Z. Cervantes, E. Morales, C. López, and I. Reyes,

“Evaluating the accuracy of manual classiﬁcation in satellite images using supervised algorithms”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 1, 2026.

Evaluating the accuracy

of manual classification

in satellite images using

supervised algorithms

ARTICLE HISTORY

Received 10 October 2025

Accepted 20 November 2025

Published 6 January 2026

Zulema Yamileth Cervantez Hernández

Universidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

zS21023196@estudiantes.uv.mx

ORCID: 0009-0000-9453-2582

Emmanuel Morales García

Universidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

emmorales@uv.mx

ORCID: 0000-0002-6837-9227

Cecilia Cruz López

Universidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

ceccruz@uv.mx

ORCID: 0000-0002-9156-5669

Itzel Alessandra Reyes Flores

Uniersidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

itreyes@uv.mx

ORCID: 0000-0003-0733-8453

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

This work is licensed under a Creative Commons

Attribution-NonCommercial-ShareAlike 4.0 International License.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 14

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January - June 2026

Evaluating the accuracy of manual classification in

satellite images using supervised algorithms

Zulema Yamileth Cervantez

Hernández

Universidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

zS21023196@estudiantes.uv.mx

Itzel Alessandra Reyes Flores

Uniersidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

itreyes@uv.mx

Emmanuel Morales García

Universidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

emmorales@uv.mx

Cecilia Cruz López

Universidad Veracruzana

Facultad de Estadística e Informática

Xalapa, Veracruz

ceccruz@uv.mx

Abstract— This research focused on evaluating the manual

classification of land cover using Sentinel-2 imagery. Supervised

algorithms were applied to validate and improve this process. Three

algorithms were selected based on their computational efficiency:

K-Nearest Neighbors (KNN), Random Forest (RF), and Support

Vector Machine (SVM). The results show that KNN achieved

optimal performance, demonstrating a solid balance between

accuracy, F1 score, and execution time compared to RF. RF, for its

part, obtained greater accuracy, indicating its superior ability to

correctly identify classes; however, it requires more computational

resources. SVM exhibited lower performance in the evaluated

metrics but achieved a shorter execution time. It was identified as

the algorithm with the greatest limitations for separating classes

within this dataset derived from the different study areas. Overall,

the comparison confirmed that the manual classifications developed

in QGIS are supported and validated by the application of these

supervised methods. The use of such algorithms contributes to

improving the accuracy, consistency, and efficiency of geospatial

classification tasks.

Keywords— Remote sensing, Machine learning, Infrared

imaging, Data science

I. INTRODUCTION

Analyzing satellite images offers benefits such as the ability

to visualize spatiotemporal patterns of the Earth, the

environment, and climate change. This type of study enables

the monitoring and understanding of these processes, leading

to greater precision [1][2],[3]. Another advantage of studying

images, such as multispectral images, is that they provide

more information about the Earth's surface and vegetation.

Currently, the Sentinel-2 remote sensor of the European

Space Agency (ESA) provides the most detailed images of

the Earth from space [4].

As noted in [5], the diversity of missions carried out by

Sentinel-2 within the Copernicus program is highly relevant

to Earth observation. Its advanced design and capabilities

have driven significant progress in land cover monitoring,

precision agriculture, natural disaster management, and

ecosystem studies. Sentinel-2 has two satellites (2A and 2B)

equipped with the Multispectral Instrument (MSI), which

captures images in thirteen spectral bands ranging from

visible to shortwave infrared. Its spatial resolution ranges

from ten to sixty meters, and its swath width is 290 km

[6],[7].

The characteristics of the remote sensor allow for continuous

and highly detailed monitoring of study areas. However,

using traditional classification methods presents

complications in terms of reproducibility and efficiency,

especially when applied to heterogeneous areas [8].

Furthermore, the spectral and spatial benefits provided by

Sentinel-2 generate a large volume of data that require

processing with advanced methods such as machine learning

or image processing techniques. In this context, supervised

algorithms have broad and optimal potential for classifying

information from satellite images. Some models used are

Support Vector Machines (SVM), Random Forests (RF), and

K-Nearest Neighbors (KNN). These algorithms can process

spectral data and identify patterns in images [9],[10],[11].

These algorithms perform well; however, their accuracy

varies depending on environmental conditions and the land

cover being studied.

Remote sensing and machine learning combined make efforts

to transform this type of data into information for decision-

making on environmental and territorial issues [12],[13].

Currently, the literature explores the processes, development,

and application of supervised algorithms in satellite image

classification. A widely cited study is that of [14], who

conducted a thorough analysis of the application of RF in

satellite image processing, demonstrating its optimal capacity

for processing large volumes of data and its robustness

against overfitting.

This background highlights the relevance of employing

algorithms such as RF, using satellite images from Sentinel-

2. To complement this information, [15] conducted a study

on the application process of remote sensing techniques,

analyzing six commonly used algorithms, such as SVM and

RF. They conclude that these types of algorithms offer the

capacity to solve high-dimensional problems. This

background reinforces the idea that comparing methods is

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 15

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

Z. Cervantes, E. Morales, C. López, and I. Reyes,

“Evaluating the accuracy of manual classification in satellite images using supervised algorithms”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 1, 2026.

beneficial for verifying operational performance in real

geospatial scenarios, as proposed in this research conducted

in Veracruz, Mexico. Based on the above, the objective of

this research is to identify which supervised machine learning

algorithms achieve the best accuracy in classifying areas in

satellite imagery. This involves evaluating the classification

of Sentinel-2 satellite images in various geospatial scenarios

in Veracruz, Mexico.

