Ecological Informatics 82 (2024) 102738 
Contents lists available at ScienceDirect
Ecological Informatics
journal homepage: www.elsevier.com/locate/ecolinf
Landsat images and GIS techniques as key tools for historical analysis of
landscape change and fragmentation
Darwin Gómez-Fernández a,b,*, Rolando Salas López a, Jhon A. Zabaleta-Santisteban a,
Angel J. Medina-Medina a, Malluri Goñas a,b, Jhonsy O. Silva-López a,c, Manuel Oliva-Cruz a,
Nilton B. Rojas-Briceño a,d
a Instituto de Investigación para el Desarrollo Sustentable de Ceja de Selva (INDES-CES), National University Toribio Rodríguez de Mendoza (UNTRM), Chachapoyas
01001, Peru
b Centro Experimental Yanayacu, Dirección de Supervisión y Monitoreo en las Estaciones Experimentales Agrarias, Instituto Nacional de Innovación Agraria (INIA),
Carretera Jaén San Ignacio KM 23.7, Jaén 06801, Cajamarca, Peru
c Laboratorio de Agrostología – AGROLAB, Facultad de Ingeniería Zootecnista, Agronegocios y Biotecnología, UNTRM, Chachapoyas 01001, Peru
d Escuela Profesional de Ingeniería Ambiental, Facultad de Ingeniería y Arquitectura, Universidad Nacional de Moquegua, Pacocha 18610, Peru
A R T I C L E I N F O A B S T R A C T
Keywords: Monitoring and evaluation of landscape fragmentation is important in numerous research areas, such as natural
Fragmentation resource protection and management, sustainable development, and climate change. One of the main challenges
LULC in image classification is the intricate selection of parameters, as the optimal combination significantly affects the
Changes accuracy and reliability of the final results. This research aimed to analyze landscape change and fragmentation
Classification
Random Forest in northwestern Peru. We utilized accurate land cover and land use (LULC) maps derived from Landsat imagery
Amazon using Google Earth Engine (GEE) and ArcGIS software. For this, we identified the best dataset based on its
Forest highest overall accuracy, and kappa index; then we performed an analysis of variance (ANOVA) to assess the
differences in accuracies among the datasets, finally, we obtained the LULC and fragmentation maps and
analyzed them. We generated 31 datasets resulting from the combination of spectral bands, indices of vegetation,
water, soil and clusters. Our analysis revealed that dataset 19, incorporating spectral bands along with water and
soil indices, emerged as the optimal choice. Regarding the number of trees utilized in classification, we deter-
mined that using between 10 and 400 decision trees in Random Forest classification doesn’t significantly affect
overall accuracy or the Kappa index, but we observed a slight cumulative increase in accuracy metrics when
using 100 decision trees. Additionally, between 1989 and 2023, the categories Artificial surfaces, Agricultural
areas, and Scrub/ Herbaceous vegetation exhibit a positive rate of change, while the categories Forest and Open
spaces with little or no vegetation display a decreasing trend. Consequently, the areas of patches and perforated
have expanded in terms of area units, contributing to a reduction in forested areas (Core 3) due to fragmentation.
As a result, forested areas smaller than 500 acres (Core 1 and 2) have increased. Finally, our research provides a
methodological framework for image classification and assessment of landscape change and fragmentation,
crucial information for decision makers in a current agricultural zone of northwestern Peru.
1. Introduction period 1990 to 2015, the total forest area of the world decreased by 3%
(Ban et al., 2019). Specifically, South America, between 2015 and 2020
On a global scale, in the last 25 years, almost 125 million hectares of lost>15million hectares of forest, becoming the secondmost deforested
forest have been deforested (Curtis et al., 2018), and a current global region in the world, with agriculture, forestry, forest fires and urbani-
rate of forest loss of 0.6% per year (Hansen et al., 2013), likewise, for the zation being the main drivers of deforestation (Chamberlain et al., 2020;
* Corresponding author at: Instituto de Investigación para el Desarrollo Sustentable de Ceja de Selva (INDES-CES), National University Toribio Rodríguez de
Mendoza (UNTRM), Chachapoyas 01001, Peru.
E-mail addresses: darwin.gomez@untrm.edu.pe (D. Gómez-Fernández), rsalas@indes-ces.edu.pe (R.S. López), jhon.zabaleta@untrm.edu.pe (J.A. Zabaleta-
Santisteban), angel.medina@untrm.edu.pe (A.J. Medina-Medina), malluri.gonas@untrm.edu.pe (M. Goñas), jhonsy.silva@untrm.edu.pe (J.O. Silva-López),
soliva@indes-ces.edu.pe (M. Oliva-Cruz), nrojasb@unam.edu.pe (N.B. Rojas-Briceño).
https://doi.org/10.1016/j.ecoinf.2024.102738
Received 22 October 2023; Received in revised form 21 July 2024; Accepted 22 July 2024
Available online 28 July 2024 
1574-9541/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC license ( http://creativecommons.org/licenses/by- 
nc/4.0/ ). 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Curtis et al., 2018). that integrates the assessment of LULC changes and associated frag-
Forest fragmentation refers to changes in forest cover and can be mentation in high Andean areas, where the expansion of agriculture and
measured using land cover maps derived from satellite products (Myr- forest resource utilization continues to advance.
oniuk et al., 2020). Anthropogenic activities, such as agricultural Landsat images are considered the standard Earth observation data
expansion and urbanization(Ellis and Ramankutty, 2008), as well as for large-scale ecological monitoring and provide unique opportunities
logging and stand burning (Haddad et al., 2015), are the main causes of to assess changes in forest cover (Wulder et al., 2012), due to their long
forest landscape loss and fragmentation (Negi et al., 2019). Under the time series, high spatial resolution and free access (Huang et al., 2010;
dual actions of climate change and land cover change, global biodiver- Shen et al., 2019; Zhu and Woodcock, 2014). In recent decades, satellite
sity and ecosystem functioning face serious threats, especially in glob- imagery collected by Landsat platforms in forestry applications has
ally recognized biodiversity hotspots (DeFries et al., 2002; Kanade and increased due to the availability of long time series of satellite obser-
John, 2018). vations of land cover and its dynamics (Gu et al., 2020). Also, many
Forests are a crucial terrestrial ecosystem, and have an indispensable studies have detected changes in forest cover and monitored forest
role in nutrient cycling and energy flow in ecological processes (Chaz- fragmentation using remote sensing data at regional and global scales
don et al., 2016). Fragmentation of landscapes results in a mixture of (Da Ponte et al., 2017; Gong et al., 2013; Vogeler et al., 2020).
land cover patches of different classes, sizes and shapes (Numata et al., On the other hand, the advent of cloud-based computing Google
2011). Thus, forest landscape fragmentation can negatively affect many Earth Engine (GEE) (Gorelick et al., 2017), in recent years has helped to
processes that occur within an ecosystem (Xun et al., 2014), in addition, facilitate large-scale studies using high-performance mapping algo-
fragmentation impacts species richness and distribution patterns of rithms (Garcia et al., 2023; Kennedy et al., 2018; Parente and Ferreira,
biodiversity (Gibson et al., 2013; Pardini et al., 2010), ecosystem ser- 2018; Zhu et al., 2019). Thus, GEE with its robust computation and
vices (Nagendra et al., 2009; Rocha-Santos et al., 2016; Uddin et al., storage capabilities, has attracted a great deal of attention (Gorelick
2015), habitat quality (Fahrig, 2003; Reddy et al., 2013), and invasive et al., 2017) and has been widely applied in vegetation monitoring
species emergence (Reddy et al., 2013; Thuiller et al., 2008). (Wang et al., 2019a; Xie et al., 2019), crop mapping (Jin-Ming et al.,
Specifically, the loss and fragmentation of forest cover is considered 2019; Wang et al., 2019b), and LULC classification (Ge et al., 2019; Tsai
the main cause of global ecosystem degradation (Newman et al., 2014). et al., 2019). Also, GEE allows easy access to different publicly available
In addition, fragmented forests may be more vulnerable to stress and datasets, including the collection of preprocessed Landsat imagery,
have lower resilience compared to connected forests (Shimizu et al., reducing the time needed to generate accurate maps (Myroniuk et al.,
2017). Specifically, the higher the degree of landscape fragmentation, 2020). In particular, Landsat data hosted on the GEE platform provide a
the lower its stability, due to the reduction of resources and therefore, unique opportunity to monitor forest cover change at high spatial res-
the survival of key species of the ecosystem is put at risk (Zhang et al., olutions, from local to global scales.
2021). Therefore, quantifying landscape changes is imperative to un- Therefore, we used Landsat satellite products and GIS tools to eval-
derstand the spatial and structural viability of land use and its associated uate forest landscape fragmentation in Northwest Peru. To do this: i) in
ecological effects (Turner, 2005). GEE we obtained accurate LULC maps for the years 1989, 2005 and
A recent study revealed that 70% of the world’s forests are within 1 2023, ii) we analyzed the changes and level of fragmentation for each
km of the forest edge, subject to the degrading effects of fragmentation period using Landscape Fragmentation Tool v2 in ArcGIS software.
