agents. After excluding the 28 misclassied PF sample units,
we selected ve VIs and calculated the VI changes before and after the
disturbance for the remaining 92 pixels and evaluated the PF condition
after different disturbance agents. The details on the selected ve VIs
were provided in Appendix B.
Among the selected VIs, we selected the Normalized Burn Ratio
(NBR) (Key and Benson, 2006) index and 0.05 as the empirical threshold
(Fig. 3). This threshold approximately separates PF disturbances
induced by strong agents (re, hurricane, landslide, and tree-cutting)
and those induced by weak agents (such as drought) or commission
error. We found that strong disturbance agents, such as re, hurricanes,
landslides, and tree-cutting, often lead to irreversible PF loss or degra-
dation, converting PF into SF or other land cover types, and corre-
sponding signicant NBR declination. In contrast, the NBR change
induced by drought is not as signicant. The PF condition changes due to
drought, such as loss of leaves, and can be detected by Landsat time
series. However, the original PFs remain and recover to their pre-
disturbance condition within a short time after the drought ends. This
means that these PFs are resilient to drought. Furthermore, the NBR
change caused by drought is similar in magnitude to that caused by
commission errors of the change detection algorithm. This suggests that
including the PF loss induced by drought or other recoverable distur-
bances may introduce considerable commission errors of PF loss
detection.
Compared to NBR, it is challenging to nd a single threshold among
other VIs that can separate the degraded PF experiencing strong
disturbances from recovered PF experiencing weak disturbances (Fig. B1
in Appendix B). Although NBR is designed to enhance the burned area,
previous studies have demonstrated its usefulness in indicating forest
recovery and resilience (Cohen et al., 2018;Kennedy et al., 2012;White
et al., 2017).
Therefore, for PF pixels that experienced disturbance, if the absolute
NBR change (ΔNBR) did not exceed 0.05 one-year after the disturbance,
the post-disturbance segment remained as PF (Fig. 4a). If the absolute
NBR change exceeded 0.05 and the post-disturbance segment was
classied as PF, we corrected it to SF (Fig. 4b).
Second, we corrected the PF classication results for segments that
previously contained non-PF classications, reclassifying them as SF
(Fig. 4c). If previous segments were classied as non-PF types, but the
current segment was classied as PF, we corrected the current segment
to SF. This was to correct the errors of misclassifying growing SF as PF.
Third, we dened the 3 ×3 minimum mapping unit (MMU) for PF
clusters to exclude the misclassied scattered PF pixels. To achieve this,
we segmented the PF pixels into multiple clusters based on the 8-
neighbor rule and converted the PF clusters that do not contain the 3
×3 PF structure to SF. The reason we selected 3 ×3 as the MMU is that
the central pixel surrounded by neighboring PF pixels is less affected by
the impact of edge effects (Laurance et al., 1997). It can serve as the core
habitat for the connected PF pixels. The line-shape PF clusters do not
include the core habitat to support biodiversity and were thus removed.
3.1.4. Further renement of the land cover map
After reducing errors in PF classication, we made further re-
nements to improve the land cover map based on the time series tra-
jectory. First, we corrected the misclassied developed pixels. For
temporal segments classied as developed but with subsequent seg-
ments consistently classied as non-developed, we replaced the segment
that was classied as developed (rst segment in Fig. 4d) with the land
cover type associated with the second highest prediction probability in
the RF model.
Second, we lled the gaps in the land cover time series to ensure the
continuity of land cover from 1996 to 2022. Data density issues can
affect the initialization and tting of the COLD algorithm, leading to
Fig. 3. Boxplots showing Normalized Burn Ratio (NBR) index changes for
different primary forest disturbance agents, which are calculated by subtracting
post-disturbance NBR from pre-disturbance NBR. The numbers on top of the
boxplots indicate the sample count for each agent. In each boxplot, the dot
point and line inside the boxplot represent the mean and median values,
respectively. The box of the boxplot represents the interquartile range (the
range between the rst and third quartiles). The whiskers represent the
maximum and minimum values of the plot data within 1.5 times of the inter-
quartile range from the rst and third quartiles, respectively. The diamond
point represent outliers that fall outside the whiskers.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
5
gaps in the land cover time series (Zhang et al., 2022). To address this
issue, we used subsequent land cover segments to ll these gaps. If the
land cover remained discontinuous at the end with no subsequent seg-
ments available, we extended the previous classication segments to ll
the gaps.
3.2. Accuracy assessment and primary forest loss driver identication
We used two sets of stratied random samples to evaluate the land
cover classication accuracy and PF loss detection accuracy. The spatial
unit is the mapped Landsat pixel (30 m ×30 m). When interpreting PF
loss sample, we also recorded the drivers causing PF loss to determine
the major drivers contributing to PF loss in Hispaniola.
For the accuracy assessment of land cover classication, we followed
the good practice workow to generate and allocate the reference
sample (Olofsson et al., 2014). We set the target standard error of the
estimated overall accuracy as 0.02 and conjectured the user’s accuracies
as 0.8 for all types. The total sample size is 400 (Eq. 13 in Olofsson et al.
(2014)). The stratum weight was calculated based on 27-year land cover
maps. For rare classes (less than 10%), fty sample units were allocated
for primary wet and dry forests stratum to reduce the estimated vari-
ance. Twenty-ve sample units were allocated to other rare strata. The
remaining reference sample was allocated proportionally to the domi-
nant stratum (more than 10%).
Four trained interpreters were involved in the interpretation. To
reduce the interpreter’s variability in the understanding of each land
cover, four interpreters interpreted the 40 mutual sample units (5
random sample units from each stratum) as the training before the
Fig. 4. Schematic of post-processing steps. Subplots (a) and (b) display the time series of the Normalized Burn Ratio (NBR) index at (17.56609 N, 71.51920 W) and
(19.02904 N, 70.97138 W), respectively. Those pixels experienced drought (a) and re (b), respectively. Subplots (a) and (b) illustrate the process of separating
recovered primary forest (PF) and secondary forest using the NBR threshold. Subplot (c) displays the time series of the NBR index at (19.00815 N, 71.18575 W). It
shows the process of correcting the misclassied PF whose previous segments were classied as non-PF. The red text shows the misclassied results from the random
forest model. Subplot (d) displays the time series of the NBR index at (18.28849 N, 71.34908 W). It shows the process of correcting the misclassied developed pixels
whose subsequent segments were classied as non-developed types. (For interpretation of the references to colour in this gure legend, the reader is referred to the
web version of this article.)
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
6
formal interpretation. When interpreting reference sample, four in-
terpreters had no prior information about the reference sample (i.e.,
“blind”interpretation) and interpreted the reference sample indepen-
dently. Google Earth high-resolution images, Landsat time series ob-
servations, and elevation information were used together to label the
reference sample. The dominant land cover type with the largest pro-
portion was assigned as the label for the reference sample. After
completing the interpretation of 40 mutual sample units, four in-
terpreters convened to discuss the scoring criteria and reached a nal
decision on the sample units with disagreement. In this way, we ensured
the understanding of each stratum is consistent among different in-
terpreters to reduce interpreter variability (Olofsson et al., 2014;Pengra
et al., 2020;Powell et al., 2004).
After completing the interpretation of the 40 mutual sample units,
four interpreters were assigned to interpret the remaining sample. In
addition to labeling the land cover types, each interpreter also recorded
the condence levels (100%, 80%, 50%, or less than 50%) for each
reference sample. After completing the initial interpretation, all in-
terpreters re-interpreted challenging sample with low condence level
(50% or less). The nal labels of these challenging sample units were
determined using majority vote and group discussion. We excluded
reference sample where interpreters could not reach an agreement and
sample that lacked reliable reference sources, such as sample with dy-
namic and mixed land cover before 2000. The excluded sample units
were approved by all interpreters.
After nishing the interpretation, we reported the user’s accuracy,
producer’s accuracy, overall accuracy, and the uncertainties at the 95%
condence level based on agreement between the reference label and
mapped land cover type (Olofsson et al., 2014). The agreement was
dened as the exact match between the reference label and mapped
type.
