Deep Learning-Based Multistage Fire Detection System and Emerging Direction

Abstract

Fires constitute a significant risk to public safety and property, making early and accurate detection essential for an effective response and damage mitigation. Traditional fire detection methods have limitations in terms of accuracy and adaptability, particularly in complex environments in which various fire stages (such as smoke and active flames) need to be distinguished. This study addresses the critical need for a comprehensive fire detection system capable of multistage classification, differentiating between non-fire, smoke, apartment fires, and forest fires. We propose a deep learning-based model using a customized DenseNet201 architecture that integrates various preprocessing steps and explainable AI techniques, such as Grad-CAM++ and SmoothGrad, to enhance transparency and interpretability. Our model was trained and tested on a diverse, multisource dataset, achieving an accuracy of 97%, along with high precision and recall. The comparative results demonstrate the superiority of the proposed model over other baseline models for handling multistage fire detection. This research provides a significant advancement toward more reliable, interpretable, and effective fire detection systems capable of adapting to different environments and fire types, opening new possibilities for environmentally friendly fire type detection, ultimately enhancing public safety and enabling faster, targeted emergency responses.

1. Introduction

Fire incidents are sudden, often catastrophic events that can occur in diverse environments, each presenting unique challenges in detection, containment, and response. These incidents involve rapid combustion of materials, resulting in intense heat, smoke, and toxic gases that spread through the air and surfaces, posing immediate and severe risks to human safety, property, and ecosystems [1,2]. The effects of fire are far-reaching; beyond threats to life; they cause extensive damage to property, destruction of infrastructure, environmental degradation, and long-term economic repercussions [3]. Fire scenarios can vary greatly depending on the environment: urban settings face risks from residential or commercial building fires, often triggered by electrical malfunctions, kitchen mishaps, or intentional acts of arson. In contrast, remote or rural areas are prone to wildfires and forest fires, commonly ignited by lightning strikes, human activity, or natural factors such as wildfire [2,4]. Therefore, a robust fire detection method with swift response is required [5].

Fire detection strategies traditionally rely on a combination of technologies, including smoke detectors, heat detectors, flame detectors, gas sensors, and CCTV monitoring, which are often integrated to improve reliability across different fire scenarios [6]. However, achieving accurate detection using these systems presents significant challenges. The main issue is the complexity involved in coordinating the network of sensors, where sensors of each type are sensitive to specific parameters, such as temperature, smoke density, and gas emissions [7]. Determining the optimal sensor placement is critical to ensure comprehensive coverage, as incorrect positioning can reduce the detection accuracy or delay alerts [8]. Furthermore, the costs associated with multisensor setups are also a concern. Integrating multiple sensor types can substantially increase expenses, both in terms of the sensors themselves and the infrastructure required to manage and interpret data from sources [9]. Additionally, environmental variations complicate the detection process; for example, smoke in cold environments may not indicate a fire, whereas elevated temperatures in desert regions may lead to frequent false alarms if not carefully recorded. These challenges collectively limit the precision, dependability, and practicality of traditional systems, which often struggle to balance sensitivity with the need to minimize false alarms [10,11]. Consequently, there is a growing demand for advanced and cost-effective detection solutions that can adapt to environmental changes and reliably distinguish the stages of fire across various settings [12].

Data from agencies such as the National Safety Council (NSC) and the National Fire Protection Association (NFPA) indicate that advancements in fire safety have led to a general decline in fire-related fatalities of 42% since 1980 which has remained a critical gap in effective fire prevention [13]. For example, 2021 marked an unexpected 33% increase in fire-related fatalities in the United States, with deaths reaching 3800, up from 2855 from the record low in 2012. This resurgence underscores the limitations of existing fire detection measures and emphasizes the pressing need for more advanced, adaptive, and precise fire detection systems capable of providing real-time insights and responses guidance. As fire events continue to cause significant economic loss and emotional problems that affect individuals and communities, the development of efficient fire detection solutions has become paramount. Diverse environments present unique challenges to traditional fire detection systems. Fires in urban settings, for example, can be exacerbated by tightly clustered buildings, increasing the potential for rapid spread, while rural and forest fires often advance uncontrollably due to vegetation, wind, and topography [14]. Addressing these complex and varied scenarios demands adaptive, multistage fire detection systems capable of real-time classification across multiple fire stages, from early smoke detection to full-blown flame—enabling targeted, environment-specific responses and enhancing public safety infrastructure across all settings.