The analysis compares the results of manual classification

with those obtained using SVM, RF, and KNN in three

distinct regions: the city of Xalapa, Pico de Orizaba, and

Cofre de Perote. These areas were selected for their unique

environmental characteristics, including urban centers,

bodies of water, diverse vegetation types, and high-altitude

snow-capped mountains, allowing for a direct evaluation of

the algorithms' performance under different topographic and

ecological conditions.

II. PROBLEM STATEMENT

One of the main challenges in satellite image analysis is

the accurate classification of land cover, particularly in

heterogeneous areas characterized by varying vegetation

density and expanding urban zones. Examples of such

environments include Xalapa, Cofre de Perote, and Pico de

Orizaba in the state of Veracruz, Mexico.

Traditional (or manual) classification methods often

present obstacles when attempting to differentiate spectrally

similar classes, such as dense and sparse vegetation, urban

areas, or arid zones, especially when using medium-resolution

imagery. These limitations can compromise the reliability of

studies focused on land-use monitoring or urban planning.

Therefore, it is essential to conduct post-hoc evaluations using

supervised algorithms to assess the accuracy of classifications

obtained through manual methods (e.g., manual class

selection in QGIS). Supervised algorithms help determine

whether classes were assigned correctly, thus strengthening

the results through greater accuracy and better generalization

across complex landscapes.

III. RELATED WORKS

In this study, the authors [16] identified and classified

greenhouses in the Anamur region of Mersin, Turkey, using

Sentinel-2 MSI medium-resolution and SPOT-7 high-

resolution images. This research focused on object-based

image analysis (OBIA) using KNN, RF, and SVM algorithms

to evaluate which of the different methods and sensors is most

effective for greenhouse classification. Multispectral images

taken on August 2, 2018, were used, along with field data and

visual validation. The methodology consisted of several

stages, including atmospheric correction for images with

cloud cover, segmentation of the Sentinel-2 MSI images

using the ESP-2 tool, extraction of spectral and textural

features, as well as the NDVI and NDWI indices, and finally,

the application and comparison of the algorithms. The results

indicate that the most accurate methods were KNN and RF

with SPOT-7 images, achieving an overall accuracy of

91.43% and a Kappa coefficient of 0.88. On the other hand,

KNN was the best classified greenhouses in Sentinel-2 MSI

images, as it had the highest accuracy (88.38%) and a Kappa

coefficient of 0.83. In conclusion, both sensors demonstrated

good effectiveness for greenhouse classification, despite

having different resolutions. Furthermore, KNN and RF

proved to be the most accurate methods.

In another study [17], the performance of Random Forest

(RF), Support Vector Machine (SVM), and a combination of

both algorithms, known as Stack, was evaluated for

classifying satellite images in rural and urban areas of

Bangladesh, specifically in the Bhola and Dhaka regions.

Images from Landsat-8, Sentinel-2, and Planet satellites were

used to determine which sensor and algorithm combination

offers the greatest accuracy for detecting land use and land

cover changes (LULC) in areas considered fragmented.

Geometric and atmospheric corrections were applied, a

training dataset was created for each region, and the objects

were classified using RStudio software. This classification

yielded six classes for Bhola: water bodies, tree vegetation,

rainfed agriculture, wetland agriculture, fallow land, and

urbanized or swampy areas. In Dhaka, only five landform

classes were identified: water bodies, tree cover, grassland or

agricultural land, urbanized areas, and landfills. The analysis

showed that the Sentinel-2 sensor and SVM were the most

accurate and highest-performing in both study areas, with an

accuracy of 0.969 in Bhola and 0.983 in Dhaka, and Kappa

coefficients of 0.948 and 0.968, respectively. However, SVM

performed better in classifying water and vegetation-related

categories, while RF and Stack were more effective at

distinguishing urbanized areas and landfills. This study

concluded that Sentinel-2 is well-suited for classifying

different areas at a small scale and that SVM offers better

results when the dataset is limited.

This study [18] compared Support Vector Machine (SVM),

Random Forest (RF), and Classification and Regression Tree

(CART) algorithms with the Google Earth Engine (GEE)

platform to map and analyze land cover changes over Lake

Urmian in Iran from 2000 to 2020. A total of 55 satellite

images from Landsat 5, 7, and 8 were used over time.

Additionally, 20,000 training points were obtained from

previous field studies and Google Earth maps, while the

control set consisted of 6,000 ground points for validation.

Before processing, the images were filtered to detect cloud

cover, as this could lead to misclassification. Subsequently,

the different algorithms were applied and evaluated using the

confusion matrix, the Kappa coefficient, and the global

accuracy metric. These algorithms were also subjected to a

spatial uncertainty analysis using Dempster-Shafer theory

(DST) with the Idrisi program. The results of this research

showed that the classified maps detected a reduction in the

lake's surface area of approximately 1000 to 1400 hectares,

and an increase in agricultural and urban areas. Furthermore,

SVM proved to be the most efficient algorithm for multi-

temporal classification with an accuracy of 92–95%,

followed by RF (82–87%) and CART (63–70%).