For the accuracy assessment of PF loss detection, we dened two
strata: PF loss and other. These were based on the accumulated PF loss
map spanning the entire period from 1996 to 2022. PF loss indicates that
a pixel was classied as PF in 1996 but changed to non-PF types between
1996 and 2022. The other stratum refers to pixels that remained stable
PF from 1996 to 2022 or were classied as other non-PF land cover
types. We followed the same workow of the land cover classication
assessment to assess the accuracy of PF loss detection. We generated 486
sample units by setting the standard error of the estimated overall ac-
curacy as 0.01 and conjecturing the user’s accuracy of PF loss and other
as 0.7 and 0.95. We assigned 100 and 386 sample units for the PF loss
and other strata, respectively. If the reference sample is labeled as PF
loss, we also recorded the driver of PF loss. We listed ve major drivers
of PF loss in Hispaniola based on prior information, which includes re,
tree-cutting, hurricane, landslide, and other. Like the process of label-
ling reference sample, we identied the drivers of PF loss using multi-
source information (including Google Earth high-resolution images,
Landsat time series observations, etc.) and group discussions from all
interpreters.
3.3. Spatiotemporal comparison of primary forest status between Haiti
and the Dominican Republic
After generating the land cover dataset that distinguishes PFs and SFs
in Hispaniola, we conducted a spatiotemporal comparison of the PF
status in Haiti and the Dominican Republic to evaluate PF conservation
status. We quantied the annual PF loss rate, identied hotspots of PF
loss, and determined the major drivers causing the hotspots of PF loss.
We also compared the PF changes inside and outside protected areas in
both countries to explore the effectiveness of protection policy on PF.
The topographic characteristics of PF were also analyzed to explore
the factors affecting PF loss. We investigated the distribution of existing
and lost primary wet and dry forests across the elevation and slope
gradients and calculated changes in mean elevation and slope of PFs
from 1996 to 2022 using SRTM and HAND datasets. Additionally, we
evaluated the fragmentation level of PFs by calculating the patch density
(number of PF patches per 100 km
2
) and mean patch area of PFs. We
segmented the PF pixels into distinct patches based on the 8-neighbor
rule, then computed the number of PF patches, patch density, and
mean patch area for Haiti and the Dominican Republic. A sparse dis-
tribution of PF patches with smaller average patch sizes indicates higher
levels of forest fragmentation, and vice versa. These analyses compre-
hensively demonstrate the PF disparity between Haiti and the Domin-
ican Republic.
4. Results
4.1. Annual land cover change in Haiti and the Dominican Republic
We generated the annual 30-m land cover map from 1996 to 2022
and determined rates of PF loss (Fig. 5). Fig. 5 shows that Haiti has less
PF (including the primary wet and dry forests) coverage and a higher
proportion of PF loss compared to the Dominican Republic. The map-
based analysis reveals that PF in Haiti decreased from 20,447 ha
(0.75%) to 12,023 ha (0.44%) with a decrease rate of 324 ha/y. The total
area of PF decreased by 41.2% from 1996 to 2022 (Fig. 6). The ndings
align with the previous study by Hedges et al. (2018) which also re-
ported less than 1% PF coverage in Haiti. In contrast, PF (including the
primary wet and dry forests) in the Dominican Republic decreased from
343,221 ha (7.14%) to 272,913 ha (5.67%) with a decrease rate of
2,704 ha/y. The total area of PF decreased by 20.5 % from 1996 to 2022.
The map-based annual land cover change results show that the
developed area in Haiti increased from 0.63% (~21,068 ha) to 1.32%
(~44,095 ha) with an expansion rate of ~717 ha/year (Fig. 6). In
contrast, the developed area of the Dominican Republic increased from
1.37% (~81,101 ha) to 1.96% (~116,119 ha) with an expansion rate of
~1,090 ha/year.
Not surprisingly, the proportion of SF increased in both countries due
to the strong disturbance of PF, which often leads to SF formation, and
the growth of SF from abandoned cropland and deforested areas. One of
the most surprising results is that more than half of the land surface in
Haiti is covered with barren, abandoned cropland, or deforested surface
("other" type with gray line in Fig. 6), a much higher proportion than in
the Dominican Republic. This is likely due to intensive and successive
agricultural activities and deforestation in the 20th century (see the
Discussion section for details) (Cohen, 1984;Hedges et al., 2018;Mar-
zelius and Droste, 2022).
We also analyzed the PF conversion pattern in Haiti and the
Dominican Republic (Fig. 7). Map-based results show that Haiti expe-
rienced the largest PF loss during the 2016–2017 period, which was
mainly caused by Hurricane Matthew (refer to Section 4.3 for more
details). Around 78.4% of lost PF converted to SF when disturbances
occurred, while around 21.6% of lost PF converted to other land cover
types. In contrast, in the Dominican Republic, around 91.3% of lost PF
transitioned to SF, and around 8.7% of lost PF converted to other land
cover types (Fig. 7). The mapping results show that both Haiti and the
Dominican Republic have experienced substantial PF loss, with the sit-
uation in Haiti being worse than that in the Dominican Republic.
4.2. Map accuracy and major primary forest loss drivers
For the accuracy assessment of land cover classication, we excluded
29 sample units that are difcult to label (Table C1). The area-based
confusion matrix shows that the overall accuracy of land cover classi-
cation is 80.5% (±5.2%) [95% condence interval] (Table 2). Specif-
ically, the user’s and producer’s accuracies for primary wet forest
classication are 83.7% (±10.5%) and 84.4% (±25.2%), respectively.
The user’s and producer’s accuracies for primary dry forest classication
are 86.0% (±9.7%) and 96.0% (±7.6%), respectively. The large vari-
ances are attributed to the limited number of reference sample and the
mixture with large proportion types, such as SF. The area adjustment
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
7
results show that both primary wet and dry forests are slightly over-
estimated based on the current land cover map, suggesting that there is
less remaining PF in Hispaniola. Shrub/Grass has the lowest accuracy
with the user’s and producer’s accuracies of 39.1% (±20.4%) and 23.4%
(±17.0%), respectively. This is because (1) the spectral signature of
shrub/grass is frequently mixed with SF and other types; (2) shrub/grass
is a rare type with the mapped proportion of 2.4%. Their accuracies can
be signicantly affected by dominant types, such as SF and other types.
For the accuracy assessment of PF loss detection, we excluded six
hard-to-interpret sample units (Table C2). The area-based confusion
matrix shows that the overall accuracy is 99.4% (±0.5%) (Table 3).The
user’s and producer’s accuracies for PF loss detection are 68.8%
(±9.3%) and 73.6% (±38%), respectively. The large variance in the
producer’s accuracy is due to the signicant area difference between the
PF loss stratum (1.05%) and the other stratum (98.95%) (Olofsson et al.,
2020). The area adjustment result shows that the current map slightly
overestimates the PF loss area with a bias of 52.3 km
2
(Table 3).
The PF loss driver analysis based on the sample interpretation shows
that re (44/67) accounts for 65.7% PF loss in Hispaniola, equivalent to
482.7 km
2
. Tree-cutting (14/67) caused 20.9% PF loss, equivalent to
153.6 km
2
. Hurricanes (9/67) caused 9.0% PF loss, equivalent to 65.8
km
2
. Landslides (2/67) caused 3.0% PF loss, equivalent to 21.9 km
2
.
Though PFs are resilient and could sustain many of the wildres if they
are not severe enough, res still contribute almost two-thirds of total PF
loss.
4.3. Hotspots of primary forest loss
We further explored the hotspots of PF loss highlighted in Fig. 5c and
displayed the PF loss year (Fig. 8). In hotspot A near the Grande Colline
Fig. 5. Land cover maps of Hispaniola in 1996 (a) and 2022 (b). Subplot (c) shows the primary forest (PF) change information. The ve rectangles in subplot (c) are
the hotspots of primary forest loss enlarged in Fig. 8.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
8
National Park and Macaya National Park in Haiti (Fig. 8), most of the PF
loss happened during 2016 and 2017. This was mainly caused by Hur-
ricane Matthew in Oct 2016, which caused widespread destruction of
housing, affecting 2.1 million people, and causing over 500 deaths in
that country (De Giorgi et al., 2021). Because of the availability of
Landsat data and cloud cover impact, conrmation of some PF loss that
happened during this event was delayed until the availability of clear
Landsat data as late as 2017. Besides the PF loss from Hurricane
Matthew, we found that subsequent PF loss in 2021 and 2022 in the
same affected region, corresponded to landslides caused by the 2021
Haiti Earthquake. The earthquake resulted in many large landslides that
destroyed PF. The 14 August 2021 earthquake is considered the largest
natural disaster on Earth for that year, causing 2,250 fatalities, and
climatic preconditioning from Hurricane Matthew exacerbated the slope
instability leading to landslides (Havenith et al., 2022).