This observation became one of our main motivations: if the system can be enhanced to identify the origin of the fire (whether in a forest, apartment, or industrial setting), firefighters can make quicker decisions and prioritize their tactics based on location [15,16,17]. The National Fire Protection Association (NFPA) in the United States estimates that a fire department receives a call to respond to a fire every 23 s. This means that every second may affect the outcome of a fire disaster. Additionally, home fires become noticeable every 93 s, home fire-related injuries occur every 47 min, and home fire-related fatalities occur approximately every 3 h and 8 min [13]. Figure 1 reveals that the majority of these fatalities are caused by house fires, underscoring the urgent need for apartment fire protection and early-stage fire detection. Despite the inherent challenge of distinguishing fires across different settings, as fire signatures often display remarkably similar visual patterns, the ability to identify fire origins in various environments can significantly improve firefighting strategies.

With advances in machine learning and deep neural networks, it has become possible to design more sophisticated and effective fire detection systems capable of addressing critical challenges during early detection, thereby reducing the potential impact of fires [18]. These systems can integrate data from multiple sources, such as images, video, and sensor data, improving detection accuracy and making fire detection less reliant on isolated sensor readings. A significant body of research has been conducted using deep learning techniques, particularly in leveraging convolutional neural networks (CNNs) for fire detection. These networks can recognize complex patterns from images and video data [19], thereby providing real-time detection of fire incidents with high precision. Furthermore, the ability of neural networks to adapt to new scenarios through training on updated data makes them suitable for a variety of environments, eliminating the dependency on specific weather or environmental conditions.

However, while deep learning offers clear advantages, previous studies have revealed several limitations that hinder its broader application in real-world fire detection systems [20]. A major issue has been the reliance on insufficient or overly simplistic datasets, which often fail to capture the diversity of fire scenarios, resulting in models that perform well under controlled conditions but struggle with generalization [21]. For example, most existing research focuses primarily on forest fire detection or binary classification (fire versus non-fire), lacking the nuanced categorization needed to distinguish between various stages or types of fire, such as smoke, apartment fires and forest fires [22]. The steps or fire-fighting apparatus for cities or forests differed slightly [23]. Furthermore, the absence of benchmark datasets that provide a balanced and extensive representation of these varied scenarios limits the accuracy and robustness of these models. Consequently, although deep learning models offer the potential for enhanced real-time fire detection, they currently face challenges in terms of data availability, scenario coverage, and the complexity needed for reliable multistage fire detection in diverse environments [24].

Following all these challenges, we created a custom dataset by combining five distinct datasets and classify the data into four categories: non-fire, smoke, flat (fire), and forest (fire). This enables the detection of a fire from its earliest stage, such as smoke, even if the fire has already begun, and, on the basis of the characteristics of the location, will alert personnel so that the appropriate plan and equipment can be applied quickly. This research paper proposes a multistage fire detection method employing a customized deep learning model based on DenseNet201 for detecting fire incidents in apartments and their surroundings. In short, the key contributions of the following research are summarized below:

1. We propose a novel real-time “situation-aware multistage fire detection” model capable of differentiating between fire types such as forest, apartment, and smoke-only scenarios. This enables responders to deploy optimized strategies and equipment based on specific fire environments.

2. Our model is built on a customized DenseNet201 architecture, enhanced with batch normalization, dropout, and ReLU activation functions. This design is intended to address overfitting, improve feature extraction, and generalize effectively across diverse fire scenarios.

3. We integrate explainable AI techniques such as Grad-CAM++ and SmoothGrad, which visually highlight regions influencing model predictions. This ensures transparency and trustworthiness in high-stakes firefighting applications.

4. We conduct an extensive performance evaluation using a comprehensive dataset, curated by merging multiple sources to include diverse fire types. We then perform a comparative analysis against baseline models, demonstrating that the proposed model outperforms existing approaches across metrics such as accuracy, precision, recall, and F1-score.

