Integrating Satellite Images Segmentation and Electrical Infrastructure

Data to Identify Possible Urban Irregularities in Power Grid

Alisson Alves

1,2 a

, Lu

ısa Souza

1,2 b

, Luiz Cho-Luck

1 c

, Raniere Lima

1 d

, Carlos Augusto

1 e

Wesley Marinho

1,2 f

, Rafael Capuano

1 g

, Bruno Costa

1 h

, Marina Siqueira

1 i

, Jesa

ıas Silva

1 j

Raul Paradeda

3 k

and Pablo Javier Alsina

2 l

SENAI Institute of Innovation in Renewable Energy, Capit

ao-Mor Gouveia Avenue, Natal-RN, Brazil

Graduate Program in Electrical and Computer Engineering, Federal University of Rio Grande do Norte, Natal-RN, Brazil

Department of Computer Science, State University of Rio Grande do Norte, Dr. Jo

ao Medeiros Filho Avenue, Natal-RN,

Brazil

Keywords:

Electrical Infrastructure, Semantic Segmentation, Deep Learning, Remote Sensing, Geospatial Analysis,

Land Use Classiﬁcation, Infrastructure Mapping.

Abstract:

Managing urban expansion and its impact on electrical infrastructure presents signiﬁcant challenges, necessi-

tating innovative methodologies to address irregular settlements and commercial losses in the electricity sector.

This paper proposes an approach integrating convolutional neural networks and geospatial data to detect ur-

ban areas lacking electrical infrastructure. High-resolution Google Earth images and low-resolution Landsat

8 data were processed using advanced semantic segmentation architectures, LinkNetB7 and D-LinkNet50, to

analyze land use patterns. The segmentation outputs were combined with data from the Brazilian Geographic

Database of the Distribution System to generate comprehensive maps of electrical infrastructure coverage.

The study focused on the SBAU substation in Sabar

a, Minas Gerais, which demonstrated commercial losses

of up to 47.5% in speciﬁc feeders. Results demonstrated the effectiveness of deep learning models in identi-

fying mismatches between urban development and infrastructure coverage, highlighting areas with potential

irregular connections. This study contributes to advancing artiﬁcial intelligence applications in urban energy

management by providing a scalable framework for analyzing land use and electrical infrastructure.

1 INTRODUCTION

In Brazil, electricity consumption increased by 12.6%

between 2014 and 2023, reaching an all-time high of

616.3 TWh in 2023, according to the Brazilian En-

ergy Research Company (EPE, 2024). The indus-

trial (36.4%) and residential (27.5%) sectors domi-

https://orcid.org/0000-0002-7999-8189

https://orcid.org/0000-0002-3707-2097

https://orcid.org/0009-0003-7053-0453

https://orcid.org/0009-0006-2677-2983

https://orcid.org/0009-0000-7708-348X

https://orcid.org/0009-0009-7101-1632

https://orcid.org/0009-0004-9627-8177

https://orcid.org/0009-0003-7980-7332

https://orcid.org/0009-0005-2446-8753

https://orcid.org/0000-0002-6586-8340

https://orcid.org/0000-0002-4031-6275

https://orcid.org/0000-0002-2882-5237

nate the country’s energy consumption proﬁle, par-

ticularly in the Southeast, which accounts for nearly

half of the national demand (EPE, 2024; Flores et al.,

2023). This rising demand underscores the impor-

tance of efﬁcient infrastructure planning in densely

populated regions. Additionally, sectors such as live-

stock, though accounting for only 5% of total elec-

tricity consumption, signiﬁcantly inﬂuence land use

and land cover changes (LULC) (Flores et al., 2023,

p. 22).

Understanding LULC changes is critical for ad-

dressing the impacts of urbanization on electrical in-

frastructure. Advances in remote sensing and deep

learning have proven instrumental in this domain, par-

ticularly in the segmentation of satellite images to

map urban and peri-urban areas (Archana and Jee-

varaj, 2024; Abujayyab et al., 2023). Convolutional

neural network (CNN) architectures excel at process-

ing multi-resolution geospatial data, enabling the de-

tection of urban patterns with high precision, even

Alves, Á., Souza, L., Cho-Luck, L., Lima, R., Augusto, C., Marinho, W., Capuano, R., Costa, B., Siqueira, M., Silva, J., Paradeda, R. and Alsina, P. J.

Integrating Satellite Images Segmentation and Electr ical Infrastructure Data to Identify Possible Urban Irregularities in Power Grid.

