Escuela Agrícola Panamericana, Zamorano 

Environmental Science and Development 

B.S. in Environmental Science and Development 

 
 

 
 

 

Special Graduation Project  

Analysis of Land Use and Land Cover Changes on the Island of Roatán 

from 2017 to 2025 

 

 

 

 

Nicholas Matthew McNab Tomé 

Advisors 

 Alexandra Manueles, Mtr. 

 José Antonio Molina, M.Sc. 

 

 

 

 

 

                                                  Honduras, august 2025 



2 

 

 

Authorities 

 

 

 

KEITH ANDREW  

President i.a. 

 

 

 

 

ANA M. MAIER ACOSTA 

Vice President and Academic Dean 

 

 

 

 

VICTORIA CORTÉS MATAMOROS 

Director of Environmental Science and Development 

 

 

 

 

JULIO NAVARRO 

Secretary General 

 

 



3 

 

 

Table of Contents 

List of Tables ........................................................................................................................................... 5 

List of Figures .......................................................................................................................................... 6 

List of Appendices ................................................................................................................................... 7 

Abstract ................................................................................................................................................... 8 

Introduction .......................................................................................................................................... 10 

Methodology ......................................................................................................................................... 12 

Study Area ............................................................................................................................................. 12 

Roatán Municipality (Western Region) ................................................................................................. 13 

José Santos Guardiola Municipality (Eastern Region) .......................................................................... 13 

Study Design ......................................................................................................................................... 14 

Spectral Index-Based Visual Analysis .................................................................................................... 15 

NDVI ...................................................................................................................................................... 16 

NDBI ...................................................................................................................................................... 17 

Supervised Land Use and Cover Classification ...................................................................................... 18 

Pre-processing phase of Supervised classification of LULC Maps......................................................... 19 

Processing phase for Supervised Classification of LULC Maps ............................................................. 22 

Postprocessing for Supervised Classification of LULC Maps ................................................................. 25 

Confusion Matrix ................................................................................................................................... 25 

Cross-Tabulation Matrix for quantification of category-level gains, losses and transitions ................. 28 

Interpreting Change Performance from Pontius Matrix ....................................................................... 30 

Results and Discussion .......................................................................................................................... 31 

Spectral-Indexes: visual based analysis ................................................................................................ 31 

Supervised Classification Land use and Land cover Maps .................................................................... 34 

Cross tabulation matrix: quantification of gains, losses, and transitions at category level ................. 47 



4 

 

 

Classification Accuracy Assessment Using Confusion Matrices ............................................................ 51 

Conclusions ........................................................................................................................................... 54 

Recommendations ................................................................................................................................ 55 

References ............................................................................................................................................ 57 

Appendices ............................................................................................................................................ 60 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

  



5 

 

 

List of Tables  

Table 1 Classification of land cover according to NDVI values ............................................................. 17 

Table 2 Land cover classification based on NDBI values ....................................................................... 18 

Table 3 Spectral bands, wavelength ranges, and spatial resolutions of Sentinel-2 MSI sensor ........... 20 

Table 4 Simplified CORINE land cover classification scheme ................................................................ 22 

Table 5 Example of Confusion Matrix format ....................................................................................... 26 

Table 6  LULC 2017 and 2025 Pontius Matrix Results Tabulated .......................................................... 47 

Table 7 Results from random forest classifier for each year showing total land cover area (ha) for each 

land cover class for the entire island .................................................................................................... 48 

Table 8  Results for LULC Gains, losses, and changes ........................................................................... 50 

Table 9 Transition trends towards urbanization ................................................................................... 51 

Table 10 2017 Random Forest Classifier Accuracy Results  .................................................................. 52 

Table 11 2025 Random Forest Classifier Accuracy Results ................................................................... 52 

 

 

 

 

 

 

 

 

 

 

 

 



6 

 

 

List of Figures 

Figure 1 Roatán Island map with municipal boundaries....................................................................... 13 

Figure 2 Data collection and processing flowchart ............................................................................... 19 

Figure 3 Spectral reflectance curves for different land surface types .................................................. 21 

Figure 4 Sentinel 2A for 2017 and 2025 respectively, followed by stacked rasters imagery ............... 21 

Figure 5 Creation of ROIs (Regions of Interest) in QGIS........................................................................ 23 

Figure 6 Simplified Random Forest classification process .................................................................... 24 

Figure 7 Cross-tabulation matrix used for land cover change analysis ................................................ 28 

Figure 8 NDVI map of Roatán for 2017 for Western portion of the island ........................................... 31 

Figure 9 NDVI map of Roatán for 2025 Western portion of the island ................................................ 32 

Figure 10 NDBI processed image 2025 central area from south to north shore .................................. 33 

Figure 11 NDBI processed image of the year 2017 central area from south to north shore ............... 33 

Figure 12 Supervised classification results for LULC Random Forest Classifier Algorithm maps 2017 and 

2025 ...................................................................................................................................................... 35 

Figure 13 LULC Maps West Bay 2017 and 2025.................................................................................... 36 

Figure 14  LULC Maps Coxen hole, Sandy Bay, north and south shore 2017 and 2025 ....................... 37 

Figure 15 LULC Maps Mud hole, north shore 2017 and 2025 .............................................................. 38 

Figure 16  LULC Maps Crawfish,  Prospera, Pristine Bay, French Key, north shore 2017 and 2025 ..... 40 

Figure 17 LULC Maps Los Fuertes, French Harbour, French Key, South Shore ..................................... 41 

Figure 18 LULC Maps José Santos Guardiola, east side, north and south shore .................................. 42 

Figure 19 LULC Maps Jose Santos Guardiola, Oak ridge, Port Royal National Park, St Elena 2017 and 

2025 ...................................................................................................................................................... 43 

Figure 20  LULC Maps St Helene mangrove system and St Helene island, eastern side 2017 and 2025

 .............................................................................................................................................................. 44 

Figure 21 LULC Maps Barbareta island, east side, 2017 and 2025 ....................................................... 46 



7 

 

 

List of Appendices 

Appendix A Satellite Imagery for Roatan Island 2017 and 2025 from Coppernicus Data Hub ............ 60 

Appendix B Pontius matrix for NDVI index classification from 2017 to 2025 (in hectares) ................. 62 

Appendix C 2017 and 2025 Pontius Matrix .......................................................................................... 63 

Appendix D  High Res Comparison Catalog, LULC Results and Google Earth Pro® ............................... 64 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 



8 

 

 

Abstract 

This study analyzes land use and land cover (LULC) changes on the island of Roatán, Honduras, 

between 2017 and 2025 using remote sensing and GIS-based techniques to categorize land use and 

quantify the gains, losses, and transitions related to shifting land use patterns associated with tourism 

and development pressures. Multi-temporal Sentinel-2A imagery was processed through supervised 

classification using the Random Forest algorithm, guided by training data corresponding to key macro 

land cover categories including vegetation, artificial surfaces, wetlands, open spaces, and water 

bodies. To complement the classification process, two spectral indices: NDVI and NDBI were applied 

as visual interpretation tools to support the identification of vegetation health, built-up density, and 

land transformation trends. This study employed a simplified Corine Land Cover (CLC)-inspired 

methodology, adapted to the tropical island context. Quantitative change analysis was conducted 

using the Pontius matrix to evaluate spatial gains, losses, and persistence across land cover classes. 

Results indicate a marked expansion of industrial, commercial and transport zones and a notable 

reduction in open areas with little or no vegetation, reflecting a trend toward intensified 

industrialization on Roatán. Forests showed both losses and localized gains, reflecting land use 

fluctuations. This research contributes to understanding the spatial development dynamics and 

environmental change in Roatán, with implications for land management, conservation planning, and 

future coastal resilience strategies. 

Keywords: Land use and land cover change (LULC), NDBI, NDVI, Pontius matrix, remote sensing 

 

 

 

 

 

 



9 

 

 

Resumen 

Este estudio analiza los cambios en el uso y la cobertura del suelo (LULC, por sus siglas en inglés) en la 

isla de Roatán, Honduras, entre 2017 y 2025, utilizando técnicas de teledetección y SIG para 

categorizar el uso del suelo y cuantificar las ganancias, pérdidas y transiciones relacionadas con los 

patrones cambiantes de uso asociados a las presiones del turismo y el desarrollo. Imágenes 

multitemporales del satélite Sentinel-2A fueron procesadas mediante clasificación supervisada 

utilizando el algoritmo Random Forest, guiada por datos de entrenamiento correspondientes a 

categorías macro clave de cobertura del suelo, incluyendo vegetación, superficies artificiales, 

humedales, espacios abiertos y cuerpos de agua. Para complementar el proceso de clasificación, se 

aplicaron dos índices espectrales: NDVI y NDBI, como herramientas de interpretación visual para 

apoyar la identificación del estado de salud de la vegetación, la densidad de áreas construidas y las 

tendencias de transformación del suelo. Este estudio empleó una metodología simplificada inspirada 

en Corine Land Cover (CLC), adaptada al contexto de una isla tropical. El análisis cuantitativo de 

cambios se realizó utilizando la matriz de Pontius para evaluar las ganancias, pérdidas y persistencia 

espacial entre clases de cobertura del suelo. Los resultados indican una marcada expansión de las 

zonas industriales, comerciales y de transporte, así como una reducción notable de las áreas abiertas 

con poca o ninguna vegetación, lo que refleja una tendencia hacia una industrialización intensificada 

en Roatán. Los bosques mostraron tanto pérdidas como ganancias localizadas, reflejando 

fluctuaciones en el uso del suelo. Esta investigación contribuye a la comprensión de las dinámicas de 

desarrollo espacial y cambio ambiental en Roatán, con implicaciones para la gestión del territorio, la 

planificación de la conservación y futuras estrategias de resiliencia costera. 

Palabras clave: Matriz de Pontius, NDBI, NDVI, sensores remotos, transformación del suelo, 

uso y cobertura del suelo (LULC) 

 

  



10 

 

 

Introduction 

Island ecosystems are highly vulnerable to land use/cover changes (LULC) due to their limited 

space and fragile biodiversity (Tuholske et al., 2017). Roatán, the largest of the Bay Islands of 

Honduras, has experienced significant LULC transformations driven by rapid urbanization, tourism 

expansion, and demographic growth. These pressures have led to deforestation, loss of mangroves, 

and degradation of coastal habitats, impacting both the environment and local communities. 

Understanding the magnitude and trends of these changes is essential for sustainable land 

management and decision-making.  