The seven parts of the paper are presented in the following order: An introduction is presented in Section 1, and evaluations of previous studies in Section 2. Section 3 details the dataset used. In Section 4, we present the technique, and in Section 5, we compare the outcomes. Section 6 presents a discussion. The paper concludes and discusses potential steps in Section 7.

2. Literature Review

2.1. Background Study

Fire poses a severe threat to both residential and commercial settings and has the potential to cause significant destruction and loss of life. Effective fire containment depends on rapid detection, underscoring the need for reliable and responsive fire detection systems. Recent advancements have introduced heuristic technologies, such as deep neural networks (DNNs) and the Internet of Things (IoT), which can autonomously detect and classify fire patterns across various scenarios [25,26]. However, the success of these IoT- and DNN-based systems depends on factors such as model design, data quality, and computation resources, as IoT systems require a well-coordinated sensor network and DNNs rely on extensive, labeled datasets. Despite their promise, current methods are often limited to specific fire indicators such as flames or smoke, making them less adaptable to different fire types and settings. Thus, identifying and addressing these limitations through further research is essential for developing flexible, multistage detection models that can provide timely and accurate alerts across diverse environments.

2.2. IoT-Based Smart City and Forest Fire Surveillance

Fire detection systems have long been a critical component of building safety, evolving from basic smoke alarms to advanced sensor-driven networks that provide early warnings and enable a rapid response. Recent advancements in technology, particularly the integration of the Internet of Things (IoT), have significantly increased the precision, efficiency, and responsiveness of these systems [27]. IoT-enabled fire detection systems utilize a network of interconnected sensors and devices to monitor fire indicators such as heat, smoke, and gas emissions. By relaying sensor data to a centralized control unit for analysis, these systems can quickly trigger alerts, sound alarms or even contact emergency services to mobilize an immediate response.

To further improve fire detection accuracy, recent research has explored multivariate calibration and validation algorithms using data from gas sensors [28]. In these systems, the authors implemented algorithms capable of detecting fires in their early stages, showing good results in enhancing accuracy and reducing false alarms. However, the reliability of these methods may be affected by the presence of other environmental chemicals that can interfere with the sensor readings. Additionally, the proposed approach is yet to be thoroughly evaluated with a broader array of sensor types, such as temperature or smoke detectors, to determine their adaptability and precision in complex, real-world settings. Another study explored optimization-based methods for enhancing outdoor temperature fire detection systems with the aim of reducing both costs and setup time. One approach proposes a mathematical model that incorporates geographic, climatic, and vegetation factors to determine the optimal sensor arrangement rather than evenly distributing sensors across an area [29]. This model not only reduces the number of sensors required but also lowers the false alarm rates. However, because the study was conducted in a single region, its applicability to other landscapes and climates remains uncertain. Moreover, although this approach improves detection efficiency, it does not offer precise information about the exact location or progression of a fire. To address forest fire prediction, other researchers have integrated artificial intelligence (AI) with IoT sensor networks, deploying IoT sensors in forested areas in combination with YOLOv5 to increase detection accuracy and reduce false alarms [7,30]. Despite these improvements, these systems lack critical data, such as fire location and intensity information, which are vital for effective firefighting responses.

Baba et al. introduced a novel fire detection and alarm system using Arduino-based multisensory technology, GSM communication, an RF module, and an Android application designed for quick and reliable fire detection in diverse settings, including residential, commercial, and public spaces [31]. While effective in detecting fires, the accuracy of the system can be affected by environmental conditions, such as wind and humidity, which influence smoke dispersion and temperature. Furthermore, its indoor-specific design limits its effectiveness in large-scale environments, such as forests and industrial facilities. In another study, Dampage et al. proposed a forest fire detection system using wireless network sensor nodes and machine learning (ML) algorithms to provide real-time monitoring of ambient temperature and humidity, aiming to detect signs of forest fires based on environmental fluctuations [32]. However, the effectiveness of this approach may be compromised by external factors, such as wind or precipitation, which can alter temperature and humidity distributions potentially leading to misclassification.