(Ganivet and Bloomberg, 2019). Frequent landscape fragmentation oc- Finally, this study provides baseline information on the current state of
curs particularly in many developing countries (Abdullah et al., 2019). landscape conservation for proper land management. This evaluation
Thus, there has been a growing need to better understand the impor- strategy will allow not only quantifying landscape fragmentation but
tance of landscape fragmentation processes (Hermosilla et al., 2019; also providing valuable information for decision-making in environ-
Hysa and Başkaya, 2017). mental management and conservation of natural resources in the region.
The level of landscape fragmentation is an important attribute of
land pattern because it has potential implications for land loss preven- 2. Materials and methods
tion and management [30]. That is, it can change environmental con-
ditions and species composition in ways that could influence forest 2.1. Study area
susceptibility (Schwartz et al., 2019). Thus, assessing the extent of
fragmented forests can help inform policy and decision making for forest The study area extends from 78◦06′ to 78◦58′ west longitude, and
management practices (Wulder et al., 2009). Therefore, forest mapping 5◦30′ to 6◦02′ south latitude, covering 2904.96 km2, has elevations from
on fragmented landscapes is essential at a regional scale (Myroniuk 351 to 3684 m.a.s.l. and encompasses the Alto Marañón II and III inter-
et al., 2020). basins, as well as the lower Utcubamba river basin, as shown in Fig. 1.
Additionally, studies have been conducted examining forest frag- This area was chosen as a case study because it is currently the most
mentation in different regions of the world. For example, in Bangladesh, affected sector by the cultivation of rice, coffee, cocoa, and extensive
between 1989 and 2021, using remote sensing data, identified changes livestock farming in the aforementioned watersheds of the Amazonas
in forest cover and forest fragmentation, highlighting the importance of and Cajamarca regions, located in the northwest Peru (Gobierno
forest restoration (Hassan et al., 2023). Another study optimized the Regional de Amazonas (GOREA) and Instituto de Investigaciones de la
VOR model to measure ecosystem health using remote sensing tech- Amazonía Peruana (IIAP), 2007; Gobierno Regional de Cajamarca
nology, showing a gradual decline in regional ecosystem health over (GRC), 2010MIDAGRI, 2021).
time (Bao et al., 2022). Similarly, a study analyzed the impact of land- On the other hand, according to the climatic classification elaborated
scape fragmentation on the provision of ecosystem services over three by SENAMHI (2020), the study area is conditioned by the C(r)B type,
decades, identifying a negative relationship between landscape frag- that is, Semi-dry with abundant humidity in all seasons of the year, and
mentation and the value of these services (Do et al., 2022). with temperate temperature efficiency, the maximum temperature
By 2030, it is estimated that the Peruvian Amazon will experience ranges from 21 to 25 ◦C, and minimum from 7 to 11 ◦C, with respect to
greater deforestation and forest degradation than any other region, precipitation, in this type of climate 700 to 2000 mm per year are
driven by agriculture, commercial mining, and land artificialization due registered.
to urbanization (Smith and Schwartz, 2015). To date, studies in the
Amazonas region have focused on LULC changes; however, landscape 2.2. Methodological flow
fragmentation levels have been scarcely studied (Rojas Briceño et al.,
2019). Therefore, it is essential to provide a methodological framework Fig. 2 presents an overview of the methodological flow employed for
2 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 1. Geographic location of study area in Northwest Peru.
the analysis of landscape fragmentation in Northwest Peru since 1989 to generated for all three years using ArcMap 10.5 and LFT v2.
2023, applying Landscape Fragmentation Tool (LFT) as GIS tools and
Landsat image classification.
We initiated the process of identifying the optimal combination of 2.3. Obtaining LULC maps in GEE
data to effectively classify Landsat 8 images. To achieve this, within the
GEE platform, we filtered the image collection removing cloudy areas 2.3.1. Data source and classification algorithm
using a cloud mask, then, we combined spectral bands, indices of LULC maps were generated in GEE (Gorelick et al., 2017), from the
vegetation, soil, water, and clusters generated, this process yielded a oldest year with adequate data (1989 in our case) to the most recent year
total of 31 unique data combinations. (2023), divided into two periods with similar number of years (1989 to
We used the Random Forest algorithm for classification, using 2005 and 2005 to 2023). For this purpose, we use Landsat images
training points for each data set. Subsequently, an evaluation was per- (Landsat 4, 5 y 8 images courtesy of the U.S. Geological Survey). The
formed using field points to determine the accuracies of the 31 data sets. data from the Landsat-4 Thematic Mapper (TM) mission were used to
An ANOVA was then used to identify significant differences between the obtain the 1989 LULC map, whereas Landsat-5 TM images were
data sets. Once the combination yielding the highest accuracies was employed for the 2005 LULC map, and finally, Landsat-8 Operational
identified, the 1989 and 2005 maps were ranked using the input data Land Imager and Thermal Infrared Sensor (OLI-TIRS) images were used
from the most accurate dataset. These maps were then reclassified into to obtain the 2023 LULC map.
forest and non-forest categories. Finally, fragmentation maps were The Landsat data used for the three years of analysis were from Level
2, Collection 2, Tier 1, this means that the surface reflectance was
3 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 2. Methodological flow used in the analysis of landscape fragmentation in Northwest Peru from 1989 to 2023.
4 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
atmospherically corrected using the Land Surface Reflectance Code; this Table 2
correction eliminated the influence of aerosol scattering, thin clouds, Datasets obtained.
and other atmospheric effects on the detection and characterization of Name Inputs Name Inputs
land surface changes (Masek et al., 2006; Vermote et al., 2016). Addi-
Dataset 1 Spectral bands (B) Dataset 17 B + VI + WI
tionally, to avoid problems associated with cloud cover, a cloud mask Dataset 2 Vegetation Indices (VI) Dataset 18 B + VI + C
procedure was conducted using the quality assessment band available in Dataset 3 Soil Indices (SI) Dataset 19 B + SI + WI
the Landsat data. We used Bitmasks 1, 2, 3, 4, and 6, which represent Dataset 4 Water Indices (WI) Dataset 20 B + SI + C
dilated clouds, high confidence cirrus, clouds, cloud shadows, and clear Dataset 5 Cluster (C) Dataset 21 B + WI + C
conditions, respectively. For more details, refer to Section 1 of the script Dataset 6 B + VI Dataset 22 VI + SI + WIDataset 7 B + SI Dataset 23 VI + SI + C
shared in the supplementary material. Dataset 8 B + WI Dataset 24 VI + WI + C
On the other hand, based on the fact that, with respect to other Dataset 9 B + C Dataset 25 SI + WI + C
classification algorithms running in GEE, Random Forest (RF) is one of Dataset 10 VI + SI Dataset 26 B + VI + SI + WI
the machine learning classifiers with the best accuracies in the classifi- Dataset 11 VI + WI Dataset 27 B + VI + SI + C
Dataset 12 VI + C Dataset 28 B + VI + WI + C
cation of satellite images to obtain LULC maps (Gómez Fernández et al., Dataset 13 SI + WI Dataset 29 B + SI + WI + C
2022; Ouma et al., 2022; Shetty, 2019; Talukdar et al., 2020), therefore, Dataset 14 SI + C Dataset 30 VI + SI + WI + C
for the classification, the RF algorithm was trained using training points Dataset 15 WI + C Dataset 31 B + VI + SI + WI + C
collected in the field and the inputs of the best dataset, obtaining LULC Dataset 16 B + VI + SI
maps with 30-m pixels, according to the spatial resolution of Landsat
missions, finally, through a post classification and visual inspection, the
products were improved. Table 3
LULC classes based on CLC methodology.
2.3.2. Datasets of classification variables Class Description Code RGB color
A total of 31 data combinations were generated (Table 2), resulting
Artificial surfaces All classes of continuous and 1 230–000-
from the combination of spectral bands, vegetation indices, soil, water, discontinuous urban fabric. 077
and clusters. Subsequently, the best dataset was selected as the combi- Agricultural areas All arable land, permanent crops, 2 255–255-
nation that accumulated the highest overall accuracy by summing the pastures and heterogenous 168
accuracies of the 9 decision trees. To determine the optimal number of agricultural areas.