We further plotted the land cover change in Macaya National Park
(Fig. 9), which is the largest national park in Haiti with the highest
proportion of PF. Map-based results show that Hurricane Matthew in
2016 and the Earthquake in 2021 caused 1,910 ha and 411 ha of PF loss
in Macaya National Park, respectively. These losses account for 22.6%
and 4.9% of the total PF loss area in Haiti between 1996 and 2022
(8,424 ha based on Fig. 6). Moreover, based on the PF areas in Haiti in
2015 (15,785 ha) and 2020 (12,823 ha), the PF losses in Macaya Na-
tional Park caused by Hurricane Matthew in 2016 and the Earthquake in
2021 represent 12.10% and 3.21% of the national PF area, respectively.
Additionally, landslides in 2021 caused signicant SF loss (~294 ha)
and expansion of barren land in Macaya National Park (Fig. 9).
Hotspot B near the Forˆ
et de Pins region in Haiti shows a more frag-
mented PF loss pattern. This region experienced dramatic deforestation
during the 20th century and the pattern of existing PF in 1996 was
already fragmented. Continuous human pressure resulted in ongoing PF
loss, reecting a fragmented and vulnerable forest landscape.
Hotspot C near the Sierra de Bahoruco National Park in the Domin-
ican Republic, shows multiple large patches of PF loss in a single year.
Map-based results show the areas of PF loss in 2000, 2007, 2013, 2015,
and 2022 were around 838 ha, 1,109 ha, 3,265 ha, 1,245 and 2,242 ha,
respectively. Landsat time series conrmed re is the major driver
causing the PF loss in these years.
Hotspot D near the Valle Nuevo National Park in the Dominican
Republic shows two large patches of PF loss that occurred in 2014 and
2015, which was due to the re in the summer of 2014. Map-based re-
sults show around 5,000 ha of PF loss, which accounted for about 1.76%
of the Dominican Republic’s total PF in 2014.
Hotspot E near the Jos´
e del Carmen Ramírez National Park in the
Dominican Republic shows two big re events in 1997 and 2005, ac-
counting for the major PF loss events. Map-based results show a loss of
approximately 6,400 ha and 10,900 ha PF in 1997 and 2005, respec-
tively. These losses accounted for about 1.9% and 3.4% of the total PF in
the Dominican Republic in the respective years.
4.4. Primary forest change within and outside the protected area
We compared the distribution and variation of PF within and outside
the protected areas in Haiti and the Dominican Republic (Fig. 10). Map-
based results show that approximately 31% of primary forests in Haiti
are not within protected areas. This is particularly notable for primary
dry forests, with about 50% located outside protected areas. In contrast,
the Dominican Republic has 86% of its PFs contained within protected
areas, which aligns with expectations given its larger number of desig-
nated protected areas.
In Haiti, PF loss inside the protected area covers a higher proportion
of the total primary forest loss for both primary wet and dry forests. Not
only is the area of PF loss greater inside these areas, but the proportion of
total PF loss is also higher. In the Dominican Republic, the loss of pri-
mary wet forest inside protected areas accounts for the majority of the
total primary wet forest loss. However, the loss of primary dry forest
outside protected areas constitutes the majority of total primary dry
forest loss. This is because most of the primary dry forest is well-
Fig. 6. Map-based land cover change in Haiti (a &c) and the Dominican Republic (b &d) from 1996 to 2022. The left and right y-axes show the area and cor-
responding percentage, respectively. For better visualization, the curves for secondary forest and other land cover types are displayed separately due to the
signicantly unbalanced proportions among the various land cover types.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
9
preserved in Jaragua National Park and Cotubanam´
a National Park in
the southern Dominican Republic.
4.5. Topographic characteristics of primary forest
We analyzed the topographic characteristics of the primary wet and
dry forests in Hispaniola (Fig. 11). There is a notable elevation distinc-
tion between primary wet and dry forests. Most primary dry forests are
in low-elevation areas (typically below 600 m) with small slope (typi-
cally less than 5 degrees). In contrast, most primary wet forests are in
high-elevation areas (typically above 600 m). In Haiti, primary wet
forests are concentrated in two zones: Zone I is around 1,500 m with a
slope of ~40 degrees while Zone II is around 2,000 m with a slope of
~10 degrees (Fig. 11a). In the Dominican Republic, the primary wet
forests are concentrated in regions with elevations ranging from 1,100 to
2,100 m and the slope ranges from 5 to 35 degrees (Fig. 11c).
We also analyzed topographic characteristics of the lost primary wet
forest to explore factors affecting its loss. Fig. 11b shows that most of the
primary wet forest loss happened in Zone II, indicating that regions with
lower slopes are more prone to primary wet forest loss. This phenome-
non is also observed in the Dominican Republic. Fig. 11d shows a rela-
tively high proportion of primary wet forest loss in relatively small-slope
regions. To study it in detail, we analyzed the mean elevation and slope
change of primary wet forests from 1996 to 2022 using the SRTM and
HAND topography datasets (Fig. 12). Our analysis revealed that, in
general, the mean SRTM elevation of primary wet forests is decreasing
while the mean HAND elevation of them is increasing. Additionally, the
mean slope of primary wet forests in Haiti and the Dominican Republic
shows increasing trends in both datasets.
The slope change is as predicted: primary wet forest has a higher
probability of surviving on land with steeper slopes because it is more
inaccessible to humans. This implies that slope emerges as an important
factor of PF loss. The mean slope declination in Haiti after 2020 is
attributed to landslide-induced PF loss, which is the major contributor to
PF loss, as PF in steeper regions is more likely to be affected by landslide
(Havenith et al., 2022). This further implies that slope plays a critical
role in affecting PF loss. The decrease in SRTM elevation (120 m in Haiti
and 40 m in the Dominican Republic) is because PF in higher elevations
but with relatively smaller slope has a greater contribution to the PF
loss. The increase in HAND elevation suggests that PFs closer to a
drainage basin are more susceptible to disturbances from human ac-
tivities and natural events, such as hurricanes (Heinrich et al., 2023).
4.6. Primary forest fragmentation in Haiti and the Dominican Republic
Fragmentation analysis of PFs shows that Haiti has a lower patch
Fig. 7. Map-based annual loss and conversion of primary forest (PF, including primary wet and dry forests) from 1996 to 2022 in Haiti (a) and the Dominican
Republic (b), respectively. Each value for a given year indicates the detected PF loss area from the previous year to the indicated year. For example, the value for the
year 2022 indicates the detected PF loss area from 2021 to 2022. The blue colour represents the area of PF converted to secondary forest (SF), while the orange colour
represents the area of PF converted to other land cover types. (For interpretation of the references to colour in this gure legend, the reader is referred to the web
version of this article.)
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
10
density than the Dominican Republic (blue lines in Fig. 13). Addition-
ally, the mean patch area of PF (orange lines in Fig. 13) in Haiti is
smaller than that of the Dominican Republic, for both primary wet and
dry forests. Those ndings indicate a higher level of PF fragmentation in
Haiti.
The patch density of primary wet forest in Haiti generally decreased
due to the ongoing human pressure from tree-cutting (Fig. 13a). There
was a sudden increase in the patch density of primary wet forests from
2016 to 2017, caused by the fragmentation of large PF patches in
Macaya National Park induced by Hurricane Matthew. The mean patch
area of the primary wet forest decreased consistently with fragmenta-
tion. For primary dry forests in Haiti, the time series of the fragmenta-
tion level change showed that the patch density and mean patch area
decreased from 1996 to 2022, indicating simultaneous clearance of
small primary dry forest patches and fragmentation of large primary dry
forest patches (Fig. 13b).
In the Dominican Republic, the mean patch area of primary wet
forest consistently decreased, while the patch density initially increased
and then decreased after peaking in 2014 (Fig. 13c). This implies that
the primary wet forest loss was initially dominated by the fragmentation
of large-patch primary wet forest into small patches. Subsequently, the
patch density was reduced by the clearance of small patches of primary
wet forest. The primary dry forest shows a different pattern compared
with the primary wet forest (Fig. 13d). The patch density decreased with
the increase in the mean patch area. This indicates that small patches of
primary dry forest were removed while the large patches were not,
corresponding to the primary dry forest in Jaragua National Park and
Cotubanam´
a National Park, the two largest areas of primary dry forest in
the Dominican Republic.
