See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/378497388

Forest Fire prediction using Machine Learning Methods: A Comparative Study

Conference Paper · May 2023

CITATIONS READS

0 64

3 authors, including:

Ferdaous Abdessemed Souheila Bouam

Computer science Université Batna 2
3 PUBLICATIONS 0 CITATIONS 25 PUBLICATIONS 105 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ferdaous Abdessemed on 27 February 2024.

The user has requested enhancement of the downloaded file.

 First National Conference in Computer Science Research and its Applications
Mai 03, 2023

Forest Fire prediction using Machine Learning Methods

Ferdaous Abdessemed1, Souheila Bouam2, Chafik Arar3

1 LAMIE laboratory, Computer Science department, University of Banta 2, Banta, Algeria
2,3 LAMIE laboratory, Computer Science department, University of Banta 2, Banta, Algeria

E-mail address: [email protected], [email protected], [email protected]

Abstract: Forest fires are one of the most disastrous natural calamities causing significant harm to both human life and property. Early
detection of forest fires is crucial to prevent their spread and minimize the damage caused. In recent years, machine learning (ML)
techniques have emerged as promising tools for forest fire detection. In this paper, we present a comparative study of four popular ML
methods - decision tree, random forest, k-nearest neighbors (KNN), and support vector machine (SVM) - for forest fire prediction, in
terms of accuracy, precision, recall, and F1 score. Our study also compared the test duration of each method, since early identification
of fire is the most critical aspect of our topic. After identifying the best-performing method, we will further optimize it to enhance its
performance. The experimental results show that the decision tree outperforms the other three algorithms, achieving an accuracy of
97.95%, a precision of 100%, a recall of 97.05%, and an F1 score of 98.5%.

Keywords: Fire Prediction, Decision Tree , Random Forest, KNN, SVM, Machine Learning, Forest Fire

relative humidity (RH), and wind speed - which have an

1. INTRODUCTION impact on the occurrence of fires. The classification
Forest fires are a major concern worldwide, particularly performed in this study involves two classes, namely “fire”
in countries with extensive forest cover. These fires not and “no fire”.
only cause significant damage to the environment, but also Our objective is to provide insights into the effectiveness
pose a threat to human lives. As stated in [1], more than of these methods in detecting forest fires based on metrics
31,513 fires resulted in the burning of over 320,409 such as accuracy, precision, recall, and F1 score. By
hectares of forest between 2008 and 2017. Additionally, in identifying the best-performing method, we hope to
2012, there were over 5,110 fires that burned an area of contribute to the development of more accurate and
more than 99,061 hectares in Algeria. According to [2], efficient forest fire detection systems.
Algeria ranks fourth in terms of forest fire impact among It should be noted that the focus of this paper is on the
the countries monitored by the European Forest Fire software implementation of these classifiers as a predictive
Information System (EFFIS). This highlights the model, which will be later implemented on hardware. The
importance of the early detection of forest fires to minimize hardware implementation and the design of the sensor node
the risk and impact of these fires. Traditionally, forest fire architecture are not included in the scope of this research.
detection relies on human observation or satellite imagery
[3]. However, these methods have limitations in terms of 2. RELATED WORKS
accuracy, speed, and cost. With the emergence of the Numerous studies have been conducted to explore the use
machine learning methods, there is now an opportunity to of data mining techniques in detecting and predicting fires.
improve the accuracy and speed of prediction of forest fires Data mining has been found to be particularly useful in
while also reducing costs. Machine learning methods have extracting potentially useful patterns and information from
shown great potential for the early detection and prediction large databases. By analyzing historical data, it is possible
of forest fires. These methods use algorithms and statistical to identify patterns and trends that can help to predict the
models to analyze data and identify patterns that indicate likelihood and severity of future fires.
the presence of a fire. By analyzing historical data on forest In [4] authors focused on the application of data mining
fires, the machine learning algorithms can learn to techniques and algorithms for wireless sensor networks
recognize patterns and predict the likelihood of a fire (WSN) in fire detection. They proposed three different
occurring in a particular location. In this paper, we experiments on WSN fire detection data to compare the
investigate the performance of four popular machine- classification accuracy of various data mining techniques
learning methods for forest fire detection and compare and algorithms to determine which algorithm has the best
them to identify the best method. The paper focuses on the performance. The study concludes that for large datasets,
utilization of three meteorological variables - temperature, the neural network classifier achieves superior prediction
 2 A.Ferdaous: Forest Fire prediction using Machine Learning Methods