DOI: 10.5220/0013437500003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 929-936

ISBN: 978-989-758-749-8; ISSN: 2184-4992

929

under challenging conditions of spatial and temporal

variation (Atas¸, 2023).

This study builds on these advances by integrat-

ing remote sensing data with georeferenced elec-

trical infrastructure datasets, speciﬁcally the Brazil-

ian Geographic Database of the Distribution Sys-

tem (BGDDS). By leveraging high-resolution Google

Earth imagery and low-resolution Landsat 8 data,

combined with advanced CNN-based segmentation

architectures, this work identiﬁes discrepancies be-

tween urban land use and electrical infrastructure cov-

erage. These insights aim to support urban energy

planning and mitigate commercial losses in Brazil’s

electricity sector.

2 RELATED WORKS

This section presents the main related works that uti-

lize machine learning and deep learning approaches

to evaluate land use and land cover in urban contexts,

highlighting their advances, challenges, and contri-

butions in processing and analyzing remote sensing

data.

Khan and Sudheer (2022) applied a combination

of Artiﬁcial Neural Networks and Cellular Automata

(ANN-CA) to model urban growth in Islamabad, Pak-

istan, and project future land use changes. The study

utilized Landsat image data from 1991 to 2021 to as-

sess the increase in urbanized areas and predict their

future expansion by 2041. The approach highlighted

the unplanned growth of the city, emphasizing the

importance of policy interventions to mitigate nega-

tive environmental impacts, such as the loss of green

spaces.

The extraction of LULC class features was also

addressed by Rani et al. (2024), who utilized the U-

Net model to perform semantic segmentation of satel-

lite images. The study demonstrated the effectiveness

of U-Net for extracting features of buildings, vegeta-

tion, and water bodies from high-resolution images,

achieving an F1 score of 92%. The application of

this automated technique enables rapid classiﬁcation

of large datasets, reducing the need for human inter-

vention and enhancing the efﬁciency of land use anal-

ysis.

The study by He et al. (2024) focused on the de-

tailed classiﬁcation of urban buildings using a deep

neural network for low-resolution satellite images.

The research proposed UB-FineNet, a model that

addresses the class imbalance problem by applying

high-resolution techniques and contrastive learning.

This approach provides a robust solution for ﬁne-

grained building classiﬁcation in large urban areas,

supporting urban planning and resource management.

Ullah et al. (2024) investigated the impact of land

use changes on land surface temperature (LST) in

Kabul, Afghanistan. Using a combination of sup-

port vector machines (SVM) and cellular automata

with logistic regression (CA-LR), the study demon-

strated a direct correlation between urbanization and

rising air temperatures. Projections for 2046 indi-

cate an increase in urbanized areas and, consequently,

higher surface temperatures, highlighting the adverse

effects of unplanned urban expansion on the local at-

mosphere.

Although numerous studies apply machine learn-

ing techniques for LULC prediction, urban detection,

and environmental impact analysis, little research ex-

plores the relationship between unplanned urban de-

velopment and its impact on electrical infrastructure.

Issues such as illegal connections and irregular set-

tlements, exacerbated by deeply rooted cultural prac-

tices, compromise the efﬁciency, safety, and planning

of the electrical system. In this context, the use of

machine learning to identify urbanized areas, analyze

energy demands and predict potential anomalies in

consumption within these regions can offer valuable

solutions for sustainable urban planning.

3 METHODOLOGY

The methodology employs deep learning techniques

to analyze urban growth through satellite image seg-

mentation, identifying land use and land cover areas.

The process involved acquiring and preprocessing im-

ages and binary masks, training segmentation algo-

rithms with labeled data, and analyzing urban density

based on the BGDDS. Additionally, an energy bal-

ance framework was used to evaluate non-technical

losses (NTL) by region and detect anomalies in feeder

measurements.

3.1 Datasets Used

3.1.1 Satellite Sensors and Image Preprocessing

To examine the impact of different resolution levels

on land use segmentation, images of varying reso-

lutions were obtained. A Python script automated

the download of high-resolution Google Earth im-

ages (50-60 cm/pixel) and low-resolution Landsat 8

images (30 m/pixel), using binary masks from Open

Buildings V3 Polygons (Sirko et al., 2021) and the

MapBiomas project (Rosa et al., 2019). Preprocess-

ing involved converting images to PNG, segmenting

them into 384 × 384 pixel patches, and balancing the

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

930

dataset to remove pairs without pixels of interest.

