You are on page 1of 5

2020 IEEE REGION 10 CONFERENCE (TENCON)

Osaka, Japan, November 16-19, 2020

Early Detection of Forest Fire using Deep


Learning
Medi Rahul, Karnekanti Shiva Saketh, Attili Sanjeet and Nenavath Srinivas Naik
Department of Computer Science and Engineering,
International Institute of Information Technology(IIIT) Naya Raipur, India
Email: medi18100@iiitnr.edu.in, shiva18100@iiitnr.edu.in, attili18100@iiitnr.edu.in, srinu@iiitnr.edu.in

Abstract—Forest fires have become a serious threat to The rest of the paper is organized in the following manner.
mankind. Besides providing shelter and protection to a large A detailed review of traditional methods used for the detec-
number of living beings, they have been a major source of tion of forest fires is presented in Section II. An overview
food, wood, and a great supply of other products. Since ancient
times forests have played an important role in social, economic, of the Proposed model is detailed in Section III. Section IV
and religious activities and have enriched human life in a gives an overview of the results. Finally, Section V concludes
variety of ways both material and psychological. To protect the paper.
our nature from these rapidly rising forest fires, we need to be
cautious enough of every decision we take which could lead to II. L ITERATURE S URVEY
a disastrous end, once and for all. So for the early detection of
forest fires, we propose an image recognition method based on Studies about fire detection are very few compared to
Convolutional Neural Networks (CNN). We have fine-tuned the general classification and detection algorithms. A method for
Resnet50 architecture and added a few convolutional layers with detecting forest fires has been proposed based on wireless
ReLu as the activation functions, and a binary classification sensor networks [8]. This is a traditional approach to the
output layer which showed a huge impact on the training and
test results when compared to the other SOTA methods like problem. The results are accurate but implementing it on a
VGG16 AND DenseNet121. We achieved a training set accuracy large scale is bound to be difficult.
of 92.27% and 89.57% test accuracy with a stochastic gradient A small scale CNN is implemented using YOLOv3 and
descent optimizer and we have avoided the underfit/overfitting the test results show that the detection rate of this algorithm
on the model with the help of the Stochastic Gradient Descent is approximately 83% [1]. The algorithm is sensitive to a
(SGD) algorithm.
Index Terms—Forest Fire Detection, CNN large forest areas and the detection of fire in less probabilistic
fire-prone areas is bad.
I. I NTRODUCTION A Faster R-CNN has been proposed to detect forest fires
The most common forest hazard is a forest fire. They [3]. To solve the ambiguity in training data non-real smoke
pose a threat not only to the richness of the forest but also images were produced by putting up real or simulative smoke
to the flora and fauna, which seriously affect biodiversity into the forest background. However, this detection algorithm
and ecology. According to the Brazilian National Institute is more prone to simulative forest smoke, and hence it
of Space Research (INPE), in Brazil, Amazonas, at the last requires further improvements in boosting the performance
official count, about 76,000 trees were burned, more than with real-time fire detection.
80% compared to the same period last year. At least 33 Wildfire’s smoke detection method has been developed to
people were killed, including 4 firefighters and more. On provide a fast, inexpensive, and reliable system for moni-
January 31, 2020, more than 11 million hectares (110,000 toring fire-sensitive areas [4]. The proposed approach uses
square kilometers) of bushes and forest parks across Australia algorithms to extract features based on image processing
were burned. techniques and classifications based on computer-aided intel-
The importance of forests should not be ignored. There is ligence strategies. In terms of space and time, the preparatory
a great dependence on our survival with forests, on the air step consists of subsampling the image sequences, then de-
we breathe, on the wood we use. Forests not only provide tecting the movement zone, which extracts the smoking zones
habitats for animals and humans, they also protect river in question. The candidate regions are then processed using
basins, prevent soil erosion, and reduce climate change. a smoke color analysis algorithm, a sharp edge detector,
With today’s world advancing in the technology era, there and detection techniques for growing and increasing areas.
has been a keen interest in the image processing technology, Neural classifiers are then used to estimate the areas that
the fire detection method based on the computer vision describe the smoking areas in different settings. It was an
system [9] reached its pinnacle and replaced many traditional effective approach that worked well even under less than
methods and is playing a vital role in our day to day life. ideal conditions. Poor visibility conditions can lead to false
Human and natural disasters are the main cause of forest positives but do not increase false alarms.
fires. CNN smoothens the image recognition in the fea- A real-time smoke detection has been proposed for the
ture extraction process by theoretically extracting deeper techniques of accumulation of foreground images and optical
functions which greatly improves the accuracy of fire im- flow [7]. This technique was used to extract the movement
age recognition. In this study, we compared the perfor- properties of smoke and flames and to distinguish them
mance of different state of the art models like VGG16 and by their color disturbances. Their method is more likely
DenseNet121 with the proposed fine-tuned ResNet50 model to suppress noise and attenuate it completely in successive