This research evaluated and compared the Random Forest

(RF), K-Nearest Neighbors (KNN), and Gaussian Mixture

Model (GMM) algorithms for generating urban land cover

maps of Quezon City in the Philippines [19]. The aim was to

determine which algorithms are the most accurate in

classifying urban and non-urbanized areas, as well as to

monitor urban sprawl in the city. For this purpose, pre-

corrected Sentinel-2A satellite images downloaded on July 1,

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 16

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January - June 2026

2021, were used. A training set was created with 70% of the

images manually labeled and 30% for validation. Spectral

bands with a resolution of 10 and 20 meters were also used.

Image processing was performed using QGIS software with

the Semi-Automatic Classification (SCP) and Dzetsaka

plugins. Three machine learning algorithms were applied and

evaluated using confusion matrices, producer accuracy, user

accuracy, and overall accuracy. The results indicated that the

algorithms achieved high classification accuracy, with RF

showing the highest at 99.32%, followed by KNN at 98.05%,

and GMM at 97.17%. However, execution times varied: RF

took 7 minutes, KNN approximately 2.5 minutes, and GMM

less than 5 seconds. In conclusion, the applied methods

proved effective for classifying urban areas, with RF

exhibiting the highest accuracy, while GMM stood out for its

faster data processing.

The potential of satellite imagery from Sentinel-1 with its

SAR sensor, Sentinel-2 in multispectral mode, and the

combination of both sensors was also studied using Linear

Regression (LR), Classification and Regression Trees

(CART), and Random Forest (RF) algorithms, along with

airborne LiDAR data. This was done to identify which

datasets best model canopy height in the Atlantic Forest of

Paraná, Brazil [20]. Additionally, raw, ind, and all features

were extracted from the satellite combinations. These

methods were evaluated using Mean Absolute Error (MAE),

Root Mean Square Error (RMSE), and R² metrics with

training and test data samples, using R software. The findings

showed that Sentinel-2 and the sensor combination are the

most suitable for modeling canopy height. While Random

Forests performed better than the other algorithms, achieving

an RMSE of 4.92 m and an R² of 0.58 using only Sentinel-1

data, and an RMSE of 4.86 m and an R² of 0.60 using

Sentinel-2 data, Sentinel-2 data demonstrates good accuracy

in estimating canopy height.

In this study, the performance of Support Vector Machine

(SVM), Decision Tree (DT), Random Forest (RF), K-Nearest

Neighbors (KNN), and Naive Bayes was compared using

high-resolution satellite imagery from Cartosat-2E, Cartosat-

3, and LISS-4 over the Jaipur area of India [21]. The objective

of this study was to compare two types of classification

techniques, object-based and pixel-based—to determine

which method offers better results in land use and land cover

(LULC) classification. The data used in the study were

multispectral and orthorectified images downloaded from the

National Remote Sensing Agency (NRSC), which were

classified using QGIS and the Orfeo Toolbox. The images

also underwent a segmentation process, separability indices

were calculated between the different classes, and the

algorithms were compared using accuracy, recovery, F1

score, and Kappa coefficient. According to the results, for the

object-based technique, the algorithm with the highest

accuracy was Decision Trees with a Kappa coefficient of

0.90. In contrast, with the pixel-based method, both KNN and

SVM performed best, especially with images related to the

Cartosat-2E and Cartosat-3 sensors.

Finally, for this research, spectral indices (NDWI, MNDWI,

AWEI_SH, AWEI_NSH, AWEI_BOTH) and machine

learning algorithms (SVM, RT, MLC, KNN) were evaluated

to detect water surfaces in Sentinel-2 images [22]. The aim

was to determine which method offers the best classification

accuracy using a pixel-based approach, as well as to identify

image characteristics that could cause erroneous

classification results. For this purpose, images from the Red

River, Sylvia Grinnell, Rivière-Rouge, and Fraser Rivers—

areas with rivers located in Canada—were used. High-

resolution images (PLÉIADES, WorldView-2, TripleSat, and

KOMPSAT-3) were also used as a validation dataset. The

images were corrected for cloud cover issues, and spectral

indices were obtained. The threshold values for the

algorithms were manually adjusted and evaluated using the

Critical Success Index (CSI) and the Root Mean Square Error

(RMSE) to estimate river width. The results showed that the

AWEI_NSH index and the SVM algorithm performed best

for classification across all study areas, demonstrating greater

consistency and accuracy. Furthermore, the study revealed

that features increasing the likelihood of misclassification are

most closely related to environmental factors such as

vegetation, urban infrastructure, water turbidity, and shadows

cast by objects.

IV. METHODOLOGY

A methodological framework was designed to guide

the development of this research and to achieve the stated

objec- tive (Figure 1).

Fig. 1: Methodological process for this research

A. Identification of the study area

For the selection of the study areas, regions with

diverse environmental scenarios—such as vegetation,

urban zones, water bodies, and barren land—were

considered in order to include a greater variety of

classes. Three regions within the state of Veracruz were

selected: Xalapa (urban area), Pico de Orizaba (protected

natural area), and Cofre de Per- ote (national park). This

environmental heterogeneity makes these areas ideal for

assessing the accuracy of land-cover classification

methods.

B. Satellite image

The images used in this study were obtained from

the Copernicus platform via the Sentinel-2 satellite. The

selected period, from February to April, was chosen to

ensure favorable weather conditions for satellite

observation characterized by lower precipitation and minimal

cloud coverthus guaranteeing higher image quality. It is

impor- tant to note that the input images were obtained in RGB

format.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 17

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

Z. Cervantes, E. Morales, C. López, and I. Reyes,

“Evaluating the accuracy of manual classification in satellite images using supervised algorithms”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 1, 2026.