Google Earth, via Google Earth Engine (Gore-

lick, 2013), provided high-resolution images from ur-

ban areas of Brazil. Thirty images (8192 × 3059 pix-

els) were processed using a cropping algorithm, gen-

erating approximately 5040 smaller images. A ﬁlter

removed pairs without white pixels (buildings), leav-

ing 1029 balanced pairs. Figure 1 illustrates an exam-

ple of an image and its corresponding binary mask,

obtained from Open Buildings V3 Polygons, which

details building outlines.

(a) Google Earth image. (b) Open Buildings mask.

Figure 1: Pair of Google Earth images of a region of Brazil

containing a cut-out of a satellite image in its original com-

position (RGB) with its respective mask obtained through

the Open Buildings V3 Polygons.

The Landsat 8 satellite (Avdan and Jovanovska,

2016) has two main sensors: OLI (Operational Land

Imager) and TIRS (Thermal Infrared Sensor). The

OLI captures multispectral images in nine bands with

a 30-meter spatial resolution, while the TIRS collects

data in two thermal bands with a 100-meter resolu-

tion, resampled to 30 meters to align with the OLI. For

this study, 30-meter multispectral images were used,

with a temporal resolution of 16 days and 510 images

in total. The Figure 2 shows an example of an RGB

image and the reference mask.

(a) Landsat 8 image. (b) MapBiomas mask.

Figure 2: Pair of images with 30 meters per pixel resolution

for neural network training. (a) RGB satellite image from

Landsat 8. (b) Groundtruth of land cover from MapBiomas.

3.1.2 Brazilian Geographic Database of the

Distribution System

The Brazilian National Electric Energy Agency

(ANEEL) mandates electricity distributors to provide

BGDDS, an annually updated geographic model that

represents Brazil’s electrical system, detailing assets

such as operational areas and infrastructure, including

high, medium, and low voltage networks, transform-

ers, poles, feeders, substations, and transmission tow-

ers. BGDDS supports grid planning, operation, and

oversight, ensuring infrastructure keeps pace with ur-

ban growth. In this study, this data was used to map

the electrical infrastructure of Minas Gerais, integrat-

ing remote sensing to identify built-up areas lacking

network coverage.

3.1.3 Integrated Framework for Monitoring the

Distribution Electrical Grid

Given the extensive number of assets managed by a

Brazilian electrical company, a tool was developed to

ensure energy balance consistency and assess com-

mercial losses at different levels (feeder, substation,

and region). The Framework integrates data from

operational measurements, telemetry-based loss as-

sessments, load ﬂow analysis, and BGDDS network

information with monthly energy load records.

The validation process begins with data extraction

from the Framework database, followed by an ETL

(Extract, Transform, Load) workﬂow. Data is pre-

processed to remove inconsistencies, handle missing

values, and align with company standards. Once val-

idated, the dataset is formatted into a JSON payload

for advanced analysis.

This approach enables ranking substations by an-

nual commercial losses, identifying problematic feed-

ers, and calculating expected consumption based on

measured data, technical losses, and distributed gen-

eration, simulating ideal conditions without commer-

cial losses.

3.2 Image Segmentation Algorithms:

Architectures and Basic

Characteristics

Semantic segmentation is a crucial technique in com-

puter vision that classiﬁes each pixel in an image

into categories such as roads, buildings, vegetation,

and water. In this context, CNN-based models have

demonstrated superior precision and computational

efﬁciency compared to traditional methods (Jiwani

et al., 2021). For this study, two semantic segmenta-

tion algorithms were applied to divide an image into

semantically meaningful regions, labeling each pixel

according to its respective class. The architectures

used for the analysis of various remote sensors were

LinkNetB7 and D-LinkNet50, they will be described

next.

Integrating Satellite Images Segmentation and Electrical Infrastructure Data to Identify Possible Urban Irregularities in Power Grid

931

3.2.1 Convolutional Neural Networks

Architectures

The ﬁrst semantic segmentation neural network used

was LinkNetB7 (Chaurasia and Culurciello, 2017),

which employs a pretrained EfﬁcientNet-B7 as its en-

coder, enhancing feature extraction and model gen-

eralization. With compound scaling for parameter

adjustment (Akyel and Arıcı, 2022), the architec-

ture combines encoder blocks (dimensionality reduc-

tion) and decoder blocks (image reconstruction), us-

ing batch normalization, ReLU, skip connections, and

class weighting to improve segmentation. Simple and

efﬁcient, the model requires less data and training

time while maintaining results comparable to other

methodologies (Figure 10 B).