The ongoing PF fragmentation in Haiti and the Dominican Republic
exacerbates edge effects, further encroaching on the remaining PFs
(Laurance et al., 1997). This fragmentation also impedes forest recovery,
as the growing distance between PFs and SFs limits natural regeneration
process (Heinrich et al., 2023). Moreover, the higher level of PF frag-
mentation in Haiti poses signicant challenges for maintaining biodi-
versity. Smaller and less connected habitats increase the risk of
biodiversity loss, further threatening ecological stability (Hedges et al.,
2018).
5. Discussion
5.1. Implications of primary forest mapping in Hispaniola
Primary forest is critical habitat for supporting biodiversity in trop-
ical regions (Gibson et al., 2011). Separating PF and SF is therefore
critical for evaluating the true impact of forest loss on biodiversity. Our
study successfully distinguished PF and SF using dense Landsat time
series data and high-quality training sample. We developed a quantita-
tive and standardized approach for continuously mapping PF and SF,
which can potentially be used to create a universal PF and SF map on a
large scale. Factors such as forest height and age may be considered for
inclusion in future mapping efforts.
When mapping PF, we considered its resilience, dened as the ability
of PF to recover to pre-disturbance levels within a short period (e.g., one
year) with unchanged forest composition, structure, and ecosystem
function (Seidl and Turner, 2022). We measured PF recovery and indi-
cated the resilience using changes in vegetation indices (VIs) before and
after the disturbance. Through sample interpretation (Sections 3.1.3),
we used the change in NBR index with a threshold of 0.05 to distinguish
PF that has quickly recovered within one year after a weak disturbance
and PF that has been permanently degraded and converted to SF after a
Table 2
Confusion matrix of land cover classication in area proportions, adjusted area, and reported accuracy at the 95% condence interval. The “Map bias”denotes the
mapped area minus the adjusted area. The “Area 95% CI”denotes the uncertainty of the adjusted area at the 95% condence interval.
Reference
Map Developed Primary wet forest Primary dry forest Secondary forest Shrub/
Grass
Water Wetland Other Total
Confusion matrix, area proportion
Developed 0.0116 0 0 0.0007 0 0 0 0.0021 0.014
Primary wet forest 0 0.0297 0 0.0058 0 0 0 0 0.036
Primary dry forest 0 0.0001 0.0060 0.0008 0 0 0 0 0.007
Secondary forest 0 0.0054 0. 0.3052 0.0161 0 0 0.0643 0.391
Shrub/Grass 0 0 0 0.0104 0.0094 0 0 0.0042 0.024
Water 0 0 0 0 0 0.0083 0.0004 0.0004 0.009
Wetland 0 0 0.0003 0.0008 0 0.0010 0.0040 0 0.006
Other 0 0 0 0.0581 0.0145 0 0.0097 0.4310 0.513
Total 0.012 0.035 0.006 0.382 0.040 0.009 0.014 0.502
Area and accuracy
Mapped area (km
2
) 29,149.5 72,126.7 14,133.7 793,648.0 48,599.4 18,402.4 12,176.6 1,042,143.2
Adjusted area (km
2
) 23,597.2 71,505.4 12,662.4 775,188.3 81,127.4 18,959.6 28,516.9 1,018,822.0
Map bias (km
2
) 5,552.3 621.2 1,471.4 18,459.6 −32,528.1 −557.2 −16,340.3 23,321.1
Area 95% CI (km
2
) 5,016.6 22,610.9 1,695.4 99,552.2 50,153.4 2,725.6 27,261.0 100,241.3
User’s accuracy (%) 81.0 ±17.2 83.7 ±10.5 86.0 ±9.7 78.1 ±9.3 39.1 ±20.4 92.0 ±10.9 66.7 ±19.3 84.0 ±7.0
Producer’s accuracy (%) 100.0 ±0.0 84.4 ±25.2 96.0 ±7.6 79.9 ±6.9 23.4 ±17.0 89.3 ±8.8 28.5 ±27.7 85.9 ±5.9
Overall accuracy (%) 80.5 ±5.2
Table 3
Confusion matrix of primary forest (PF) loss detection in area proportions,
adjusted area, and reported accuracy at the 95% condence interval. The “Map
bias”denotes the mapped area minus the adjusted area. The “Area 95% CI”
denotes the uncertainty of the adjusted area at the 95% condence interval.
Reference
Map Other PF loss Total
Confusion matrix, area proportion
Other 0.9870 0.0026 0.9895
PF loss 0.0033 0.0072 0.0105
Total 0.9902 0.0098
Area and accuracy
Mapped area (km
2
) 74,411.9 787.3
Adjusted area (km
2
) 74,464.2 735.1
Map bias −52.3 52.3
Area 95% CI (km
2
) 386.8 386.8
User’s accuracy (%) 99.7 ±0.5 68.8 ±9.3
Producer’s accuracy (%) 99.7 ±1.0 73.6 ±38
Overall accuracy (%) 99.4 ±0.5
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
11
strong disturbance. Strong disturbances, such as tree-cutting, can
permanently alter the PF structure and signicantly reduce the NBR
index (Fig. 4b). However, the impact of disturbances on PF structure and
the NBR index can also depend on the resilience of PF. Weak disturbance
can signicantly reduce the NBR index in PF with low resilience. Frag-
mentation of PF can reduce its resilience and intensify the impact of
disturbances on forest structure and the corresponding spectral response
(Hansen et al., 2020). There are various ways to consider forest resil-
ience, and the denition of resilience varies among studies (Poorter
et al., 2016). In our study, spectral change measured by Landsat time
series offers an effective way to consider resilience in large-scale PF
mapping (White et al., 2017). Combining this with other contextual
information, such as the eld measurement reports, can improve resil-
ience measurement (Kennedy et al., 2012).
Our study applied the dense Landsat time series to continuously map
PF and SF in Hispaniola. We conducted a comprehensive evaluation of
PF disparity between Haiti and the Dominican Republic, a comparison
not previously undertaken by other studies. We compared the PF con-
version pattern, PF status inside and outside protected areas, and PF
fragmentation levels (Fig. 7,Fig. 10 &Fig. 13). Our comparison results
demonstrate that there is less PF in Haiti and it is more fragmented than
in the Dominican Republic, indicating a need for more urgent conser-
vation action for PF in Haiti. Our analysis also highlights widespread PF
loss inside and outside protected areas of Hispaniola. To accurately
assess the effectiveness of protection policy on PF, it is important to
consider PF outside of the protected areas with similar spatial and
climatic characteristics (Geldmann et al., 2019; Liang et al., 2023).
However, the widespread PF loss within protected areas revealed in our
study raises concerns for policymakers who need to improve protection
strategies (John and Yolanda, 2019;Sangermano et al., 2015b).
Additionally, we found SFs are increasing in the two countries due to
the conversion of PF and forest regrowth, resulting in an overall increase
in total forest area. This resolves the contradictory ndings from pre-
vious studies on whether the tree cover is increasing or decreasing in
Haiti and the Dominican Republic, which often conated PF and SF
(Churches et al., 2014;John and Yolanda, 2019;Pauleus and Aide,
2020;Rodrigues-Eklund et al., 2021;Sangermano et al., 2015a). Our
map results supported previous work (e.g., Hedges et al. (2018))
showing that PF often occupies a small portion of the FAO-dened
“forest”, which is “land spanning more than 0.5 hectares with trees
higher than 5 meters and a canopy cover of more than 10 percent, or
trees able to reach these thresholds in situ”(FAO, 2015). Conating PF
and SF could have disastrous consequences for biodiversity conserva-
tion. More analysis and mapping of PF are needed globally to mitigate
this problem.
The land cover maps also show that more than half of the land sur-
face in Haiti is covered with barren, cropland (active and abandoned), or
other deforested surface. This is surprising because various types of
vegetation (e.g., SF) can grow on poor soil, so why isn’t it vegetated?
This is likely because of intensive and successive agricultural activities
in the 20th century (Hedges et al., 2018;Marzelius and Droste, 2022).
The large charcoal market, which provides energy for cooking, probably
Fig. 8. Five examples of hotspots of primary forest (PF) loss in Hispaniola. Subplots (a) through (e) correspond to the hotspots A through E in Fig. 5c, respectively.