To acknowledge the human dimensions of global change, the International Geosphere-

Biosphere Program (IGBP) and the International Human Dimensions Program (IHDP) initiated the 

“Land Use/Cover Change (LUCC)” project in 1995 (Guan et al., 2011). One of the most effective 

methods for assessing LULC changes is remote sensing, which utilizes satellite imagery to monitor and 

analyze environmental transformations over time. This technology has become increasingly valuable 

for detecting land cover shifts, quantifying deforestation, and assessing urban expansion. Accurate 

and up-to-date land use and land cover (LULC) maps are essential for numerous applications, including 

urban and regional planning, disaster and hazard monitoring, natural resource and environmental 

management, and ensuring food security (Zhang & Li, 2022). Modern satellite missions such as Landsat 

and Sentinel provide high-resolution, multi-temporal data that enable detailed change detection 

analyses, supporting long-term environmental monitoring efforts. Satellite imagery has become a 

crucial tool for observing and analyzing transformations in both natural landscapes and human-made 

environments (Hemati et al., 2021). 

Roatán has diverse topography and ecosystems, ranging from forested highlands to coastal 

mangroves and wetlands. A ridge runs along the island’s center, reaching a maximum elevation of 244 

m, with rock formations composed of serpentinite, amphibolite, biotite, chlorite schist, and gneiss 

(McBirney & Bass, 1969; Sutton, 2015). The northwestern coastline features rocky outcrops, alluvial 



11 

 

 

deposits, and limestone formations, while several natural and artificial beaches line the island's shores  

(Sutton, 2015). Tuholske et al., 2017 documented significant tourism-related land cover changes on 

Roatán over a 30-year period, identifying substantial habitat loss driven by development pressures. 

Their results indicate that urban area increased by 982.8 ha (227.7%), with 224.1 ha (19.1%) of 

mangroves being converted to urban land. This study highlighted the extent of environmental 

transformation associated with tourism growth; however, its temporal scope ended in 2015, leaving 

a critical gap in understanding more recent land use and land cover (LULC) dynamics on the island. 

There is an urgent need for updated data to inform land management strategies.  

A periodic assessment of LULC changes is essential for informed decision-making. Forested 

land use is associated with better coral health indicators, whereas human-driven land uses such as 

urbanization and agriculture contribute to reductions in coral cover, species diversity, colony size, and 

structural complexity. (Carlson et al., 2019). The rapid tourism explosion since the 1990s of over 2.41 

million tourists in 2024, according to the Roatan Tourism Bureau (2024), which has driven 

infrastructure development, calling for regular monitoring of land cover changes. The findings of this 

study could be valuable for policymakers, municipal authorities, and local organizations such as 

ZOLITUR (Zona libre turistica islas de la bahia), RTMP (Roatan marine park), and BICA (Bay islands 

conservation agency), aiding in sustainable planning and conservation efforts.  

This research aims to analyze land use and land cover (LULC) changes on Roatán from 2017 to 

2025 using remote sensing and geo graphical information-based techniques, which constitutes the 

general objective of the study. To achieve this, the study seeks to classify and quantify land cover 

changes; specifically gains, losses, and transitions between categories—through supervised 

classification and change detection analysis using multi-temporal satellite imagery. Additionally, it 

aims to identify which land cover categories experienced the greatest increases and decreases in area 

over the study period, providing a clearer understanding of the spatial dynamics and land 

transformation trends on the island. 



12 

 

 

Methodology 

Study Area 

Roatán, the largest of Honduras’s Bay Islands, located in the Caribbean Sea, approximately 60 

km north of mainland Honduras. It extends about 59 km in length and 8 km at its widest point, covering 

an area of approximately 130 km². The island is part of the Mesoamerican Reef System (MAR), the 

largest coral reef system in the Western Hemisphere, which stretches over 1,000 km from Mexico to 

Honduras. Recognized as a United Nations Educational, Scientific and Cultural world heritage site, the 

MAR plays a vital role in coastal protection, marine biodiversity, and local fisheries, while also serving 

as a key attraction for the island’s growing tourism industry (Burke & Maidens, 2004; Mcfield et al., 

2018). However, coastal development, deforestation, and pollution increasingly threaten this delicate 

marine ecosystem. 

Roatán has a tropical climate, receiving an average annual rainfall of 2,000 mm. The rainy 

season occurs from October to November, while the dry season (January to June) sees significantly 

lower precipitation, often below 100 mm. Rainfall distribution varies, with the eastern part of the 

island receiving less precipitation than the west, leading to drier vegetation types, such as mixed 

shrubs and grasslands (Stonich, 1998; Tomczyk, 2010).  

Administratively, Roatán is divided into two municipalities, Roatan and José Santos Guardiola, 

each with distinct land use patterns (Figure 1).  

 

 

 

 

 



13 

 

 

Figure 1 

Roatán Island map with municipal boundaries 

 Note. Own elaboration based on geoportal data  (Instituto de Conservacion Forestal, n.d). 

Roatán Municipality (Western Region) 

The western region has experienced significant urban expansion, largely due to tourism-driven 

infrastructure projects. This area is home to the Roatán International Airport and two major cruise ship 

terminals—Port of Roatán and Mahogany Bay—which have contributed to rapid land conversion from 

natural habitats to urbanized areas (Tuholske et al., 2017). Major population centers such as Coxen 

Hole, French Harbour, and West End have seen substantial increases in commercial and residential 

development, often at the expense of forests, wetlands, and mangrove ecosystems (Stonich, 1998). 

José Santos Guardiola Municipality (Eastern Region) 

In contrast, the eastern municipality of José Santos Guardiola remains less urbanized, with 

lower population density and greater ecological conservation efforts. This region contains several 

protected areas, including Port Royal National Park, which supports biodiversity conservation and 



14 

 

 

sustainable land management (Tomczyk, 2010). Additionally, Santa Elena Wetland System recognized 

as a RAMSAR (Convention of wetlands) site in 2018 (1,542.8 ha) helps safeguard mangroves and 

marine habitats. While small urban centers such as Oak Ridge and Punta Gorda have grown modestly, 

land conversion has been much slower compared to the western region.  

Study Design  

This is a descriptive, non-experimental, and longitudinal study. It is a descriptive study because 

the analysis focuses on classifying, describing, and comparing LULC changes in Roatán between 2017 

and 2025. There is no manipulation of variables in a controlled setting, making it non-experimental. It 

is a longitudinal study as it examines LULC dynamics over two different time periods to assess changes 

over nearly a decade. 

This study uses Sentinel-2A satellite imagery from the years 2017 and 2025 to analyze land 

use and land cover (LULC) changes in Roatán. A supervised classification approach will be employed 

to generate land cover maps using a simplified version of the CORINE Land Cover (CLC) methodology. 

This process will involve assigning land cover categories—such as urban areas, vegetation, and 

wetlands—based on manually selected training samples and spectral signatures.  

In addition to supervised classification, a set of spectral indices will be applied to support 

visual interpretation of landscape dynamics. These indices do not constitute an unsupervised 

classification method but rather serve as analytical tools to highlight specific land cover features and 

facilitate a more intuitive understanding of spatial patterns. Specifically, the Normalized Difference 

Vegetation Index (NDVI) (Tucker, 1979) and the Normalized Difference Built-up Index (NDBI) (Zha et 

al., 2003), will be used to generate thematic maps that reveal vegetation health, urban density, and 

built-up distribution. In line with the main objective to categorize and classify the different land uses 

and land covers.  

 



15 

 

 

This combined use of supervised classification and index-based visualization enables a more 

comprehensive and nuanced analysis of land cover transformations across the study period. All image 

processing and spatial analysis will be carried out using QGIS 3.42.3®. A confusion matrix will be 

generated to assess the accuracy of the supervised classification results for each year. This matrix 

compares the predicted land cover classes from the classification output against a set of reference 

data, allowing for the identification of both correctly classified and misclassified area. Finally, the 

Pontius matrix will be used to quantify land cover change by estimating the spatial gains, losses, and 

transitions between land cover categories. For this study´s secondary objective to quantify and 

describe the trends exhibited from the LULC maps that have been generated from the main objective.  

Spectral Index-Based Visual Analysis 

In addition to the supervised classification performed using the Random Forest algorithm, this 

study employs spectral indices as complementary tools for interpreting land cover conditions and 

change patterns. These indexes are not used as unsupervised classification techniques, but rather as 

diagnostic visual layers that enhance the interpretation of thematic features such as vegetation health 

and built-up density, across the landscape. By applying targeted index formulas to Sentinel-2 imagery, 

it is possible to generate raster outputs that highlight specific land cover characteristics and support 

a clearer, more informed understanding of the transformations that occurred between 2017 and 

2025. 

The selected indexes; NDVI and NDBI are widely recognized in the remote sensing literature 

for their effectiveness in delineating vegetation, artificial surfaces, and mixed-use urban 

environments, respectively. Each index was computed using appropriate spectral bands from Sentinel-

2 imagery and reclassified into descriptive value ranges. These outputs were used not as classification 

layers, but as visual interpretation tools to validate patterns, identify outliers, and complement the 

creation of training data for the supervised classification.  

 



16 

 

 

NDVI 

The Normalized Difference Vegetation Index (NDVI) is one of the most fundamental and 

enduring remote sensing indices used to evaluate vegetative cover and condition. Originally 

developed by Tucker (1979), NDVI was designed to detect live green vegetation by exploiting the 

unique spectral reflectance properties of vegetation: strong absorption of red light due to chlorophyll 

and high reflectance in the near-infrared (NIR) spectrum due to the internal structure of healthy plant 

leaves. The NDVI formula is expressed in Equation 1. 

 
                                                𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 = (𝑁𝑁𝑁𝑁𝑁𝑁−𝑅𝑅𝑅𝑅𝑅𝑅)

(𝑁𝑁𝑁𝑁𝑁𝑁+𝑅𝑅𝑅𝑅𝑅𝑅)
                                                                 [1] 

 
Tucker’s pioneering study utilized satellite-derived radiance data to monitor vegetation 

dynamics in semi-arid regions of Senegal. His findings demonstrated that NDVI could reliably quantify 

vegetation biomass, phenology, and canopy structure over time, even in areas with low vegetation 

density. This marked the beginning of spectral index-driven remote sensing for land monitoring. More 

than four decades later, NDVI remains a central tool in land use and land cover (LULC) classification, 

due to its simplicity, robustness, and interpretability. 