C. Analysis tools

For the initial process—manual classification of the image

classes—QGIS version 3.40.3 was used. This stage involved

area delimitation, user-assigned classifications, and the cre-

ation of a database containing pixels and their corresponding

classes. For the supervised classification algorithms, Python

was employed, as it allows efficient handling of large data

volumes.

D. Image processing

In this phase, the images were preprocessed to correct

atmospheric noise, cloud cover, and shadows. These

corrections were performed using QGIS. The objective of this

step was to optimize the accuracy of spectral identification. It

is worth noting that the true-color (RGB) images were

converted into false-color composites using the B08, B04, and

B03 bands, as the B08 band enhances vegetation, facilitating

classification in each image. Following this, spectral features

were obtained from the B2, B3, B4, B8, B11, and B12 bands,

chosen for their ability to distinguish land cover.

E. Manual classification of images

Manual classification was performed, in which the user

assigns classes based on predefined criteria image content,

identifying various types of land cover in each study area. A

total of six classes were defined: urban areas, water bodies,

dense vegetation, sparse vegetation, no vegetation, and snow.

Five classes were identified in Xalapa and Pico de Orizaba,

while six categories were present in Cofre de Perote. After

classification, a comma-delimited (CSV) file containing the

pixels of each image and their corresponding class was

exported.

F. Supervised algorithms

RF is a supervised algorithm that combines multiple

decision trees to generate a more robust model. Developed by

American statistician Leo Breiman in 2001, it is based on the

principle of bagging (bootstrap aggregation), where multiple

decision trees are built and trained with random subsets of data

and features. This approach reduces correlation between trees

and improves model extension. At each node, a random subset

of variables is selected to determine the optimal split,

increasing the diversity of the forest. Final predictions are

derived from the results of all trees. While each individual tree

represents a weak classifier, their combination results in a

robust, stable, and low-variance ensemble model [23], [24].

KNN is based on the principle that a new data point can be

classified or predicted by analyzing its k nearest neighbors

within the feature space. The algorithm's performance

depends heavily on the distance metric used and the

appropriate selection of the hyperparameter k. In

classification, the class of a new data point is determined by

the average of its nearest neighbors, while in regression, the

predicted value corresponds to the average of the neighbors'

outputs. It is important to note that selecting an appropriate

value for k is crucial, as choosing a value that is too high can

increase the prediction error and negatively impact the model's

performance [25].

SVM is used in classification and prediction. Its main

function is to find the optimal hyperplane that separates data

points of different classes with the maximum margin; that is,

the greatest possible distance between the hyperplane and the

nearest data points of each class, known as support vectors.

SVM combines the maximum margin principle, which

improves the model's generalizability, with the kernel method,

which allows the algorithm to handle nonlinearly separable

data by projecting it onto a higher-dimensional feature space

where linear separation becomes possible [26], [27].

G. Characteristics of the models and their process

This section describes how the classes used to generate the

classification were heterogeneously distributed, which

directly influenced the model performance. The category

with the most data was vegetation, followed by urban areas,

and then the remaining classes.

Hyperparameters:

• For RF, 200 trees were used.

• For KNN, k=5 and Euclidean distance were used.

• For the SVM, an RBF kernel with C=1.0 and

gamma=0.1 was used.

For all three methods, cross-validation was used to

strengthen the results; a k-fold = 5 was performed.

H. Metrics for the evaluation of algorithms

Confusion Matrix

Table I shows the characteristics of a confusion matrix.

TABLE I. Example of the confusion matrix

Prediction

Positive

Negative

Observation

Positive

True Positive (TP)

False Positive (FP)

Negative

False Negative (FN)

True Negative (TN)

Where:

•

TP: the model correctly predicts a positive case.

•

TN: the model correctly predicts a negative case.

•

FP: the model predicts positive when it is negative

(Type I error).

•

FN: the model predicts negative when it is positive

(Type II error).

Other important metrics:

• F1 score: An F1 score close to 1 indicates a good model,

while a value of 0 indicates poor predictive

performance [29].

• Accuracy: Measures the proportion of positive cases

correctly predicted [28].

• Recall: Identifies the proportion of actual positive cases

[28].

• ROC curve: Helps with the overall model evaluation; it

is expressed by the AUC (area under the curve). A

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 18

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January - June 2026

value close to 1 represents a good model [30].

V. RESULTS

A. . Images analyzed

The main results of this research are presented below,

covering the process from image acquisition to the evaluation

of manual classifications. To perform the classification,

Sentinel-2 satellite images were cropped to delimit the study

area. The images used correspond to true color composites

(RGB with bands B4, B3, and B2). Within the delimited

areas, objects in each image were manually classified using

QGIS.

Subsequently, the original image was transformed into a

false-color composite by incorporating the near-infrared band

in place of one of the visible bands. This approach highlights

vegetation and other elements with higher reflectance in the

infrared spectrum.

Fig. 2. Study areas cut out and classified

Figure 2 presents the set of satellite images corresponding to

the three study areas: Pico de Orizaba, Xalapa, and Cofre de

Perote. The top row displays the color composites of each

area, highlighting differences in land cover: vegetation

(reddish tones), urban areas (bluish tones), and sparse

vegetation (gray tones). It is important to note that these

images were generated using infrared bands to enhance the

accuracy of manual classification.