978-1-7281-8455-5/20/$31.00 ©2020 IEEE 1136

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 07,2022 at 13:22:34 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. Overview of proposed model.

frames. Restrictions were limited to light-sensitive objects. trained to classify into two classes fire and non-fire. The
For example, lights turned on and off, etc. But their technique above model is then run on the test set to provide the results.
knew how to distinguish the other fire-like objects. The proposed model uses ResNet50 with transfer learning
A combination of motion detection and color detection to train the model. The images are resized to 224x224 pixels.
had improved detection speed and eliminated excess events The dataset is split into train and test data with 80% and 20%
[16]. Second, the motion detection was based on the im- respectively. Fine-tuning plays a major role in the reliability
age frame difference method. Finally, pretreatment can be of the proposed model. First, the layers in the proposed
combined with advanced flame extraction to reduce time model are frozen. Besides, we proposed some of the layers
complexity. Subsequently, important visual properties of the with ReLu function and softmax for classification.
flame were used, such as color, movement, change of area The proposed layers include one convolutional layer, one
and shape, and a new algorithm were offered to stabilize pooling layer, a Batch Normalization, a ReLu activation layer
the center of gravity based on space and time. Finally, the with a Dropout (to avoid overfitting), and finally, a softmax
support vector machine gives satisfactory results, according layer for the image classification. The optimizer chosen was
to the experimental results, the accuracy was almost 95%, SGD as the test cases prove that SGD gives a good global
but the algorithm has certain limits, for example, when minimum. By using the optimal learning rate, the proposed
personnel dressed in crimson clothing is blocked by devices model was able to achieve a training set accuracy of 92.27%
or when workers have a pick-up within a certain range by and a test set accuracy of 89.57%.
the introduced advanced features can lead to false alarms In figures 2 and 3, the detection of images is classified
in these conditions. Therefore, optimistic methods must be into two classes (fire and non-fire). First, the input image
implemented to solve the problems of the real environment. is convolved with different feature detectors to get a feature
map and then pooled to get a max feature pooled map. After
III. P ROPOSED M ODEL
that, it is flattened and passed on to the input features. All
Forests play a significant role in preserving the ecosystem. this working is taken care of by ResNet50 architecture. The
As a result, there has been considerable work done on forest image is resized to 224x224 and passed onto the ResNet50
fire detection. Most of the literature does not focus on trying model and once the probabilities have crossed the threshold
to achieve high detection rates across multiple databases. then it gets classified with the help of softmax into the classes
In this paper, we use transfer learning to develop a 0 (fire) and 1 (non-fire).
model that addresses this challenge. The system involves pre-
processing the image data and applying data augmentation Parameters Value
Batch Size 16
such as shearing, flipping, etc. to generate more images. We Dropout 0.5
then use models like VGG16, ResNet50, and DenseNet121 Activation Function Softmax
for the classification of images. Optimizer SGD with learning rate 0.001
Loss Function Binary cross entropy
Figure 1 states the explanation of the model proposed. Epochs 100
The model initially divides the train and test sets in 80%
and 20% respectively. Next, the data is prepared to enter TABLE I
PARAMETERS
the pre-processing phase and gets augmented to reduce the
overfitting in the model (Augmentation is a regularization
method, which used to prevent overfitting). Finally, it is Two types of parameters are used to compile the model:

1137

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 07,2022 at 13:22:34 UTC from IEEE Xplore. Restrictions apply.
Fig. 2. Detecting Images with Class: Fire

Fig. 3. Detecting Images with Class: No Fire

learning rate and momentum. Learning rates were dependent


on the type of optimizers used (Adam/SGD) while momen-
tum is initially 0.9. Fine-tuning was done to make sure the
model does not generalize the pattern found in images. The
batch size is initialized as 16 and the model is trained over
100 epochs.
The parameters listed in Table I are finalized after per-
forming multiple test cases. To decrease the training time,
the batch size is finalized as 16. To classify images based
on probabilities, the activation function used was softmax.
The optimal learning rate (0.001) is used so that while we
iterate over the number of epochs, the weights get updated
in backpropagation and reach its global minimum.