In the present study, NDVI was computed from Sentinel-2 imagery using Band 8 (NIR) and 

Band 4 (Red). Its application enabled a rapid and effective unsupervised classification of the study area 

into vegetation and non-vegetation zones, without the need for training data. The resulting NDVI 

raster was segmented into land cover categories based on value thresholds (Table 1), allowing 

differentiation of dense vegetation, sparse cover, and anthropogenic surfaces. 

This approach is well-supported by recent literature. Da Silva et al. (2020) used NDVI to classify 

a complex rural landscape in southern Brazil, demonstrating high accuracy in distinguishing native 

forests, fallow land, and water bodies. Similarly, Sohail (2020) applied NDVI in the arid urban 

environment of Karachi, successfully mapping orchards, mangroves, and vegetation pockets. Akbar et 



17 

 

 

al. (2019) further operationalized NDVI thresholds for automatic classification of vegetation density in 

Pakistan’s forested regions, providing a clear range of values for land cover interpretation. 

The NDVI range classification used in this study is adapted from Akbar et al. (2019), as shown 

in Table 1. 

Table 1 

Classification of land cover according to NDVI values 

NDVI Range Land Cover Description 
-0.28-0.015 Water 
0.015 – 0.14 Artificial surfaces 

0.14-0.27 Open spaces with little or no 
vegetation 

0.27-0.36 Sparse vegetation 
0.36-0.74 Dense vegetation 

Note. Adapted from Akbar et al. (2019) 

NDBI 

The Normalized Difference Built-up Index (NDBI) was first proposed by Zha et al. (2003), as a 

simple yet effective method for identifying built-up areas in satellite imagery, especially in rapidly 

urbanizing environments. The index utilizes the spectral properties of urban materials, which tend to 

reflect strongly in the short-wave infrared (SWIR) region and less in the near infrared (NIR), a pattern 

opposite to that of vegetation. The NDBI is calculated using Equation 2. 

NDBI =
(SWIR - NIR) 
(SWIR + NIR)

                                                             [2] 

In this study, NDBI was derived from Sentinel-2 imagery using Band 11 (SWIR) and Band 8 

(NIR). This index was chosen to isolate and classify built-up areas such as residential, commercial, and 

transportation infrastructure without the need for ground-truth data or training samples. The index 

values typically range between -1 and +1, where positive values generally indicate urban surfaces and 

negative values correspond to vegetation or water (Table 2). 



18 

 

 

Table 2 

Land cover classification based on NDBI values 

NDBI Range Land Cover Description 
< 0.00 Vegetation or water 
0.00 – 0.20 Transitional surfaces 
> 0.20 Built-up / urban features 

Note. Adapted from Zha et al. (2003) 

Supervised Land Use and Cover Classification  

A supervised classification method was applied, utilizing a simplified version of the Corine 

Land Cover (European Environment Agency, 2019), methodology to categorize different land cover 

types such as vegetation, urban areas, water bodies, and wetlands. The classification accuracy will be 

evaluated using a confusion matrix, which compares classified results with reference data to assess 

errors of commission and omission. Change detection analysis will then be performed by comparing 

classified maps from both years, allowing for the identification of land cover trends such as 

deforestation, urban expansion, or mangrove loss.  

The process begins with a pre-processing phase and concludes with a post-processing phase 

(Fig. 2). This approach has been widely used in similar studies, demonstrating its effectiveness in 

monitoring environmental changes and supporting sustainable land management strategies 

(Alqurashi & Kumar, 2013).  



19 

 

 

Figure 2 

Data collection and processing flowchart 

 

Note. Own elaboration based on supervised classification workflow using and Sentinel-2 imagery. 

 

Pre-processing phase of Supervised classification of LULC Maps  

The pre-processing phase ensures that the satellite imagery is prepared and standardized for 

analysis. Sentinel-2A Level-1C imagery corresponding to the study area of Roatán was downloaded 

from the Copernicus Open Access Hub for the years 2017 and 2025. Images were selected based on 

criteria of minimal cloud coverage and seasonal comparability (March to May) to reduce climatic 

variability and ensure consistency in land cover interpretation. To focus solely on the study area, all 

images were clipped to the geographic boundaries of Roatán. 

Each Sentinel-2 image includes 13 spectral bands, out of which six (Table 3) were selected for 

processing based on their spatial resolution and relevance to land use and land cover (LULC) 

classification. These included Band 2 (Blue), Band 3 (Green), Band 4 (Red), and Band 8 (Near Infrared) 

each at 10-meter resolution; and Band 11 (SWIR 1) and Band 12 (SWIR 2), both at 20-meter resolution. 

These bands were chosen for their ability to distinguish between key land cover types such as 

vegetation, bare soil, urban surfaces, and wetlands, as supported by literature (Iurist et al., 2016). 



20 

 

 

Table 3 

Spectral bands, wavelength ranges, and spatial resolutions of Sentinel-2 MSI sensor 

Band Number Band Description Wavelength Range (nm) Resolution (m) 
B2 Blue 458–523 10 
B3 Green 543–578 10 
B4 Red 650–680 10 
B8 Near infrared 

(NIR) 
785–900 10 

B11 Shortwave 
infrared 1 (SWIR1) 

1,565–1,655 20 

B12 Shortwave 
infrared 2 (SWIR2) 

2,100–2,280 20 

Note. Adapted from Drusch et al. (2012). 

Band 2 (Blue, 458–523 nm) is especially sensitive to surface reflectance from sand, cement, 

and artificial surfaces, which show high reflectance in this range, while vegetation and water bodies 

show minimal reflectance due to chlorophyll and water absorption. Band 3 (Green, 543–578 nm) is 

effective for detecting vegetation vigor and identifying shallow aquatic environments. Band 4 (Red, 

650–680 nm) provides strong contrast between vegetation and non-vegetated surfaces, making it 

essential for basic vegetation discrimination and for generating indices such as NDVI in combination 

with Band 8. Band 8 (Near-Infrared, 785–900 nm) captures the high reflectance of plant canopies and 

is critical for analyzing vegetation density and health. 

Bands 11 and 12 (SWIR1: 1,565–1,655 nm and SWIR2: 2,100–2,280 nm) are particularly 

sensitive to surface and vegetation moisture. Their inclusion allows the detection of wetland extents 

and the differentiation of built-up or dry zones from moist vegetative or water-saturated areas. 

 The spectral reflectance curves presented in Figure 3 further illustrate how different land 

cover types (e.g., green vegetation, dry soil, water, concrete) respond across these wavelengths. For 

example, vegetation exhibits high reflectance in the NIR band and dips in the red and blue bands, 

while built-up surfaces and bare soils show relatively higher reflectance across the visible spectrum.  



21 

 

 

Figure 3 

Spectral reflectance curves for different land surface types 

Note. Adapted from Hou et al. (2016). 

The selected bands were clipped and stacked using the Semi-Automatic Classification Plugin 

(SCP) in QGIS® to produce a single composite raster for each year. These stacked images were used to 

create a true color base and enhance visualization for the selection of training samples. They also 

served as the foundation for defining Region of Interest (ROI) polygons corresponding to major land 

cover categories such as vegetation, built-up areas, wetlands, and open spaces.  The resulting 

composite images for 2017 and 2025 are shown below in Figure 4. 

Figure 4 

Sentinel 2A for 2017 and 2025 respectively, followed by stacked rasters imagery 

  



22 

 

 

  

Note.  Coppernicus Satellite Hub Satellite Sentinel 2A imagery 

Processing phase for Supervised Classification of LULC Maps 

Land use and cover was categorized into three main classes following a simplified Corine Land 

Cover (CLC) classification structure into subclasses (Table 4): (1) Artificial Surfaces, subdivided into 

Urban Fabric and Industrial, Commercial and Transport Units; (2) Forest and Seminatural Areas, 

including Forests and Open Spaces with Little or No Vegetation; and (3) Wetlands, with emphasis on 

Maritime Wetlands.  

Table 4 

Simplified CORINE land cover classification scheme 

Level 1 Level 2 
1. Artificial Surfaces 1.1 Urban Fabric 

 1.2 Industrial, commercial, and transport units 
3. Forest and seminatural areas 3.1 Forests 

 3.3 Open spaces with little or no vegetation 
4. Wet-lands 4.2 Maritime wetlands 

Note. Simplified adaptation from the CORINE Land Cover classification scheme European Environment Agency (2019). 

 
The classification process begins with the creation of training samples, also referred to as 

Regions of Interest (ROIs) or Training Areas. ROIs are polygons manually drawn over spectrally 

homogeneous areas of the image that contain pixels belonging to a single land cover class (Figure 5). 

These samples define the spectral signature for each class and serve as the reference input for the 

classification algorithm. For this study, approximately 200 ROIs were selected proportionally and 



23 

 

 

evenly distributed across the defined classes and years (2017 and 2025), ensuring representative 

training data for both temporal datasets. 

Figure 5 

Example of ROIs (Regions of Interest) in QGIS® 

Note. Screenshot of polygon-based ROI definition in QGIS® during supervised classification preparation. 

Using these ROIs, a supervised classification was performed through the Random Forest 

Mapping (RFM) algorithm, a robust machine learning method widely adopted for land cover 

classification used by Tuholske et al. (2017). Machine learning in remote sensing refers to a class of 

algorithms that learn patterns from training data to predict classifications over new data. The process 

involves iteratively splitting the dataset into a training set, validation set, and test set, allowing the 

model to calibrate, tune, and evaluate its accuracy independently (European Space Agency Lab, 2019). 

The Random Forest algorithm (Figure 6) was used for supervised classification in this study 

due to its high execution speed, robustness to overfitting, and proven effectiveness in handling high-

dimensional data. Random Forests, introduced by Tin Kam Ho (1995), are based on constructing 

multiple decision trees trained on randomly selected subspaces of the feature space. This ensemble 



24 

 

 

approach leverages the diversity among trees to produce more stable and accurate classifications, 

especially on unseen data. Unlike traditional decision trees, which often suffer from overfitting and 

poor generalization, Random Forests generalize better by combining the predictions of multiple 

diverse trees, each trained on different random subsets of the data and feature space. 

A key advantage of this method lies in its ability to expand classification capacity without a 

significant increase in overfitting. Ho (1995) demonstrated that classification accuracy can be 

monotonically improved by increasing the number of decision trees, as long as the trees are 

sufficiently diverse. In this study, Random Forests provided a reliable framework for generating land 

cover maps, as they capture complex patterns in spectral data while maintaining high classification 

performance across different land cover classes. 