decision trees, the accuracies of the 31 datasets were summed, and the Forest All broad-leaved, coniferous and 3 128–255-mixed forest 000
decision tree with the highest overall accuracy accumulated was chosen. Scrub and/or All natural grasslands, Moors and 4 204–242-
The input data were i) spectral bands: Ultra blue, Blue, Green, Red, herbaceous heathland, Sclerophyllous 077
Near Infrared (NIR), Shortwave Infrared 1 (SWIR-1) and Shortwave vegetation vegetation and Transitional
Infrared 2 (SWIR-2), ii) Vegetation Indices: Normalized Difference woodland-shrub
Open spaces with All sand, bare rocks, sparsely 5 230–230-
Vegetation Index (NDVI) and Enhanced vegetation index (EVI), iii) Soil little or no vegetated and burnt areas 230
Indices: Soil Adjusted Vegetation Index (SAVI) and Bare Soil Index (BSI), vegetation
iv) Water Indices: Modified Normalized Difference Water Index Water bodies All water courses and water bodies 6 128–242-
(MNDWI) and Normalized Difference Moisture Index (NDMI), and v) 230
Clusters. No data No data 999 –
The aforementioned spectral indices were calculated according to
the formulas in Table 1. In order to obtain clusters, a superpixel clus- 2.3.4. Sampling size and validation areas
tering based on Simple Non-Iterative Clustering (SNIC), was performed On this occasion, to define the training and testing samples, we relied
in GEE using the function ee.Algorithms.Image.Segmentation.SNIC. For on Chuvieco (2020), who argues that for categorical variables, at least
more details, refer to Section 2 of the script shared in the supplementary 196 test points should be considered. Furthermore, Chuvieco (2020),
material. also mentions that, at times, this number can be very small when
Table 2 shows the obtained combinations of spectral bands, vege- compared to the total number of pixels in the image. Therefore, we
tation, soil, water indices, and clusters. The combinations were gener- explored new sampling strategies that combine the previous scientific
ated online (https://es.planetcalc.com/3757/), and to obtain the 31 foundation with the current capabilities of GEE.
datasets, combinations of size one, two, three, four, and five were used. Taking into account Congalton (1991) and Hay (1979), who suggest
at least 50 pixels per thematic class, we decided to implement an
2.3.3. Definition of LULC classes alternative strategy. Instead of strictly distributing 196 training areas
The LULC classes defined in this research were aligned with the among the 6 thematic classes, we chose to include pure training areas.
Coordination of Information on the Environment (CORINE) Land Cover These well-identified areas should contain at least 50 pixels in total and
(CLC) methodology. Table 3 below shows the LULC classes present in the prominently represent the class of interest. This distribution was carried
study area and their respective descriptions. out through simplified random sampling (Chuvieco, 2020), allowing us
to effectively capture variability within each class and improve the
Table 1
Spectral indices calculated for each collection.
Index Formula Reference
NDVI (NIR − RED)/(NIR+ RED) (Rouse et al., 1974)
EVI C*[(NIR − RED)/(NIR+ C1*RED − C2*BLUE+ L) ] (Gao et al., 2003)
SAVI [(NIR − RED)/(NIR+ RED+ L) ]*1+ L (Huete, 1988)
BSI [(RED+ SWIR) − (NIR+ BLUE) ]/[(RED+ SWIR) + (NIR+ BLUE) ] (Rikimaru et al., 2002)
MNDWI (GREEN − SWIR1)/(GREEN+ SWIR1) (McFeeters, 2007)
NDMI (NIR − SWIR1)/(NIR+ SWIR1) (Wilson and Sader, 2002)
*Where: C = 2.5; C1 = 6; C2 = 7.5; L = 0.5 and Blue, Red, Green, NIR, SWIR depend on each Landsat. Mission band.
5 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
representativeness of our samples. 2.5. Fragmentation measurement
Finally, for the validation areas, a minimum of 25% of the training
areas was considered. Therefore, in this research, we used 40 training In the ArcGIS 10.5 software, the Landscape Fragmentation Tool
areas, comprising 2463 pixels, and, additionally, 10 validation areas, (LFT) (Uddin et al., 2015; Vogt et al., 2006), was utilized to generate
comprising 866 pixels. For more details, please refer to Section 3.1 of the fragmentation maps. Initially, the LULC maps were converted into a
script shared in the supplementary material. binary format (foreground/background). The Forest and Scrub/Herba-
ceous vegetation classes were designated as the foreground, while the
2.3.5. Measuring the accuracy of LULC maps remaining classes (Artificial surfaces, Agricultural areas, Open spaces
The assessment of accuracy can be approached from various per- with little or no vegetation and Water bodies) comprised the
spectives, utilizing expert judgment (visual inspection) or external sta- background.
tistical sources, and this process must be carried out to ensure the As for the input parameters of the LFT, the reclassified raster was
highest reliability of the results with the ground truth (Blissag et al., assigned two classes: 1 = nonforest (background), and 2 = forest
2024). In this research, the measurement of accuracy was based on the (foreground), ultimately with an edge width of 100 m, as argued by
validation points. Using this approach, the confusion matrix was Uddin et al. (2015) and Vogt et al. (2006). The fragmentation maps
calculated, allowing us to obtain the Overall Accuracy (OA) and Kappa generated had the following forest patterns and/or fragmentation cat-
Index (KI), coefficients necessary for the evaluation of classification egories: The “Core” is situated at a considerable distance from the
accuracy (Thomlinson et al., 1999). For this purpose, Eqs. (1) and (2) boundary between forest and non-forest areas, while the “Patch” con-
were followed. sists of cohesive forest regions that are too limited in size to encompass
Total number of pixel classified correctly the core forest. The “Perforated” establishes the limits between the core
OA = *100 (1) forest and relatively small perforations, and the “Edge” encompasses
total number of pixels interior boundaries with relatively large perforations, as well as the
∑n ∑n external boundaries of core forest regions (Vogt et al., 2006).
N ai,i − (Ti + Fi)
KI i=1 i=1= n (2)∑ 2.5.1. Determination of transitions and risk levels
N2 − (Ti + Fi) Due to fragmentation causing a loss of connectivity within the
i=1 ecosystem, it’s considered one of the main drivers of landscape degra-
Where: i represent the class number. N is the total number of clas- dation (Jaramillo et al., 2023), by intersecting and summing the codes of
sified values compared to truth-values. ai,i is the number of values of the fragmentation categories for the years 1989, 2005, and 2023, in ArcGIS
truth class i classified also as class i, which is the values found along the Pro version 3.0.1, we generated transition matrices for the periods
diagonal of the confusion matrix. Fi represents the total number of 1989–2005, 2005–2023 and 1989–2023.
predicted values of the i class and Ti is the total number of truth-values In Table 5, we present the base transition matrix, with the 6 cate-
of the i class (Foody, 2020; Stehman, 1997a; Stehman, 1997b). gories of fragmentation, this allowed us to identify the changes from one
Additionally, to make a decision based on the two-accuracy metrics, category to another in the aforementioned periods.
Table 4 shows the classification accuracy according to the Kappa index, On the other hand, in Table 6 shows the risk levels that each tran-
and the operator’s decision on the classification. sition represents, based on the definitions of each forest pattern (Vogt
Finally, in order to select a specific dataset, an ANOVA of one way of et al., 2006).
the 31 datasets was performed in Google Colaboratory, considering a
significance level of 95%. The null hypothesis was that “all the means of 3. Results
the dataset accuracies are equal,” while the alternative hypothesis was
that “at least one mean is different from the others”. This analysis was 3.1. Dataset accuracy metrics
conducted in relation to the number of repetitions, taking into account
each accuracy obtained with different numbers of decision trees. We obtained 31 datasets and calculated their OA and KI. In Fig. 3
displays the variation of OA and KI along the Y-axis, corresponding to
the number of decision trees on the X-axis.
2.4. Measurement of LULC change It can be observed that OA and KI exhibit an increasing trend up to a
certain point. Afterward, they stabilize or begin to decrease. This phe-
We calculated the areas (in km2) of each LULC class for the years nomenon is associated with the number of decision trees utilized. In
1989, 2005, and 2023. This enabled us to ascertain the gains and losses some dataset increasing the number of decision trees contributes to
of each category over the evaluated years. Subsequently, we determined enhanced accuracies, while in others, the opposite occurs. The accu-
the rate of change using Eq. (3), as described in (Food and Agriculture racies of the 31 datasets were highly variable, in the case of OA, it
Organization of the United Nations, 1993; Puyravaud, 2003). ranged from 0.33 to 0.92, while KI varied between 0.2 and 0.90.
( ) ( )
1 A2 Five datasets were identified that exceeded 0.9 for overall accuracy
r = *ln (3)
t2 − t1 A1 and 0.87 for Kappa index, of which Dataset 19 was selected as the most
suitable due to its highest cumulative OA value. Fig. 4 shows the cu-
Where: r is the rate of change per year, t represents year and A the mulative overall accuracy of the 31 datasets as a function of the 9 types
area covered by the class, subscripts 2 and 1 represent the final and of decision trees used, identifying Dataset 19 as the best dataset with
initial period, respectively. 8.32 and 0.92 for cumulative and average OA, respectively.
Fig. 4 confirms the positive impact of spectral indices on classifica-
Table 4 tion accuracy. For example, in Dataset 1, which only utilized spectral
Classification accuracy according kappa index. bands, an acceptable accuracy was achieved. However, by adding water
Kappa Index Accuracy Decision indices, this accuracy slightly improved, and by also including soil
> 0.90 Very high Acceptable indices, the highest classification accuracy was reached, as observed in
0.80–0.90 High Acceptable the stacked bar of Dataset 19. On the other hand, a noticeable aspect is
0.60–0.79 Moderate Depend of application the low accuracy of Dataset 5, which only used the cluster band. This
< 0.50 Low Not acceptable low accuracy negatively affected the accuracy of datasets that included
Note: Table data adapted from (Cohen, 1960; Landis and Koch, 1977). this band. Therefore, caution is recommended when combining it for a
6 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Table 5
Transition matrix of fragmentation categories.