Recent studies have thoroughly assessed advanced fire detection sensors, including smoke, heat, gas, and flame detectors, providing insights into their functionality, applications, and specific limitations. These advancements in sensor technology have substantially enhanced the accuracy of fire detection systems, offering improved reliability for public safety and reducing the risk of property damage [33]. Despite these advancements, IoT-based fire detection systems present inherent challenges. While they enable comprehensive data collection and real-time monitoring, their high installation and maintenance costs, combined with the complexities of managing multisensor data streams, often limit scalability and practicality in various environments [34]. Additionally, IoT-based systems can be sensitive to environmental changes, which may affect the sensor accuracy and lead to increased false alarms [11]. Thus, although IoT integration holds promise for enhanced detection capabilities, these systems require further refinement to overcome the cost and adaptability limitations in real-world applications.

2.3. Deep Learning-Based Approach for Multimedia Surveillance During Fire Emergencies

In recent years, there has been interest in fire detection systems that utilize deep learning techniques. This is because they have the ability to achieve remarkable levels of accuracy and resilience in detecting fires [35,36].

Accordingly, this paper presents YOLOv5s-CCAB, an enhanced forest fire detection model that extends the capabilities of YOLOv5s. The model seeks to address the issue of low detection accuracy resulting from the diverse scales and geometries of forests fires [37]. The empirical results demonstrate that the model presented in this study can achieve a mean average precision (mAP) of 87.7%. Although the YOLOv5s-CCAB model improves accuracy by addressing thescale and geometry diversity in forest fires, it is constrained by its binary classification approach, which limits its adaptability to various fire stages or types. Fire detection was conducted using the six-layer CNN architecture presented in [38]. The accuracy of the selected dataset analyzed via the aforementioned arrangement was greater at 89.64% than that of the alternative approaches. However, its high error rates raise concerns regarding its effectiveness in real-world settings. The limited dataset of the model may not capture the full diversity of residential fires, suggesting that larger, more varied types of data collection are necessary to improve reliability in deployment scenarios.

The proposed system uses picture luminance and a novel convolutional neural network that incorporates an enhanced YOLOv4 model with a convolutional block attention module [39]. The YOLOv4-based system with a convolutional block attention module excels in terms of recall and precision. It is primarily trained on internal flame images. This narrow scope of the dataset limits its effectiveness in outdoor or large-scale fire environments, where flame patterns and visual cues differ significantly from those in indoor environments. Abdusalomov AB et al. introduced an innovative approach to categorize forest fires by employing an improved version of the Detectron2 platform and implementing advanced deep learning techniques [40]. The suggested model can accurately identify small fires from long distances, regardless of whether they are in the daytime or nighttime. The experimental findings revealed that the proposed strategy for detecting forest fires achieved a 99.3% increase in precision in detecting fires. However, it relies heavily on a highly preprocessed custom-labeled dataset. This dependency can reduce the adaptability of the model to new or unforeseen forests under fire conditions.

This study utilized indoor closed-circuit television (CCTV) monitoring to create a computer vision-based model for detecting fires at an early stage (EFDM) [41]. EFDM has been proven to have a shorter response time than smoke and heat detectors for fire detection. This offers rapid detection, which is limited by its sensitivity to specific materials involved in fire incidents. Certain substances may delay visible signs (e.g., flames or smoke) that the model relies on for detection, potentially affecting the response time and undermining early detection in scenarios with slow-burning materials. Another study introduced a forest firefighting system that utilized unmanned aerial vehicles (UAVs) in conjunction with artificial intelligence (AI) to consistently monitor and detect fires in forest areas [42]. They suggested a transfer learning technique that exploits the VGG19 model to increase the accuracy. The method achieved an average classification accuracy of 95%. While the forest firefighting system leveraging UAVs and AI achieved high accuracy with the DeepFire dataset, a notable limitation is the dependency on favorable weather conditions. UAVs face operational challenges in adverse weather, such as strong winds, rain, or fog, which can limit the visibility and the effectiveness of aerial monitoring and detection. This study introduced YOLOv4, a fire detection system that utilizes convolutional neural networks (CNNs) and is specifically designed for low-power devices with limited resources [43]. Their approach involves training a deep detection network and subsequently eliminating the less significant convolutional filters to decrease the computational burden while maintaining the initial performance. They showed that it is feasible to remove as many as 83.88% of the network parameters and decrease the computational cost (measured in BFLOPs) by up to 83.60%, all while preserving network performance. By pruning convolutional filters to minimize the computational load of the model, the system may lose nuanced detection capabilities, potentially compromising its accuracy in complex fire scenarios where details are crucial.