Figure 6 

Simplified Random Forest classification process 

 

 

Note. Adapted from Azhari et al. (2019). 

This supervised classification was performed separately for both 2017 and 2025, producing 

two LULC raster maps. These outputs served as the core inputs for the detection and quantification of 

land use/cover changes across the study period through the change detection matrix.  

  



25 

 

 

Postprocessing for Supervised Classification of LULC Maps  

Post-processing in this study involved visual inspection, manual refinement, and accuracy 

validation of the classified land cover maps for the years 2017 and 2025. Rather than relying solely on 

automated tools such, the classification outputs were reviewed through on-screen comparison with 

high-resolution satellite imagery and known reference features obtained from Google Earth Pro® and 

local information sources. This approach enabled manual re-selection and adjustment of misclassified 

areas based on visual interpretation and contextual knowledge of the terrain. 

Misclassifications were identified by detecting discrepancies between expected land cover 

patterns and the outputs generated by the supervised classification process. When inconsistencies 

were found—such as isolated urban pixels within vegetated zones or fragmented class boundaries, 

manual edits were made to improve spatial coherence and ensure that the maps more accurately 

represented the actual land cover distribution. These refinements were performed directly in QGIS 

3.42.3® by editing raster values or updating vector-based classification masks. 

Confusion Matrix 

Also known as an error matrix, is a widely accepted tool in remote sensing for accuracy 

assessment due to its ability to detect both commission and omission errors across individual land 

cover classes (Congalton, 1991). The matrix (Table 5) compares the predicted classes from the 

classification output (rows) with known reference classes (columns), allowing for a detailed evaluation 

of classification performance. The structure highlights both correct and incorrect classifications: 

diagonal elements (e.g., a, f, k, p) indicate correctly classified pixels where the predicted and actual 

land cover types match while off-diagonal values represent misclassifications. Row totals reflect the 

area assigned to each class by the classifier, whereas column totals represent the actual number of 

reference area per class. The total number of evaluated area is denoted as N. 



26 

 

 

Table 5 

Example of Confusion Matrix format 

Classified \ Reference Forest (F) Urban (U) Water (W) Bare Soil (B) Row Total 
Forest (F) a b c d R₁ 
Urban (U) e f g h R₂ 
Water (W) i j k l R₃ 
Bare Soil (B) m n o p R₄ 
Column Total C₁ C₂ C₃ C₄ N 

Note: Self generated with adaptations from Congalton (1991). 

In this study, the confusion matrices were generated using the Accuracy Assessment tool 

available in the SCP plugin in QGIS®. This tool allows the user to input validation data usually ground-

truth reference points and calculates several important statistical measures.  

Among them, user’s accuracy (Equation 7) indicates how often the class on the map 

represents the real-world class; it reflects the probability that an area classified into a given category 

truly belongs to that category. Producer’s accuracy (Equation 8), on the other hand, measures the 

probability that a reference area has been correctly identified by the classification; it relates to how 

well real features are represented on the map. Overall accuracy (Equation 9) expresses the proportion 

of the total number of correctly classified areas over the total number of reference areas, providing a 

general sense of map reliability. Kappa (Equation 10) adjusts the proportion of agreement between 

the classification and the ground truth (Equation 11) by subtracting the agreement that would be 

expected by random chance (Equation 12). A Kappa value of 1 indicates perfect agreement, 0 means 

no better than random, and negative values indicate worse than random. 

This coefficient is particularly useful when evaluating classification quality across multiple 

classes and is widely considered a more conservative indicator than overall accuracy alone (Foody, 

2002). 

User Accuracy UA 

𝑈𝑈𝑈𝑈 = 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑖𝑖𝑖𝑖 𝑎𝑎 𝑟𝑟𝑟𝑟𝑟𝑟
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑖𝑖𝑖𝑖 𝑡𝑡ℎ𝑎𝑎𝑎𝑎 𝑟𝑟𝑟𝑟𝑟𝑟

= 𝑎𝑎
𝑅𝑅1

, 𝑓𝑓
𝑅𝑅2

, …                                         [7] 

 



27 

 

 

Producers Accuracy PA 

𝑃𝑃𝑃𝑃 = 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑖𝑖𝑖𝑖 𝑎𝑎 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 𝑖𝑖𝑖𝑖 𝑡𝑡ℎ𝑎𝑎𝑎𝑎 𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐

= 𝑎𝑎
𝐶𝐶1

, 𝑓𝑓
𝐶𝐶2

, …                                      [8] 

 

Overall Accuracy OA 

                                                                        𝑂𝑂𝑂𝑂 = 𝑎𝑎+𝑓𝑓+𝑘𝑘+𝑝𝑝
𝑁𝑁

                                                                            [9] 

 

Kappa Coefficient 

 𝐾𝐾� = 𝑃𝑃𝑜𝑜−𝑃𝑃𝑒𝑒
1−𝑃𝑃𝑒𝑒

                                                                                 [10] 

 

Observed agreement  

𝑃𝑃𝑜𝑜 = ∑ 𝑥𝑥𝑥𝑥𝑥𝑥𝑘𝑘
𝑖𝑖=1
𝑁𝑁

                                                                                 [11] 

 

Expected agreement 

                                                                         𝑃𝑃𝑒𝑒 = ∑ (𝑅𝑅𝑖𝑖∗𝐶𝐶𝑖𝑖
𝑁𝑁2

𝑘𝑘
𝑖𝑖=1 )                                                                           [12] 

 
In the context of this study, all accuracy metrics were calculated using area-based confusion 

matrices derived from the Semi-Automatic Classification Plugin in QGIS®. This approach ensures that 

accuracy values reflect the actual spatial extent of each class, enhancing the relevance of the 

evaluation in real-world land use terms. The combination of visual inspection and confusion matrix 

analysis ensured that the final land cover maps were both spatially consistent and statistically reliable. 

Once the classifications were finalized and validated, corrected raster layers were generated to 

represent the land use and land cover categories for 2017 and 2025. These validated maps served as 

the foundation for the second phase of the study: change analysis. 

 



28 

 

 

 

Cross-Tabulation Matrix for quantification of category-level gains, losses and transitions 

The cross-tabulation matrix (Figure 7), also known as a transition matrix, was used to analyze 

land use and land cover changes between 2017 and 2025 by comparing classifications from two time 

periods. This method distinguishes areas that remained stable (persistence) from those that changed, 

allowing for the identification of specific transitions between land cover categories. It also quantifies 

key components of change, including gross gains, losses, net change, and swap, providing a more 

detailed understanding of landscape dynamics beyond overall net change. This approach was chosen 

over other change detection techniques because it accounts for both the magnitude and direction of 

land transitions, while explicitly highlighting persistence, a common and often overlooked feature in 

land cover analyses. It is particularly useful for identifying systematic patterns of change that may be 

masked by total change summaries. The method, as formalized by Pontius et al. (2004), assumes 

consistent classification schemes and sufficient accuracy across time points to ensure that observed 

changes are not the result of misclassification. One limitation is that it does not directly account for 

spatial configuration or distance of change, but it lays a solid foundation for further spatial pattern 

analysis if needed. 

Figure 7 

Cross-tabulation matrix used for land cover change analysis 

Note. Adapted from Pontius et al. (2004). 
 
Rows (Time 1 Categories): Represent the initial land cover classes at Time 1. These are the 

actual or reference categories. 



29 

 

 

Columns (Time 2 Categories): Represent the land cover classes at Time 2. These are the 

predicted categories. 

Cells (Pij): Each cell contains a value Pij, which represents the proportion (or number of 

pixels/area) of the landscape that was classified as category i at Time 1 and changed to category j at 

Time 2. 

Diagonal Values (Pii**): These values (e.g., P11, P22, P33) represent the areas that remained 

in the same land cover category from Time 1 to Time 2, indicating no change or persistence. 

Off-Diagonal Values (Pij where i ≠ j**): These values indicate transitions between different 

land cover categories. For example, P12 represents the proportion of land that changed from 

category 1 at Time 1 to category 2 at Time 2. These highlight areas of change. 

Key Components and Derived Metrics 

The matrix includes totals and derived metrics that provide further insights into the nature 

of land cover change: 

Total Time 1 (Pi+): The sum of the values in each row (Pi+) represents the total proportion 

(or area) of each land cover category at Time 1. This is calculated by summing across the row. 

Total Time 2 (P+j**): The sum of the values in each column (P+j) represents the total 

proportion (or area) of each land cover category at Time 2. This is calculated by summing down the 

column. 

Gain (P+j - Pjj**): Gain represents the net increase in the total amount of a category from 

Time 1 to Time 2. It is calculated as the column total (P+j) minus the diagonal value (Pjj). It shows 

which categories have expanded. 

Loss (Pi+ - Pii**): Loss represents the net decrease in the total amount of a category from 

Time 1 to Time 2. It is calculated as the row total (Pi+) minus the diagonal value (Pii). It shows which 

categories have shrunk. 

 



30 

 

 

Interpreting Change Performance from Pontius Matrix  

Stability Assessment (High Diagonal Values): If a particular land cover category has high values 

on the main diagonal (high Pii), it suggests that most of its area remained stable over the time period. 

This indicates a strong persistence of that land cover type. 

Transition Analysis (High Off-Diagonal Values): If a particular land cover category has high 

values in its off-diagonal cells (i.e., large Pij values where i ≠ j), this indicates that the category 

experienced significant transitions to other categories. This suggests high change dynamics. 

Gain vs. Loss Analysis: Comparing the Gain and Loss values for each category helps identify 

which land cover types are increasing (gaining) or decreasing (losing) over the study period. This 

reveals trends in landscape transformation. 

 

 

 



31 

 

 

Results and Discussion 

Spectral-Indexes: visual based analysis  

The NDVI image for 2017 (Figure 8) reveals a strong presence of healthy vegetation across 

most of Roatán, particularly in the central and western interior zones, represented by darker green 

tones indicating high NDVI values (> 0.5). The coastal periphery, especially along the southern 

shoreline and near urbanized areas such as Coxen Hole and French Harbour, shows lower NDVI values 

ranging from 0.1 to 0.3, consistent with built-up land or degraded vegetation. Sparse vegetation and 

open areas with exposed soils are visible along roads and ridge lines, forming linear patches of lower 

reflectance. 

Figure 8 

NDVI image of Roatán for 2017 for Western portion of the island  

Note. NDVI map generated from Sentinel-2 imagery in QGIS®. 