Table 6
Considered transitions and risk levels.
definitive classification. influence the overall accuracy of the classification. However, it is
Therefore, to solidify the results and demonstrate significant differ- noteworthy that when 100 decision trees were used, the overall accu-
ences among the datasets, the results of the ANOVA conducted are racy was slightly higher than the others.
shown in Table 7. Additionally, Fig. 7 shows the distribution of the overall accuracies
Table 7 displays the principal results of ANOVA, indicating a highly of the 9 decision trees, indicating that the highest density of cases is
significant p-value, which is less than the conventional significance level concentrated in the accuracy range between 0.6 and 1. This suggests a
of 0.05, additionally, F-statistic is greater than the critical value (0.05). common trend towards high levels of overall accuracy in the classifi-
Therefore, there is compelling evidence to reject the null hypothesis that cations performed by all types of trees.
the average of overall accuracy and kappa index of all datasets are equal.
Instead, we accept the alternative hypothesis, suggesting that at least 3.2. LULC maps
one average of OA and KI among the datasets differs from the others.
For more details regarding the global accuracy, Fig. 5 illustrates the As Dataset 19 obtained higher accuracies, its inputs were used to
differences and similarities between groups derived from the Tukey obtain the LULC map of 1989 and 2005, and 100 decision trees for each
Honestly Significant Difference (HSD) test, considering a Family-Wise year. Fig. 8 displays the RGB and LULC maps for 1989, 2005, and 2023.
Error Rate (FWER) of 0.05 to control the overall risk of type I error in
the statistical analysis. In 69% of the comparisons, there are significant 3.2.1. LULC changes
differences, meaning that the means of OA aren’t equal. On the other Fig. 9 shows the square kilometers covered by the different LULC
hand, in the remaining 31%, statistically the means of global accuracies categories and Table 8 shows gains and losses for the three years of
are equal, as depicted in Fig. 5, where the blue color indicates similarity analysis.
between means and gray the opposite. From Fig. 9 and Table 8, a shift towards more intensive land use and
Then, after selecting the best dataset, we proceeded to identify the a loss of forested areas is evident. Over the analyzed periods, there is an
most effective decision tree. On this occasion, in sub-Fig. 6a, the cu- acceleration in the conversion of land to agricultural and urban uses at
mulative and average overall accuracy of each decision tree is presented, the expense of natural and semi-natural areas, reflecting a clear pattern
and in sub-Fig. 6b, the same overall accuracy but standardized. This of urbanization and expansion of agricultural areas.
standardization was carried out because the differences between the Between 1989 and 2005, artificial surfaces increased from 16.2 km2
trees were minimal and not clearly appreciated in sub-Fig. 6a. to 33.4 km2, following the same trend as agricultural areas, which rose
As seen in Fig. 6, there aren’t differences in the cumulative overall from 571.2 km2 to 676.5 km2. Conversely, forest cover decreased from
accuracies, which was also statistically confirmed, obtaining a p-value of 556.2 km2 to 470.36 km2, as did open spaces with little or no vegetation,
0.99 in an ANOVA performed. Therefore, with the parameters consid- which shrank from 973.9 km2 to 796.9 km2.
ered in this research, the number of decision trees doesn’t directly From 2005 to 2023, artificial areas continued to expand, reaching 44
7 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 3. Variation of OA and KI among the 31-dataset obtained for 2023, as influenced by the number of decision trees utilized.
8 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 4. Variation of the cumulative overall accuracy of the 31 datasets according to the 9 types of decision trees considered.
km2, while agricultural areas expanded by 383.00 km2. The accumu-
Table 7 lated loss of forest cover was 145.70 km2, and open spaces with little or
Summary of the analysis of variance performed. no vegetation decreased by 447.34 km2.
F-Statistic P-value While Table 8 provides substantial information, Table 9 below shows
Overall accuracy 161.36 5.14e− 145 the rate of change for the periods 1989–2005, 2005–2023 and
Kappa index 158.93 3.06e− 144 1989–2023.
As shown in Table 9, artificial surfaces grew at an annual rate of
km2, and agricultural areas grew to 954.2 km2. Forest cover continued to 2.94%, and agricultural areas at a rate of 1.51% per year. Forest cover
decline to 410 km2, and open spaces with little or no vegetation also decreased at an annual rate of − 0.89%, while shrub and/or herbaceous
decreased to 526.56 km2. vegetation grew at a rate of 0.71% per year. Open spaces with little or no
Overall, from 1989 to 2023, artificial surfaces increased by 27.80 vegetation decreased at a rate of − 1.81% per year, and water bodiesincreased at an annual rate of 0.49%. These trends confirm the ongoing
Fig. 5. Differences and similarities between groups considering OA values.
9 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 6. Cumulative and average overall accuracy (a) and its standardization (b) of the 31 datasets based on each decision tree used.
Fig. 7. Distribution of the overall accuracy of the 31 datasets based on each decision tree used.
urbanization and agricultural expansion over time, as well as the Below, Fig. 12 displays the transitions of fragmentation categories
continuous decrease in forests and open spaces. for the periods 1989–2005, 2005–2023, and 1989–2023. Additionally,
Table 8 provides detailed information on the risk levels and the covered
area (km2) for each transition.
3.3. Fragmentation maps Table 10 consolidates the changes in the four risk levels for the three
analysis periods, showing a trend towards increased fragmentation and
Fig. 10 shows the fragmentation levels for the years 1989, 2005 and a decrease in core forest areas. Specifically, from 1989 to 2005, the
2023. It can be seen that in 1989 the study area presented a greater transition from Core to Patch-Edge covered an area of 71.96 km2, rep-
forest core, while in 2023 the appearance of patches and edges is more resenting 66.76% of the total high-risk transitions, and increased to
frequent. 86.93 km2 by 2023. Similarly, the transition from Edge to Patch and
From Fig. 10 it can be inferred that the area covered by patches has from Perforated to Patch covered 28.75 km2 in the first period and
increased, as have the edges, which is not the case for fragmentation increased to 41.92 km2 by 2023.
category Core 3 (Cores >500 acres). For this reason, Fig. 11 shows the Regarding medium-risk transitions, the transition from Core to
area covered calculated in km2 of the different fragmentation categories, Perforated covered an area of 132.34 km2 (44.48%) from 1989 to 2005
showing that, effectively since 1989 to 2023, patches increased 21.81 2
km2
but decreased to 118.50 km by 2023. On the other hand, moderate-risk
, and edges increased 156.95 km2, while Core 3 has a decreasing
2 transitions showed more stable patterns. For example, the transitionstrend, losing a total of 322.24 km for the period of analysis, affecting
2 from Perforated to Cores covered 35.12 km
2 from 1989 to 2005, then
the increase of Core 1 (Cores <250 acres), which increased 49.95 km . increased to 75.54 km2 by 2023, highlighting a trend towards the
In summary, Figs. 10 and 11 show that the appearance of patches and consolidation of fragmented areas.
edges (as a result of agricultural activities or other factors) has brought Finally, low-risk transitions showed minimal changes. Transitions of
with it the fragmentation of forests, that is, the increase of level 1 and 2 the Core 3 unchanged covered the largest area within this level,
cores.
10 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 8. RGB Compositions and LULC maps obtained for 1989, 2005 and 2023. In the RGB compositions the white pixels correspond to cloud masked, that will be
considered as NoData in the calculation of areas.
encompassing 96.29% in 1989–2005 and 99.18% of the low-risk tran- 4. Discussion
sitions throughout the study period, indicating the stability of core forest
areas despite the surrounding fragmentation dynamic. Similarly to Cho et al. (2022), who used the spectral bands of Landsat
products and derived vegetation indices, obtaining 17 predictors for
classification, we generated 31 datasets, considering spectral bands,
11 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 9. Coverage in square kilometers of LULC classes for 1989, 2005 and 2023.
Table 8
LULC changes for 1989, 2005 and 2023.
1989 2005 2023 Gains and losses (Km2)
Km2 % Km2 % Km2 % 1989 to 2005 2005 to 2023 1989 to 2023
Artificial surfaces 16.20 0.56 33.40 1.15 44.00 1.51 17.20 10.60 27.80
Agricultural areas 571.20 19.66 676.50 23.29 954.20 32.85 105.30 277.70 383.00
Forest 556.2 19.15 470.36 16.19 410.50 14.13 − 85.84 − 59.86 − 145.70
Scrub and/or herbaceous vegetation 739.71 25.46 894.50 30.79 940.80 32.39 154.79 46.30 201.09
Open spaces with little or no vegetation 973.90 33.53 796.90 27.43 526.56 18.13 − 177.00 − 270.34 − 447.34
Water bodies 20.60 0.71 22.70 0.78 24.30 0.84 2.10 1.60 3.70
NoData 27.15 0.93 10.60 0.36 4.60 0.16
Total 2904.96 100 2904.96 100 2904.96 100
Table 9
Rate of change for the periods 1989–2005, 2005–2023 and 1989–2023.