The black boundary denotes the protected area boundary with the name annotated. The numbers indicate the top ve years with the most signicant PF loss. The
major events/drivers causing the PF loss in specic years were provided. PF loss in the remaining years is categorized as “other”and is displayed using gray pixels.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
12
explains some of this. Charcoal can be made from small stems (seed-
lings) and therefore plants never have a chance of growing. Addition-
ally, continued deforestation over many decades (or centuries) can cause
desertication and severe erosion, whereby the land can no longer
support vegetation (Cohen, 1984).
Furthermore, we quantied the drivers of PF loss in Hispaniola and
Fig. 9. Impacts of 2016 Hurricane Matthew and the 2021 Haiti earthquake on the land cover change in Macaya National Park (located in Hotspot A, Fig. 5c). (a)
Map-based area (left y-axes) and percentage (right y-axes) of primary wet forest, secondary forest and other types in Macaya National Park from 1996 to 2022. (b)–
(d) Land cover maps of Macaya National Park in 2014, 2017, and 2022, respectively.
Fig. 10. Map-based primary forests (including primary wet and dry forests) area variation inside and outside protected areas. (a) Primary wet forest change in Haiti;
(b) Primary dry forest change in Haiti; (c) Primary wet forest change in the Dominican Republic; (d) Primary dry forest change in the Dominican Republic.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
13
identied the PF loss hotspots. The analysis of PF loss hotspots indicates
that extreme events such as res, hurricanes, and landslides caused by
earthquakes signicantly contribute to PF loss in Hispaniola (Fig. 8).
The impacts of extreme events can be compounded by anthropogenic
activities. These large-scale res (Fig. 8), most of which are likely
human-sparked wildres that happened in pine-dominated PF (S.B.H.
and J.T. personal observations), are much more destructive than those
caused by nature (Balch et al., 2017;Hantson et al., 2022). With
increased human activities encroaching into remote primary forests, PF
loss caused by human-induced res is expected to intensify in the future.
The forest fragmentation caused by human activities likely has accel-
erated PF loss during hurricanes and landslides. Extensive fragmentation
of forests on the island exposes otherwise protected forests to wind
damage and biomass loss (Laurance et al., 1997), preconditioning the
forests to damage from hurricanes. In turn, deforestation reduces soil
integrity created by roots, thus preconditioning the forests for further
loss from landslides during earthquakes (Havenith et al., 2022). While
earthquakes are natural events, hurricanes are believed to be strength-
ening due to human-caused climate change (Knutson et al., 2010). The
ndings underscore the urgent need for effective conservation policies
and interventions to mitigate the impacts of both anthropogenic and
natural disturbances on PFs in Hispaniola.
5.2. Uneven Landsat image density in Hispaniola
Landsat density, i.e., the frequency of clear Landsat observations, can
affect the tting of the COLD algorithm and change detection (Tollerud
et al., 2023;Zhang et al., 2021). Hispaniola has fewer clear observations
compared to data-rich regions, such as the United States (Zhang et al.,
2022). The Landsat data for Hispaniola is unevenly distributed in space
and time (Fig. 14), which brings uncertainties in land cover mapping
and disturbance detection. The eastern Dominican Republic has fewer
clear Landsat observations than other regions in Hispaniola due to
frequent cloud cover. This might lead to the underestimation of distur-
bance in those regions. Fortunately, the eastern Dominican Republic
covers only a small proportion of primary dry forest, which has limited
impacts on PF loss monitoring.
The temporal density of Landsat observations is inuenced by the
number of operational Landsat satellites. Before 2000, Landsat 5 was the
dominant satellite in operation, but a single satellite can only provide
limited clear observations. When Landsat 7 became operational after
2000, the average of clear observations per year increased slightly from
5.8 images/year to 7.3 images/year (Fig. 14). This was due to the
Landsat 7 Scan Line Corrector (SLC) failure and the limited capacity of
USGS to store those historical Landsat images. Fortunately, the situation
improved signicantly with the launch of Landsat 8 in 2013. Further-
more, the release of Landsat 9 data in October 2021 ensures sufcient
clear observations for monitoring disturbance (Fig. 14d).
To cope with the uneven distribution of limited Landsat clear ob-
servations and capture all potential land disturbances, especially for PFs,
which are sensitive to subtle disturbances, we employed a sensitive
parameter setting for COLD (Zhu et al., 2020). This might cause some
commission errors in the primary wet forest disturbance detection, but
the NBR threshold can greatly reduce the negative impact of commission
error (Fig. 3). More commission errors might also be created for other
land disturbances. However, the land cover classication might still be
the same as the pre-disturbance classication, which can eliminate the
impact of more commission errors in land change detection.
Fig. 11. Topographic characteristics of primary wet and dry forest distribution and loss. Subplot (a) shows the topographic distribution of primary forests (PFs) in
Haiti as of 1996. Subplot (b) shows the topographic distribution of lost PF in Haiti between 1996 and 2022. Subplots (c) and (d) correspond to subplots (a) and (b),
respectively, for the Dominican Republic (DR). The orange and green colour bars represent the primary dry and wet forests, respectively. The values indicate point
density, where larger values and darker colors correspond to higher point density. (For interpretation of the references to colour in this gure legend, the reader is
referred to the web version of this article.)
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
14
6. Conclusion
Distinguishing PF and SF using remote sensing observations is crit-
ical for evaluating the impact of forest loss on biodiversity, which is
rarely done in tropical regions. Our study used dense Landsat time series
and the COLD algorithm to map annual PF changes in Haiti and the
Dominican Republic from 1996 to 2022. We successfully distinguished
PF and SF with an overall accuracy of 80.5% (±5.2%) for the land cover
Fig. 12. Mean elevation (blue lines) and slope (orange lines) change of primary wet forest from 1996 to 2022. Subplots (a) and (b) display the results with SRTM
(Shuttle Radar Topography Mission) and HAND (Height Above the Nearest Drainage) elevation datasets in Haiti, respectively. Subplots (c) and (d) are the same as (a)
and (b), but for the Dominican Republic, respectively. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of
this article.)
Fig. 13. Fragmentation level analysis for primary forest in Hispaniola. Subplot (a) shows the patch density (patch numbers per 100 km
2
) and mean patch area of the
primary wet forest in Haiti. Subplots (b), (c) and (d) are the same as subplot (a), but for primary dry forest in Haiti, primary wet forest in the Dominican Republic, and
primary dry forest in the Dominican Republic, respectively. The blue curves and left y-axis denote the patch density. The orange curves and right y-axis denote the
mean patch area. (For interpretation of the references to colour in this gure legend, the reader is referred to the web version of this article.)
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
15
classication. The major conclusions include: (1) From 1996 to 2022,
both Haiti and the Dominican Republic experienced rapid PF loss, inside
and outside protected areas, indicating that those areas are ineffective.
Map-based results indicate that Haiti’s PF decreased from 0.75% to
0.44% and the Dominican Republic’s PF decreased from 7.14% to
5.67%. Haiti has a higher proportion of PF directly converted to non-
forest surface. (2) Haiti exhibits signicantly greater PF fragmentation
compared to the Dominican Republic. This higher degree of fragmen-
tation further threatens Haiti’s remaining PF and its associated biodi-
versity. (3) Slope plays a crucial role in the PF loss. PF in regions with
steeper slopes tends to be less accessible to humans, thus experiencing
less PF loss. (4) The major driver of PF loss in Hispaniola is re. This is
followed by tree-cutting, hurricanes, and landslides. The generated land
cover map can be valuable in guiding the PF and biodiversity conser-
vation efforts. With training data collected from other locations, this
approach could be applied to other regions and at larger scales.
CRediT authorship contribution statement
Falu Hong: Writing –original draft, Validation, Software, Method-
ology, Investigation, Formal analysis, Data curation, Conceptualization.
S. Blair Hedges: Writing –review &editing, Supervision, Resources,
Project administration, Methodology, Funding acquisition, Data cura-
tion, Conceptualization. Zhiqiang Yang: Writing –review &editing,
Supervision, Conceptualization. Ji Won Suh: Writing –review &edit-
ing, Validation. Shi Qiu: Writing –review &editing, Validation. Joel
Timyan: Writing –review &editing, Data curation. Zhe Zhu: Writing –
review &editing, Validation, Supervision, Resources, Project adminis-
tration, Methodology, Funding acquisition, Data curation,
Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgment
This work was supported by a grant from the NSF Biodiversity on a
Changing Planet (BoCP) program (2326013 to S.B.H. and 2326014 to Z.
Z.).