In contrast, the 2025 NDVI image (Figure 9) shows a subtle but notable shift in vegetation 

distribution. An expansion of high NDVI zones is observed in several inland areas, suggesting 

vegetative regrowth or increased canopy density, particularly in the northwest. However, there is also 

a marked increase in low-NDVI zones (red to orange), especially surrounding urban areas and 

transportation corridors. This indicates vegetation loss and land transformation in peri-urban zones, 

with previously vegetated areas being replaced or fragmented, likely due to development pressure. 



32 

 

 

Overall, the 2025 map reflects a landscape experiencing both vegetation recovery in certain natural 

areas and urban encroachment in formerly green zones. 

Figure 9 

NDVI image of Roatán for 2025 Western portion of the island  

 
Note. NDVI map generated from Sentinel-2 imagery in QGIS®. 

The NDBI images for Roatán show a noticeable increase in built-up areas from 2017 to 2025. 

The 2025 image (Figure 10) exhibits a broader distribution and greater intensity of higher NDBI values 

(represented by darker red tones), particularly concentrated around the central and western sectors 

of the island. These areas correspond to urban centers and zones of tourism-driven development, 

such as near the airport and along the coastal corridor. 



33 

 

 

Figure 10 

NDBI processed image 2025 central area from south to north shore  

Note. NDVI map generated from Sentinel-2 imagery in QGIS®. 

In contrast, the 2017 NDBI image (Figure 11) shows a lower intensity of built-up signatures, 

indicating that artificial surfaces were less prevalent or less spatially extensive at that time. The lighter 

tone ominance in the 2017 image suggests more vegetation or natural land cover in comparison to 

2025. 

Figure 11 

NDBI processed image of the year 2017 central area from south to north shore 

Note: Self made with Sentinel 2 Imagery and the Raster Calculator from QGIS 3.42.3®. 

The spatial expansion and densification of built-up areas over the eight-year period aligns with 

trends previously observed through the supervised classification and NDVI results. Notably, regions 



34 

 

 

that showed declining NDVI values in 2025 coincide with areas that present increased NDBI, 

reinforcing the interpretation that vegetation loss is being driven by urban expansion. 

This transition is particularly evident in the southern and central zones near the airport and 

urban core, where infrastructure and development have intensified. The visual evidence from NDBI 

analysis supports the overall conclusion that Roatán has experienced significant land transformation, 

characterized by increasing urbanization and a reduction in vegetated land cover between 2017 and 

2025. 

Supervised Classification Land use and Land cover Maps  

Figure 12 presents the supervised classification results of land use and land cover (LULC) for 

the island of Roatán in 2017 and 2025 using the Random Forest algorithm. The maps illustrate 

noticeable increases in artificial surfaces particularly urban fabric and industrial/commercial areas 

along the island’s western and southern coastal zones. In contrast, natural covers such as forests and 

mangroves show signs of reduction or fragmentation. This spatial trend reflects broader patterns of 

land conversion observed in previous studies across Roatán, where rapid tourism-driven urbanization 

has replaced natural habitats, often beyond protected areas (Canty et al., 2018; Tuholske et al., 2017). 



35 

 

 

Figure 12 

Supervised classification results for LULC Random Forest Classifier Algorithm maps 2017 and 2025  

Note. 2017 top image, 2025 bottom image supervised classification performed with Random Forest using SCP Plugin for QGIS 3.42.3®. 

 



36 

 

 

Figure 13 shows a focused comparison of land use and land cover (LULC) in the West Bay area 

of Roatán between 2017 (left) and 2025 (right). A significant intensification of urban fabric and 

industrial/commercial zones is evident across this high-value tourism corridor. Natural covers such as 

forests and open vegetated spaces appear increasingly fragmented or replaced. This pattern aligns 

with findings by Tuholske et al. (2017), who identified West Bay as a critical hotspot for mangrove-to-

urban transitions driven by unregulated coastal development. The spatial clustering of artificial 

surfaces in this region highlights the pressures exerted by tourism infrastructure expansion. 

Figure 13 

LULC images West Bay 2017 and 2025 

  
Note. 2017 left image, 2025 right image 

In the Coxen Hole–Sandy Bay corridor (Figure 14), substantial urban expansion is evident 

between 2017 and 2025, especially around the international port, airport, and adjacent residential 

communities. This growth reflects the spatial pressures of tourism-driven infrastructure development 

concentrated around Roatán’s primary gateway. According to the Roatan Tourism Bureau (2024), the 

island became the most visited cruise destination in Central America in 2024, receiving over 1.7 million 



37 

 

 

cruise passengers across more than 447 ship arrivals a record that underscores the island’s centrality 

in regional tourism circuits. This surge has reinforced the need for ports, transport networks, lodging, 

and associated services concentrated in and around Coxen Hole.  

Such intense and sustained growth aligns with regional patterns described by the 

Environment Department Papers (2001), who emphasize that tourism in the Caribbean frequently 

contributes to land conversion by stimulating induced settlement, increasing demand for 

infrastructure, and placing disproportionate stress on energy and water systems. Tourist-serving 

infrastructure often spreads beyond designated zones, especially in port hubs, accelerating land 

transformation in environmentally sensitive coastal corridors. The observed urbanization in Coxen 

Hole illustrates these dynamics, as the area has absorbed much of the island’s tourism-related 

development, becoming a hotspot for informal and formal urban expansion as Roatán’s cruise tourism 

continues to break arrival records.  

Figure 14  

LULC images Coxen hole, Sandy Bay, north and south shore 2017 and 2025  

Note. 2017 top image, 2025 bottom image 

 



38 

 

 

The North Shore of Roatán (Figure 15) exhibits a marked increase in built-up areas between 

2017 and 2025, particularly in zones previously dominated by open vegetation. This spatial shift 

corresponds with the expansion of key infrastructure, including the development of the landfill site, 

the installation of the RECO solar park, and widespread vegetation clearing in preparation for large-

scale tourism investments such as the planned Margaritaville resort. The construction of the coastal 

road has further accelerated land conversion along this corridor. These transformations reflect a 

broader trend toward high-intensity land use, with long-term implications for ecological stability. As 

Canty et al. (2018) emphasize, such fragmentation of forest and mangrove ecosystems undermines 

the integrity of coastal habitats and presents enforcement challenges, especially when land-use 

changes outpace legal protections or institutional oversight. 

Figure 15 

LULC images Mud hole, north shore 2017 and 2025 

 
Note. 2017 bottom image, 2025 top image 

 



39 

 

 

Classification maps for the Crawfish and Pristine Bay area from the north shore (Figure 16) 

show a notable surge in urban fabric and open areas between 2017 and 2025. This transformation 

corresponds with major development initiatives such as the ZEDE Prospera enclave, the expansion of 

high-end resorts like Pristine Bay, and real estate expansion adjacent to Coral View. Previously 

forested and vegetated lands have been increasingly fragmented, with visible conversion into 

residential and commercial land uses. 

These changes reflect the intensifying commodification of coastal land in Roatán driven by 

private investment enclaves and tourism-linked real estate speculation. As highlighted by the 

Environment Department Papers (2001), such tourism-induced land transformation imposes 

environmental pressures through increased infrastructure demands and settlement sprawl, often in 

areas lacking sufficient planning oversight. Without stricter ecological zoning and sustainable tourism 

governance, continued growth in this region risks undermining the very landscapes that attract 

investment and visitors. 



40 

 

 

Figure 16  

LULC images Crawfish, Prospera, Pristine Bay, French Key, north shore 2017 and 2025 

Note. 2017 bottom image, 2025 top image  
 

The LULC classification highlights significant land transformation in the Los Fuertes and French 

Harbour region between 2017 and 2025 (Figure 17). Built-up areas have expanded substantially, 

particularly along the southern coastline, where residential development and tourism-driven 

infrastructure have intensified. Notable changes include the construction of RECO’s solar park near 

the French Key area, which converted previously open and vegetated land into industrial 

infrastructure. In addition, the Los Fuertes community has expanded considerably, reflecting a lack of 

oversight and insufficient territorial planning—an issue commonly associated with unregulated 

tourism-driven growth in the Caribbean. Further, the progressive pavementation and extension of 

road infrastructure have facilitated urban sprawl inland, increasing pressure on surrounding 

ecosystems. 



41 

 

 

Figure 17 

LULC images Los Fuertes, French Harbour, French Key, South Shore 2017 and 2025 

 
Note. 2017 bottom image, 2025 top image 

 
The LULC analysis for José Santos Guardiola (Figure 18) reveals a contrasting trend compared 

to the western and central zones of Roatán. Between 2017 and 2025, there is evidence of moderate 

expansion of urban fabric and infrastructure, particularly along coastal corridors and road networks, 

but these changes are significantly less intense than in West Bay or Coxen Hole. Notably, there is 

visible reforestation in areas previously categorized as “open areas with little or no vegetation,” 

indicating land recovery processes or reduced urbanization pressure in certain inland sectors. This 

pattern suggests a de-intensification of urban development on the eastern side of the island, aligning 

with observations that development remains concentrated in the western zones where tourism and 

commercial hubs dominate (Tuholske et al., 2017). 



42 

 

 

While this sector has remained relatively less urbanized, the persistence of mangrove belts 

and vegetated zones highlights the ecological importance of maintaining current conservation 

strategies. (Canty et al., 2018) emphasize that these eastern ecosystems serve as critical buffers 

against environmental risks, and their preservation contrasts with the heavy transformation observed 

in Roatán’s west. 

Figure 18 

LULC images José Santos Guardiola, east side, north and south shore  

Note. 2017 top image, 2025 bottom image 

 

 



43 

 

 

This region of José Santos Guardiola (Figure 19) exhibits a markedly different pattern from the 

intense urbanization observed in the western zones like West Bay or Coxen Hole. The LULC maps (2017 

vs. 2025) reveal a subdued urban footprint only modest growth concentrated along coastal corridors 

and road networks. Simultaneously, large patches of dense vegetation persist, with pockets of 

reforestation evident where “open areas with little or no vegetation” have regenerated. These 

dynamics suggest a de-intensification trend, where development pressures exist but remain limited 

amidst broader island transformations. Yet, the incremental encroachment through smaller roads and 

scattered infrastructure underscores the vulnerability of this transitional zone to future growth. 

Echoing Tuholske et al. (2017) findings, this pattern signals the urgent need for proactive spatial 

planning to preserve ecological integrity, especially given that the eastern municipality, José Santos 

Guardiola, houses rural communities with limited governance capacity and remain significantly less 

developed than the tourism-driven western half. 