The second neural network used was D-

LinkNet50 (Zhou et al., 2018), originally designed for

road segmentation in satellite images but adaptable to

other domains with parameter and label adjustments.

It employs ResNet50 as its encoder, a 50-layer net-

work suitable for complex tasks (Li and Liu, 2022).

Its architecture, illustrated in Figure 10 (A), is divided

into three blocks: A (encoder, which generates a com-

pact latent representation), B (uses dilated convolu-

tions to expand the receptive ﬁeld while preserving

spatial details), and C (decoder, which restores im-

age resolution via transposed convolutions, similar to

LinkNet) Zhou et al. (2018).

3.2.2 Training Parameters and Environment

The training hyperparameters were deﬁned based on

the literature (Zhou et al., 2018; Akyel and Arıcı,

2022) and previous experiments, including an initial

learning rate of 0.001, a binarization threshold of 0.3,

and a batch size of 8. Two loss functions were tested:

BCE IoU, which combines Binary Cross Entropy and

IoU to penalize discrepancies and overlaps (Batchkala

and Ali, 2021), and Dice BCE, which balances lo-

cal and global accuracy in segmentation (Montazerol-

ghaem et al., 2023). To avoid overﬁtting (Afaq and

Rao, 2020), training was limited to 60 epochs with

early stopping after ﬁve epochs without signiﬁcant

improvement in the loss curve.

Data augmentation techniques, such as rotation,

ﬂipping, and shifting, were applied, generating 12

variations per training image to improve model gener-

alization. Non-augmented satellite images were used

for validation, ensuring artiﬁcial diversiﬁcation and

robust learning. The algorithms were implemented

in Python using PyTorch, running on a machine with

an AMD Ryzen Threadripper 3970X x64 processor

and an NVIDIA GeForce RTX 4090 GPU (24 GB),

accelerated by the CUDA toolkit, complemented by

256 GB of RAM.

3.3 Detection of Urban Areas Covered

by Electrical Infrastructure

The methodology is based on the principle that ar-

eas with high urban occupation but lacking electrical

infrastructure are susceptible to irregular settlements

and illegal connections. To identify them, a map

of the electrical infrastructure was created using the

’PONNOT’ layer from BGDDS, ﬁltering only poles

in urban areas, which represent the highest capillarity

of the grid (Figure 4). Image segmentation identiﬁed

occupied areas not covered by the map, generating in-

sights into regions prone to illegal connections. This

information enables operational teams to act more as-

sertively in combating commercial losses.

The PONNOT buffer, which simulates the cover-

age of electrical infrastructure from poles, was cre-

ated using mathematical morphology techniques. The

closing operation, composed of dilation (boundary

expansion) and erosion (boundary reﬁnement) with

the same structuring element (SE), was applied to

smooth contours, ﬁll gaps, and connect nearby areas,

ensuring continuous coverage (Gonzalez and Woods,

2000; Solomon and Breckon, 2013). This approach

eliminated gaps between buffers, resulting in a homo-

geneous area (Figure 4). The ﬁnal electrical infras-

tructure coverage map is presented in Section 4.

4 RESULTS AND DISCUSSIONS

The study area was selected based on two criteria:

urbanization and supply by a substation with high

commercial losses. The SBAU substation, located in

Sabar

a, Minas Gerais, met these criteria, recording a

commercial loss of 32.094 MWh in 2023, equivalent

to 26% of all measured energy—well above the na-

tional average of 6.7% (ANEEL, 2025).

SBAU serves ∼ 30, 000 low and medium-voltage

customers across ∼ 85 km

through four feeders:

SBAU02, SBAU03, SBAU04, and SBAU05. Figure

5 shows their historical measurement data. SBAU05

(Figure 5 d) had the highest commercial loss rate

(47.5%), aligning with studies linking non-technical

losses to socioeconomic factors (ANEEL, 2025). No-

tably, its data consistency suggests it as a key case for

loss mitigation studies.

Other feeders also exhibit relevant patterns.

SBAU02 (Figure 5 a) had the lowest loss (6%) in

2023, serving the fewest customers. SBAU03 (Fig-

ure 5 b) showed stable losses with three notable vari-

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

932

Figure 3: (a) Architecture of the D-LinkNet50 Deep Neural Network, (B) LinkNet Deep Neural Network Architecture.

