Unsupervised Image Classification Algorithms Applied to Fire-Prone

Area Detection

I. Rahimi

1,2 a

, L. Duarte

1,2 b

and A. C. Teodoro

1,2 c

Department of Geosciences, Environment and Spatial Planning Faculty of Sciences of the University of Porto, Portugal

Institute of Earth Sciences, FCUP Pole, Porto, Portugal

Keywords: Remote Sensing, Wildfire, Classification, Zagros Mountains.

Abstract: Remote sensing data has become critical in identifying fire-prone areas, providing essential insights through

satellite imagery and various geospatial inputs. These data sources allow for real-time monitoring, mapping

fire susceptibility, and assessing factors such as vegetation, fuel moisture, land use, and environmental

conditions. Numerous supervised and unsupervised models combined with remote sensing data have shown

great potential in predicting fire-prone regions, offering accurate and timely information for early warning

systems and resource allocation. This study focuses on applying two unsupervised methods—PCA, and K-

means—using inputs like Sentinel-2 imagery, elevation, and the Zagros Grass Index (ZGI) to identify fire-

prone areas in the Kurdo-Zagrosian forests, an area increasingly vulnerable to wildfires. Among the two

methods evaluated, PCA demonstrated superior performance in predicting fire-susceptible areas, accurately

classifying 80% of the burned regions from 2021 to 2023 as moderate to high-risk zones.

1 INTRODUCTION

Forest fires' increasing frequency and intensity

worldwide is an escalating concern, driven by natural

and human-induced factors such as extreme weather

conditions, shifting land use patterns, and rapid urban

expansion (Zema, 2020; Bowman, 2017). These

factors, especially under the growing influence of

climate change, exacerbate the risk of wildfires. This

significant loss of forest cover highlights the urgent

need for effective wildfire monitoring and prevention

strategies.

In recent years, integrating geospatial and remote

sensing (RS) data/technologies has provided

invaluable insights into wildfire risk factors

(Teodoro, 2013). Through RS data, researchers can

monitor and analyze variables such as land cover,

temperature, and vegetation phenology (Duarte,

2018). This spatial data, combined with Geographic

Information Systems (GIS), allows for the continuous

monitoring of large areas and provides timely

information about the likelihood of fire incidents

(Mishra, 2024). RS data enables researchers to track

environmental changes in real time and analyze

https://orcid.org/0009-0002-7411-8637

https://orcid.org/0000-0002-7537-6606

https://orcid.org/0000-0002-8043-6431

critical variables such as fuel moisture content,

temperature trends, and human activity patterns, all of

which contribute to the increased risk of fires.

A key application of this technological

advancement is the development of Forest Fire

Susceptibility Maps (FSMs). These maps are crucial

for identifying areas at high risk of wildfires, enabling

authorities to allocate resources efficiently and

implement mitigation strategies in advance. By

locating and assessing fire-prone regions, FSMs play

a central role in reducing the vulnerability of

ecosystems and communities while supporting

informed ecological and urban planning decisions

(Ghorbanzadeh, 2019). The effectiveness of FSMs

has been significantly enhanced by integrating GIS,

RS, and image classification algorithms, allowing

researchers to predict fire susceptibility better and

improve early warning systems (Rihan, 2023).

The growing popularity of image classification

algorithms in fire susceptibility mapping is due to

their ability to process large-scale, high-dimensional

datasets and model complex non-linear relationships

more effectively than traditional statistical methods

(Kantarcioglu, 2023). These models perform well at

136

Rahimi, I., Duarte, L. and Teodoro, A. C.

Unsupervised Image Classiﬁcation Algorithms Applied to Fire-Prone Area Detection.

DOI: 10.5220/0013201800003935

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

In Proceedings of the 11th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2025), pages 136-141

ISBN: 978-989-758-741-2; ISSN: 2184-500X

integrating diverse data types, from climatic

conditions and vegetation indices to human activity

metrics, which allows for more precise fire risk

assessments. In recent years, the development of FSM

has been advanced through both single-model

applications and ensemble methods, frequently

showing improved performance over more

conventional techniques (Piao, 2022; Saha, 2023).