Four pretrained ResNets, namely, ResNet18, ResNet50, ResNet101, and InceptionResNetV2, were utilized in this investigation as documented in another study [44]. A support vector machine (SVM) classifier was used to classify the data. ResNet models, utilizing an SVM classifier and a 10-fold cross-validation methodology, achieved accuracies ranging from 98.91 to 99.15%. The main constraint of the proposed model is the insufficient sample size of images used during its development. Despite achieving high accuracy, this ResNet-based approach suffers from an insufficient sample size, which limits the generalizability of the model. Another suggested alternative strategy for voting ensembles involves integrating the YOLO architecture with two weights using the CNN architecture [45]. The classification model achieves an F1-score of 0.95. The model uses transfer learning to categorize the data. The evaluation of the detector model showed positive outcomes, with a mean average precision (mAP) score of 0.85 for the smoke detection model and 0.76 for the combined model. The ensemble model integrating YOLO with CNN weights achieved high accuracy for smoke detection but had a lower mean average precision (mAP) for combined fire scenarios. This limitation suggests that although the model is effective for smoke, it may struggle with fire instances involving complex visual cues, potentially limiting its reliability across diverse fire types.

2.4. Overall Findings

Most of the resources we examined focused on forest fire detection, although we did identify a few that were developed with cities in mind. We further broadened our literature review to include more aspects to distinguish between how flat and forest fires are handled. Fighting blazes on a flat and in a forest has certain commonalities, but the environment and goals require different approaches.

In the case of a fire in a multifamily building, firefighters must act quickly to contain flames and preserve their lives [46]. Firefighters often use equipment such as hoses, ladders, and extinguishers in addition to a water source such as a fire hydrant to douse blazes [47]. The construction of the structure makes matters more difficult in several ways, such as narrow hallways and stairwells and the possibility of building collapse. The primary goal of forest firefighting is to contain fire [48]. This is because forest fires may swiftly spread out of control, wreaking havoc on valuable ecosystems and threatening human lives. Some of the tactics utilized by firefighters include digging fire lines, conducting backburns, and using aircraft drop water or fire retardants from the sky. Forest firefighting may be difficult because of factors such as steep terrain, variable wind patterns, and high temperature and humidity. Because the equipment used to extinguish a fire may differ depending on whether it is in a city or a forest, it is crucial to identify the fire site characteristics in real-time, which is absent in all the literature we reviewed.

After reviewing the relevant literature, we concluded that sensor-based fire detection offers several advantages, such as the potential to respond more rapidly in certain situations, the need for an optimal combination and placement of sensors, the requirement of well-trained people to monitor the chain of sensors, and the additional cost associated with a larger number of sensors being in place to keep an eye on a single location. In contrast to the limitations of IoT-based systems, deep learning models have shown great promise for fire detection owing to their ability to understand complex patterns and attributes from large datasets. Because these models can automatically extract information from images or videos to distinguish fires from non-fires, they are accurate and reliable. The benefits that may be realized by using a deep learning model for fire detection include high accuracy, adaptability to different scenarios, real-time alerts, a reduction in false alarms, and integration with other technologies. We have demonstrated, however, that the vast majority of research relies on binary classification and that utilizing a benchmark dataset with a large sample size is out of vogue. In addition, the majority of previous studies on deep learning systems have focused on forest fire detection. Regrettably, there is a lack of extensive research on real-time multistage fire detection which focuses on identifying the exact location of fire incidents and resolving classification challenges in a sufficiently large benchmark dataset. To achieve high accuracy in multistage fire detection, it is necessary to utilize a suitable deep learning technique in conjunction with a strong dataset. Therefore, we were encouraged to perform this multistage categorization of fire detection which considers the user’s current position. To solve this multiclass problem, we presented a specialized DenseNet201-based model and evaluated it using a reference dataset.