Categories 1989 2005 2023 Rate of Change
Km2 Km2 Km2 1989 to 2005 2005 to 2023 1989 to 2023
Artificial surfaces 22.73 20.82 24.41 4.52% 1.53% 2.94%
Agricultural areas 682.88 720.18 857.13 1.06% 1.91% 1.51%
Forest 565.12 327.63 398.80 − 1.05% − 0.76% − 0.89%
Scrub and/or herbaceous vegetation 589.13 766.17 703.09 1.19% 0.28% 0.71%
Open spaces with little or no vegetation 982.02 716.84 444.86 − 1.25% − 2.30% − 1.81%
Water bodies 14.23 32.82 44.03 0.61% 0.38% 0.49%
NoData 48.84 320.50 432.64
Total 2904.96 2904.96 2904.96
vegetation, soil and water indices and clusters, with the objective of Nowadays, analyses of variance are very useful to identify sources of
determining which combination yields the highest accuracies. variability in the data, and to determine if these sources have a signifi-
We run Random Forest in GEE, as it is one of the best machine cant effect on the accuracy of the maps, for example, Lin et al. (2015),
learning classifiers (Gómez Fernández et al., 2022), in turn, supported analyzed the impact of atmospheric correction and pansharpening on
by Talukdar et al. (2020), through an evaluation of six LULC classifiers the accuracy of LULC classification using a two-factor factorial design
determined that Random Forest is the best classifier but still needs to be and determined that atmospheric correction was statistically insignifi-
tested in different morphoclimatic conditions. cant, while pansharpening and the interaction with atmospheric
The accuracies of the maps vary according to the input parameters, correction were statistically significant. On the other hand, Hua (2017),
and various factors must be considered, in our case, on average the 31 analyzed land cover changes and their impact on water quality, and
datasets had OA between 0.33 and 0.92 and KI between 0.2 and 0.90, through an ANOVA between LULC and water quality data concluded
while Rwanga et al. (2017), who classified Landsat 8 images, consid- that the built-up area significantly contaminated water quality.
ering similar LULC classes as our research, but obtained OA and KI of In this research, through an ANOVA we determined the significant
0.817 and 0.722, respectively. On the other hand, Aghababaei et al. differences of the 31 datasets, because the accuracies were almost high,
(2021) in multitemporal image classification obtained a kappa index of concluding that, if there were significant differences with a confidence
0.74 and an overall accuracy of 0.81. level of 95%, it was decided to use the inputs of the most accurate
12 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 10. Fragmentation’s levels for 1989, 2005 and 2023.
Fig. 11. Coverage of fragmentation classes for 1989, 2005 and 2023.
dataset. per year, along with Agricultural areas (1.51%). However, Forests
The rates of changes of LULC categories vary based on the analysis decreased − 0.89% per year, as shown in Table 9 and Fig. 9, Forest losses
period (Arunyawat and Shrestha, 2016; Talukdar et al., 2020), for and Open spaces with little or no vegetation transformations are linked
example, Traore et al. (2021), determined that, the Bare Ground dras- to the rise of Agricultural areas, and artificial surfaces.
tically decreased (− 155.83), while, the Artificialized and Forested area Globally, due to fragmentation, forests have decreased in size and
increased (137.06 and 8. 44%, respectively, since 1986 to 2017), in our cover, while the total number of forest fragments has increased (Encisa-
case, since 1989 to 2023: Open spaces with little or no vegetation Garcia et al., 2020; Taubert et al., 2018), this phenomenon brings about
decreased − 1.81% per year, while Artificial surfaces increased 2.94% changes in the composition of plant communities, with these changes
13 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Fig. 12. Transitions of fragmentation categories for the 3 periods of analysis.
Table 10
Risk levels and covered area (km2) for each fragmentation category transition.
are more pronounced in highly fragmented habitats compared to less changes, being the built-up areas and bare soils those areas that pre-
fragmented ones (Collins et al., 2017), in addition, fragmentation can sented higher temperatures with respect to the vegetation and bodies of
also affect seed dispersal (Dener et al., 2021). water.
Deforestation often brings with it various changes, for example Identifying the drivers of deforestation is feasible through co-
variation in surface temperature, Traore et al. (2021), quantified the occurrence and supply chain analysis, as in the case of Cho et al.
LULC changes and their relationship with the surface temperature (2022), who analyzed deforestation and its relationship with rubber
14 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
production, concluding that rubber production may be driving defor- order to improve readability and language. After using this tool, the
estation, which may have a negative impact on the environment. Simi- authors reviewed and edited the content as needed and take) full re-
larly, future research should assess deforestation and its relationship sponsibility for the content of the publication.
with the advancement of the agricultural frontier in northwest Peru.
It should be noted that changes in land use aren’t always negative, Data availability
many times the expansion of agricultural land has a positive impact on
the economy of farmers, for example, Appelt et al. (2022), through a The code of classification, ANOVA, and the results of the Tukey’s
systematic review, evaluated the socioeconomic results, the possible Test are available in the following web repository: https://github.
compensations and synergies of changes in the use of agricultural land in com/dargofer/Fragmentation_Peru.
Southeast Asia, finding positive impacts for income and employment,
negative impacts for food security, equality of gender and economic, and Acknowledgements
mixed impacts in the health sector.
As shown in Fig. 1, the study area is located in a tropical zone with The authors acknowledge and thank INDES-CES of the Universidad
altitudes between 351 and 3684m.a.s.l, which leads to high cloud cover, Nacional Toribio Rodríguez de Mendoza de Amazonas (UNTRM) for its
but thanks to filter and masking functions in GEE it was possible to support. And the Yanayacu Experimental Center, Direction of Supervi-
obtain mosaics with minimal cloud cover. In addition, the calculation of sion and Monitoring in Agricultural Experimental Stations, National
spectral indices and the generation of clusters was fast and flexible due Institute for Agrarian Innovation (INIA).
to the strength of GEE, so that this research, despite the limitations,
presents substantial information for the study area. References
In synthesis, the present research provides substantial information
for land management, reporting LULC changes, as well as landscape Abdullah, A.Y.M., Masrur, A., Gani Adnan, M.S., Al Baky, M.A., Hassan, Q.K., Dewan, A.,
fragmentation levels, thus becoming a starting point for future research 2019. Spatio-temporal patterns of land use/land cover change in the heterogeneouscoastal region of Bangladesh between 1990 and 2017. Remote Sens. 11 (7) https://
such as, identification of deforestation drivers, impact of deforestation doi.org/10.3390/rs11070790.
on local economy, environmental quality of agricultural zone, evalua- Aghababaei, M., Ebrahimi, A., Naghipour, A.A., Asadi, E., Verrelst, J., 2021. Vegetation
tion of map accuracies using other classifiers, relationship of LULC types mapping using multi-temporal Landsat images in the Google earth engineplatform. Remote Sens. 13 (22), 4683. https://doi.org/10.3390/RS13224683.
changes and climatic anomalies etc. Appelt, J.L., Garcia Rojas, D.C., Verburg, P.H., van Vliet, J., 2022. Socioeconomic
outcomes of agricultural land use change in Southeast Asia. Ambio 51 (5),
5. Conclusions 1094–1109. https://doi.org/10.1007/S13280-022-01712-4/TABLES/3.Arunyawat, S., Shrestha, R.P., 2016. Assessing land use change and its impact on
ecosystem Services in Northern Thailand. Sustainability 8 (8), 768. https://doi.org/
We identified the optimal dataset for generating accurate LULCmaps 10.3390/SU8080768.
for the years 1989, 2005, and 2023, based on statistical tests applied to Ban, N.C., Gurney, G.G., Marshall, N.A., Whitney, C.K., Mills, M., Gelcich, S., Bennett, N.J., Meehan, M.C., Butler, C., Ban, S., Tran, T.C., Cox, M.E., Breslow, S.J., 2019. Well-
the results from 31 datasets and 9 types of decision trees. Our analysis being outcomes of marine protected areas. Nat. Sustain. 2 (6), 524–532. https://doi.
determined that using between 10 and 400 decision trees in Random org/10.1038/s41893-019-0306-2.
Forest classification doesn’t significantly affect overall accuracy or the Bao, Z., Shifaw, E., Deng, C., Sha, J., Li, X., Hanchiso, T., Yang, W., 2022. Remote
sensing-based assessment of ecosystem health by optimizing vigor-organization-
Kappa index. However, we observed a slight cumulative increase in resilience model: A case study in Fuzhou City, China. Ecol. Inform. 72, 101889
accuracy metrics when using 100 decision trees. Therefore, we recom- https://doi.org/10.1016/J.ECOINF.2022.101889.
mend using 100 decision trees for the Random Forest classifier and Blissag, B., Yebdri, D., Kessar, C., 2024. Spatiotemporal change analysis of LULC using
selecting Dataset 19 (spectral bands, soil, and water indices) as input remote sensing and CA-ANN approach in the Hodna basin, NE of Algeria. Phys.Chem. Earth Parts A/B/C 133, 103535. https://doi.org/10.1016/J.
bands for areas similar to our study area. PCE.2023.103535.
We show graphically and numerically the changes, fragmentation Chamberlain, J.L., Darr, D., Meinhold, K., 2020. Rediscovering the contributions of
and rate of change of the LULC categories, being Artificial surfaces forests and trees to transition global food systems. Forests 11 (10), 1098. https://doi.org/10.3390/F11101098.