The bottom row shows the final classification results. This

representation enables a clearer comparison of the spatial

distribution of vegetation cover versus urban land area in

each of the study areas. Table II below summarizes the

classifications produced in QGIS.

TABLE II. Land Cover Classes and Coding

Color

Land cover

classes

Coding

Urban area

UA (0)

Water bodies

WB (1)

Dense

vegetation

DV (2)

Sparse

Vegetation

SV (3)

No vegetation

NV (4)

Snow

S (5)

Table II presents the classifications obtained manually in

QGIS. Six general groups were defined: Urban Zone (UZ),

representing built-up areas; Water Bodies (CA), correspond-

ing to rivers, lakes, or dams; Dense Vegetation (VD), associ-

ated with areas of high vegetation cover; Sparse Vegetation

(PV), referring to areas with intermediate or degraded cover;

No Vegetation (NV), representing arid surfaces; and Snow

(N), comprising areas covered by ice and snow. This last

category appears only in the images of Pico de Orizaba and

Cofre de Perote.

B. Evaluation of manual classification

The database used to run the algorithms corresponds to the

classified pixels from the three satellite images (Xalapa:

352,338; Pico de Orizaba: 572,412; Cofre de Perote: 359,827

classified pixels in each image). These datasets were exported

from QGIS in CSV format. Each record represents a pixel

with its spectral values per band, along with its coordinates

and the assigned class label (e.g., Urban Area, Water Bodies,

Snow, etc.).

For algorithm training and evaluation, each pixel set was

divided into two subsets: 80\% of the data were used for

training, and the remaining 20\% for testing. Subsequently,

5-fold cross-validation was applied to reduce bias and

increase the robustness of the algorithm results. The

outcomes are summarized in Table III.

TABLE III. Classification Metrics by Zone and Algorithm

Zones

Metrics

SVM

KNN

Xalapa

Accuracy

0.9808

0.9293

0.9811

F1 Score

0.9808

0.9288

0.9811

Pico de

Orizaba

Accuracy

0.9806

0.7254

0.9776

F1 Score

0.9805

0.7151

0.9776

Cofre de

Perote

Accuracy

0.9874

0.8855

0.9828

F1 Score

0.9873

0.8746

0.9827

The evaluation of the metrics shows that, for the

classifications of the Xalapa image, both Random Forest (RF)

and KNN achieved nearly identical accuracy, with an F1

score of 0.98, indicating optimal and balanced performance.

In contrast, SVM produced more variable results (0.70–0.93),

reflecting a higher number of classification errors compared

to the other algorithms. For Pico de Orizaba and Cofre de

Perote, RF and KNN also achieved higher accuracy, while

SVM still provided acceptable performance.

These results were obtained through 5-fold cross-validation,

which reduces overfitting and provides more reliable

performance estimates. The execution times of each

algorithm for the study areas are shown below.

Fig. 3: execution time in algorithms

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 19

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

Z. Cervantes, E. Morales, C. López, and I. Reyes,

“Evaluating the accuracy of manual classification in satellite images using supervised algorithms”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 1, 2026.

Another key factor in evaluating the algorithms was

execution time. Random Forest (RF) demonstrated consistent

and robust performance across the previously discussed

metrics.

However, it was also the algorithm with the highest

computational cost for image classification, with execution

times ranging from 13 to 30 minutes. In contrast, KNN and

SVM required substantially less time (between 1 and 5

minutes), making them more practical when the goal is to

accelerate classification and reduce processing time (Figure

3).

Overall, these results indicate that KNN provides a more

favorable balance between efficiency and accuracy,

supporting its use as a reliable complement to manual image

evaluation. Conversely, RF remains a strong alternative when

the priority is to maximize classification accuracy, regardless

of execution time.

C. Visualization of the confusion matrix and ROC curve

Below are some of the results obtained from the different

classification algorithms. Figures 4, 5, and 6 illustrate the

performance of Random Forest in Xalapa, SVM in Pico de

Orizaba, and KNN in Cofre de Perote. Although all three

algorithms were applied to each study area, only

representative cases—ranging from lower to higher

classification performance—are presented.

Fig.4. ROC curve and confusion matrix for the classification

of Xalapa (RF)

In the case of Xalapa, the Random Forest (RF) model

achieved an AUC value close to 1, indicating an outstanding

ability to distinguish between classes. The confusion matrix

further confirms this accuracy, as most observations are

concentrated along the main diagonal with minimal

classification errors. These results demonstrate the robustness

of the algorithm in a heterogeneous urban–vegetation

environment (Figure 4).

Fig.5. ROC curve and confusion matrix for the classification

of Pico de Orizaba (SVM)

Another example is Pico de Orizaba, where the SVM

algorithm exhibited more variable performance. Some

classes, such as Urban Areas and Sparse Vegetation, achieved

AUC values greater than 0.99, whereas other classes,

including Dense Vegetation and No Vegetation, showed

values between 0.80 and 0.88, indicating difficulties in

discriminating among land-cover types. The confusion

matrix, in turn, reveals notable misclassifications between

Dense Vegetation and other land-cover types, as well as

between No Vegetation and Sparse Vegetation (Figure 5).