3. Dataset

To accurately identify smoke levels and determine the exact locations of fire incidents, a dataset that includes categories beyond only “fire” and “non-fire” is necessary. To develop a fire detection system capable of accurately differentiating between various types of flames, for example, forest fires and apartment fires, we classified the fire classes separately. This would allow emergency services to respond promptly and appropriately in each scenario. Forest fires present considerable difficulties because of their extensive scope and capacity for swift propagation, whereas apartment fires demand immediate action to avert casualties and the destruction of property in densely inhabited regions. While there is a feature available to determine the coverage area of the cameras, properties located outside the town are encompassed by trees or a small woodland. Therefore, a solitary camera can capture both the scenarios within a certain location. This incentive emphasizes the importance of precisely categorizing various fire types for efficient disaster control and mitigation techniques. In addition, we had to amalgamate and integrate the data from several sources. The choice to combine various datasets to form a “multistage” fire detection dataset was motivated by the necessity of providing a thorough and varied dataset that encompasses a broad spectrum of fire situations and environmental circumstances. Every dataset can possess distinct attributes, such as disparities in image quality, lighting circumstances, and types of fires, which can enhance the training data and resilience of the model. By combining datasets from several sources, we can increase the ability of the model to make predictions for new data and adapt to changing real-world situations. In addition, the combined datasets enable us to utilize the advantages of each individual dataset and minimize any potential biases or limits that may exist in a single dataset. In Table 1, we provide specific information regarding the datasets we have gathered.

To achieve this goal, we combined sections from several databases. The first was a reference dataset consisting of forest fire and non-fire photos. A significant number of complete data points were used in this manner. For the second dataset, we only used smoke training. The images for the forest and flat categories were extracted so that we could use the fire data points from the third and fourth datasets. We narrowed the fifth dataset to fire photographs only. Figure 2 and Table 2 illustrate how we classified the images in our final dataset as “non-fire”, “smoke”, “apartment” (fire), or “forest” (fire). Here, the dataset is divided into 70% for training, 20% for validation, and 10% for testing, which is a widely accepted approach in machine learning, chosen to ensure that the model has adequate data for learning and provides a robust basis for evaluation. Assigning 70% of the data to training allows the models to learn complex patterns across various fire stages, such as smoke, apartment fires, and forest fires, which is essential for an accurate multistage classification. The 20% allocated for validation serves to monitor and tune the model during training, helping to detect issues such as overfitting and allowing adjustments to the hyperparameters to improve generalization. Finally, the 10% data portion is dedicated to testing and is used exclusively for the final evaluation, ensuring an unbiased measure of model performance on unseen data. This supports both model accuracy and reliability, aligning with standard practices in deep learning studies, particularly for applications requiring high precision, such as fire detection.

The term “neutral” refers to images of the woods, apartment blocks and roads that were not affected by the fire. A smoke class is necessary for early-stage fire detection when several smoke images from various locations are available. If smoke is detected quickly enough, the fire may be contained and, more significantly, lives can be saved. The remaining types are used to spot fires and collect broad contextual data, such as where the blaze is located (forest or structure). This may be especially significant in the suburbs, where many houses are surrounded by greenery. Decisions on firefighting equipment and, more importantly, the sequencing of approaches to battle flames may be made more quickly with the help of this tool.

4. Methodology

The Keras and TensorFlow libraries were used to train all the baseline models in Google Colab, and the models’ input sizes, epoch batch sizes, augmentation parameters, learning rates, optimizer activation functions, etc., varied widely among the datasets. The steps below are the details of the framework of our proposed strategy, which are also illustrated in Figure 3.

4.1. Image Pipeline

4.1.1. Data Augmentation

Data augmentation plays a pivotal role in enhancing the ability of the model to generalize the training data, particularly in image classification tasks. For the fire detection dataset, we employed a series of augmentation techniques designed to simulate various environmental conditions and perspectives that the model might encounter in real world scenarios. The augmentation parameters are as follows:

• Rescaling the pixel values to a normalized range of 0–1 improves the model conversion to during training.

• Random width and height shifts of up to 20%, introduce spatial variability positioning of fire and smoke in the images.

• Images were rotated within a range of 2 degrees to mimic slight angular variations in camera positioning.

• A zoom range of 10%, simulating varying distances from the fire sources.