Figure 19 

LULC images Jose Santos Guardiola, Oak ridge, Port Royal National Park, St Elena 2017 and 2025 

 

 

 

 

 



44 

 

 

The St. Helene mangrove system (Figure 20), located on the eastern tip of Roatán, 

demonstrates significant ecological resilience between 2017 and 2025. The LULC classification results 

show minimal change, with dense mangrove coverage and surrounding forest remaining largely intact. 

This stability reflects its designation as part of the Bay Islands Marine National Park and its inclusion 

in the Ramsar Convention list of wetlands of international importance. Such legal frameworks have 

provided a protective buffer against the rapid urbanization and infrastructure expansion observed in 

western Roatán. The persistence of these mangroves is critical, as they serve not only as biodiversity 

hotspots but also as key buffers against coastal erosion, saltwater intrusion, and climate-driven 

impacts like storm surge. However, the analysis also underscores the importance of continued 

vigilance: although large-scale alterations are absent, incremental encroachment along coastal fringes 

or increased boat traffic linked to tourism could gradually erode these conservation gains if 

management efforts weaken. 

Figure 20  

LULC images St Helene mangrove system and St. Helene island, eastern side 2017 and 2025  

 



45 

 

 

 
Note. 2017 top image, 2025 bottom image 

 

Barbareta (Figure 21) presents a contrasting case of proactive conservation, where decades-

long private management has safeguarded its ecological integrity. The LULC results indicate a virtually 

unchanged landscape dominated by dense forest and natural vegetation between 2017 and 2025, 

with no significant evidence of urban fabric or large-scale land clearing. This outcome reflects a 

deliberate strategy of private stewardship that has prioritized environmental protection while limiting 

development pressures. Such stability stands out in comparison to other areas of Roatán, where 

unregulated expansion has fragmented habitats. According to historical management reports, this 

privately managed island has maintained an excellent state of vegetation cover for over thirty years, 

positioning it as an ideal candidate for eco-tourism projects designed under sustainability principles 

(Valade et al., 2002). The island’s combination of well-preserved ecosystems and archaeological 

findings enhances its potential for low-impact tourism that supports conservation objectives rather 

than undermining them. 

 

 

 



46 

 

 

Figure 21 

LULC images Barbareta island, east side, 2017 and 2025  

 

 

Note. 2017 top image, 2025 bottom image 

 

 

 

 

 



47 

 

 

Cross tabulation matrix: quantification of gains, losses, and transitions at category level 

Based on the results of the Pontius Matrix (Table 6) from the random forest classification LULC 

from 2017 to 2025, the total losses for each land category were the following: Urban with 146.18 ha, 

Industrial with 160.30 ha, Forest with 1,264.37 ha, Open with 1,302.47 ha, and Mangrove with 541.86 

ha. On the other hand, the total gains for each land category were the following: Urban with 196.96 

ha, Industrial with 1,262.8 ha, Forest with 880.52 ha, Open with 667.44 ha, and Mangrove with 407.46 

ha. As a result, the category that had the most losses of land by 2025 was the Open category, and the 

category that had most gains of land by 2025 was the Industrial category.  

Table 6  

LULC 2017 and 2025 Pontius Matrix Results Tabulated in hectares 

T. 2017 | T. 
2025 Urban Industrial Forest Open Mangrove Total T1 

2017 Losses 

Urban 26.59 64.15 7.95 73.69 0.39 172.77 146.18 
Industrial 17.42 156.19 32.78 44.53 65.57 316.49 160.3 
Forest 59.34 390.34 7,891.35 311.51 303.17 9,155.72 1,264.37 
Open 103.5 254.03 614.75 529.67 29.99 1,832.14 1,302.47 
Mangrove 16.7 56.28 225.04 998.43 104.96 1,540.29 5,41.86 
Total T2 
2025 223.55 1,418.99 8,771.87 1,197.11 1,405.09 13,017.41  

Gains 196.96 1,262.8 880.52 667.44 407.46   
Note. Pontius matrix showing land cover transitions for LULC classification from 2017 to 2025 generated through SCP Plug-in 

Table 7 summarizes the total extent of each land cover class for 2017 and 2025 across Roatán. 

The results reveal significant land use shifts over the eight-year period. Industrial land cover displays 

the most dramatic growth, increasing from 316.49 ha in 2017 to 1,418.99 ha in 2025—a more than 

fourfold rise—reflecting the rapid development of tourism-related infrastructure, energy projects, 

and road connectivity. This sharp growth may also be partially influenced by spectral similarity 

between urban fabric and industrial/commercial surfaces, which can cause overestimation in 

classification outputs when using pixel-based approaches. 

 



48 

 

 

Urban areas also show a noticeable increase, from 172.77 ha to 223.55 ha, highlighting the 

continued spread of settlements and service facilities linked to population growth and tourism 

activity. In contrast, natural covers declined during the same period. Forests decreased from 9,155.72 

ha to 8,771.87 ha, a net reduction of 383.85 ha, likely due to clearing for infrastructure and tourism 

development in areas such as French Harbour, Pristine Bay, and the North Shore. Open areas dropped 

from 1,832.14 ha to 1,197.11 ha, indicating that previously undeveloped or transitional land has been 

converted into built-up areas. Finally, mangrove ecosystems fell from 1,540.29 ha to 1,405.09 ha, 

although smaller than past decades, remains ecologically critical given their role as coastal buffers. 

These losses point toward intensified land conversion for tourism and residential developments, 

particularly in zones adjacent to the north shore road, French Harbour, and West Bay corridors areas 

identified as hotspots for resort and infrastructure growth. Such transitions erode natural buffers, 

reduce aquifer recharge zones, and amplify risks of coastal erosion and saltwater intrusion, as 

emphasized in regional conservation studies (Canty et al., 2018). 

Table 7 

Results from random forest classifier for each year showing total land cover area (ha) for each land 

cover class for the entire island 

Class T1 2017 (ha) T2 2025 (ha) 
Urban 172.77 223.55 
Industrial 316.49 1,418.99 
Forest 9,155.72 8,771.87 
Open 1,832.14 1,197.11 
Mangrove 1,540.29 1,405.09 

Note. Tabulated with data from Pontius Matrix 2017 2025  

 

 

 



49 

 

 

The gains, losses, and changes results (Table 8) highlight pronounced contrasts in land-use 

dynamics on Roatán between 2017 and 2025. Industrial land cover exhibited the most dramatic 

increase, with a net gain of 1,102.5 ha, representing a 348.3% change, with the highest among all 

categories. This surge underscores the rapid pace of infrastructure development tied to large-scale 

tourism and energy projects, such as the Roatan Electric Company´s solar park and resort complexes 

in Pristine Bay and the North Shore corridor, West Bay, etc. The considerable increase also reflects 

potential spectral confusion between industrial and dense urban fabric, as both surfaces share similar 

reflectance in high-albedo bands, potentially inflating the classified industrial category under the 

Random Forest algorithm. 

Urban areas also recorded a substantial percentage increase (29.4%), rising from 172.77 ha to 

223.55 ha, consistent with incremental settlement growth and road expansion in peri-urban and 

tourism-intensive zones. These patterns align with trends observed by Tuholske et al. (2017), who 

reported that urbanization on Roatán between 1985 and 2015 expanded by over 227%, strongly 

correlated with tourism-driven economic growth. 

Conversely, open areas experienced the greatest proportional decline, shrinking by 34.7% 

(from 1,832.14 ha to 1,197.11 ha). This reduction is consistent with the functional role of these areas 

as transitional landscapes often cleared in preparation for future development. Their rapid conversion 

into urban or industrial categories reflects the early stages of construction processes that precede full 

infrastructure establishment. This trend emphasizes how open lands serve as “buffers” that absorb 

the initial impacts of tourism-related expansion, as noted in regional land-change models. 

Natural covers also registered notable losses. Forest cover decreased by 4.2%, translating to 

a net loss of 383.85 ha, primarily through encroachment along road corridors and tourism enclaves. 

Mangroves declined by 8.8%, dropping from 1,540.29 ha to 1,405.09 ha. While smaller in absolute 

terms compared to forests and open areas, this reduction is ecologically significant, given the critical 

role of mangroves in coastal resilience and carbon sequestration. Tuholske et al. (2017) documented 



50 

 

 

similar dynamics, reporting annual mangrove losses of nearly 1% in unprotected areas—primarily 

linked to tourism infrastructure and settlement expansion—underscoring the vulnerability of these 

habitats despite national and international protective designations. 

Table 8  

Results for LULC Gains, losses, and changes 

Class Gains (ha) Losses(ha) Net change (ha) % Change 
Urban 196.96 146.18 50.78 29.3916768 
Industrial 1,262.8 160.3 1,102.5 348.3522386 
Forest 880.52 1,264.37 -383.85 -4.192461106 
Open 667.44 1,302.47 -635.03 -34.66056087 
Mangrove 407.46 541.86 -135.2 -8.777567861 

Note. Tabulated from Pontius Matrix 2017 and 2025 

Table 9 details the primary land cover transitions contributing to urbanization and artificial 

surface expansion between 2017 and 2025. The largest single contributor is forest-to-artificial surface 

conversion, totaling 449.68 ha, indicating that forested areas remain the primary source for 

infrastructure and settlement expansion. This reflects patterns observed in tourism-driven 

development corridors such as West Bay, Coxen Hole, and French Harbour, where road connectivity 

and resort construction have fragmented natural habitats. 

Open areas to artificial surfaces account for 357.53 ha, underscoring the role of open or 

sparsely vegetated lands as transitional zones that are often cleared first in preparation for 

construction. This aligns with the substantial 34.6% decline in open areas shown in previous results, 

confirming their function as buffers in the development process. 

Although smaller in magnitude, mangrove-to-artificial surface transitions amount to 72.98 ha, 

which is ecologically significant given the limited extent and critical ecosystem services provided by 

mangroves. This reinforces concerns raised by Rull (2023), who discussed that coastal urbanization is 

a key driver of mangrove loss in Roatán and the wider Caribbean. 

Additionally, the matrix reveals a noteworthy forest-to-open conversion of 614.75 ha, which 

often precedes full urbanization. This suggests a two-stage clearing process: first reducing forest to 



51 

 

 

open land, then transitioning to artificial surfaces. Combined, these changes contribute to 880.19 ha 

of new artificial surfaces and 644.74 ha of open area gains, reflecting ongoing landscape restructuring 

toward built-up land. 