Fig.5. ROC curve and confusion matrix for the classification

of Cofre de Perote (KNN)

Finally, for the Cofre de Perote image, the KNN algorithm

achieved high AUC values, and the confusion matrix shows

that, despite the overall accuracy, some misclassifications

occurred between areas such as Dense Vegetation and No

Vegetation (Figure 6).

VI. DISCUSSION

According to the results obtained in this research, supervised

algorithms are essential for evaluating manual land cover

classifications derived from satellite imagery. Consistent

with the findings of [16], [19], and [20], the RF algorithm

demonstrated the highest accuracy in distinguishing the

classes created during the experimental phase. Its robustness,

ability to handle spectral variability, and efficiency in

managing complex pixel data confirm its excellent

performance in validating image classifications in

heterogeneous regions such as Xalapa, Cofre de Perote, and

Pico de Orizaba, areas characterized by urban zones, dense

vegetation, and bodies of water.

The K-Nearest Neighbors (KNN) algorithm also performed

well, yielding results comparable to those of RF. However,

the main difference lies in the execution time, with KNN

being considerably faster. As noted in [19], KNN stands out

for its computational efficiency, which was also confirmed in

the experimental phase of this study. Therefore, this

algorithm can be considered an optimal alternative when

speed and accuracy are required.

Conversely, the Support Vector Machine (SVM) algorithm

showed lower accuracy in this study compared to the other

two classifiers, unlike the results reported by [17] and [18],

where the SVM achieved superior performance. This

discrepancy can be attributed to the distinctive characteristics

of the study areas in this research (high variability, cloud

cover, and dense vegetation), which can reduce the spectral

separation of the categories, thus improving the performance

of the SVM algorithm, as well as the high spatial variability.

Finally, compared to the other two algorithms used, SVM is

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026 20

DOI:

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

https://doi.org/10.33333/lajc.vol13n1.01

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January - June 2026

governed by a linear function and is therefore sensitive to a

lack of homogeneity, unlike RF and KNN.

VII. CONCLUSION

The utility of supervised algorithms allows for the validation

and reinforcement of classifications performed manually in

QGIS, corroborating their objectivity and computational

efficiency to support decision-making in land cover analysis.

While user-assigned classifications depend on their judgment

and experience (supported by the QGIS software), supervised

algorithms provide a quantitative and reproducible

framework that minimizes subjectivity in the process.

Therefore, the use of RF, KNN, and SVM not only produces

accurate results but also offers a scientific basis for

confirming visually delineated boundaries, thus improving

the reliability of the resulting classifications and their

applicability in remote sensing studies.

REFERENCES

[1]

Ouchra, H., Belangour, A., & Erraissi, A. (2023). Machine

learning algorithms for satellite image classification using

Google Earth Engine and Landsat satellite data: Morocco case

study. IEEE Access, 11, 71127–71142.

https://doi.org/10.1109/ACCESS.2023.3293828.

[2]

Wang, Y., Sun, Y., Cao, X., Wang, Y., Zhang, W., and Cheng,

X. (2023). A review of regional and Global scale Land Use/Land

Cover (LULC) mapping products generated from satellite

remote sensing. ISPRS Journal of Photogrammetry and Remote

Sensing, 206, 311-334.

[3]

Joshi, P., Saxena, P., & Sharma, A. (2024). Trends and

Advancements in Satellite Image Classification: From

Traditional Methods to Machine Learning Approaches.

ResearchGate. 10.13140/RG.2.2.13239.23201.

[4]

European Space Agency. (n.d.). Introducing Sentinel-2. ESA.

https://www.esa.int/Applications/Observing the

Earth/Copernicus/Sentinel- 2/Introducing Sentinel-2.

[5]

Phiri, D., Simwanda, M., Salekin, S., Nyirenda, V. R.,

Murayama, Y., and Ranagalage, M. (2020). Sentinel-2 data for

land cover/use mapping: A review. Remote sensing, 12(14),

2291.

[6]

Segarra, J., Buchaillot, M. L., Araus, J. L., & Kefauver, S. C.

(2020). Remote sensing for precision agriculture: Sentinel-2

improved features and applications. Agronomy, 10(5), 641.

[7]

Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez, V.,

Gascon, F., ... & Bargellini, P. (2012). Sentinel-2: ESA’s optical

high- resolution mission for GMES operational services.

Remote sensing of Environment, 120, 25-36.

[8]

Fragoso-Campo´n, L., QUIRO´ S, R. E., & GUTIE´RREZ, G. J. (2020).

Clasificacio´n supervisada de ima´genes PNOA-NIR y fusio´n con

datos LiDAR-PNOA como apoyo en el inventario forestal. Caso

de estu- dio: Dehesas. CUADERNOS DE LA SOCIEDAD

ESPAN˜ OLA DE

CIENCIAS FORESTALES: Sociedad Espanola de Ciencias

Forestales, 45(3), 77-96.

[9]

Zhang, C., Liu, Y., & Tie, N. (2023). Forest land resource

information acquisition with sentinel-2 image utilizing support

vector machine, k- nearest neighbor, random forest, decision

trees and multi-layer percep- tron. Forests, 14(2), 254.

[10]

Yuh, Y. G., Tracz, W., Matthews, H. D., & Turner, S. E. (2023).

Application of machine learning approaches for land cover

monitoring in northern Cameroon. Ecological informatics, 74,

101955.

[11]

Thanh Noi, P., & Kappas, M. (2017). Comparison of random

forest, k- nearest neighbor, and support vector machine

classifiers for land cover classification using Sentinel-2

imagery. Sensors, 18(1), 18.

