International Journal of Research Publication and Reviews, Vol 3, no 6, pp 3434-3438, June 2022

International Journal of Research Publication and Reviews

Journal homepage: www.ijrpr.com ISSN 2582-7421

Detection of Forest fire using Image Classification

Shiyam Prasad Sa
KPR Institute of Engineering and Technology, Coimbatore 641407, India

ABSTRACT:

The paper describes the idea used to detect forest fire based on video processing. The basic idea is to detect fire in forest areas where frequent attention is needed.
Using cameras in forests with video processing will help to detect fire and can be used as a caution system for fire fighters. Based on the dataset that has been
provided, the algorithm uses supervised learning method to train a model and detect fire in the input image using the trained model. The system also provides the
accuracy of the prediction ranges from 0 to 100. Experimental results show that the developed method can achieve fully automatic surveillance of forest fire.
These kind of system provide more accuracy than traditional fire detection sensors and can cover larger areas using a single camera. The accuracy of the model
can also be improved by providing more number of input images that can be trained for classification.

Keywords: Image classification, Forest fire, Tensorflow, Neural Network

Introduction:

As technology grows everyday artificial intelligence is being an essential part in everyday tasks as it has more potential than others. The
fields like machine learning and deep learning and their subsets have an immense growth. The training of the models using the algorithms and
classifiers requires huge amount of data to analyze them and to extract features. The focus of classification process is to group all the pixels from a
digital image into classes. Normally the numerical data in each pixel is used for classification of an image. The goal of image classification is to
identify the unique features occurring in image in terms of object that actually represent on the ground. Classifying between the objects is a difficult
task and therefore image classification is a crucial task in the field of digital image processing.
The term image classification refers to labelling of images into fixed classes. There can be n number of classes in which the input image can be
categorized. Classifying the image manually may be a hardest task when there is a huge amount of data to be classified but that can be handed over to a
machine that can automate the entire process using computer vision. This approach can be used in any fields where humans need to spend more time
supervising a work. The development in the field of autonomous driving is a great example of application of image classification.
From our case study forest fire doesn’t only affect the forest but it also affects the flora and fauna that poses a serious threat to the bio
diversity. There are two main causes for forest fire. One is man-made causes and the other is natural causes. Both of them not only affects the forests
but also the ecology. Existing fire detection system uses the combination of infrared and ultraviolet sensors to detect fire and there are some researches
done by many scholars. A paper named early fire-detection method based on image processing describes the use of sensors to detect fire. The paper
describes the use RGB values that has been converted from the given input video feed to detect fire but sometimes RGB values can differ and it may
not detect the fire accurately. Forest fires should be taken in a serious manner to avoid loss of valuable timbers and to avoid global warming. Our
proposed method uses supervised learning process in which the machines are trained using well "labelled" training data, and on basis of that data,
machines predict the output by creating a supervised model. The labelled data are those which has the training and testing images of forest fire and
images of green forest that has no fire.

Image Classification

In machine learning Supervised learning is a type where the models are trained using data that is labelled for training and testing purpose.
With using the data the machine will predict the output for the given input data. A labelled data is nothing but the data that is marked with correct
output. In this method the data that is used as training data will work as a controller that tells the machine to predict the correct output. It is defined as a
the construction of building where the workers work under the supervision of the builder. It is a overall process of providing the data as input with the
correct output data to the classification model. The main focus of supervised learning is to find a mapping function to map the (x) that is the input
function and (y) that is the output function.

In this method of learning the models are trained using dataset that are labelled before, Then the model learns about the data using the labels. After the
completion of the training using the labelled data the model is tested on the test data. There are two types of image classification, Single label image
classification and Multi label image classification.
 International Journal of Research Publication and Reviews, Vol 3, no 6, pp 3434-3438, June 2022 3435

2.1 Single Labelclassification

Single label image classification is the most common method of supervised learning for image classification. In this method a only one label
or an annotation is present for each image and the output of the model will be a single output for each input image. The output from our trained model
is a vector of length that is equal to the number of classes and some values denoting the score that the image belongs to this class. The softmax function
is used to add the score to one and the totall score is taken from the output. Softmax helps to bound the output between zero and one. Single label
classification can predict multiple classes where there are more than one or two classes or binary classes, in which there will be two classes.

2.2 Multi Label Classification

Multi label classification is the supervised learning method in which two or more number of labels can be in a image. Multi llabel
classification is still used in the medical field where the xx-ray shows the number of disease that a patient have.

2.3 Convolutional Neural Network

Convolutional neural network is also known as CNN is an artificial neural network that is most popularly used for classifica
classification problems.
Image classification mostly uses CNN but other types of data can also be classified. CNN forms a neural network and it is special in finding the patterns
in the data and use to predict output. Those patterns are very useful for image classification.

Fig 1 – Convolutional Neural Network

Convolutional neural network have hidden layers called convolutional layers and these layers make the CNN. It may not have convolutional layers but
the base of CNN is the convolutional layers. The convolutional layer receives the input and transforms it into another form and a sends the data to the
next layer. Getting and transforming the input into another form is called convolutional operation.. As discussed earlier convolutional neural network
detect patterns form images precisely so that it is used to predict output. In each convolutional llayer
ayer we should apply filters to detect patterns. Patterns
are nothing but the features in the image like the edges, shapes
shapes, corners etc arere used with filters. A filter is nothing but a small matrix and the row and
column of the matrix will be determined byy the user. The matrix is first initialized with random digits. Now the convolutionallayer
convolutionallayer with a filter will be a
3x3 filter. When the input is given to the convolutional layer the filter slides through the dimension of the image with the 3x3 filter and thist process is
known as convolving.