Table 9 

Transition trends towards urbanization 

Class Area (ha) 
Forest to As 449.68 
Open to As 357.53 
Mangrove to As 72.98 
Forest to Open 614.75 
Mangrove to Open 29.99 
Total Artifical Surface Gain 880.19 
Total Open areas Gain 644.74 

Note. Information tabulated from Pontius Mtrix LULC results 

Classification Accuracy Assessment Using Confusion Matrices 

The performance of the supervised land use/land cover (LULC) classifications for the years 

2017 and 2025 was evaluated using area-based confusion matrices generated through the Semi-

Automatic Classification Plugin (SCP) in QGIS®. These matrices compare the classified output with a 

set of reference data and report various accuracy metrics, including Overall Accuracy (OA), Producer’s 

Accuracy (PA), User’s Accuracy (UA), and the Kappa coefficient Congalton (1991). The results confirm 

the effectiveness of the Random Forest classifier (RFM) in achieving high reliability in LULC mapping, 

even in complex coastal environments like Roatán. 

For the 2017 classification (Table 10), the model achieved an Overall Accuracy of 97.23% and 

a Kappa coefficient of 0.9413, indicating a very high agreement between the classified map and the 

reference data. Class-wise performance showed excellent reliability for Forests (PA = 99.56%, UA = 

99.95%), Open Spaces with Little or No Vegetation (PA = 99.71%, UA = 88.65%), and Mangroves (PA = 

100%, UA = 96.32%). Industrial, Commercial, and Transport Units also performed well (PA = 75.07%, 

UA = 91.16%). However, Urban Fabric exhibited the weakest performance with a Producer’s Accuracy 



52 

 

 

of 33.45% and User’s Accuracy of 63.70%, suggesting a significant degree of confusion with other 

developed land types or sparse vegetation. 

Table 10 

2017 Random Forest Classifier Accuracy Results  

Class Urban Fabric Industrial Forest Open Mangrove 
PA [%] 33.4451 75.0676 99.5604 99.706 100 
UA [%] 63.7011 91.1628 99.9486 88.651 96.3185 
Kappa hat 0.6276 0.9089 0.9982 0.8703 0.9584 

Overall accuracy [%] = 97.2343 
Kappa hat classification = 0.9413 

Note. Generated through precision assesment available in SCP Plug in for QGIS® 3.42 

For the 2025 classification (Table11), the Overall Accuracy was slightly lower at 94.41%, with 

a Kappa coefficient of 0.8896. The classifier continued to perform strongly for Forests (PA = 98.11%, 

UA = 100%), Open Spaces (PA = 99.77%, UA = 63.88%), and Mangroves (PA = 99.84%, UA = 88.48%). 

Industrial, Commercial, and Transport Units showed improvement in both PA (77.24%) and UA 

(97.08%). However, Urban Fabric again remained the most problematic class, with a PA of 47.02% and 

UA of 59.05%, reflecting persistent spectral confusion—likely with surrounding built-up or bare areas. 

Table 11 

2025 Random Forest Classifier Accuracy Results  

Class Urban Fabric Industrial Forest Open Mangrove 
PA [%] 47.0242 77.2411 98.111 99.7704 99.8416 
UA [%] 59.0551 97.0793 100 63.8835 88.4832 
Kappa hat 0.5815 0.9662 1 0.6162 0.8726 

Overall accuracy [%] = 94.4132 
Kappa hat classification = 0.8896 

Note. Generated through precision assesment available in SCP Plug in for QGIS® 3.42 

The high accuracy achieved across most classes reaffirms the suitability of the Random Forest 

classifier in remote sensing applications, particularly for coastal and heterogeneous landscapes. As 

demonstrated by Rodriguez-Galiano et al. (2012), RF is a robust ensemble learning method that 

consistently outperforms traditional classifiers in terms of accuracy, resistance to overfitting, and 

tolerance to noisy or reduced training datasets. Their study reported an average Kappa coefficient of 



53 

 

 

0.92 in complex Mediterranean landscapes using multiseasonal Landsat imagery, validating the 

reliability of RF-based approaches in diverse environments. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

  



54 

 

 

Conclusions 

To compare and analyze the Land Use and Land Cover (LULC) change between 2017 and 2025, 

remote sensing and GIS-based techniques were applied to five different land use categories: Urban, 

Industrial, Forest, Open, and Mangrove. A clear difference between development and encroachment 

from areas previously vegetated to open and artificial surfaces. The categories that lost the most land 

coverage from 2017 to 2025 were Open areas with little or no vegetation, with 1,302.47 ha, and Forest 

with 1,264.37 ha. Both categories suffered great loss of their land for different purposes, like 

industrialization and urbanization. 

The categories that had the most gains for land coverage from 2017 to 2025 were Industrial 

with 1,262.8 ha, and Forest with 880.52 ha. Both categories were positively affected in terms of having 

more land by 2025. The tendency for LULC was mostly inclined toward high expansion for the 

Industrial category and a great decrease for the Open category, followed by the Forest category, 

showing that industrialization had more presence on Roatán by 2025 having a negative impact over 

Open and Forest categories.  

These observed land cover changes provide a baseline for evaluating how future land use 

decisions may affect ecological functions across the island. The loss of natural covers, particularly 

Open and Forest categories, suggests reduced buffering capacity and highlights the importance of 

integrated spatial planning. Continued monitoring of these patterns can support efforts to strengthen 

land governance and improve long-term resilience in developing zones. 

 

 

 

 

 

 

 
 



55 

 

 

Recommendations 

Several recommendations are proposed to improve future land use and land cover (LULC) 

analyses in Roatán and similar island environments. First, it is essential to ensure a balanced and well-

distributed selection of training data (ROIs) across all classes. Overrepresentation of certain 

categories, such as artificial surfaces, can lead to model bias and reduce classification accuracy. Future 

efforts should prioritize achieving representative sampling that captures the heterogeneity of land 

cover types on the island. 

The integration of higher-resolution or multisensor data is also recommended. Sentinel-2 

imagery provided a solid foundation, but limitations in spatial resolution made it difficult to distinguish 

between spectrally similar classes, such as dense vegetation and urban infrastructure. Incorporating 

higher-resolution optical imagery (e.g., Maxar) may improve class separability and enhance overall 

map quality. Additionally, the application of spectral indices such as NDVI, NDBI prior to classification 

can serve as effective filters to delineate vegetation, built-up areas, and water bodies, while also 

offering valuable insights into ecological trends such as vegetation recovery and urban expansion. 

Optimizing classification parameters within the Random Forest model such as increasing the 

number of trees, enabling balanced class weights, and applying cross-validation can significantly 

improve accuracy and generalization. The classification algorithms complexity is limited by processing 

power and could have a more precise classification if the hardware allows it. Furthermore, the use of 

post-classification comparison methods, including the Pontius matrix, proved useful in analyzing 

category-level transitions and should be standard in future change detection studies. When possible, 

the inclusion of field-based validation or participatory mapping could strengthen classification 

accuracy and build local capacity for environmental monitoring. 

Lastly, the observed changes in land cover patterns underscore the need for more proactive 

land management. It is recommended that municipal authorities use the outputs of this study to 

inform sustainable zoning regulations, manage urban growth, and protect remaining forested and 



56 

 

 

wetland areas. Expanding the temporal scope of future studies such as, using more frequent imagery 

to monitor seasonal trends would also contribute to a more dynamic understanding of Roatán’s 

rapidly changing landscape and directly relate different factors involved in these trends.  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 



57 

 

 

References 

Akbar, T. A., Hassan, Q. K., Ishaq, S., Batool, M., Butt, H. J., & Jabbar, H. (2019). Investigative Spatial 
Distribution and Modelling of Existing and Future Urban Land Changes and Its Impact on 
Urbanization and Economy. Remote Sensing, 11(2), 105. https://doi.org/10.3390/rs11020105 

Alqurashi, A. F., & Kumar, L. (2013). Investigating the Use of Remote Sensing and GIS Techniques to 
Detect Land Use and Land Cover Change: A Review. Advances in Remote Sensing, 02(02), 193–
204. https://doi.org/10.4236/ars.2013.22022 

Azhari, M., Alaoui, A., Achraoui, Z., Ettaki, B., & Zerouaoui, J. (Eds.) (2019). Adaptation of the random 
forest method. https://dl.acm.org/doi/proceedings/10.1145/3368756 

Burke, L. M., & Maidens, J. (2004). Reefs at risk in the Caribbean. World Resources Institute. 
http://pdf.wri.org/reefs_caribbean_full.pdf  

Canty, S. W., Preziosi, R. F., & Rowntree, J. K. (2018). Dichotomy of mangrove management: A review 
of research and policy in the Mesoamerican reef region. Ocean & Coastal Management, 157, 
40–49. https://doi.org/10.1016/j.ocecoaman.2018.02.011 

Carlson, R. R., Foo, S. A., & Asner, G. P. (2019). Land Use Impacts on Coral Reef Health: A Ridge-to-Reef 
Perspective. Frontiers in Marine Science, 6, Article 562. 
https://doi.org/10.3389/fmars.2019.00562 

Congalton, R. G. (1991). A review of assessing the accuracy of classifications of remotely sensed data. 
Remote Sensing of Environment, 37(1), 35–46. https://doi.org/10.1016/0034-4257(91)90048-
B 

Da Silva, V. S., Salami, G., Da Silva, M. I. O., Silva, E. A., Monteiro Junior, J. J., & Alba, E. (2020). 
Methodological evaluation of vegetation indexes in land use and land cover (LULC) 
classification. Geology, Ecology, and Landscapes, 4(2), 159–169. 
https://doi.org/10.1080/24749508.2019.1608409 

Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V., Gascon, F., Hoersch, B., Isola, C., 
Laberinti, P., Martimort, P., Meygret, A., Spoto, F., Sy, O., Marchese, F., & Bargellini, P. (2012). 
Sentinel-2: ESA's Optical High-Resolution Mission for GMES Operational Services. Remote 
Sensing of Environment, 120, 25–36. https://doi.org/10.1016/j.rse.2011.11.026 

Environment Department Papers. (2001). Tourism and environment in the Caribbean: An Economic 
Framework. World Bank. Environmental Economic Series. 
https://www.researchgate.net/publication/255619697_Tourism_and_the_Environment_in_
the_Caribbean 

European Environment Agency. (2019). Corine Land Cover 2018 (vector), Europe, 6-yearly - version 
2020_20u1, May 2020. https://doi.org/10.2909/71c95a07-e296-44fc-b22b-415f42acfdf0 