[12]

Knudby, A. (2021). Accuracy assessment. Pressbooks.

https://ecampusontario.pressbooks.pub/remotesensing/chapter/ch

apter- 7-accuracy-assessment/

[13]

Tehsin, S., Kausar, S., Jameel, A., Humayun, M., & Almofarreh, D.

K. (2023). Satellite image categorization using scalable deep

learning. Applied Sciences, 13(8), 5108.

https://doi.org/10.3390/app13085108.

[14]

Belgiu, M., & Dragut, L. (2016). Random forest in remote

sensing: A review of applications and future directions. ISPRS

Journal of Photogrammetry and Remote Sensing, 114, 24–31.

[15]

Maxwell, A. E., Warner, T. A. & Fang, F. (2018).

”Implementation of machine-learning classification in remote

sensing: An applied review.” International Journal of Remote

Sensing, 39(9), 2784-2817.

[16]

Balcik, F. B., Senel, G., & Goksel, C. (2020). Object-Based

Classifi- cation of Greenhouses Using Sentinel-2 MSI and

SPOT-7 Images: A Case Study from Anamur (Mersin), Turkey.

IEEE Journal Of Selected Topics In Applied Earth Observations

And Remote Sensing, 13, 2769- 2777.

https://doi.org/10.1109/jstars.2020.2996315

[17]

Rahman, A., Abdullah, H. M., Tanzir, M. T., Hossain, M. J.,

Khan, B. M., Miah, M. G., & Islam, I. (2020). Performance of

different machine learning algorithms on satellite image

classification in rural and urban setup. Remote Sensing

Applications Society And Environment, 20, 100410.

https://doi.org/10.1016/j.rsase.2020.100410

[18]

Feizizadeh, B., Omarzadeh, D., Garajeh, M. K., Lakes, T., &

Blaschke, T. (2021). Machine learning data-driven approaches

for land use/cover mapping and trend analysis using Google

Earth Engine. Journal Of Environmental Planning And

Management, 66(3), 665-697.

https://doi.org/10.1080/09640568.2021.2001317

[19]

Santiago, R. M., Gustilo, R., Arada, G., Magsino, E., & Sybingco,

E. (2021). Performance Analysis of Machine Learning

Algorithms in Generating Urban Land Cover Map of Quezon

City, Philippines Using Sentinel-2 Satellite Imagery. 2021 IEEE

13th International Conference on Humanoid, Nanotechnology,

Information Technology, Communi- cation and Control,

Environment, and Management (HNICEM), 1–6.

https://doi.org/10.1109/hnicem54116.2021.9731856

[20]

Torres, C., Jéssica Gerente, Rodrigo, Caruso, F., Providelo,

L. A., Marchiori, G., & Chen, X. (2022). Canopy Height

Map- ping by Sentinel 1 and 2 Satellite Images, Airborne

LiDAR data and machine learning. Remote Sensing, 14(16),

4112-4112. https://doi.org/10.3390/rs14164112.

[21]

Kumar, A., Kumar, G., Patil, D. S., &Gupta, R. (2024). Evaluating

Machine Learning Classifiers for IRS High Resolution Satellite

Images Using Object-Based and Pixel-Based Classification

Techniques. Journal Of The Indian Society Of Remote Sensing.

https://doi.org/10.1007/s12524-024-02084-w.

[22]

Purwanto, A. D., Wikantika, K., Deliar, A., & Darmawan, S. (2022).

Decision Tree and Random Forest Classification Algorithms for

Mangrove Forest Mapping in Sembilang National Park, Indonesia.

Remote Sensing, 15(1), 16. https://doi.org/10.3390/rs15010016.

[23]

Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5–32.

https://doi.org/10.1023/a:1010933404324.

[24]

Farhadi, Z., Bevrani, H., Feizi-Derakhshi, M., Kim, W., & Ijaz, M. F.

(2022). An Ensemble Framework to Improve the Accuracy of

Prediction Using Clustered Random-Forest and Shrinkage Methods.

Applied Sciences, 12(20), 10608.

https://doi.org/10.3390/app122010608.

[25]

Tamamadin, M., Lee, C., Kee, S., \& Yee, J. (2022). Regional

Typhoon Track Prediction Using Ensemble k-Nearest Neighbor

Machine Learning in the GIS Environment. Remote Sensing, 14(21),

5292. https://doi.org/10.3390/rs14215292.

[26]

Adugna, T., Xu, W., & Fan, J. (2022b). Comparison of Random Forest

and Support Vector Machine Classifiers for Regional Land Cover

Mapping Using Coarse Resolution FY-3C Images. Remote Sensing,

14(3), 574. https://doi.org/10.3390/rs14030574.

[27]

Du, K., Jiang, B., Lu, J., Hua, J., & Swamy, M. N. S. (2024b).

Exploring Kernel Machines and Support Vector Machines: Principles,

Techniques, and Future Directions. Mathematics, 12(24), 3935.

https://doi.org/10.3390/math12243935.

[28]

S. Visa, B. Ramsay, A. Ralescu, and E. Van Der Knaap, “Confusion

matrix-based feature selection,” CEUR Workshop Proc., vol. 710, pp.

120-127, 2011.

[29]

Pikabea, I. (2022). Review of learning models in the field of question

answering [Undergraduate thesis, University of the Basque Country,

Faculty of Informatics]. University of the Basque Country.