Adapted from: Chaurasia and Culurciello (2017); Zhou et al. (2018).

Figure 4: The image illustrates the PONNOT points (yel-

low) representing electrical infrastructure poles and their

corresponding buffers (green), which simulate the coverage

area of electrical infrastructure.

ations, while SBAU04 (Figure 5 c) had pronounced

ﬂuctuations, possibly due to operational maneuvers or

irregularities.

Following this analysis, the next step was to assess

the pole infrastructure coverage in the feeder-supplied

regions to identify buildings not connected to the en-

ergy grid.

After training and validating the satellite image

segmentation models for land use and land cover de-

tection, performance metrics such as accuracy, f1-

score, and IoU were evaluated. Additionally, Table

1 presents the number of epochs required for each

model to reach the stopping criterion, as well as the

corresponding training time.

Deep learning techniques, such as U-Net, achieve

up to 95% accuracy in land use classiﬁcation

(

Agton de Oliveira, 2020). However, the choice of ar-

chitecture and loss function depends on image resolu-

tion and application context. High-resolution images

are ideal for detailed building identiﬁcation, whereas

lower-resolution images are better suited for large-

scale detection tasks.

For Google Earth images, BCEIoU loss yielded

the best performance for LinkNetB7, reaching opti-

mal results in 30 epochs. In contrast, DICE BCE

achieved the highest overall accuracy, requiring 36

epochs and a training time of 4 hours and 38 min-

utes. For Landsat 8, LinkNetB7 combined with DICE

BCE achieved an impressive accuracy of 0.97 in just

12 epochs.

Table 1 highlights the signiﬁcant impact of ar-

chitecture and loss function on segmentation perfor-

mance. BCE IoU proved superior for high-resolution

images, while DICE BCE outperformed in lower-

resolution scenarios. The variation in training time

and required epochs underscores the importance of

careful model selection based on the speciﬁc task and

data characteristics.

Figures 6 and 8 provide qualitative insights into

the model’s performance, showcasing its ability to

handle both detailed and large-scale features. Addi-

tionally, Figures 7 and 9 present the prediction masks

for Google Earth and Landsat 8 images, respectively,

further illustrating the model’s adaptability across dif-

ferent resolutions.

Google Earth images enable individual building

block differentiation, whereas Landsat 8 images re-

veal only occupation patches. High-resolution images

are preferred for precise urban monitoring, as they

differentiate individual lots, crucial for detailed analy-

sis. In contrast, low-resolution images merge multiple

lots, reducing clarity but remaining useful for track-

ing rapid urban growth. Sensor selection should align

with monitoring needs: low-resolution images suit

dense or rapidly urbanizing areas, where changes are

more noticeable at the edges, while high-resolution

images are essential for low-density regions, such as

rural areas, to detect subtle land use changes.

Figure 10 (A and B) present maps generated

through semantic segmentation of land use and oc-

Integrating Satellite Images Segmentation and Electrical Infrastructure Data to Identify Possible Urban Irregularities in Power Grid

933

(a) SBA02 (b) SBA03

Figure 5: Historical series of feeder measurements, expected consumption, and NTL for the SBAU substation.

Table 1: Performance of the Segmentation Neural Network trained with images from different sensors.

Satellite

Architecture

Loss

Accuracy F1-Score IoU Epochs

Training

Sensor Function Time

LinkNetB7 BCEIoU 0,86 0,78 0,77 30 10 h 19 min

Google Earth LinkNetB7 DICE BCE 0,85 0,77 0,76 6 2 h 19 min

(∼ 0,60 m / pixel) D-LinkNet50 BCEIoU 0,84 0,75 0,74 28 9 h 22 min

D-LinkNet50 DICE BCE 0,84 0,75 0,73 27 9 h 01 min

LinkNetB7 BCEIoU 0,96 0,96 0,93 46 7 h 26 min

Landsat 8 LinkNetB7 DICE BCE 0,97 0,96 0,94 12 2 h 03 min

(30 m / pixel) D-LinkNet50 BCEIoU 0,94 0,92 0,88 19 3 h 04 min

D-LinkNet50 DICE BCE 0,95 0,94 0,90 16 2 h 36 min

(a) (b) (c) (d) (e)

Figure 6: Prediction results overlaid on the original for the Google Earth satellite image. (a) Original image; (b) LinkNetB7

BCEIoU; (c) LinkNetB7 DICE BCE; (d) D-LinkNet50 BCEIoU; (e) D-LinkNet50 DICE BCE.

cupation, overlaid with electrical infrastructure cov-

erage (in green). Urban areas are highlighted (in red),

emphasizing densely occupied regions, particularly

those without electrical grid access.