European Space Agency Lab (Ed.). (2019). A Machine Learning glossary. https://philab.esa.int/a-
machine-learning-glossary/ 

Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote Sensing of 
Environment, 80(1), 185–201. https://doi.org/10.1016/S0034-4257(01)00295-4 



58 

 

 

Guan, D., Li, H., Inohae, T., Su, W., Nagaie, T., & Hokao, K. (2011). Modeling urban land use change by 
the integration of cellular automaton and Markov model. Ecological Modelling, 222(20-22), 
3761–3772. https://doi.org/10.1016/j.ecolmodel.2011.09.009 

Hemati, M., Hasanlou, M., Mahdianpari, M., & Mohammadimanesh, F. (2021). A Systematic Review 
of Landsat Data for Change Detection Applications: 50 Years of Monitoring the Earth. Remote 
Sensing, 13(15), 2869. https://doi.org/10.3390/rs13152869 

Ho, T. K. (1995). Random decision forests. In Tin Kam Ho (Ed.), Proceedings of 3rd International 
Conference on Document Analysis and Recognition (pp. 278–282). IEEE Comput. Soc. Press. 
https://doi.org/10.1109/ICDAR.1995.598994 

Hou, W., Wang, J., Xu, X., Reid, J. S., & Han, D. (2016). An algorithm for hyperspectral remote sensing 
of aerosols: 1. Development of theoretical framework. Journal of Quantitative Spectroscopy 
and Radiative Transfer, 178, 400–415. https://doi.org/10.1016/j.jqsrt.2016.01.019 

Instituto de Conservacion Forestal. (n.d). Geoportal. Retrieved August 6, 2025, from 
https://geoportal.icf.gob.hn/geoportal/main 

Iurist, N., Stătescu, F., & Lateş, I. (2016). Analysis of Land Cover and Land Use Changes Using Sentinel-
2 Images. Present Environment and Sustainable Development, 10(2), 161–172. 
https://doi.org/10.1515/pesd-2016-0034 

McBirney, A. R., & Bass, M. N. (1969). Geology of Bay Islands, Gulf of Honduras. 
https://www.semanticscholar.org/paper/Geology-of-Bay-Islands%2C-Gulf-of-Honduras-
McBirney-Bass/c191f607f83d6bf299153a411094845610bf602c 

Mcfield, M., Kramer, P., Álvarez, F., Drysdale, I., Rueda Flores, M., Giró Petersen, A., & Soto, M. 
(2018). 2018 Mesoamerican Reef Report Card. https://doi.org/10.13140/RG.2.2.19679.36005 

Pontius, R. G., Shusas, E., & McEachern, M. (2004). Detecting important categorical land changes 
while accounting for persistence. Agriculture, Ecosystems & Environment, 101(2-3), 251–268. 
https://doi.org/10.1016/j.agee.2003.09.008 

Roatan Tourism Bureau. (2024). Roatan: The Most Visited Cruise Destination in Central America in 
2024. https://roatantourismbureau.com/community-updates/roatans-record-cruise-arrivals-
2024 

Rodriguez-Galiano, V. F., Ghimire, B., Rogan, J., Chica-Olmo, M., & Rigol-Sanchez, J. P. (2012). An 
assessment of the effectiveness of a random forest classifier for land-cover classification. 
ISPRS Journal of Photogrammetry and Remote Sensing, 67, 93–104. 
https://doi.org/10.1016/j.isprsjprs.2011.11.002 

Rull, V. (2023). Rise and fall of Caribbean mangroves. The Science of the Total Environment, 885, 
163851. https://doi.org/10.1016/j.scitotenv.2023.163851 

Sohail, U. (2020). Analysis the Potential of Vegetation Indices (NDVI) for Land Use / Cover Classification 
in Karachi by Landsat 8 Data. International Journal of Biology and Biotechnology, 17(2), 359–
366. https://www.researchgate.net/publication/353909568 

Stonich, S. C. (1998). Political ecology of tourism. Annals of Tourism Research, 25(1), 25–54. 
https://doi.org/10.1016/S0160-7383(97)00037-6 



59 

 

 

Sutton, D. (2015). Structural and Geophysical Analysis on Roatan Island, Honduras, Western 
Caribbean. 
https://www.researchgate.net/publication/282669831_STRUCTURAL_AND_GEOPHYSICAL_
ANALYSIS_ON_ROATAN_ISLAND_HONDURAS_WESTERN_CARIBBEAN 

Tin Kam Ho (Ed.) (1995). Proceedings of 3rd International Conference on Document Analysis and 
Recognition. IEEE Comput. Soc. Press. https://doi.org/10.1109/ICDAR.1995.598994 

Tomczyk, T. (2010). Nature's refuage at Port royal: A 31-year effort of creating land protected area on 
the east side of Roatan, begins to bear fruit. Bay Island Voice. 
http://www.bayislandvoice.com/ 

Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoring vegetation. 
Remote Sensing of Environment, 8(2), 127–150. https://doi.org/10.1016/0034-
4257(79)90013-0 

Tuholske, C., Tane, Z., López-Carr, D., Roberts, D., & Cassels, S. (2017). Thirty years of land use/cover 
change in the Caribbean: Assessing the relationship between urbanization and mangrove loss 
in Roatán, Honduras. Applied Geography, 88, 84–93. 
https://doi.org/10.1016/j.apgeog.2017.08.018 

Valade, I., Grelot, J., & Laurent, J. M. (2002). Proyecto manejo ambiental de las Islas de la Bahia. 
https://rsis.ramsar.org/RISapp/files/36545686/documents/HN2456_ECD190603.pdf 

Zha, Y., Gao, J., & Ni, S. (2003). Use of normalized difference built-up index in automatically mapping 
urban areas from TM imagery. International Journal of Remote Sensing, 24(3), 583–594. 
https://doi.org/10.1080/01431160304987 

Zhang, C., & Li, X. (2022). Land Use and Land Cover Mapping in the Era of Big Data. Land, 11(10), 1692. 
https://doi.org/10.3390/land11101692 

 

 

 

 

 

 

 

 

 

 

 



60 

 

 

Appendices 

Appendix A 

Satellite Imagery for Roatan Island 2017 and 2025 from Coppernicus Data Hub 

 

 

 



61 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 



62 

 

 

Appendix B 

Pontius matrix for NDVI index classification from 2017 to 2025 (in hectares) 

T. 2017 | 
T. 2025 

Water Artificial 
surface 

Open 
spaces 

with little 
or no 

vegetation 

Sparse 
vegetation 

Dense 
vegetation 

Total T1 
2017 

Losses 

Water 135.48 11.85 1.7 0.18 0.06 149.27 13.79 

Artificial 
surface 

146.24 185.57 76.79 19.69 14.5 442.79 257.22 

Open 
spaces 
with little 
or no 
vegetation 

6.77 99.6 228.61 155.13 207.52 697.63 469.02 

Sparse 
vegetation 

2.92 32.87 113.19 188.36 715.83 1,053.17 864.81 

Dense 
vegetation 

5.83 70.41 147.59 236.38 10,214.47 10,674.68 460.21 

Total T2 
2025 

297.24 400.3 567.88 599.74 11,152.38 13,017.54  

Gains 161.76 214.73 339.27 411.88 937.91   

Note. NDVI Pontius matrix showing land cover transitions and area gains/losses between 2017 and 2025. 

 

 

 

 

 

 

 

 

 



63 

 

 

Appendix C 

2017 and 2025 Pontius Matrix 

T. 2017 | 
T. 2025 

Urban Industrial Forest Open Mangrove Total T1 
2017 

Losses 

Urban 26.59 64.15 7.95 73.69 0.39 172.77 146.18 
Industrial 17.42 156.19 32.78 44.53 65.57 316.49 160.3 
Forest 59.34 390.34 7,891.35 311.51 303.17 9,155.72 1,264.37 
Open 103.5 254.03 614.75 529.67 29.99 1,832.14 1,302.47 
Mangrove 16.7 56.28 225.04 998.43 104.96 1,540.29 541.86 
Total T2 
2025 

223.55 1,418.99 8,771.87 1,197.11 1,405.09 13,017.41  

Gains 196.96 1262.8 880.52 667.44 407.46   
Note. Pontius matrix showing land cover transitions for LULC classification from 2017 to 2025. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 



64 

 

 

Appendix D  

High Res Comparison Catalog, LULC Results and Google Earth Pro® 

 

 

 

 

LULC 2017 vs 2025
West bay, West Side

High resolution images from Google Earth
Pro: 2017 and 2023



65 

 

 

 

 

LULC 2017 vs 2025 East Side: Port Royal National Park and Camp Bay
High resolution images from Google Earth Pro: 2016 and 2023 

LULC 2017 vs 2025 Mud Hole, North Shore High resolution images from Google
Earth Pro: 2017 and 2024 



66 

 

 

 

    

 

 

 

 

LULC 2017 vs 2025 Los Fuertes, French Harbour, French Key, South Shore High resolution images from Google Earth Pro: 2016 and 2023 

LULC Crawfish, Prospera, French Key, North Shore
High resolutionimages from
GoogleEarth Pro: 2016 and
2023 



67 

 

  

LULC 2017 vs 2025 Jose Santos Guardiola
High resolution images from Google Earth Pro: 2016 and 2023 

LULC 2017 VS 2025 Santa Helena High resolution images from Google Earth Pro: 2016 and 2024

LULC 2017 vs 2025 Barbarretta
High resolution images from Google Earth Pro: 2019 and 2024


	List of Tables
	List of Figures
	List of Appendices
	Abstract
	Introduction
	Methodology
	Study Area
	Roatán Municipality (Western Region)
	José Santos Guardiola Municipality (Eastern Region)

	Study Design
	Spectral Index-Based Visual Analysis
	NDVI
	NDBI

	Supervised Land Use and Cover Classification
	Pre-processing phase of Supervised classification of LULC Maps
	Processing phase for Supervised Classification of LULC Maps
	Postprocessing for Supervised Classification of LULC Maps
	Confusion Matrix

	Cross-Tabulation Matrix for quantification of category-level gains, losses and transitions
	Interpreting Change Performance from Pontius Matrix


	Results and Discussion
	Spectral-Indexes: visual based analysis
	Supervised Classification Land use and Land cover Maps
	Cross tabulation matrix: quantification of gains, losses, and transitions at category level
	Classification Accuracy Assessment Using Confusion Matrices

	Conclusions
	Recommendations
	References
	Appendices