[30]

Nahm, F. (2022). Receiver Operating Characteristic Curve: Overview

and Practical Use for Clinicians [Epub 2022 Jan 18]. Korean Journal

of Anesthesiology, 75 (1), 25-36. https://doi.org/10.4097/kja.21209

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

AUTHORS

Estudiante de octavo semestre de la Licenciatura en Estadística en la

Facultad de Estadística e Informática de la Universidad Veracruzana. Su

formación académica se ha orientado al análisis de datos, modelación

estadística y computación aplicada, con un interés particular en el uso

de tecnologías emergentes para la solución de problemas complejos.

Actualmente desarrolla su tesis de investigación en el área de

percepción remota, enfocándose en el procesamiento de imágenes

satelitales, Machine Learning y técnicas de detección de objetos.

Su trabajo integra métodos estadísticos con algoritmos de visión

por computadora para la identiﬁcación automatizada de patrones

espaciales, con aplicaciones potenciales en monitoreo ambiental

y análisis territorial. Ha participado en actividades académicas

relacionadas con programación en Python y R, así como en cursos

sobre minería de datos, aprendizaje supervisado y análisis espacial. Su

formación combina fundamentos teóricos sólidos con competencias

prácticas que fortalecen su perﬁl como futura profesionista en ciencia

de datos y estadística aplicada.

Licenciado en Ciencias y Técnicas Estadísticas y Especialista en

Métodos Estadísticos por la Universidad Veracruzana, con Maestría

en Ciencias de la Información Geoespacial por el Centro Geo, CDMX

(Centro CONAHCYT), actualmente cursando el Doctorado en Ciencias

de la Computación en la Universidad Veracruzana. Profesor en la

Licenciatura en Estadística, en la Especialidad en Métodos Estadísticos

y la Maestría en Economía y Sociedad de China y América Latina en

la misma universidad, donde he dirigido 15 tesis de licenciatura y 4 de

especialidad. También tengo experiencia como Analista Estadístico

en la Oﬁcina del Programa de Gobierno del Estado de Veracruz.

Mis líneas de investigación incluyen metodologías de cómputo,

programación estadística, estadística multivariada, análisis espacial,

ciencia de datos y modelos estadísticos, con aplicaciones en biología,

medicina, ciencias administrativas y sociales. He participado en

diversos congresos nacionales e internacionales.

Zulema Yamileth Cervantez

Emmanuel Morales García

Z. Cervantes, E. Morales, C. López, and I. Reyes,

“Evaluating the accuracy of manual classiﬁcation in satellite images using supervised algorithms”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 1, 2026.

ISSN:1390-9266 e-ISSN:1390-9134 LAJC 2026

LATIN-AMERICAN JOURNAL OF COMPUTING (LAJC), Vol XIII, Issue 1, January 2026

AUTHORS

Profesora de tiempo completo en la Facultad de Estadística e

Informática de la Universidad Veracruzana (UV). Es Doctora en

Investigación Educativa por la UV, Maestra en Ciencias con especialidad

en Estadística Aplicada por el ITESM Campus Monterrey, así como

Especialista en Métodos Estadísticos y Licenciada en Estadística

por la UV. Actualmente coordina la Especialización en Métodos

Estadísticos y pertenece al Sistema Nacional de Investigadores (SNII)

como Candidata.

Sus líneas de investigación abarcan la Educación Estadística, la

Metodología Estadística Aplicada y la integración de la estadística con

tecnologías emergentes como Machine Learning, Ciencia de Datos y

Análisis Espacial. Ha colaborado en proyectos sobre sustentabilidad,

alfabetización digital y actitudes hacia la estadística en estudiantes

latinoamericanos.

Entre sus publicaciones recientes destacan trabajos sobre aprendizaje

supervisado, formación en consultoría estadística y aplicaciones

multivariantes, además de capítulos sobre alfabetización digital y

redes sociales. Ha dirigido más de 70 tesis, combinando docencia,

investigación e impulso al uso estratégico de la estadística para el

desarrollo sostenible.

Cecilia Cruz López

Z. Cervantes, E. Morales, C. López, and I. Reyes,

“Evaluating the accuracy of manual classiﬁcation in satellite images using supervised algorithms”,

Latin-American Journal of Computing (LAJC), vol. 13, no. 1, 2026.

Es Licenciada en Informática, Maestra en Sistemas Interactivos

Centrados en el Usuario y Doctora en Ciencias de la Computación, con

formación completa en la Universidad Veracruzana. Se desempeña

como Técnica Académica de Tiempo Completo en la Facultad de

Estadística e Informática de la UV, donde colabora en actividades

de docencia, investigación, acompañamiento académico, gestión

académica y apoyo en el análisis de procesos administrativos de

la universidad. Posee experiencia sólida en desarrollo de software,

diseño de interfaces de usuario y experiencia de usuario (UX/UI),

diseño de cursos e-learning y desarrollo de aplicaciones móviles,

además de contar con certiﬁcaciones en programación de software

que fortalecen su práctica profesional y amplían su campo de acción.

Es autora y coautora de artículos de investigación en Interacción

Humano-Computadora, Inteligencia Artiﬁcial, Trabajo Colaborativo

Asistido por Computadora y Tecnología Educativa, contribuyendo

de manera signiﬁcativa al desarrollo académico y al avance del

conocimiento en estas áreas.

Itzel Alessandra Reyes Flores