These advancements underscore FSM's potential to

surpass traditional frameworks, particularly by

leveraging methods that incorporate broader

environmental and spatial analysis techniques.

Unsupervised classification methods, such as

Principal Component Analysis (PCA) and K-means

clustering, have proven valuable for extracting

meaningful patterns from satellite imagery. PCA

reduces large datasets into key components, capturing

the most significant variations (Sharma, 2021), which

is crucial for identifying fire susceptibility factors like

vegetation, moisture, and land use patterns. K-Means,

on the other hand, clusters satellite image data based

on spectral similarity, enabling the identification of

regions with similar environmental characteristics

(Theodoridis, 2009). These methods are especially

effective in handling large-scale, heterogeneous

datasets from RS data (Celik, 2009).

The objective of this study is to apply two

unsupervised image classification techniques—PCA

and K-means—using inputs like Sentinel-2 imagery,

elevation, and the Zagros Grass Index (ZGI) (Rahimi,

2024) to identify fire-prone areas in the Kurdo-

Zagrosian forests. This region, vulnerable to

increasing wildfire incidents, presents a complex

environmental landscape where these methods can

highlight patterns indicating higher fire susceptibility.

By utilizing unsupervised algorithms, the study aims

to provide a data-driven approach to wildfire

susceptibility mapping, improving prevention and

mitigation strategies in fire-prone regions.

2 METHODS

2.1 Study Area

The method is designed for areas with Semi-Arid

(SA) and Semi-Mediterranean (SM) climates,

exemplified by the forests of Marivan and Sarvabad

in the Kurdistan Province, western Iran, bordering

Iraq. These regions have experienced frequent fire

incidents over the past decade. Geographically, the

study area spans between longitudes 45°57'50" E and

46°46'41" E and latitudes 35°1'1" N and 35°49'51" N.

Located in the northern Zagros mountains, Marivan

sits at an average elevation of 1,287 meters above sea

level and features varied topography, from mountains

to valleys, shaped by its cross-border proximity with

Iraq (Rahimi, 2023).

Figure 1: Study area: Marivan and Sarvabad, Kurdistan

province, Iran.

The climate in this region is SM, characterized by

cold winters and hot summers. Annual rainfall

averages 991 mm, with considerable variability (±235

mm). Humidity averages around 54%, which is

insufficient to sustain green pastures during summer,

leaving behind dry grass that elevates fire risks

(Rahimzadeh, 2008).

2.2 Data Collection and Preparation

Table 1 summarizes the data used in this study.

Sentinel-2 satellite imagery and related products

provided high-resolution multispectral data essential

for environmental analysis and fire risk assessment.

Table 1: Data sources used.

Data Type, Projection

System, Spatial

Resolution (m)

Time Period

Data

Source

Sentinel 2

(13 bands), UTM

, 10

13/05/2020 and

15/09/2020

ESA

Sentinel Burned Area

Product, UTM, 10

2021 -2023

ZGI, UTM, 10 2020

(Rahimi,

2024)

DEM

SRTM

, UTM, 10 - USGS

1. Universal Transverse Mercator

2. Digital Elevation Model/Shuttle Radar Topography Mission

3. European Space Agency

4. United States Geolo

ical Surve

2.3 Methodology

The collected data layers were first stacked using

Python 3.12, ensuring spatial alignment and

consistency across all layers. To integrate Sentinel-2

and ZGI layers, the datasets were clipped to the

boundaries of the study area using a defined polygon.

Unsupervised Image Classiﬁcation Algorithms Applied to Fire-Prone Area Detection

137

The clipped layers were then aligned to ensure

consistent spatial resolution and extent. Finally, the

processed layers were stacked into a multi-band raster,

enabling the combination of spectral information and

vegetation index data for subsequent fire susceptibility

analysis. This approach ensured a precise and

comprehensive dataset for modeling and assessment.

These layers were integrated into a matrix consisting

of all raster layers, which included Sentinel-2 bands

(13 bands), elevation, and the ZGI. The resulting

matrix had dimensions of 15×8020×6557, meaning

that 15 different values represented each pixel in the

study area. Figure 2 illustrates the methodology and

process flow of this approach.

Figure 2: Methodological framework.