results than other classifiers. However, for small datasets, A. Brief review of machine learning methods
classifiers such as OneR (a one-level decision tree Decision Tree is a supervised learning algorithm used for
algorithm) or FURIA (a fuzzy rule induction algorithm) classification and regression, it is a visual representation of
produce satisfactory outcomes. choices and their possible outcomes in the form of a tree-
The authors in [5] utilized a data mining technique to like structure. The nodes in the tree correspond to events or
decrease the size of data in a Wireless Sensor Network decisions, while the edges represent the decision rules or
(WSN) that included several sensors, such as temperature, conditions that lead to different outcomes. Each tree is
humidity, smoke, and light sensors. They employed the composed of nodes and branches, where each node
naive Bayes classifier for detecting forest fires, and the represents a group of attributes to be classified and each
study reported an accuracy of approximately 94%. branch represents a possible value that the node can take
In [6] the authors proposed a solution using the [10].
convergence of the Internet of Things (IoT) and You Only Random Forest is an ensemble learning method that
Look Once Version 5 (YOLOv5) for real-time monitoring combines multiple decision trees. It constructs multiple
of fire disasters. The experimental results showed that IoT decision trees at training time and outputs the class that is
devices were able to validate some of the falsely detected the mode of the classes of the individual trees.
fires or undetected fires that YOLOv5 reported. The paper
Support-vector machines are a type of supervised
also compared the performance of the proposed method
learning models and algorithms used in machine learning
with other fire detection approaches using performance
for analyzing data and classification/regression analysis.
metrics. SVMs can effectively perform non-linear classification
Heffeda et al. introduced a Wireless Sensor Network using the kernel trick to map inputs into high-dimensional
(WSN) for forest fire detection in their study [7]. Their feature spaces. SVM is particularly useful when the
system is based on the Fire Weather Index (FWI) system number of features is much larger than the number of
components, namely the Fine Fuel Moisture Code (FFMC) samples [10].
and the FWI, which are calculated using the weather
conditions. The detection system is modeled as a k- K-Nearest Neighbor is a classification method that
coverage model, and the calculated FFMC and FWI values involves a database with data points categorized into
are transmitted to a processing center for further action. multiple classes, and the algorithm attempts to classify a
given sample data point as part of a classification problem.
The classification is determined by the majority class of the
In [8] , the authors proposed a decision tree-based system k nearest neighbors in the feature space [11].
for forest fire prediction, with the goal of integrating the
decision tree classifier into a smart sensor node B. Steps in Forest Fire Prediction using Machine
architecture to allow for automated and intelligent fire Learning
prediction without human intervention. The prediction is In this study, we aimed to compare the performance of
based on meteorological data including temperature, several machine learning methods for detecting forest fires
relative humidity, and wind speed, which are critical using a publicly available dataset of the Algerian Forest
factors in the forest fire occurrence. The proposed system Fires Dataset [8]. The dataset contains valuable
achieved an accuracy of about 82.92% in the software information about the meteorological variables that can
implementation. The aim of this system is to mitigate the affect the occurrence of forest fires, such as temperature,
threat of forest fires, which can cause economic, humidity, wind speed, and rain. We trained the machine
ecological, and human life damage. learning models on 70% of the data and used the remaining
30% to evaluate their performance. To ensure that the
This paper [9] provides a review that explores the use of models were effective, we used four metrics namely
data mining and algorithms to predict and detect forest accuracy, precision, recall, and F1-score. The accuracy
fires. The review identified several common factors that measures the proportion of correctly classified samples,
influence forest fires, such as temperature, relative while precision measures the proportion of true positives
humidity, precipitation, wind speed, month, heat, smoke, among all positive predictions. Recall measures the
and atmospheric gases. proportion of true positives among all actual positives, and
the F1-score is the harmonic mean of precision and recall.
3. METHODOLOGY
By comparing the performance of the different machine
This section outlines our forest fire prediction approach learning methods using these metrics, we were able to
using machine learning techniques. We provide a concise identify the most effective method for detecting forest fires
introduction to the machine learning methods utilized as a based on the meteorological variables present in the
predictive model, which include decision tree, random Algerian Forest Fires dataset.
forest, KNN, and SVM. Following this, we describe the In addition to evaluating the methods based on these
steps involved in our work. metrics, our study also examined the test duration of
First National Conference in Computer Science Research and its Applications 3

