Encyclopedia
Please note this is an old version of this entry, which may differ significantly from the current revision.
Wildfires are major natural disasters that can cause extensive damage to ecosystems and threaten human lives. It is an uncontrollable and destructive fire that rapidly spreads through vegetation, grasslands, or other flammable areas. Wildfires are typically triggered by a combination of factors, including the presence of abundant dry vegetation and favorable weather conditions like high temperatures, low humidity, and strong winds. The sources of ignition for wildfires are diverse and can range from natural causes like lightning strikes to human activities such as campfires, careless disposal of cigarettes, or even intentional acts of arson. Besides the destructive nature of wildfires, the smoke from wildfires can have severe human health risks and environmental consequences as it can contribute to air quality degradation, disrupt the balance of ecosystems, and even impact the behavior and survival of wildlife. Therefore, early fire and smoke detection are crucial.
- wildfires
- fire and smoke detection
- computer vision
- deep learning
- database
- image processing technique
- artificial neural network
1. Introduction
Early fire and smoke detection are crucial; some recent methods and tools focus on smoke and fire detectors [1] due to their low cost. However, they have some limitations, such as poor performance in open and wide areas and the detection time, which is dependent on the distance between the detector and the smoke source [2]. Recent monitoring systems such as remote sensing technologies, satellite imagery, and ground-based sensors rely on artificial intelligence-based algorithms in detecting and monitoring the outbreak of fires and their spread. In those systems, vision software runs at its core and automatically detects the presence of an event.
In the last years, like other fields, computer vision-based fire and smoke detection methods have attracted the attention of researchers [3,4,5,6]. Compared to conventional methods [7,8,9,10,11], they have many advantages like the large area coverage, the real-time response, and the high accuracy.
Some of the existing approaches integrated various combinations of hand-crafted traditional techniques to detect smoke image features like shape, color, and texture [12,13,14]. At the same time, others exploit the moving nature of smoke to extract the motion characteristics [13,15,16] (motion value, direction, energy, convex hull), achieving good results considering some particular cases.
In recent years, deep learning-based algorithms, especially CNNs [17,18,19,20,21], have gained so much attention, and many researchers investigated those models and achieved good precision results. Hence, the use of those methods is still challenging since they rely on supervised learning and require huge databases for training, and their effectiveness depends on various factors, including the strong hardware systems for implementation and the quality and diversity of the dataset used for training. A database is one of the bases of a deep learning model, and the provided results largely depend on the comprehensiveness and the authenticity of training data. Industries and institutions involved in deep learning spend a lot of time and energy collecting datasets. The significant investment required to acquire these datasets adds value to deep learning training and underscores their importance. The availability of public smoke and fire databases with their ground truth labels is one of the biggest challenges related to deep learning models. In addition, smoke and wildfire images captured from fixed cameras, drones in motion, or other imaging devices are often affected by noise, artifacts, or other imperfections that can affect their interpretation and analysis. To address these challenges, Image processing techniques, such as noise reduction, image enhancement, and image restoration, are employed to improve the quality of acquired images, making them more suitable for interpretation.
2. Recent Methods Related to Fire and Smoke Detection
Computer vision and deep neural networks have emerged as powerful strategies for studying the behavior of fire and smoke. Numerous methods have been introduced in the literature to leverage these technologies. Kaabi et al. [22] introduced a method for smoke detection. First, a motion-based feature extraction with the GMM algorithm was employed; then, a trained DBN classifier based on the feature of smoke was used to detect the smoke region in videos. Xu et al. [23] proposed an end-to-end method for smoke detection based on a deep saliency network; the framework was used to extract the smoke saliency map via a pixel level and object level salient CNN. Yifan et al. [24] proposed a real-time detector network for fire and smoke based on a light YOLOv4. The model is based mainly on three modules: the MobileNetv3, the BiFPN, and a new feature extraction technique with depthwise separable convolution and attention block. The method attends good performance with a minimum of trainable parameters. Cao et al. [25] proposed a bidirectional LSTM network for forest fire and smoke detection in videos, and it consists of bidirectional learning of the discriminative spatiotemporal features. Ali Khan et al. [26] exploit the advantages of transfer learning to train the VGG19 model to localize wildfires; then, they build a network composed of unmanned aerial vehicles communicated to an assistance center to simplify and increase the data transmission process. Frizzi et al. [27] proposed a new CNN architecture for fire and smoke segmentation and classification. The network is composed of coding and decoding paths and achieves better accuracy compared to similar methods with low false positives (clouds, haze) and in optimal segmentation time. Aymen et al. [28] integrate the non-linear adaptive level set method with an artificial neural network model to track the forest fire regions in wildfire videos; the method consists of estimating and localizing the possible fire contours by analyzing the chromatic and statistical features with the linear discriminant analysis combined with an artificial neural network. Then, a level-set algorithm was applied to refine the segmentation results. The method gives good results in terms of speed and accuracy. An interesting approach proposed by [29] concentrates and studies the architecture and the training of networks; it combines the self-attention mechanism with a multi-scale feature connection for real-time fire and smoke detection. The authors first fused the feature maps of the network into a radial connection; then, they applied a permutation attention mechanism to gather the relevant information, and designed a feature fusion block to increase the detection efficiency. The method gives good results compared to standard proposed methods.
Recent technologies based on advanced artificial neural network architecture achieve good results, but they still have limitations, especially when trying to implement them for daily use because of the need for huge databases for training and the strong/expensive computation resources. From this, many methods attempt to detect fire and smoke by building a pipeline integrating the image processing techniques and by exploiting the smoke textural features (Table 1). Ref. [30] built a statistical model combined with an optical flow algorithm for real-time fire and smoke detection. The method first extracts the smoke and fire-like regions with frame differential steps, a color model of fire and smoke, and a foreground accumulation technique. A motion feature discriminating model with the optical flow was applied to the first resulting image to extract the final fire and smoke regions. In [31], the fire presence decision in video frames is reached by analyzing the color variations and periodic behavior of the flame with the temporal and spatial wavelet transform algorithm. Ref. [32] proposed a fire detection system consisting of modeling the color information in the CIE L*a*b color space and detecting the motion of fire pixels with a background subtraction technique.
This entry is adapted from the peer-reviewed paper 10.3390/electronics12183888
This entry is offline, you can click here to edit this entry!
