УДК 004

Overview and comparison of convolutional neural networks in fire detection

Абдуразак Куанышбек Абдуразакович – магистрант Казахстанско-Британского технического университета (Республика Казахстан, Алматы).

Дуйсек Бермагамбет Ерикулы – магистрант Казахстанско-Британского технического университета (Республика Казахстан, Алматы).

Abstract: This paper represents an overview and comparison of machine learning algorithms and techniques for fire detection. Machine learning is used to calculate the images to identify the pixels that might be the source of a fire. Identifying system is trained on the input data that are considered as the features and the resulting forecast is derived from input data that passed through machine learning approaches on a normalized dataset. This research involves various methods of implementing the identifying system based on different approaches.

The extraction of complex relationships between data is performed by several machine learning techniques consisting of the convolutional neural network (CNN) and other CNN models. Those learning algorithms are commonly used as classification, statistical, and finally predictive models, which are used for data forecasting. Comparison of the performance of the applied algorithms can give useful insights for optimizing the overall performance of such models.

Аннотация: Эта статья представляет собой обзор и сравнение алгоритмов машинного обучения и методов обнаружения пожара. Машинное обучение используется для расчета изображений, чтобы определить пиксели, которые могут быть источником возгорания. Система идентификации обучается на входных данных, которые рассматриваются как функции, а результирующий прогноз получается на основе входных данных, которые прошли через подходы машинного обучения к нормализованному набору данных. Это исследование включает в себя различные методы реализации системы идентификации, основанные на различных подходах. Извлечение сложных взаимосвязей между данными выполняется с помощью нескольких методов машинного обучения, состоящих из сверточной нейронной сети (CNN) и других моделей CNN. Эти алгоритмы обучения обычно используются в качестве классификационных, статистических и, наконец, прогностических моделей, которые используются для прогнозирования данных. Сравнение производительности применяемых алгоритмов может дать полезную информацию для оптимизации общей производительности таких моделей.

Keywords: convolutional neural network, neural network, fire detection, machine learning, deep learning.

Ключевые слова: сверточная нейронная сеть, нейронная сеть, обнаружение пожаров, машинное обучение, глубокое обучение.

Introduction

Preventing wildfires is a complex task, considering the catastrophic consequences on vegetation, animals, the environment, property, etc. However, with the advance of modern technology, there are plenty of ways to prevent such disasters from happening or at least reduce the damage. Satellite imagery is one of the tools that may help to fire prevention efforts. By using satellite imagery, it is possible to understand places that are prone to fire, spot early causes of a fire starting, and take preventative action to lower the danger of a fire. Satellite images provide the viewer with a unique perspective by giving a bird's-eye vision over the immense area. These graphics aid in locating areas that are at risk from fires since they are based on different factors, including vegetation density, terrain, weather patterns, and historical fire data. Experts can prioritize fire prevention measures like stricter land management practices, creating fire breaks, or employing controlled burns in strategic spots by first identifying fire-prone areas in satellite images. Satellite photos also help in the early discovery of fires. Satellites can detect heat signatures connected to flames thanks to their capacity to collect thermal and infrared radiation. Fire detection systems can instantly identify the location and size of fires by tracking these photos in real time, enabling a prompt reaction. Emergency services can organize resources, evacuate impacted areas, and dispatch firefighting professionals to contain the fire before it gets out of control thanks to this early warning.

CNN

In the study, we decided to choose CNN, which stands for Convolutional Neural Network. It can be described as a group of machine learning approaches and techniques that are mostly used in image and video recognition tasks. Images and other variants of data with a grid-like pattern can be interpreted by a CNN for future manipulations. CNN has a great application in situations that require the implementation of some sort of computer vision functionality, as the result of very high performance on a range of visual identification tasks, such as object detection, image categorization, facial recognition, and scene interpretation.

The main idea of the initial development of the structure of CNN has its roots in the organization of the sensory cortex in the human brain. The neural network has several layers, including convolutional, pooling, and fully connected layers. These layers are designed to derive valuable features from the initial input data and then utilize them to provide forecasting results.

The most crucial component of CNN is the convolutional layer. This layer applies very small filters or kernels to the incoming data and performs convolution operations in an effort to extract regional patterns or features. These filters traverse the input data, compute dot products at each location, and generate feature maps. Convolutional layers can record local spatial linkages and hierarchical patterns, making them valuable in applications for picture identification.

Pooling layers are used to decrease the sample size of the feature maps generated by the convolutional layers. They decrease the spatial dimensions of the data while maintaining the most important components. Two popular pooling operations are max and average pooling.

The thick layers sometimes called fully linked layers, are in charge of making predictions based on the learned attributes. They take the output from the convolutional and pooling layers and transform it into a vector of class probabilities or regression results, depending on the job at hand.

CNNs train themselves to recognize and classify patterns in the data by changing the weights of the connections between neurons. In this training, large labeled datasets and optimization methods like gradient descent are typically used.

Methodology

In this research, the wildfire dataset is retrieved from The official website of the Government of Canada [1]. Other models are taken from “Active fire detection in Landsat-8 imagery: A large-scale dataset and a deep-learning study”[2] research paper as well as some of the datasets from the United States Geological Survey website [3]. The datasets have images and images with layers that use different masks to extract specific layers.

Figure 1. Example of image with wildfire from Canada.ca dataset.

It was decided to form a convolutional neural network (Figure 2) with 3 convolutional layers, and 3 pooling layers.

Figure 2. Structure of CNN developed by authors.

The resulting output CNN and accuracy were compared to other CNN models, often used in such tasks. The structure of our newly developed CNN is shown in Figure 2., CNN layers are 2D convolutional layer (first, third, fifth), max pooling 2D (second, fourth, sixth), dropout (seventh, tenth, twelfth), flatten (eighth), dense (ninth, eleventh, thirteenth).

Table 1. Comparison of accuracy and loss between Murphy and Kumar Roy.

Results and conclusion

By assigning layers in this particular way, shapes are depicted in Figure 2., while it results are depicted in Table 1. Even though the new CNN performed a bit worse in terms of accuracy, approximately 1 percent of the difference between our CNN and the best performing one, Murphy CNN [4], our CNN achieved a significantly lower Loss, 59.4 percent. Moreover, it is worth mentioning that with a gradual increase in training data, the resulting metrics can be improved further, outperforming Murphy CNN and Kumar Roy CNN [5].

References

1. Government of Canada, Open Government, https://open.canada.ca/en.
2. Gabriel Henrique de Almeida Pereira, Andre Minoro Fusioka, Bogdan Tomoyuki Nassu, and Rodrigo Minetto, Active fire detection in Landsat-8 imagery: A large-scale dataset and a deep-learning study, ISPRS Journal of Photogrammetry and Remote Sensing.
3. United States government,S. Geological Survey, https://www.usgs.gov.
4. John H. Murphy, An Overview of Convolutional Neural Network Architectures for Deep Learning, Microway Inc.
5. Pradeep Kumar Roy, Abhinav Kumar, Convolutional Neural Network for Text: A Stepwise Working Guidance, Proceedings of the Yukthi 2021- The International Conference on Emerging Trends in Engineering – GEC Kozhikode, Kerala, India.