Fig. 5. VGG16 Loss

them, ResNet50 performed better compared to VGG16 and


DenseNet121.
Results generated using VGG16 are shown in Table II.
The first test condition was running Adam optimizer without
momentum and the results were close to 75%. Next, when
the test condition was added with a momentum of 0.9, the
accuracy was close to 83% only. Even when the optimizer
was changed to Stochastic Gradient Descent and with a
momentum of 0.9 the accuracy was still 85%. This made
us move onto ResNet50 and DenseNet121.
Results generated using ResNet50 are shown in Table
Fig. 4. VGG16 Accuracy III. The first test condition was running Adam optimizer
without momentum and the results were close to 87%. Next,
when the test condition was added with a momentum of 0.9,
IV. RESULTS the accuracy was close to 89% only. When the optimizer
Results have been generated using CNN based mod- was changed to Stochastic Gradient Descent and with a
els like VGG16, ResNet50, and DenseNet121. Among momentum of 0.9 the accuracy achieved was 92.27%.

1138

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 07,2022 at 13:22:34 UTC from IEEE Xplore. Restrictions apply.
Results generated using DenseNet121 are shown in Table
IV. The first test condition was running Adam optimizer
without momentum and the results were close to 83%. Next,
when the test condition was added with a momentum of
0.9, the accuracy was still close to 85% only. When the
optimizer was changed to Stochastic Gradient Descent and
with a momentum of 0.9 the accuracy achieved was 89%
with traces of overfitting.

TABLE II
VGG16 T EST C ASE R ESULTS

Test Condition Result (Accuracy)


Adam optimizer without momentum 74.52%
Adam optimizer with momentum=0.9 82.73%
SGD optimizer with momentum=0.9 85.21%

Fig. 7. ResNet50 Loss

TABLE III
R ES N ET 50 T EST C ASE R ESULTS decent explanation for vanishing gradient descent problem
when compared to VGG16.
Test Condition Result (Accuracy) Figure 6 shares information on ResNet50 architecture’s
Adam optimizer without momentum 87.37%
Adam optimizer with momentum=0.9 89.92% accuracy versus Epochs (100 epochs). The relevant re-
SGD optimizer with momentum=0.9 92.27% sults generated are shown in Table III. ResNet50 consumes
less memory when compared to VGG16. The accuracy of
ResNet50 is the highest among all the models built.
In figure 7, the information shows us that the training
TABLE IV
D ENSE N ET 121 T EST C ASE R ESULTS
set error and test set error are well under control with low
bias and low variance(ideal condition). It gradually decreases
Test Condition Result (Accuracy) with an increase in the number of epochs. This proves that
Adam optimizer without momentum 83.24%
Adam optimizer with momentum=0.9 84.59%
ResNet50 is a reliable model.
SGD optimizer with momentum=0.9 89.04% In figures 8 and 9, the plot describes the Model accuracy
against the number of epochs. We can observe that the
training set accuracy is very good, but the test set error gets
very less on further epochs. Whereas in the loss vs epochs
the training set error is very minute with low bias in training,
but in test set error it is increasing. This creates a condition
of high variance or overfitting.

Fig. 6. ResNet50 Accuracy

Figure 4 displays the model accuracy and the results gener-


ated to prove that this model is not reliable. At tenth epoch,
the difference between train accuracy and test accuracy is
Fig. 8. DenseNet121 Accuracy
huge, and the training accuracy plummets to 60%. This
information shows us that the model is facing difficulty to
classify the images. V. C ONCLUSION
In figure 5, the graph of the loss function (categorical We proposed transfer learning-based approach for early
cross-entropy) and epochs are shown. The predictions in detection of forest fires. The proposed model successfully
our VGG16 model’s test set error leads to overfitting. Due classifies most of the images into two classes, fire, and no
to which we have moved to ResNet50. ResNet50 offers a fire. Out of VGG16, ResNet50, and DenseNet121, ResNet50