each method since early prediction and detection are

the most critical aspects of our research topic.
The testing time is important to consider because it
represents the time it takes to evaluate the model's
performance on new, unseen data. This is particularly
important when the model is deployed in a real-world
setting where it will be making predictions in real time.
Once we compared various methods and selected the
most suitable one for our topic, we adjusted its
parameters to further enhance its accuracy.
4. RESULTS AND DISCUSSION
A. comparison of diffirent methods
In this section, we compared the performance of the
previously mentioned methods. “Table.Ⅰ” displays a
Figure 1. Machine Learning Algorithm Performance Comparison
comparison of their accuracy, recall, precision, and F1- Graph
score measures, which were calculated for both the 'fire'
and 'no fire' classes. The Table also includes a comparison B. Optimizing Decision Tree Parameters for Higher
of the testing duration for each method, which is a crucial Accuracy
aspect of our study in achieving early fire prediction.
After evaluating and comparing various methods, we
Accuracy F1 Testing
found that the decision tree is the optimal method for the
model Precision Recall classification task. However, the accuracy of the decision
(%) score time
Decision tree model can be further improved by optimizing its
95.91 1.00 0.94 0.96 0.002521
tree parameters. By adjusting parameters like the minimum
Random number of samples required to create a split, the maximum
93.87 1.00 0.91 0.95 0.693289
forest
number of features, and other parameters, the accuracy of
KNN 91.83 1.00 0.88 0.93 0.011033 the model can be greatly enhanced. Optimizing these
SVM 69.38 0.69 0.98 0.81 0.018270 parameters enables the decision tree to generalize more
effectively to new data.
Table I. PERFORMANCE COMPARISON OF MACHINE LEARNING The Table below provides a summary of the decision tree
ALGORITHMS parameter values utilized in the study. These values
Our results showed that decision trees outperformed the comprise the minimum number of samples required for a
other machine learning methods for forest fire detection, leaf node, the minimum number of samples required for
achieving an accuracy of 95.91% compared to 93.83% for splitting an internal node, the maximum number of
random forests, 69.38% for SVM, and 91.83% for KNN. features to consider for each split, the random seed value
Decision trees also had the highest precision and recall utilized for reproducibility, as well as the metric used to
scores, indicating that it was able to correctly identify evaluate the split's quality.
forest fires with fewer false positives and false negatives
Parameter Value
compared to the other methods. In addition, in comparison
Min samples leaf 1
to the other methods, the decision tree approach requires Min samples split 2
the least amount of time for testing, which can aid in Max features None
predicting fires earlier. Random state 75
Criterion Entropy

Table II. DECISION TREE PARAMETER VALUES USED IN THE STUDY

By utilizing the parameter values listed in “Table. Ⅱ”, the

study was able to attain improved accuracy of 97%, along
with 100% precision, 97% recall, and 98% F1-score as
shown in “Table.Ⅲ”.
 4 A.Ferdaous: Forest Fire prediction using Machine Learning Methods