(2.94%) and Agricultural areas (1.51%) the ones with the highest pos- Chazdon, R.L., Brancalion, P.H.S., Laestadius, L., Bennett-Curry, A., Buckingham, K.,
itive rate, on the other hand, Forest (− 0.89%) and Open spaces with Kumar, C., Moll-Rocek, J., Vieira, I.C.G., Wilson, S.J., 2016. When is a forest a forest?
little or no vegetation ( 1.81%) present negative rates. Finally, it was Forest concepts and definitions in the era of forest and landscape restoration. Ambio− 45 (5), 538–550. https://doi.org/10.1007/s13280-016-0772-y.
determined that the coverage of level 3 forests cores decreased, and Cho, K., Goldstein, B., Gounaridis, D., Newell, J.P., 2022. Hidden risks of deforestation in
consequently, the coverage by patches and level 1 cores increased, thus global supply chains: A study of natural rubber flows from Sri Lanka to the United
showing a fragmentation of the landscape in the study area. States. J. Clean. Prod. 349, 131275 https://doi.org/10.1016/J.JCLEPRO.2022.131275.
Chuvieco, E., 2020. Fundamentals of Satellite Remote Sensing : An Environmental
CRediT authorship contribution statement Approach (Third Edit). CRC Press. https://doi.org/10.1201/9780429506482.
Cohen, J., 1960. A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 20
(1), 37–46. https://doi.org/10.1177/001316446002000104.
Darwin Gómez-Fernández: Conceptualization, Data curation, Collins, C.D., Banks-Leite, C., Brudvig, L.A., Foster, B.L., Cook, W.M., Damschen, E.I.,
Formal analysis, Investigation, Methodology, Software, Visualization, Andrade, A., Austin, M., Camargo, J.L., Driscoll, D.A., Holt, R.D., Laurance, W.F.,
Writing – original draft. Rolando Salas López: Investigation, Project Nicholls, A.O., Orrock, J.L., 2017. Fragmentation affects plant community
composition over time. Ecography 40 (1), 119–130. https://doi.org/10.1111/
administration, Resources, Software, Supervision, Visualization. Jhon ECOG.02607.
A. Zabaleta-Santisteban: Data curation, Formal analysis, Investigation, Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely
Software, Visualization. Angel J. Medina-Medina: Data curation, sensed data. Remote Sens. Environ. 54 (5), 593–600. https://doi.org/10.1016/0034-
Formal analysis, Investigation, Software, Visualization. Jhonsy O. 4257(91)90048-B.Curtis, P.G., Slay, C.M., Harris, N.L., Tyukavina, A., Hansen, M.C., 2018. Classifying
Silva-López: Investigation, Software, Validation, Writing – review & drivers of global forest loss. Science 361 (6407), 1108–1111. https://doi.org/
editing. Manuel Oliva-Cruz: Investigation, Project administration, Re- 10.1126/SCIENCE.AAU3445/SUPPL_FILE/AAU3445_CURTIS_SM.PDF.
sources. Nilton B. Rojas-Briceño: Investigation, Software, Validation, Da Ponte, E., Roch, M., Leinenkugel, P., Dech, S., Kuenzer, C., 2017. Paraguay’s AtlanticForest cover loss – satellite-based change detection and fragmentation analysis
Writing – review & editing. between 2003 and 2013. Appl. Geogr. 79, 37–49. https://doi.org/10.1016/J.
APGEOG.2016.12.005.
Declaration of Generative AI and AI-assisted technologies in the DeFries, R.S., Houghton, R.A., Hansen, M.C., Field, C.B., Skole, D., Townshend, J., 2002.Carbon emissions from tropical deforestation and regrowth based on satellite
writing process observations for the 1980s and 1990s. Proc. Natl. Acad. Sci. 99 (22), 14256–14261.
https://doi.org/10.1073/PNAS.182560099.
During the preparation of this work the authors used Chat GPT in
15 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Dener, E., Ovadia, O., Shemesh, H., Altman, A., Chen, S.C., Giladi, I., 2021. Direct and Huang, C., Goward, S.N., Masek, J.G., Thomas, N., Zhu, Z., Vogelmann, J.E., 2010. An
indirect effects of fragmentation on seed dispersal traits in a fragmented agricultural automated approach for reconstructing recent forest disturbance history using dense
landscape. Agric. Ecosyst. Environ. 309, 107273 https://doi.org/10.1016/J. Landsat time series stacks. Remote Sens. Environ. 114 (1), 183–198. https://doi.org/
AGEE.2020.107273. 10.1016/J.RSE.2009.08.017.
Do, A.N.T., Tran, H.D., Ashley, M., Nguyen, A.T., 2022. Monitoring landscape Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25 (3),
fragmentation and aboveground biomass estimation in can Gio mangrove biosphere 295–309. https://doi.org/10.1016/0034-4257(88)90106-X.
reserve over the past 20 years. Eco. Inform. 70, 101743 https://doi.org/10.1016/J. Hysa, A., Başkaya, F.A.T., 2017. Landscape fragmentation assessment utilizing the matrix
ECOINF.2022.101743. green toolbox and corine land cover data. J. Digit. Landscape Architect. 2017 (2),
Ellis, E.C., Ramankutty, N., 2008. Putting people in the map: anthropogenic biomes of 54–62. https://doi.org/10.14627/537629006.
the world. Front. Ecol. Environ. 6 (8), 439–447. https://doi.org/10.1890/070062. Jaramillo, J.J., Rivas, C.A., Oteros, J., Navarro-Cerrillo, R.M., 2023. Forest fragmentation
Encisa-Garcia, J.O., Pulhin, J.M., Cruz, R.V.O., Simondac-Peria, A.C., Ramirez, M.A.M., and landscape connectivity changes in Ecuadorian mangroves: some hope for the
De Luna, C.C., 2020. Land use/land cover changes assessment and forest future? Applied Sciences (Switzerland) 13 (8), 5001. https://doi.org/10.3390/
fragmentation analysis in the baroro river watershed, La Union, Philippines. APP13085001/S1.
J. Environ. Sci. Manag. SI-2, 14–27. Jin-Ming, Y., Li-Gang, M., Cheng-Zhi, L., Yang, L., Jian-li, D., Sheng-Tian, Y., 2019.
Fahrig, L., 2003. Effects of habitat fragmentation on biodiversity, 34, 487–515. https:// Temporal-spatial variations and influencing factors of lakes in inland arid areas from
doi.org/10.1146/ANNUREV.ECOLSYS.34.011802.132419. 2000 to 2017: a case study in Xinjiang. Geomat. Nat. Haz. Risk 10 (1), 519–543.
Food and Agriculture Organization of the United Nations, 1993. Forest Resources https://doi.org/10.1080/19475705.2018.1531942.
Assessment 1990: Tropical Countries. Food and Agriculture Organization of the Kanade, R., John, R., 2018. Topographical influence on recent deforestation and
United Nations. https://www.fao.org/3/t0830e/T0830E00.htm. degradation in the Sikkim Himalaya in India; implications for conservation of east
Foody, G.M., 2020. Explaining the unsuitability of the kappa coefficient in the Himalayan broadleaf forest. Appl. Geogr. 92, 85–93. https://doi.org/10.1016/J.
assessment and comparison of the accuracy of thematic maps obtained by image APGEOG.2018.02.004.
classification. Remote Sens. Environ. 239, 111630 https://doi.org/10.1016/J. Kennedy, R.E., Yang, Z., Gorelick, N., Braaten, J., Cavalcante, L., Cohen, W.B., Healey, S.,
RSE.2019.111630. 2018. Implementation of the LandTrendr algorithm on Google earth engine. Remote
Ganivet, E., Bloomberg, M., 2019. Towards rapid assessments of tree species diversity Sens. 10 (5), 1–10. https://doi.org/10.3390/rs10050691.
and structure in fragmented tropical forests: a review of perspectives offered by Landis, J.R., Koch, G.G., 1977. The measurement of observer agreement for categorical
remotely-sensed and field-based data. For. Ecol. Manag. 432 (April 2018), 40–53. data. Biometrics 33 (1), 159. https://doi.org/10.2307/2529310.
https://doi.org/10.1016/j.foreco.2018.09.003. Lin, C., Wu, C.C., Tsogt, K., Ouyang, Y.C., Chang, C.I., 2015. Effects of atmospheric
Gao, X., Huete, A.R., Didan, K., 2003. Multisensor comparisons and validation of MODIS correction and pansharpening on LULC classification accuracy using WorldView-2
vegetation indices at the semiarid jornada experimental range. IEEE Trans. Geosci. imagery. Inform. Proc. Agriculture 2 (1), 25–36. https://doi.org/10.1016/J.
Remote Sens. 41 (10 PART I), 2368–2381. https://doi.org/10.1109/ INPA.2015.01.003.