This study employed PCA and K-means to detect

fire-prone areas effectively. To evaluate the

effectiveness of these models, the Leave-One-Out

(LOO) method was implemented, allowing for a

thorough assessment of how different data layers

contributed to fire susceptibility mapping. The

validation of the results with post-2020 data (from

2021–2023) demonstrated the reliability of these

methods for accurately predicting fire-prone areas

over time, showcasing their potential for long-term

wildfire risk assessment and management.

K-means and PCA were chosen as they are

common unsupervised methods suitable for studies

with limited field data. Unlike supervised learning

approaches, which require extensive datasets and

computational resources, these methods provide a

practical and efficient framework for fire

susceptibility classification. The components were

retained to classify fire susceptibility into high,

average, and low categories. This decision reflects the

study area's vegetation and land cover characteristics

alongside the limited field data available. This

approach balances capturing sufficient data variance

with the need for interpretability, ensuring robust and

practical classification.

To validate the results, the post-2020 burned area

was used. They derived from sentinel 2 by applying

the Normalize Burned Ratio (NBR) on data from

2021 to 2023 (Giddey, 2021). The NBR index proved

to be a valuable tool for identifying burned areas by

leveraging spectral differences between vegetation

and charred surfaces. This methodology ensured

reliable validation of the predictive models and

reinforced the importance of integrating remote

sensing indices into fire susceptibility research.

3 RESULTS

Figure 3 shows the results of applying PCA and K-

means. The results revealed varying levels of effective-

ness (High, Average, and Low) in predicting fire-prone

areas in the study region using 2020 data. The black

polygons are the burned areas calculated from Sentinel

2 data from 2021 to 2023, using the NBR index.

These patterns emphasize the spatial variability of

fire-prone areas, which is influenced by factors such

as vegetation type, fuel density, and topographic

conditions, as suggested by earlier studies (Rahimi et

al., 2024). The classification levels in this study were

established based on the number of components

selected for each of the unsupervised methods, PCA

and K-means, with three components defined for

optimal model performance in both cases. The trade-

off between preserving critical information and

minimizing computational complexity informed this

selection process for the number of components. By

reducing the dimensionality of the data, both PCA

and K-means enhanced interpretability while

maintaining the integrity of key predictive features

(Giddey, 2021).

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

138

Figure 3: Maps generated by a) PCA and b) K-means

according to data in 2020. The black polygons demonstrate

the burned areas from 2021 to 2023.

The visual representation in Figure 3 highlights

distinct patterns of fire susceptibility zones, with clear

differentiation between regions of varying risk levels.

Among the methods, PCA produced the most

accurate and reliable results, as shown in Figure 3a,

compared to K-means (Figure 3b). The red areas

show high-fire susceptible areas, while the yellow

and green colors show the average and low-fire

susceptible areas. The stark contrast between the red

zones and the remaining areas underscores the ability

of PCA to highlight regions requiring urgent

attention. This distinction aligns with practical

applications in fire management, where prioritizing

high-risk zones can significantly enhance resource

allocation and mitigation efforts.

The validation process compared high-

susceptibility areas identified by the models with

actual burned areas in subsequent years. PCA

exhibited a higher overlap between its predicted high

fire-susceptible zones and the areas affected by post-

2020 wildfires, with approximately 41%. In contrast,

this value was 31.3% for the K-means (Figure 4). This

significant difference in overlap suggests that PCA's

ability to identify subtle patterns in the data gives it a

clear edge over K-means. The higher accuracy of

PCA aligns with its proven effectiveness in

environmental modeling and its capacity to extract

critical components from high-dimensional datasets.

Regarding High and Average fire-prone areas

together, both methods offered almost the same

results. This indicates that PCA outperformed the K-

means models in accurately detecting fire-prone

regions, highlighting its robustness in mapping fire

susceptibility. However, the similar performance of

the two methods in identifying combined High and

Average susceptibility areas suggests that K-means

may still hold utility in broader-scale applications

where granular accuracy is less critical. This

underscores the need for selecting methods based on

specific use-case requirements.

Figure 4: Detection of fire-prone areas for the post-2020

years by 2020’s data using PCA and K-means.