Appendix A. Statistical results and examples of collected training sample
Table A1
Statistical results of collected training sample.
Land cover Count of digitized polygons Count of pixels Count of selected training sample
Developed 59 26,620 1,206
Primary wet forest 46 25,476 1,341
Primary dry forest 6 47,221 752
Secondary forest 63 11,606 7,973
Shrub / Grass 106 41,890 1,823
Water 34 70,518 768
Wetland 38 10,171 771
Other 198 108,511 10,449
Sum 550 342,013 25,083
Fig. 14. The mean number of clear observations per year used as input for time series analysis in Hispaniola. (a) from 1996 to 1999 with Landsat 4 and 5 mainly in
operation. (b) from 2000 to 2012 with Landsat 5 and 7 mainly in operation. (c) from 2013 to 2021 with Landsat 7 and 8 mainly in operation. (d) In 2022 with Landsat
8 and 9 mainly in operation. The numbers in the lower right corner represent the mean value for all of Hispaniola.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
16
Fig. A1. Examples showing high-resolution images of training sample. (a) Developed; (b) Primary wet forest; (c) Primary dry forest; (d) Secondary forest; (e) Shrub/
Grass; (f) Water; (g) Wetland; (h) Other-barren; (i) Other-cropland.
Appendix B. Vegetation index changes for different disturbance agents
We selected ve vegetation indices to compare their capability of distinguishing between the recovered PF and SF: Normalized Difference
Vegetation Index (NDVI) (Tucker, 1979), kNDVI (Camps-Valls et al., 2021), Normalized Burn Ratio (NBR) (Key and Benson, 2006), Enhanced
Vegetation Index (EVI) (Huete et al., 2002), and Normalized Difference Fraction Index (NDFI) (Bullock et al., 2020;Souza et al., 2005).
The equations to calculate the selected vegetation indices are given as follows.
(1) Normalized Difference Vegetation Index (NDVI) (Tucker, 1979)
NDVI =NIR −Red
NIR −Red (2)
where NIR and Red represent the surface reectance of near infrared and red bands of Landsat.
(2) kNDVI (Camps-Valls et al., 2021)
We used the simplied version of kNDVI index as displayed in Eq. (3).
kNDVI =tanhNDVI2(3)
(3) Normalized Burn Ratio (NBR) (Key and Benson, 2006)
NBR =NIR −SWIR2
NIR +SWIR2(4)
where SWIR2 represents the surface reectance of Shortwave Infrared (SWIR) 2 band of Landsat.
(4) EVI (Huete et al., 2002)
EVI =2.5* NIR −Red
NIR +6*Red −7.5*Blue +1(5)
where Blue represents surface reectance of blue band of Landsat.
(5) Normalized Difference Fraction Index (NDFI) (Bullock et al., 2020;Souza et al., 2005)
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
17
NDFI =GVshade − (NPV +Soil)
GVshade + (NPV +Soil)(6)
where GV
shade
is calculated as:
GVshade =GV
1−Shade (7)
The green vegetation (GV), non-photosynthetic vegetation (NPV), Soil, Shade parameters represent the fraction of endmembers. The endmembers
are calculated as the linear combination of different Landsat bands. The coefcients are provided in Bullock et al. (2020). The fraction of the end-
members is calculated by spectral unmixing following the simple linear mixture model.
The boxplots show the VI changes before and after different disturbance agents (Fig. 3 &Fig. B1). The results show that VI changes caused by re,
hurricanes, landslides, and tree-cutting are more signicant compared to those caused by drought. Compared with NBR, it is challenging for NDVI,
kNDVI, EVI, and NDFI to nd a single threshold to separate the strong disturbances and weak disturbances. Therefore, we used NBR and the threshold
of 0.05 in the PF post-processing.
Fig. B1. Boxplots showing changes in vegetation indices (VIs) for different disturbance agents, which are calculated by subtracting post-disturbance VIs from pre-
disturbance VIs. (a) Normalized Difference Vegetation Index (NDVI); (b) kNDVI; (c) Enhanced Vegetation Index (EVI); (d) Normalized Difference Fraction Index
(NDFI). The meanings of each component in the boxplot are the same as those in Fig. 3.
Appendix C. Sample-based confusion matrices of land cover classication and primary forest loss detection
Table C1
Confusion matrix of land cover classication in sample counts. PWF, PDF, and SF indicate primary wet forest, primary dry forest, and secondary forest, respectively.
Reference
Map Developed PWF PDF SF Shrub/Grass Water Wetland Other Total Weight (%)
Developed 17 0 0 1 0 0 0 3 21 1.44
PWF 0 41 0 8 0 0 0 0 49 3.55
PDF 0 1 43 6 0 0 0 0 50 0.70
SF 0 1 0 57 3 0 0 12 73 39.09
Shrub/Grass 0 0 0 10 90 0 4 23 2.39
Water 0 0 0 0 0 23 1 1 25 0.91
Wetland 0 0 1 3 0 4 16 0 24 0.60
Other 0 0 0 12 3 0 2 89 106 51.33
Total 17 43 44 97 15 27 19 109 371
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
18
Table C2
Confusion matrix of primary forest (PF) loss detection in sample counts.
Reference
Map Other PF loss Total Weight (%)
Other 383 1 384 98.95
PF loss 30 66 96 1.05
Total 413 67 480
Data availability
The generated land cover map is freely available at https://doi.
org/10.6084/m9.gshare.28100408 and Google Earth Engine
(https://gers.users.earthengine.app/view/hispaniola-lc). The corre-
sponding code is available at https://github.
com/faluhong/hispaniola_land_cover_mapping.
References
Aide, T.M., Clark, M.L., Grau, H.R., L´
opez-Carr, D., Levy, M.A., Redo, D., Bonilla-
Moheno, M., Riner, G., Andrade-Nú˜
nez, M.J., Mu˜
niz, M., 2013. Deforestation and
reforestation of Latin America and the Caribbean (2001-2010). Biotropica 45,
262–271.
Alroy, J., 2017. Effects of habitat disturbance on tropical forest biodiversity. Proc. Natl.
Acad. Sci. 114, 6056–6061.
´
Alvarez-Berríos, N., Redo, D., Aide, T., Clark, M., Grau, R., 2013. Land change in the
Greater Antilles between 2001 and 2010. Land 2, 81–107.
Ar´
evalo, P., Olofsson, P., Woodcock, C.E., 2020. Continuous monitoring of land change
activities and post-disturbance dynamics from Landsat time series: A test
methodology for REDD+reporting. Remote Sens. Environ. 238, 111051.
Balch, J.K., Bradley, B.A., Abatzoglou, J.T., Nagy, R.C., Fusco, E.J., Mahood, A.L., 2017.
Human-started wildres expand the re niche across the United States. Proc. Natl.
Acad. Sci. 114, 2946–2951.
Barlow, J., Gardner, T.A., Araujo, I.S., ´
Avila-Pires, T.C., Bonaldo, A.B., Costa, J.E.,
Esposito, M.C., Ferreira, L.V., Hawes, J., Hernandez, M.I.M., Hoogmoed, M.S.,
Leite, R.N., Lo-Man-Hung, N.F., Malcolm, J.R., Martins, M.B., Mestre, L.A.M.,
Miranda-Santos, R., Nunes-Gutjahr, A.L., Overal, W.L., Parry, L., Peters, S.L.,
Ribeiro-Junior, M.A., da Silva, M.N.F., da Silva Motta, C., Peres, C.A., 2007.
Quantifying the biodiversity value of tropical primary, secondary, and plantation
forests. Proc. Natl. Acad. Sci. 104, 18555–18560.
Belgiu, M., Dr˘
agut
¸, L., 2016. Random forest in remote sensing: A review of applications
and future directions. ISPRS J. Photogramm. Remote Sens. 114, 24–31.
Bernier, P.Y., Par´
e, D., Stinson, G., Bridge, S.R.J., Kishchuk, B.E., Lempri`
ere, T.C.,
Thiffault, E., Titus, B.D., Vasbinder, W., 2017. Moving beyond the concept of
“primary forest”as a metric of forest environment quality. Ecol. Appl. 27, 349–354.
Brooks, T.M., Mittermeier, R.A., Mittermeier, C.G., Da Fonseca, G.A.B., Rylands, A.B.,
Konstant, W.R., Flick, P., Pilgrim, J., Oldeld, S., Magin, G., Hilton-Taylor, C., 2002.