1139

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 07,2022 at 13:22:34 UTC from IEEE Xplore. Restrictions apply.
[12] A. Koubaa and B. Qureshi, “Dronetrack: Cloud-based real-time object
tracking using unmanned aerial vehicles over the internet,” IEEE
Access, vol. 6, pp. 13 810–13 824, 2018.
[13] D. Kolaric, K. Skala, and A. Dubravic, “Integrated ´ system for forest
fire early detection and management,” Periodicum Biologorum, vol.
110, no. 2, pp. 205–211, 2008.
[14] J. R. Martinez-de Dios, B. C. Arrue, A. Ollero, L. Merino, and
F. Gomez-Rodrıguez, “Computer vision techniques for forest fire
perception,” Image and Vision Computing, vol. 26, no. 4, pp. 550–562,
2008.
[15] N. A. Che-Bin Liu, “Vision based fire detection,” in International
Conference in Pattern Recognition, 2004.
[16] Faming Gong, Chuantao Li, Wenjuan Gong, Xin Li, Xiangbing
Yuan, Yuhui Ma and Tao Song, ”A real-time fire detection method
from video with Multifeature Fusion,” Computational Intelligence and
Neuroscience, Hindawi, 2019.

Fig. 9. DenseNet121 Loss

proves to be a reliable model. ResNet50 predicts the type of


image that is fed to the model and successfully classifies most
of the images with high train and test accuracies as 92.27%
and 89.57% respectively. Fine-tuning helped the model to
avoid overfitting. Altogether, this model can be applied in
real-time to avoid forest fires disrupting the ecosystem. The
future work includes getting a better dataset, to optimize the
model using Single Shot Detection (SSD), and to apply the
model in a real-time scenario using Unmanned Aerial Vehicle
(UAV).

R EFERENCES
[1] Zhentian Jiao1, Youmin Zhang, Jing Xin, Lingxia Mu, Yingmin Yi,
Han Liu and Ding Liu, ”A Deep Learning based forest fire detection
approach using UAV and YOLOv3,” 1st International Conference on
Industrial Artificial Intelligence (IAI), 2019
[2] Qingjie Zhang, Jiaolong Xu, Liang Xu and Haifeng Guo, ”Deep
Convolutional Neural Networks for forest fire detection,” in Proceed-
ings of the 2016 International Forum on Management, Education and
Information Technology Application, Atlantis Press, 2016.
[3] Qi-xing ZHANG, Gao-hua LIN, Yong-ming ZHANG, Gao XU, Jin-jun
WANG, ”Wildland forest fire smoke detection on Faster R-CNN using
synthetic smoke images,” 8th International Conference on Fire Science
and Fire Protection Engineering(on the Development of Performance-
based Fire Code), 2017.
[4] Genovese, Angelo and Labati, Ruggero and Piuri, Vincenzo and Scotti,
Fabio, ”Wildfire smoke detection using computational intelligence
techniques,” IEEE International Conference on Computational Intelli-
gence for Measurement Systems and Applications Proceedings, 2011.
[5] C. Yuan, Z. X. Liu, and Y. M. Zhang, “UAV-based forest fire detection
and tracking using image processing techniques,” in 2015 International
Conference on Unmanned Aircraft Systems. IEEE, 2015, pp. 639–643.
[6] C. Yuan, Z. X. Liu, and Y. M. Zhang, “Aerial images based forest fire
detection for firefighting using optical remote sensing techniques and
unmanned aerial vehicles,” Journal of Intelligent & Robotic Systems,
vol. 88,no. 2-4, pp. 635–654, 2017.
[7] X. Z. Chunyu Yu, Zhibin Mei, “A real-time video fire flame and smoke
detection algorithm,” in Asia-Oceania Symposium on Fire Science and
Technology, 2013.
[8] YongMin Liu, YiMing Liu, HongLei Xu, Kok Lay Teo, ”Forest fire
monitoring, detection and decision making systems by wireless sensor
network,” IEEE Chinese Control And Decision Conference (CCDC),
2018.
[9] A. Tariq, S. Osama, and A. Gillani, “Development of a low cost and
light weight UAV for photogrammetry and precision land mapping
using aerial imagery,” in 2016 International Conference on Frontiers
of Information Technology. IEEE, 2016, pp. 360–364.
[10] T.-H. Chen, P.-H. Wu, and Y.-C. Chiou, “An early fire detection method
based on image processing,” in 2004 International Conference on
Image Processing, 2004.
[11] Pulkit Chugh, Eric Tom Mathews, G. Barath Kumar, ”Forest fire
detection through UAV imagery using CNNs,” unpublished.

1140

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on September 07,2022 at 13:22:34 UTC from IEEE Xplore. Restrictions apply.

You might also like