Moreover, the analysis indicated that certain

Sentinel-2 spectral bands could be excluded without

substantially reducing the model's accuracy. This

Unsupervised Image Classiﬁcation Algorithms Applied to Fire-Prone Area Detection

139

finding highlights the efficiency of PCA in

dimensionality reduction, where less critical spectral

bands are discarded while retaining the most

informative features. Such optimization reduces

computational demands and facilitates faster data

processing, which is crucial for large-scale studies

(Kantarcioglu, 2023). This suggests that PCA

effectively reduced dimensionality, preserving only

the most informative features while maintaining high

predictive accuracy. The ability to maintain

predictive accuracy while reducing data complexity

makes PCA an attractive option for fire susceptibility

mapping, particularly in resource-constrained

settings where computational efficiency is essential.

The fact that the PCA model consistently provided

reliable predictions across multiple years

demonstrates its potential for long-term application in

fire susceptibility mapping.

This consistency underscores PCA’s robustness

and adaptability, especially in evolving

environmental conditions and fire patterns. Its

reliability across diverse temporal datasets reaffirms

its suitability for integration into long-term fire

management frameworks.

4 CONCLUSIONS

In conclusion, this study highlights the effectiveness

of using RS data combined with unsupervised image

classification techniques to identify fire-prone areas

in the Kurdo-Zagrosian forests. Among the two

methods tested, PCA performed better in accurately

predicting fire-susceptible zones, with 80% of the

burned areas from 2021 to 2023 correctly classified

as moderate to high-risk. Its ability to reduce

dimensionality while preserving critical information,

especially when excluding less relevant Sentinel-2

bands, further solidifies PCA as a robust tool for long-

term fire risk mapping. In contrast, K-means showed

moderate success, identifying around 50% of the

burned areas. These findings emphasize the potential

of PCA for improving fire management strategies,

while K-means may require further refinement to

achieve similar predictive accuracy.

In the future, more unsupervised classification

methods will be compared to further evaluate their

effectiveness and explore potential improvements in

fire risk prediction.

ACKNOWLEDGMENTS

The work is funded by national funds through FCT –

Fundação para a Ciência e Tecnologia, I.P., in the

framework of the UIDB/04683/2020 (https://doi.org/

10.54499/UIDB/04683/2020) and UIDP/04683/2020

633 (https://doi.org/10.54499/UIDP/04683/2020) –

Instituto de Ciências da Terra programs.

REFERENCES

Bowman, M.J.S.D.; Moreira-Munoz, A.; Kolden, C.A.;

Chavez, R.O.; Munoz, A.A.; Salinas, F.; Gonzalez-

Reyes, A.; Rocco, R.; de la Barrera, F.; Williamson,

G.J.; et al. Human- Environmental Drivers and Impacts

of the Globally Extreme 2017 Chilean Fires. Ambio

2018, doi:10.1007/s13280-018-1084-1

Celik, N. T. (2009). Unsupervised Change Detection in

Satellite Images Using Principal Component Analysis

and $k$-Means Clustering. IEEE Geoscience and

Remote Sensing Letters, 6(4), 772–776.

https://doi.org/10.1109/lgrs.2009.2025059

Duarte, L., Teodoro, A., Monteiro, A.T., Cunha, C., &

Gonçalves, H., "PhenoMetrics: An open source

software application to assess vegetation phenology

metrics," Computers and Electronics in Agriculture

148, 82–94 (2018); https://doi.org/10.1016/j.compag.2

018.03.007.

ESA European Space Agency (ESA), "Copernicus Open

Access Hub, Sentinel-2 Level-2A products,"

https://scihub.copernicus.eu.

Ghorbanzadeh, O.; Valizadeh, K.K.; Blaschke, T.; Aryal,

J.; Naboureh, A.; Einali, J.; Bian, J. Spatial prediction

of wildfire susceptibility using field survey GPS data

and machine learning approaches. Fire 2019, 2(3), 43.

https://doi.org/10.3390/fire2030043.

Giddey, B. L., Baard, J. A., & Kraaij, T. (2021).