Habitat loss and extinction in the hotspots of biodiversity. Conserv. Biol. 16,
909–923.
Brown, J.F., Tollerud, H.J., Barber, C.P., Zhou, Q., Dwyer, J.L., Vogelmann, J.E.,
Loveland, T.R., Woodcock, C.E., Stehman, S.V., Zhu, Z., Pengra, B.W., Smith, K.,
Horton, J.A., Xian, G., Auch, R.F., Sohl, T.L., Sayler, K.L., Gallant, A.L., Zelenak, D.,
Reker, R.R., Rover, J., 2020. Lessons learned implementing an operational
continuous United States national land change monitoring capability: The Land
Change Monitoring, Assessment, and Projection (LCMAP) approach. Remote Sens.
Environ. 238, 111356.
Bullock, E.L., Woodcock, C.E., Olofsson, P., 2020. Monitoring tropical forest degradation
using spectral unmixing and Landsat time series analysis. Remote Sens. Environ.
238.
Camps-Valls, G., Campos-Taberner, M., Moreno-Martínez, ´
A., Walther, S., Duveiller, G.,
Cescatti, A., Mahecha, M.D., Mu˜
noz-Marí, J., García-Haro, F.J., Guanter, L.,
Jung, M., Gamon, J.A., Reichstein, M., Running, S.W., 2021. A unied vegetation
index for quantifying the terrestrial biosphere. Sci. Adv. 7, eabc7447.
Churches, C.E., Wampler, P.J., Sun, W., Smith, A.J., 2014. Evaluation of forest cover
estimates for Haiti using supervised classication of Landsat data. Int. J. Appl. Earth
Obs. Geoinf. 30, 203–216.
Cohen, W.B., 1984. Environmental degradation in Haiti: an analysis of aerial
photography. US Agency for International Development/Haiti, Port-au-Prince, Haiti.
Cohen, W.B., Yang, Z., Healey, S.P., Kennedy, R.E., Gorelick, N., 2018. A LandTrendr
multispectral ensemble for forest disturbance detection. Remote Sens. Environ. 205,
131–140.
De Giorgi, A., Solarna, D., Moser, G., Tapete, D., Cigna, F., Boni, G., Rudari, R.,
Serpico, S.B., Pisani, A.R., Montuori, A., Zoffoli, S., 2021. Monitoring the recovery
after 2016 Hurricane Matthew in Haiti via markovian multitemporal region-based
modeling. Remote Sens. 13, 3509.
Decuyper, M., Ch´
avez, R.O., Lohbeck, M., Lastra, J.A., Tsendbazar, N., Hackl¨
ander, J.,
Herold, M., Vågen, T.-G., 2022. Continuous monitoring of forest change dynamics
with satellite time series. Remote Sens. Environ. 269, 112829.
Donchyts, G., Winsemius, H., Schellekens, J., Erickson, T., Gao, H., Savenije, H., van de
Giesen, N., 2016. Global 30m height above the nearest drainage. HAND 1000.
FAO, 2015. Global Forest Resources Assessments 2015: Terms and Denitions. In: FAO
Report.
Fern´
andez, E., 2007. Hispaniola: A Photographic Journey Through Island Biodiversity,
Biodiversidad a Trav´
es de Un Recorrido Fotogr´
aco. Harvard University Press/
Belknap Press, Cambridge, MA.
Friedl, M.A., Woodcock, C.E., Olofsson, P., Zhu, Z., Loveland, T., Stanimirova, R.,
Arevalo, P., Bullock, E., Hu, K.-T., Zhang, Y., Turlej, K., Tarrio, K., McAvoy, K.,
Gorelick, N., Wang, J.A., Barber, C.P., Souza, C., 2022. Medium spatial resolution
mapping of global land cover and land cover change across multiple decades from
Landsat. Front. Remote Sens. 3.
Geldmann, J., Manica, A., Burgess, N.D., Coad, L., Balmford, A., 2019. A global-level
assessment of the effectiveness of protected areas at resisting anthropogenic
pressures. Proceedings of the National Academy of Sciences 116, 23209–23215.
Gibson, L., Lee, T.M., Koh, L.P., Brook, B.W., Gardner, T.A., Barlow, J., Peres, C.A.,
Bradshaw, C.J.A., Laurance, W.F., Lovejoy, T.E., Sodhi, N.S., 2011. Primary forests
are irreplaceable for sustaining tropical biodiversity. Nature 478, 378–381.
Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.A., Tyukavina, A.,
Thau, D., Stehman, S.V., Goetz, S.J., Loveland, T.R., Kommareddy, A., Egorov, A.,
Chini, L., Justice, C.O., Townshend, J.R.G., 2013. High-resolution global maps of
21st-century forest cover change. Science 342, 850–853.
Hansen, M.C., Wang, L., Song, X.-P., Tyukavina, A., Turubanova, S., Potapov, P.V.,
Stehman, S.V., 2020. The fate of tropical forest fragments. Sci. Adv. 6, eaax8574.
Hantson, S., Andela, N., Goulden, M.L., Randerson, J.T., 2022. Human-ignited res result
in more extreme re behavior and ecosystem impacts. Nat. Commun. 13, 2717.
Havenith, H.B., Guerrier, K., Schl¨
ogel, R., Braun, A., Ulysse, S., Mreyen, A.S., Victor, K.
H., Saint-Fleur, N., Cauchie, L., Boisson, D., Pr´
epetit, C., 2022. Earthquake-induced
landslides in Haiti: analysis of seismotectonic and possible climatic inuences. Nat.
Hazards Earth Syst. Sci. 22, 3361–3384.
Hedges, S.B., Cohen, W.B., Timyan, J., Yang, Z., 2018. Haiti’s biodiversity threatened by
nearly complete loss of primary forest. Proc. Natl. Acad. Sci. 115, 11850–11855.
Heinrich, V.H.A., Vancutsem, C., Dalagnol, R., Rosan, T.M., Fawcett, D., Silva-Junior, C.
H.L., Cassol, H.L.G., Achard, F., Jucker, T., Silva, C.A., House, J., Sitch, S., Hales, T.
C., Arag˜
ao, L.E.O.C., 2023. The carbon sink of secondary and degraded humid
tropical forests. Nature 615, 436–442.
Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview
of the radiometric and biophysical performance of the MODIS vegetation indices.
Remote Sens. Environ. 83, 195–213.
ITTO, 2002. ITTO guidelines for the restoration, management and rehabilitation of
degraded and secondary tropical forests. Internaitonal Tropical Timber
Organization.
John, D.L., Yolanda, M.L., 2019. Forest change within and outside protected areas in the
Dominican Republic, 2000-2016. bioRxiv, 558346.
Kennedy, R.E., Yang, Z., Cohen, W.B., Pfaff, E., Braaten, J., Nelson, P., 2012. Spatial and
temporal patterns of forest disturbance and regrowth within the area of the
Northwest Forest Plan. Remote Sens. Environ. 122, 117–133.
Key, C., Benson, N., 2006. Landscape assessment: remote sensing of severity, the
normalized burn ratio and ground measure of severity, the composite burn index.
FIREMON: Fire effects monitoring and inventory system Ogden. USDA Forest Service,
Rocky Mountain Res. Station, Utah.
Knorn, J.A.N., Kuemmerle, T., Radeloff, V.C., Keeton, W.S., Gancz, V., BiriS¸, I.-A.,
Svoboda, M., Grifths, P., Hagatis, A., Hostert, P., 2013. Continued loss of temperate
old-growth forests in the Romanian Carpathians despite an increasing protected area
network. Environ. Conserv. 40, 182–193.
Knutson, T.R., McBride, J.L., Chan, J., Emanuel, K., Holland, G., Landsea, C., Held, I.,
Kossin, J.P., Srivastava, A.K., Sugi, M., 2010. Tropical cyclones and climate change.
Nat. Geosci. 3, 157–163.
Kormos, C.F., Mackey, B., DellaSala, D.A., Kumpe, N., Jaeger, T., Mittermeier, R.A.,
Filardi, C., 2018. Primary forests: Denition, status and future prospects for global
conservation. Encycloped. Anthropoc. 31–41.
Laurance, W.F., Laurance, S.G., Ferreira, L.V., Rankin-de Merona, J.M., Gascon, C.,
Lovejoy, T.E., 1997. Biomass collapse in Amazonian forest fragments. Science 278,
1117–1118.