Fig 2 – Matrix example

 International Journal of Research Publication and Reviews, Vol 3, no 6, pp 3434-3438, June 2022 3436

At first the input image is taken from the input layer and then the filter slides through every pixels and forms a dot product with the matrix formed. The
matrix of the dot product will be the output of the matrix. This matrix will be passed to the next layer as input and the same process with the filter will
be done in the next layer and forms the output. The filters now can be mentioned as pattern detectors as it creates patterns from the input.

Proposed Method

Our proposed methodis based on single-label classification and uses Tensorflow to make the classification of fire. Tensorflow is an open source end to
end platform for machine learning. The dataset that we use has 1000 images of forest fire and 1000 images of green forest in the dimension of640x480
pixels and in three RGB channels to train the model and we will do the classification in 2 categories. Below are the examples for images that have
been used for training.

Fig 3 – Forest Fire Dataset

Our dataset contains 1000 images of forest fire used for training the sequential model.

Fig 4 –Green Forest Dataset

Necessary libraries are like Tensorflow and Numpy are imported at first. The data is then splitted into training and testing data. The data is divided like
80% of data is used to train the model and 20% of the data is used to test the trained model. As we have only two classes to classify the image for
detection of fire in forest or not so the labels for the model is 2, First label is Fire_Detected and the second label is No_Fire_Detected. The dataset is
first configured for performance so that buffered prefetching is used to so the data can be yielded from the disk without blocking the I/O blocking. The
RGB values of each pixels are standardized between 0 to 1 as the range from 0 to 255 is not ideal for a convolutional neural network. Our model
 International Journal of Research Publication and Reviews, Vol 3, no 6, pp 3434-3438, June 2022 3437

consists of three convolutional blocks with max pooling layer in each of the convolutional block and there will be a fully coconnected
nnected layer with 128 units
and that is activated
vated by a ReLU activation function. Our model is tuned to provide high accuracy. Adam optimizer is used to loss function is used to
find the metrics. After successfully configuring the model it is now trained with epochs of 10 and the training results ar
aree plotted.

Fig 5 – Validation Accuracy

Now the results shows that our accuracy is at 60% with the validation of the data data. From the above plot we can see that the training accuracy is
increasing linearly over time but validation accuracy is settled at 60%. When we use small amount of data for training the model
m will learn from the
noise and unwanted details from the given data ta which may lead to decreased in performance while testing the model
model.. This situation is known as
overfitting. To overcome over fitting we will use data augmentation to produce more training data as overfitting occurs only when there is less number
of training
ning data. Data augmentation produces more training data from the given set of data by augmenting them using random transform
transformation methods
that will output some realistic images like the input. This will help our model to train on more data and overcome the
the overfitting phenomenon. After
data augmentation is completed the model is again trained with epochs of 15 and the results of our new model is plotted.

Fig 6 –Validation accuracy after data augmentation

After data augmentation we can see significant improvement in the validation accuracy of the model. The validation accuracy is
i now gained additional
10 to 15% that will be enough
ough for the model for predicting on new data.

Result and Discussion

After the training of model, it is then tested among various images to check if it detects fire accurately. The plots shows tthe 75% accuracy
on training the model. But it can be tested only on real scenario. So for testing 2 images are loaded and classifie
classifiedd using our trained model to classify
fire. Below are the results

Fig 7 – Fire detected by our model

 International Journal of Research Publication and Reviews, Vol 3, no 6, pp 3434-3438, June 2022 3438

Fig 7 shows that our model successfully detected fire in the image and prints accordingly on the image. As our training data contains images of forest
fire so our model can detect fire in testing image.

Fig 8 – Fire not detected by our model

In the above figure Fig 8 our model has not detected any fire in the image so if prints that no_fire_detected. Our dataset contains images of green forest
and trained on those images so it detected that there is no fire on the give image. As of now we are loading the image one by one but we can direct a
webcam feed to our model to detect fire in a forest on a real-time basis.

CONCLUSION:

The research proposes a forest fire detection system. By using the Convolutional Neural Network and Image classification we did a
classifier that can detect forest fire with great accuracy. Usage of sensors can give a false signal over a normal fire even on a match stick, but our model
is capable of detecting only forest fire. As of now only 1000 images is used to train the model. In future more number of images can be used for
training for more accurate results. By fixing a camera over a distance and performing the algorithm will result in detection of forest fire in great way.

REFERENCES:

1. Alkhatib, Ahmad AA. "A review on forest fire detection techniques." International Journal of Distributed Sensor Networks 10.3 (2014): 597368.
2. Littell, Jeremy S., et al. "A review of the relationships between drought and forest fire in the United States." Global change biology 22.7 (2016): 2353-
2369.
3. Wang, Fei, et al. "Residual attention network for image classification." Proceedings of the IEEE conference on computer vision and pattern
recognition. 2017.
4. Lu, Dengsheng, and Qihao Weng. "A survey of image classification methods and techniques for improving classification performance." International
journal of Remote sensing 28.5 (2007): 823-870.
5. Boiman, Oren, Eli Shechtman, and Michal Irani. "In defense of nearest-neighbor based image classification." 2008 IEEE conference on computer
vision and pattern recognition. IEEE, 2008.
6. Çetin, A. Enis, et al. "Video fire detection–Review." Digital Signal Processing 23.6 (2013): 1827-1843.
7. Alkhatib, Ahmad AA. "A review on forest fire detection techniques." International Journal of Distributed Sensor Networks 10.3 (2014): 597368.
8. Yu, Liyang, Neng Wang, and Xiaoqiao Meng. "Real-time forest fire detection with wireless sensor networks." Proceedings. 2005 International
Conference on Wireless Communications, Networking and Mobile Computing, 2005.. Vol. 2. Ieee, 2005