TGRS.2003.813840. Masek, J.G., Vermote, E.F., Saleous, N.E., Wolfe, R., Hall, F.G., Huemmrich, K.F., Gao, F.,
Garcia, N., Alírio, J., Silva, D., Campos, J.C., Duarte, L., Arenas-Castro, S., Pôças, I., Kutler, J., Lim, T.K., 2006. A landsat surface reflectance dataset for North America,
Sillero, N., Teodoro, A.C., 2023. MontObEO, Montesinho biodiversity observatory: 1990-2000. IEEE Geosci. Remote Sens. Lett. 3 (1), 68–72. https://doi.org/10.1109/
an earth observation tool for biodiversity conservation. In: Earth Resources and LGRS.2005.857030.
Environmental Remote Sensing/GIS Applications XIV, 12734, pp. 335–345. https:// McFeeters, S.K., 2007. The use of the normalized difference water index (NDWI) in the
doi.org/10.1117/12.2678524. delineation of open water features. Int. J. Remote Sens. 17 (7), 1425–1432. https://
Ge, Y., Hu, S., Ren, Z., Jia, Y., Wang, J., Liu, M., Zhang, D., Zhao, W., Luo, Y., Fu, Y., doi.org/10.1080/01431169608948714.
Bai, H., Chen, Y., 2019. Mapping annual land use changes in China’s poverty- MIDAGRI, 2021. Atlas de la superficie agrícola del Perú. https://siea.midagri.gob.pe/po
stricken areas from 2013 to 2018. Remote Sens. Environ. 232, 111285 https://doi. rtal/informativos/superficie-agricola-peruana.
org/10.1016/J.RSE.2019.111285. Myroniuk, V., Kutia, M., Sarkissian, J., Bilous, A., Liu, S., 2020. Regional-scale Forest
Gibson, L., Lynam, A.J., Bradshaw, C.J.A., He, F., Bickford, D.P., Woodruff, D.S., mapping over fragmented landscapes using global Forest products and Landsat time
Bumrungsri, S., Laurance, W.F., 2013. Near-complete extinction of native small series classification. Remote Sens. 12 (1), 187. https://doi.org/10.3390/
mammal fauna 25 years after forest fragmentation. Science 341 (6153), 1508–1510. rs12010187.
https://doi.org/10.1126/SCIENCE.1240495/SUPPL_FILE/GIBSON.SM.PDF. Nagendra, H., Paul, S., Pareeth, S., Dutt, S., 2009. Landscapes of protection: forest
Gobierno Regional de Amazonas (GOREA), Instituto de Investigaciones de la Amazonía change and fragmentation in northern West Bengal, India. Environ. Manag. 44 (5),
Peruana (IIAP), 2007. Zonificación ecológica y económica del departamento de 853–864. https://doi.org/10.1007/S00267-009-9374-9.
amazonas, p. 199. http://iiap.org.pe/Archivos/publicaciones/PUBL520.pdf. Negi, V.S., Pathak, R., Rawal, R.S., Bhatt, I.D., Sharma, S., 2019. Long-term ecological
Gobierno Regional de Cajamarca (GRC), 2010. Zonificacion Ecologica y Economica monitoring on forest ecosystems in Indian Himalayan region: criteria and indicator
Como Base Para El Ordenamiento Territorial Del Departamento De Cajamarca, approach. Ecol. Indic. 102 (July 2018), 374–381. https://doi.org/10.1016/j.
p. 281. https://siar.regioncajamarca.gob.pe/download/file/fid/46644. ecolind.2019.02.035.
Gómez Fernández, D., Salas López, R., Rojas Briceño, N.B., Silva López, J.O., Oliva, M., Newman, M.E., McLaren, K.P., Wilson, B.S., 2014. Assessing deforestation and
2022. Dynamics of the Burlan and Pomacochas Lakes using SAR data in GEE, fragmentation in a tropical moist forest over 68 years; the impact of roads and legal
machine learning classifiers, and regression methods. ISPRS Int. J. Geo Inf. 11 (11), protection in the cockpit country, Jamaica. For. Ecol. Manag. 315, 138–152. https://
534. https://doi.org/10.3390/IJGI11110534/S1. doi.org/10.1016/J.FORECO.2013.12.033.
Gong, C., Yu, S., Joesting, H., Chen, J., 2013. Determining socioeconomic drivers of Numata, I., Cochrane, M.A., Souza, C.M., Sales, M.H., 2011. Carbon emissions from
urban forest fragmentation with historical remote sensing images. Landsc. Urban deforestation and forest fragmentation in the Brazilian Amazon. Environ. Res. Lett. 6
Plan. 117, 57–65. https://doi.org/10.1016/J.LANDURBPLAN.2013.04.009. (4) https://doi.org/10.1088/1748-9326/6/4/044003.
Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017. Ouma, Y., Nkwae, B., Moalafhi, D., Odirile, P., Parida, B., Anderson, G., Qi, J., 2022.
Google earth engine: planetary-scale geospatial analysis for everyone. Remote Sens. Comparison of machine learning classifiers for multitemporal and multisensor
Environ. 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031. mapping of urban LULC features. Int. Arch. Photogramm. Remote. Sens. Spat. Inf.
Gu, C., Zhao, P., Chen, Q., Li, S., Li, L., Liu, L., Zhang, Y., 2020. Forest cover change and Sci. XLIII-B3-2, 681–689. https://doi.org/10.5194/isprs-archives-XLIII-B3-2022-
the effectiveness of protected areas in the Himalaya since 1998. Sustainability 12 681-2022.
(15), 6123. https://doi.org/10.3390/SU12156123. Pardini, R., de Bueno, A.A., Gardner, T.A., Prado, P.I., Metzger, J.P., 2010. Beyond the
Haddad, N.M., Brudvig, L.A., Clobert, J., Davies, K.F., Gonzalez, A., Holt, R.D., fragmentation threshold hypothesis: regime shifts in biodiversity across fragmented
Lovejoy, T.E., Sexton, J.O., Austin, M.P., Collins, C.D., Cook, W.M., Damschen, E.I., landscapes. PLoS One 5 (10), e13666. https://doi.org/10.1371/JOURNAL.
Ewers, R.M., Foster, B.L., Jenkins, C.N., King, A.J., Laurance, W.F., Levey, D.J., PONE.0013666.
Margules, C.R., Townshend, J.R., 2015. Habitat fragmentation and its lasting impact Parente, L., Ferreira, L., 2018. Assessing the spatial and occupation dynamics of the
on Earth’s ecosystems. Sci. Adv. 1 (2) https://doi.org/10.1126/SCIADV.1500052/ Brazilian pasturelands based on the automated classification of MODIS images from
SUPPL_FILE/E1500052_SM.PDF. 2000 to 2016. Remote Sens. 10 (4) https://doi.org/10.3390/rs10040606.
Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.A., Tyukavina, A., Puyravaud, J.P., 2003. Standardizing the calculation of the annual rate of deforestation.
Thau, D., Stehman, S.V., Goetz, S.J., Loveland, T.R., Kommareddy, A., Egorov, A., For. Ecol. Manag. 177, 593–596. https://doi.org/10.1016/S0378-1127(02)00335-3.
Chini, L., Justice, C.O., Townshend, J.R.G., 2013. High-resolution global maps of Reddy, C.S., Sreelekshmi, S., Jha, C.S., Dadhwal, V.K., 2013. National assessment of
21st-century forest cover change. Science (New York, N.Y.) 342 (6160), 850–853. forest fragmentation in India: landscape indices as measures of the effects of
https://doi.org/10.1126/SCIENCE.1244693. fragmentation and forest cover change. Ecol. Eng. 60, 453–464. https://doi.org/
Hassan, M.M., Duveneck, M., Southworth, J., 2023. The role of the refugee crises in 10.1016/J.ECOLENG.2013.09.064.
driving forest cover change and fragmentation in Teknaf, Bangladesh. Ecol. Inform. Rikimaru, A., Roy, P.S., Miyatake, S., 2002. Tropical forest cover density mapping. Trop.
74, 101966 https://doi.org/10.1016/J.ECOINF.2022.101966. Ecol. 43 (1), 39–47.
Hay, A.M., 1979. Sampling designs to test land-use map accuracy. Photogramm. Eng. Rocha-Santos, L., Pessoa, M.S., Cassano, C.R., Talora, D.C., Orihuela, R.L.L., Mariano-
Remote. Sens. 45, 529–533. https://api.semanticscholar.org/CorpusID:131241764. Neto, E., Morante-Filho, J.C., Faria, D., Cazetta, E., 2016. The shrinkage of a forest:
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Pickell, P.D., Bolton, D.K., 2019. landscape-scale deforestation leading to overall changes in local forest structure.
Impact of time on interpretations of forest fragmentation: three-decades of Biol. Conserv. 196, 1–9. https://doi.org/10.1016/J.BIOCON.2016.01.028.
fragmentation dynamics over Canada. Remote Sens. Environ. 222 (September 2018), Rojas Briceño, N.B., Barboza Castillo, E., Maicelo Quintana, J.L., Oliva Cruz, S.M., Salas
65–77. https://doi.org/10.1016/j.rse.2018.12.027. López, R., 2019. Deforestación en la Amazonía peruana: índices de cambios de
Hua, A.K., 2017. Land use land cover changes in detection of water quality: a study based cobertura y uso del suelo basado en SIG. Boletín de La Asociación de Geógrafos
on remote sensing and multivariate statistics. J. Environ. Public Health 2017, 12. Españoles 81, 1–34. https://doi.org/10.21138/bage.2538a.
https://doi.org/10.1155/2017/7515130.