Liang, M., Gonz´
alez-Roglich, M., Roehrdanz, P., Tabor, K., Zvoleff, A., Leitold, V.,
Silva, J., Fatoyinbo, T., Hansen, M., Duncanson, L., 2023. Assessing protected area’s
carbon stocks and ecological structure at regional-scale using GEDI lidar. Global
Environmental Change 78, 102621.
F. Hong et al. Remote Sensing of Environment 318 (2025) 114590
19
Martin, P.H., Sherman, R.E., Fahey, T.J., 2004. Forty years of tropical forest recovery
from agriculture: Structure and oristics of secondary and old-growth riparian
forests in the Dominican Republic. Biotropica 36, 297–317.
Marzelius, M., Droste, N., 2022. Livelihoods matter –A comparative political ecology of
forest use on Hispaniola. Forest Policy Econ. 141, 102765.
Mikol´
aˇ
s, M., Ujh´
azy, K., Jasík, M., Wiezik, M., Gallay, I., Pol´
ak, P., Vysoký, J., ˇ
Ciliak, M.,
Meigs, G.W., Svoboda, M., Trotsiuk, V., Keeton, W.S., 2019. Primary forest
distribution and representation in a Central European landscape: Results of a large-
scale eld-based census. For. Ecol. Manag. 449, 117466.
Mikol¯
aˇ
s, M., Piovesan, G., Ahlstr¨
om, A., Donato, D.C., Gloor, R., Hofmeister, J.,
Keeton, W.S., Muys, B., Sabatini, F.M., Svoboda, M., Kuemmerle, T., 2023. Protect
old-growth forests in Europe now. Science 380, 466.
Myers, N., Mittermeier, R.A., Mittermeier, C.G., da Fonseca, G.A.B., Kent, J., 2000.
Biodiversity hotspots for conservation priorities. Nature 403, 853–858.
Nobre, A.D., Cuartas, L.A., Hodnett, M., Renn´
o, C.D., Rodrigues, G., Silveira, A.,
Waterloo, M., Saleska, S., 2011. Height Above the Nearest Drainage –a
hydrologically relevant new terrain model. J. Hydrol. 404, 13–29.
Olofsson, P., Foody, G.M., Herold, M., Stehman, S.V., Woodcock, C.E., Wulder, M.A.,
2014. Good practices for estimating area and assessing accuracy of land change.
Remote Sens. Environ. 148, 42–57.
Olofsson, P., Ar´
evalo, P., Espejo, A.B., Green, C., Lindquist, E., McRoberts, R.E., Sanz, M.
J., 2020. Mitigating the effects of omission errors on area and area change estimates.
Remote Sens. Environ. 236, 111492.
Palahí, M., Valbuena, R., Senf, C., Acil, N., Pugh, T.A.M., Sadler, J., Seidl, R., Potapov, P.,
Gardiner, B., Hetem¨
aki, L., Chirici, G., Francini, S., Hl´
asny, T., Lerink, B.J.W.,
Olsson, H., Gonz´
alez Olabarria, J.R., Ascoli, D., Asikainen, A., Bauhus, J.,
Berndes, G., Donis, J., Fridman, J., Hanewinkel, M., Jactel, H., Lindner, M.,
Marchetti, M., Maruˇ
s´
ak, R., Sheil, D., Tom´
e, M., Trasobares, A., Verkerk, P.J.,
Korhonen, M., Nabuurs, G.-J., 2021. Concerns about reported harvests in European
forests. Nature 592, E15–E17.
Pauleus, O., Aide, T.M., 2020. Haiti has more forest than previously reported: land
change 2000–2015. PeerJ 8, e9919.
Pengra, B.W., Stehman, S.V., Horton, J.A., Dockter, D.J., Schroeder, T.A., Yang, Z.,
Cohen, W.B., Healey, S.P., Loveland, T.R., 2020. Quality control and assessment of
interpreter consistency of annual land cover reference data in an operational
national monitoring program. Remote Sens. Environ. 238.
Poorter, L., Bongers, F., Aide, T.M., Almeyda Zambrano, A.M., Balvanera, P., Becknell, J.
M., Boukili, V., Brancalion, P.H.S., Broadbent, E.N., Chazdon, R.L., Craven, D., de
Almeida-Cortez, J.S., Cabral, G.A.L., de Jong, B.H.J., Denslow, J.S., Dent, D.H.,
DeWalt, S.J., Dupuy, J.M., Dur´
an, S.M., Espírito-Santo, M.M., Fandino, M.C.,
C´
esar, R.G., Hall, J.S., Hernandez-Stefanoni, J.L., Jakovac, C.C., Junqueira, A.B.,
Kennard, D., Letcher, S.G., Licona, J.-C., Lohbeck, M., Marín-Spiotta, E., Martínez-
Ramos, M., Massoca, P., Meave, J.A., Mesquita, R., Mora, F., Mu˜
noz, R.,
Muscarella, R., Nunes, Y.R.F., Ochoa-Gaona, S., de Oliveira, A.A., Orihuela-
Belmonte, E., Pe˜
na-Claros, M., P´
erez-García, E.A., Piotto, D., Powers, J.S., Rodríguez-
Vel´
azquez, J., Romero-P´
erez, I.E., Ruíz, J., Saldarriaga, J.G., Sanchez-Azofeifa, A.,
Schwartz, N.B., Steininger, M.K., Swenson, N.G., Toledo, M., Uriarte, M., van
Breugel, M., van der Wal, H., Veloso, M.D.M., Vester, H.F.M., Vicentini, A., Vieira, I.
C.G., Bentos, T.V., Williamson, G.B., Rozendaal, D.M.A., 2016. Biomass resilience of
Neotropical secondary forests. Nature 530, 211–214.
Powell, R.L., Matzke, N., de Souza, C., Clark, M., Numata, I., Hess, L.L., Roberts, D.A.,
2004. Sources of error in accuracy assessment of thematic land-cover maps in the
Brazilian Amazon. Remote Sens. Environ. 90, 221–234.
Rodrigues-Eklund, G., Hansen, M.C., Tyukavina, A., Stehman, S.V., Hubacek, K.,
Baiocchi, G., 2021. Sample-based estimation of tree cover change in Haiti using
aerial photography: Substantial increase in tree cover between 2002 and 2010.
Forests 12.
Sabatini, F.M., Bluhm, H., Kun, Z., Aksenov, D., Atauri, J.A., Buchwald, E.,
Burrascano, S., Cateau, E., Diku, A., Duarte, I.M., Fern´
andez L´
opez, ´
A.B.,
Garbarino, M., Grigoriadis, N., Horv´
ath, F., Keren, S., Kitenberga, M., Kiˇ
s, A.,
Kraut, A., Ibisch, P.L., Larrieu, L., Lombardi, F., Matovic, B., Melu, R.N., Meyer, P.,
Midteng, R., Mikac, S., Mikol´
aˇ
s, M., Mozgeris, G., Panayotov, M., Pisek, R.,
Nunes, L., Ruete, A., Schickhofer, M., Simovski, B., Stillhard, J., Stojanovic, D.,
Szwagrzyk, J., Tikkanen, O.-P., Toromani, E., Volosyanchuk, R., Vrˇ
ska, T.,
Waldherr, M., Yermokhin, M., Zlatanov, T., Zagidullina, A., Kuemmerle, T., 2021.
European primary forest database v2.0. Sci. Data 8, 220.
Sangermano, F., Bol, L., Galvis, P., Gullison, R.E., Hardner, J., Ross, G.S., 2015a. Forest
baseline and deforestation map of the Dominican Republic through the analysis of
time series of MODIS data. Data in Brief 4, 363–367.
Sangermano, F., Bol, L., Galvis, P., Gullison, R.E., Hardner, J., Ross, G.S., 2015b. Habitat
suitability and protection status of four species of amphibians in the Dominican
Republic. Appl. Geogr. 63, 55–65.
Seidl, R., Turner, M.G., 2022. Post-disturbance reorganization of forest ecosystems in a
changing world. Proc. Natl. Acad. Sci. 119, e2202190119.
Souza, C.M., Roberts, D.A., Cochrane, M.A., 2005. Combining spectral and spatial
information to map canopy damage from selective logging and forest res. Remote
Sens. Environ. 98, 329–343.
Tollerud, H.J., Zhu, Z., Smith, K., Wellington, D.F., Hussain, R.A., Viola, D., 2023.
Toward consistent change detection across irregular remote sensing time series
observations. Remote Sens. Environ.