Accuracy F1 5. CONCLUSION
Precision Recall
(%) score
Optimized This study highlights the importance of early detection and
Decision Tree 97.95 1.00 0.97 0.98 prediction of forest fires in preventing their spread and
parameters minimizing the damage caused. Machine learning
techniques have emerged as promising tools for forest fire
Table III. PEEFORMANCE METRICS OF IMPROCED DECISION TREE
MODEL
prediction, and the study compares the performance of
four popular ML methods for predicting forest fires. The
C. Comparison with existing works results show that the decision tree algorithm outperforms
The most important factor in this study is to accurately the other methods in terms of accuracy, precision, recall,
predict positive instances (the occurrence of fire). When and F1 score. Additionally, when compared to other
predicting the existence of fire, recall would be the most methods, decision tree algorithms are able to predict the
important evaluation metric among the ones mentioned. probability of forest fire occurrence in a shorter amount of
This is because predicting the existence of fire when it is time, enabling forest managers to take proactive measures
actually present (true positive) is more critical than to prevent fires. The findings of this study provide
predicting the existence of fire when it is not actually valuable insights for developing advanced forest fire
present (false positive). Therefore, we compared our detection and prediction systems using machine learning
model to existing models based on their recall techniques to protect forests and human lives.
performance. Additionally, we also evaluated the F1-
score, which takes into account both recall and precision REFERENCES
to determine how robust the classifier is in correctly [1] B. A, " Les feux de forêts en Algérie analyse et perspectives.
Technical report " Ministère de l’agriculture et de
classifying instances while not missing a considerable
développement rurale, et de la pèche, direction des forets
number of them. 2018.
In this study, a comparison between existing DT-based fire [2] F.-N. Robinne et al., "Global fire challenges in a warming world,"
prediction models and the proposed model is conducted; 2018: IUFRO.
[3] N. Bach and N. Sirimongkalertkal, "Satellite data for detecting trans-
this comparison is made to evaluate the performance of the
boundary crop and forest fire dynamics in Northern
proposed model against existing models. The results are Thailand," International Journal of Geoinformatics, vol. 7,
presented in “Table. Ⅳ”. no. 4, p. 47, 2011.
[4] M. Maksimović and V. Vujović, "Comparative analysis of data
mining techniques applied to wireless sensor network data for
fire detection," JITA-APEIRON, vol. 6, no. 2, 2013.
References Accuracy (%) Recall F1-score
[5] M. Saoudi, A. Bounceur, R. Euler, and T. Kechadi, "Data mining
techniques applied to wireless sensor networks for early forest
fire detection," in Proceedings of the International
This work 97.95 0.97 0.98 Conference on Internet of things and Cloud Computing, 2016,
pp. 1-7.
[6] K. Avazov, A. E. Hyun, A. A. Sami S, A. Khaitov, A. B.
Abdusalomov, and Y. I. Cho, "Forest Fire Detection and
[12] Stojanova et al. 81.2 0.81 0.81 Notification Method Based on AI and IoT Approaches,"
Future Internet, vol. 15, no. 2, p. 61, 2023.
[7] M. Hefeeda and M. Bagheri, "Wireless sensor networks for early
[8] Abid et al. 82.89 0.92 0.85 detection of forest fires," in 2007 IEEE International
Conference on Mobile Adhoc and Sensor Systems, 2007, pp.
1-6: IEEE.
[8] F. Abid and N. Izeboudjen, "Predicting forest fire in algeria using
[4] Maksimović et al. Not given 0.96 0.96 data mining techniques: Case study of the decision tree
algorithm," in Advanced Intelligent Systems for Sustainable
Development (AI2SD’2019) Volume 4-Advanced Intelligent
[13] Giuntini et al. Not given 0.83 0.86 Systems for Applied Computing Sciences, 2020, pp. 363-370:
Springer.
[9] A. Hooda, A. Asthana, and V. Patkar, "A Comprehensive Review on
Prediction and Detection of Forest Fires Using Machine
Table IV. A COMPARISON OF PROPOSED AND EXSITING DECISION TREE Learning Algorithms," EasyChair Preprint, no. 6334, 2021.
METHODS FOR FOREST FIRE PREDICTION [10] B. Mahesh, "Machine learning algorithms-a review,"
International Journal of Science and Research
Our proposed approach showed better performance (IJSR).[Internet], vol. 9, pp. 381-386, 2020.
compared to the models we compared it with. It is [11] S. Ray, "A quick review of machine learning algorithms," in 2019
important to note that this comparison is only indicative International conference on machine learning, big data,
because of the differences in the datasets used in previous cloud and parallel computing (COMITCon), 2019, pp. 35-39:
IEEE.
studies reported in [4], [12], and [13]. Nonetheless, our [12] D. Stojanova, P. Panov, A. Kobler, S. Džeroski, and K. Taškova,
obtained results are promising. We utilized the same "Learning to predict forest fires with different data mining
dataset as [8] and our results indicate the effectiveness of techniques," in Conference on data mining and data
our proposed model.
 First National Conference in Computer Science Research and its Applications 5

warehouses (SiKDD 2006), Ljubljana, Slovenia, 2006, pp.

255-258.
[13] F. T. Giuntini, D. M. Beder, and J. Ueyama, "Exploiting self-
organization and fault tolerance in wireless sensor networks:
A case study on wildfire detection application," International
Journal of Distributed Sensor Networks, vol. 13, no. 4, p.
1550147717704120, 2017.

View publication stats