16 
D. Gómez-Fernández et al. Ecological Informatics 82 (2024) 102738 
Rouse, J., Hass, R., Schell, J., Deering, D.W., 1974. Monitoring vegetation systems in the Uddin, K., Chaudhary, S., Chettri, N., Kotru, R., Murthy, M., Chaudhary, R.P., Ning, W.,
great plains with ERTS. NASA Spec. Publ. 351 (1), 309. Shrestha, S.M., Gautam, S.K., 2015. The changing land cover and fragmenting forest
Rwanga, S.S., Ndambuki, J.M., Rwanga, S.S., Ndambuki, J.M., 2017. Accuracy on the roof of the world: a case study in Nepal’s Kailash sacred landscape. Landsc.
assessment of land use/land cover classification using remote sensing and GIS. Int. J. Urban Plan. 141, 1–10. https://doi.org/10.1016/J.LANDURBPLAN.2015.04.003.
Geosci. 8 (4), 611–622. https://doi.org/10.4236/IJG.2017.84033. Vermote, E., Justice, C., Claverie, M., Franch, B., 2016. Preliminary analysis of the
Schwartz, N.B., Budsock, A.M., Uriarte, M., 2019. Fragmentation, forest structure, and performance of the Landsat 8/OLI land surface reflectance product. Remote Sens.
topography modulate impacts of drought in a tropical forest landscape. Ecology 100 Environ. 185, 46–56. https://doi.org/10.1016/J.RSE.2016.04.008.
(6). https://doi.org/10.1002/ecy.2677. Vogeler, J.C., Slesak, R.A., Fekety, P.A., Falkowski, M.J., 2020. Characterizing over four
SENAMHI, 2020. Mapa climático del Perú. https://www.senamhi.gob.pe/?p=mapa decades of forest disturbance in Minnesota, USA. Forests 11 (3), 362. https://doi.
-climatico-del-peru. org/10.3390/F11030362.
Shen, W., Li, M., Huang, C., Tao, X., Li, S., Wei, A., 2019. Mapping annual Forest change Vogt, P., Riitters, K.H., Estreguil, C., Kozak, J., Wade, T.G., Wickham, J.D., 2006.
due to afforestation in Guangdong Province of China using active and passive remote Mapping spatial patterns with morphological image processing. Landsc. Ecol. 22 (2),
sensing data. Remote Sens. 11 (5), 490. https://doi.org/10.3390/RS11050490. 171–177. https://doi.org/10.1007/S10980-006-9013-2.
Shetty, S., 2019. Analysis of Machine Learning Classifiers for LULC Classification on Wang, S., Azzari, G., Lobell, D.B., 2019a. Crop type mapping without field-level labels:
Google Earth Engine. Master’s thesis. University of Twente. http://essay.utwente.nl random forest transfer and unsupervised clustering techniques. Remote Sens.
/83543/1/shetty.pdf. Environ. 222, 303–317. https://doi.org/10.1016/J.RSE.2018.12.026.
Shimizu, K., Ahmed, O.S., Ponce-Hernandez, R., Ota, T., Win, Z.C., Mizoue, N., Wang, Y., Ziv, G., Adami, M., Mitchard, E., Batterman, S.A., Buermann, W., Schwantes
Yoshida, S., 2017. Attribution of disturbance agents to Forest change using a Landsat Marimon, B., Marimon Junior, B.H., Matias Reis, S., Rodrigues, D., Galbraith, D.,
time series in tropical seasonal forests in the Bago Mountains, Myanmar. Forests 8 2019b. Mapping tropical disturbed forests using multi-decadal 30 m optical satellite
(6), 218. https://doi.org/10.3390/f8060218. imagery. Remote Sens. Environ. 221, 474–488. https://doi.org/10.1016/J.
Smith, J., Schwartz, J., 2015. Deforestation in Perú. World Wild Life. https://www.world RSE.2018.11.028.
wildlife.org/magazine/issues/fall-2015/articles/deforestation-in-peru. Wilson, E.H., Sader, S.A., 2002. Detection of forest harvest type using multiple dates of
Stehman, S.V., 1997a. Estimating standard errors of accuracy assessment statistics under Landsat TM imagery. Remote Sens. Environ. 80 (3), 385–396. https://doi.org/
cluster sampling. Remote Sens. Environ. 60 (3), 258–269. https://doi.org/10.1016/ 10.1016/S0034-4257(01)00318-2.
S0034-4257(96)00176-9. Wulder, M.A., White, J.C., Andrew, M.E., Seitz, N.E., Coops, N.C., 2009. Forest
Stehman, S.V., 1997b. Selecting and interpreting measures of thematic classification fragmentation, structure, and age characteristics as a legacy of forest management.
accuracy. Remote Sens. Environ. 62 (1), 77–89. https://doi.org/10.1016/S0034- For. Ecol. Manag. 258 (9), 1938–1949. https://doi.org/10.1016/j.
4257(97)00083-7. foreco.2009.07.041.
Talukdar, S., Singha, P., Mahato, S., Shahfahad, Pal S., Liou, Y.A., Rahman, A., 2020. Wulder, M.A., Masek, J.G., Cohen, W.B., Loveland, T.R., Woodcock, C.E., 2012. Opening
Land-use land-cover classification by machine learning classifiers for satellite the archive: how free data has enabled the science and monitoring promise of
observations—a review. Remote Sens. 12 (7), 1135. https://doi.org/10.3390/ Landsat. Remote Sens. Environ. 122, 2–10. https://doi.org/10.1016/j.
RS12071135. rse.2012.01.010.
Taubert, F., Fischer, R., Groeneveld, J., Lehmann, S., Müller, M.S., Rödig, E., Xie, Z., Phinn, S.R., Game, E.T., Pannell, D.J., Hobbs, R.J., Briggs, P.R., McDonald-
Wiegand, T., Huth, A., 2018. Global patterns of tropical forest fragmentation. Nature Madden, E., 2019. Using Landsat observations (1988–2017) and Google earth engine
554 (7693), 519–522. https://doi.org/10.1038/nature25508. to detect vegetation cover changes in rangelands - a first step towards identifying
Thomlinson, J.R., Bolstad, P.V., Cohen, W.B., 1999. Coordinating methodologies for degraded lands for conservation. Remote Sens. Environ. 232, 111317 https://doi.
scaling Landcover classifications from site-specific to global: steps toward validating org/10.1016/J.RSE.2019.111317.
global map products. Remote Sens. Environ. 70 (1), 16–28. https://doi.org/ Xun, B., Yu, D., Liu, Y., Hao, R., Sun, Y., 2014. Quantifying isolation effect of urban
10.1016/S0034-4257(99)00055-3. growth on key ecological areas. Ecol. Eng. 69, 46–54. https://doi.org/10.1016/j.
Thuiller, W., Albert, C., Araújo, M.B., Berry, P.M., Cabeza, M., Guisan, A., Hickler, T., ecoleng.2014.03.041.
Midgley, G.F., Paterson, J., Schurr, F.M., Sykes, M.T., Zimmermann, N.E., 2008. Zhang, Y., Liu, X., Yang, Q., Liu, Z., Li, Y., 2021. Extracting frequent sequential patterns
Predicting global change impacts on plant species’ distributions: future challenges. of forest landscape dynamics in fenhe river basin, northern China, from landsat time
Perspect. Plant Ecol. Evol. System. 9 (3–4), 137–152. https://doi.org/10.1016/J. series to evaluate landscape stability. Remote Sens. 13 (19) https://doi.org/
PPEES.2007.09.004. 10.3390/rs13193963.
Traore, M., Lee, M.S., Rasul, A., Balew, A., 2021. Assessment of land use/land cover Zhu, Z., Woodcock, C.E., 2014. Continuous change detection and classification of land
changes and their impacts on land surface temperature in Bangui (the capital of cover using all available Landsat data. Remote Sens. Environ. 144, 152–171. https://
Central African Republic). Environ. Challeng. 4, 100114 https://doi.org/10.1016/J. doi.org/10.1016/J.RSE.2014.01.011.
ENVC.2021.100114. Zhu, Z., Wulder, M.A., Roy, D.P., Woodcock, C.E., Hansen, M.C., Radeloff, V.C.,
Tsai, Y.H., Stow, D., An, L., Chen, H.L., Lewison, R., Shi, L., 2019. Monitoring land-cover Healey, S.P., Schaaf, C., Hostert, P., Strobl, P., Pekel, J.F., Lymburner, L.,
and land-use dynamics in Fanjingshan National Nature Reserve. Appl. Geogr. 111, Pahlevan, N., Scambos, T.A., 2019. Benefits of the free and open Landsat data policy.
102077 https://doi.org/10.1016/J.APGEOG.2019.102077. Remote Sens. Environ. 224, 382–385. https://doi.org/10.1016/j.rse.2019.02.016.
Turner, M., 2005. Landscape ecology in North America: special feature. Ecology 86 (8),
1967–1974.
17