As shown in Figure 10 (A and B), regions near the

SBAU05 and SBAU04 feeders exhibit evident urban

occupation but lack electrical infrastructure coverage

in the ofﬁcial data. This absence, combined with

the high percentage of commercial losses recorded

in these feeders, suggests that these areas may corre-

spond to irregular settlements or illegal connections.

Furthermore, the spatial detail provided by the

segmentation of Google Earth images enabled a more

precise identiﬁcation of urban occupation areas, while

the Landsat 8 images, with their lower resolution,

posed challenges in detecting smaller clusters. There-

fore, the combination of segmentation data with the

analysis of electrical infrastructure proved effective in

identifying potentially problematic regions, providing

valuable insights for better management and monitor-

ing efforts.

5 CONCLUSIONS

The combination of CNNs with geospatial analysis

techniques has proven effective in detecting irregu-

lar occupations in urban areas, integrating popula-

tion, infrastructure, and satellite data. However, the

study has limitations, such as geographic restrictions

(Sabar

a, Minas Gerais), reliance on speciﬁc satellite

data (Google Earth and Landsat 8), and the lack of

real-time integration with utility operational systems.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

934

(a) (b) (c) (d) (e)

Figure 7: Prediction results for the Google Earth satellite image dataset. (a) Ground truth; (b) LinkNetB7 BCEIoU; (c)

LinkNetB7 DICE BCE; (d) D-LinkNet50 BCEIoU; (e) D-LinkNet50 DICE BCE.

(a) (b) (c) (d) (e)

Figure 8: Prediction results overlaid on the original for the Landsat 8 satellite image dataset. (a) Original satellite image; (b)

LinkNetB7 BCEIoU; (c) LinkNetB7 DICE BCE; (d) D-LinkNet50 BCEIoU; (e) D-LinkNet50 DICE BCE.

(a) (b) (c) (d) (e)

Figure 9: Prediction results for the Landsat 8 satellite image dataset. (a) Ground truth; (b) LinkNetB7 BCEIoU; (c) LinkNetB7

DICE BCE; (d) D-LinkNet50 BCEIoU; (e) D-LinkNet50 DICE BCE.

Figure 10: Land Use Prediction Using (A) Google Earth Satellite Images , (B) Landsat 8 Satellite and CEMIG Electrical Grid

Coverage.

To advance, future research should expand the ge-

ographic scope, incorporate real-time and diverse data

(socioeconomic and urban expansion), and explore

more advanced neural architectures or hybrid models.

Additionally, the development of a decision-support

framework for utilities, with urban monitoring tools

and demand forecasting, can enhance energy plan-

ning. Studies on the socioeconomic impacts of irreg-

ular settlements are also necessary for more compre-

hensive policies.

This research paves the way for artiﬁcial intel-

ligence applications in urban energy management,

promoting sustainable development and equitable re-

source distribution.

ACKNOWLEDGEMENTS

We thank CEMIG, ANEEL (PD-04950-0664/2023),

CNPq, and CAPES for their ﬁnancial support and col-

laboration, which were essential for this study and

for advancing research, development, and innovation

(R&D&I) in the electrical sector.

Integrating Satellite Images Segmentation and Electrical Infrastructure Data to Identify Possible Urban Irregularities in Power Grid

935

REFERENCES

Abujayyab, S. K., Almajalid, R., Wazirali, R., Ahmad, R.,

Tas¸o

glu, E., Karas, I. R., and Hijazi, I. (2023). In-

tegrating object-based and pixel-based segmentation

for building footprint extraction from satellite images.

Journal of King Saud University-Computer and Infor-

mation Sciences, 35(10):101802.

Afaq, S. and Rao, S. (2020). Signiﬁcance of epochs on

training a neural network. Int. J. Sci. Technol. Res,

9(06):485–488.

Akyel, C. and Arıcı, N. (2022). Linknet-b7: noise re-

moval and lesion segmentation in images of skin can-

cer. Mathematics, 10(5):736.

ANEEL (2025). Energy loss reports - light in the tariff.

Accessed: January 3, 2025.