Verification of the differenced Normalised Burn Ratio

(dNBR) as an index of fire severity in Afrotemperate

Forest. South African Journal of Botany, 146, 348–353.

https://doi.org/10.1016/j.sajb.2021.11.005

Kantarcioglu, O.; Schindler, K.; Kocaman, S. Forest fire

susceptibility assessment with machine learning

methods in north-east Türkiye. Int. Arch. Photogramm.

Remote Sens. Spatial Inf. Sci. 2023, XLVIII-M-1-2023,

161–167. https://doi.org/10.5194/isprs-archives-xlviii-

m-1-2023-161-2023.

Mishra, M.; Guria, R.; Baraj, B.; Nanda, A.P.; Santos,

C.A.G.; Da Silva, R.M.; Laksono, F.A.T. Spatial

analysis and machine learning prediction of forest fire

susceptibility: A comprehensive approach for effective

management and mitigation. Sci. Total Environ. 2024,

926, 171713. https://doi.org/10.1016/j.scitotenv.2024.1

71713.

Piao, Y.; Lee, D.; Park, S.; Kim, H.G.; Jin, Y. Forest fire

susceptibility assessment using Google Earth Engine in

Gangwon-do, Republic of Korea. Geomatics Nat.

GISTAM 2025 - 11th International Conference on Geographical Information Systems Theory, Applications and Management

140

Hazards Risk 2022, 13, 432–450. https://doi.org/

10.1080/19475705.2022.2030808.

Rahimi, I., Duarte, L., & Teodoro, A. C. (2024). Zagros

Grass Index—A new vegetation index to enhance fire

fuel mapping: A case study in the Zagros Mountains.

Sustainability, 16(10), 3900. https://doi.org/10.3390/su

16103900

Rahimzadeh Bajgiran, P., Darvishsefat, A., Khalili, A., &

Makhdoum, M., (2008). "Using AVHRR-based vegeta-

tion indices for drought monitoring in the Northwest of

Iran," Journal of Arid Environments 72, 1086–1096

Rihan, M.; Bindajam, A.A.; Talukdar, S.; Shahfahad;

Naikoo, M.W.; Mallick, J.; Rahman, A. Forest fire

susceptibility mapping with sensitivity and uncertainty

analysis using machine learning and deep learning

algorithms. Advances in Space Research 2023, 72(2),

426–443. https://doi.org/10.1016/j.asr.2023.03.026.

Saha, S.; Bera, B.; Shit, P.K.; Bhattacharjee, S.; Sengupta,

N. Prediction of forest fire susceptibility applying

machine and deep learning algorithms for conservation

priorities of forest resources. Remote Sensing

Applications: Society and Environment 2023, 29,

100917. https://doi.org/10.1016/j.rsase.2022.100917.

Sharma, D., Sharma, T., & Singhai, J. (2021). Extraction of

Water Bodies from Visible Color Satellite Images

Using PCA Feature Map. 2021 IEEE International

India Geoscience and Remote Sensing Symposium

(InGARSS), 1–4. https://doi.org/10.1109/ingarss515

64.2021.9791857

Teodoro, A.C., Duarte, L., (2013).Forest fire risk maps: a

GIS open source application a case study in Norwest of

Portugal, International Journal of Geographical

Information Science 27(4), DOI:10.1080/13658816.20

12.721554.

Theodoridis, S., & Koutroumbas, K. (2009). Clustering

Algorithms III: Schemes based on Function

Optimization. In Elsevier eBooks (pp. 701–763).

https://doi.org/10.1016/b978-1-59749-272-0.50016-5

USGS EROS Archive - Digital Elevation - Shuttle Radar

Topography Mission (SRTM), https://www.usgs.gov/

centers/eros/science/usgs-eros-archive-digital-

elevation-shuttle-radar-topography-mission-srtm-1.

Zema, D.A.; Nunes, J.P.; Lucas-Borja, M.E, (2020).

Improvement of Seasonal Runoff and Soil Loss

Predictions by the MMF (Morgan-Morgan-Finney)

Model after Wildfire and Soil Treatment in

Mediterranean Forest Ecosystems. Catena, 188,

104415, doi: 10.1016/j.catena.2019.104415

Unsupervised Image Classiﬁcation Algorithms Applied to Fire-Prone Area Detection

141