Archana, R. and Jeevaraj, P. E. (2024). Deep learning mod-

els for digital image processing: a review. Artiﬁcial

Intelligence Review, 57(1):11.

Atas¸,

I. (2023). Performance evaluation of jaccard-dice co-

efﬁcient on building segmentation from high resolu-

tion satellite images. Balkan Journal of Electrical and

Computer Engineering, 11(1):100–106.

Avdan, U. and Jovanovska, G. (2016). Algorithm for

automated mapping of land surface temperature us-

ing landsat 8 satellite data. Journal of sensors,

2016(1):1480307.

Batchkala, G. and Ali, S. (2021). Real-time polyp segmen-

tation using u-net with iou loss.

Chaurasia, A. and Culurciello, E. (2017). Linknet: Exploit-

ing encoder representations for efﬁcient semantic seg-

mentation. In 2017 IEEE visual communications and

image processing (VCIP), pages 1–4. IEEE.

EPE (2024). Balanc¸o Energ

etico Nacional 2024. Acesso

em: 30 set. 2024.

Flores, J. L. O. et al. (2023). Consumo de energia e cresci-

mento econ

omico no brasil.

Gonzalez, R. C. and Woods, R. E. (2000). Processamento

de imagens digitais. Editora Blucher.

Gorelick, N. (2013). Google earth engine. In EGU gen-

eral assembly conference abstracts, volume 15, page

11997. American Geophysical Union Vienna, Austria.

He, Z., Yao, W., Shao, J., and Wang, P. (2024). Ub-ﬁnenet:

Urban building ﬁne-grained classiﬁcation network for

open-access satellite images. ISPRS Journal of Pho-

togrammetry and Remote Sensing, 217:76–90.

Jiwani, A., Ganguly, S., Ding, C., Zhou, N., and Chan,

D. M. (2021). A semantic segmentation network

for urban-scale building footprint extraction using rgb

satellite imagery. arXiv preprint arXiv:2104.01263.

Khan, A. and Sudheer, M. (2022). Machine learning-based

monitoring and modeling for spatio-temporal urban

growth of islamabad. The Egyptian Journal of Remote

Sensing and Space Science, 25(2):541–550.

Li, S. and Liu, X. (2022). Multi-type road extraction and

analysis of high-resolution images with d-linknet50.

In 2022 3rd International Conference on Geology,

Mapping and Remote Sensing (ICGMRS), pages 244–

248. IEEE.

Montazerolghaem, M., Sun, Y., Sasso, G., and Haworth, A.

(2023). U-net architecture for prostate segmentation:

the impact of loss function on system performance.

Bioengineering, 10(4):412.

Rani, K. P., Rao, K. S., Sreedevi, P., Subbaraju, G. V.,

Dastagiriah, C., and Padmavathi, J. (2024). Land

cover extraction from satellite imagery using u-net

model. In Proceedings of the 11th International Con-

ference on Computing for Sustainable Global Devel-

opment, pages 1361–1367.

Rosa, M., Shimbo, J. Z., and Azevedo, T. (2019).

Mapbiomas-mapeando as transformac¸

oes do territ

orio

brasileiro nas

ultimas tr

es d

ecadas. VIII Simp

osio de

Restaurac¸

ao Ecol

ogica, pages 95–100.

Sirko, W., Kashubin, S., Ritter, M., Annkah, A., Bouchareb,

Y. S. E., Dauphin, Y., Keysers, D., Neumann, M.,

Cisse, M., and Quinn, J. (2021). Continental-scale

building detection from high resolution satellite im-

agery. arXiv preprint arXiv:2107.12283.

Solomon, C. and Breckon, T. (2013). Fundamentos de

processamento digital de imagens-Uma abordagem

atica com exemplos em MatLab.

Ullah, S., Qiao, X., and Abbas, M. (2024). Addressing the

impact of land use land cover changes on land surface

temperature using machine learning algorithms. Sci-

entiﬁc Reports, 14(1):18746.

Zhou, L., Zhang, C., and Wu, M. (2018). D-linknet: Linknet

with pretrained encoder and dilated convolution for

high resolution satellite imagery road extraction. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition workshops, pages 182–

186.

Agton de Oliveira, P. (2020). Deep learning na segmentac¸

autom

atica de imagens de sat

elite. 24 pages. Tra-

balho de Conclus

ao de Curso (Graduac¸

ao) - Engen-

haria Agr

ıcola e Ambiental.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

936