• Random horizontal and vertical flips, representing different orientations of the captured scenes.

These augmentations not only diversify the training set but also aim to mimic real world variations, thereby boosting the model’s robustness and performance under diverse conditions.

4.1.2. Preprocessing and Generators

The fire detection dataset was organized into distinct sets for training, validation, and testing. To facilitate efficient image handling and preprocessing, we utilized the TensorFlow Keras ImageDataGenerator. This generator serves as a powerful tool for real-time data feeding to a neural network without the need to store augmented images in the memory.

4.1.3. Training Set

For the training set, which comprised 70% of the total data, ImageDataGenerator was configured with the data augmentation parameters outlined earlier. It performs the following operations:

• Load images from the specified training directory.

• Resize images to $224\times 224$ pixels to ensure uniformity in input size.

• The defined augmentation transformations are applied randomly to each batch.

• The dataset is shuffled to ensure that each training batch has a random distribution of images.

The batch size was set to 16, balancing the trade-off between the training speed and memory usage.

4.1.4. Validation and Test Sets

The validation and test sets constituted 20% and 10% of the total data, respectively, and the ImageDataGenerator performed minimal processing:

• Only rescales the pixel values to normalize them.

• Loads and resizes images to $224\times 224$ pixels.

• It does not apply any augmentation transformations, ensuring that the evaluation of the model is based on unaltered data.

Shuffling is disabled for these sets to maintain consistency in the evaluation metrics and testing outcomes.

4.1.5. Dataset Categories and Class Count

The dataset is categorized into four classes: ‘apartment’, ‘forest’, ‘non-fire’, and ‘smoke’. These categories are crucial for the model to identify and differentiate between various types of fire and smoke scenarios, as well as normal conditions without fire. The identification of these classes by the training generator is essential for configuring the softmax output layer of the neural network. The dataset comprised four classes, corresponding to these defined categories.

The structured approach to data augmentation and preprocessing ensures that the model is trained on a rich, varied dataset, enhancing its ability to classify images in real-world fire detection scenarios.

4.2. DenseNet201

DenseNet, particularly the DenseNet201 variant, represents a major advancement in deep learning architectures. As described by Huang et al. [54], DenseNet201 is renowned for its dense connectivity patterns. This architecture employs a mechanism in which every layer receives inputs from all the layers that come before it, thereby improving the capacity of the network to acquire feature representations. DenseNet201 is structured such that:

(1)$x_l=H_l\left([x_0,x_1,\dots ,x_{l-1}]\right)$

where $x_l$ is the output of the $l^{th}$ layer, and $H_l$ represents a series of operations (Batch Normalization, ReLU, Convolution), and $[x_0,x_1,\dots ,x_{l-1}]$ denote the concatenation of all previous layer outputs. This structure facilitates an improved gradient propagation and reduces the vanishing gradient problem, making the network more efficient and easier to train.

4.3. Custom Layer

To adapt the DenseNet201 model to specific tasks, several custom layers were integrated. These layers are designed to process and refine the feature maps from DenseNet201 for targeted classification tasks. The custom layers include the following:

4.3.1. Global Average Pooling Layer (GAP)

Global average pooling (GAP) is critical for reducing the spatial dimensions of feature maps, thereby concentrating the network’s focus on the most salient features. The GAP layer functions as follows:

(2)$GAP\left(x\right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N{x}_i$

This layer simplifies the model by reducing the total number of parameters, thereby helping combat overfitting.

4.3.2. Dense Layer with L2 Regularization

Dense layers in the network are crucial for learning high-level features and patterns. Each of these layers includes L2 regularization to penalize large weights, thereby encouraging the model to learn more generalized features:

(3)$f\left(x\right)=\sigma (Wx+b)+\lambda \sum _{i=1}^n{w}_i^2$

The regularization term $\lambda {\sum }_{i=1}^n{w}_i^2$ adds a penalty proportional to the magnitude of the weights, thereby reducing overfitting.

4.3.3. Batch Normalization and Dropout

Batch normalization and dropout were integrated to improve the generalizability of the model capabilities. Batch normalization ensures that the output of each layer is standardized, whereas dropout selectively deactivates a portion of the activations by setting them to zero, thus preventing the network from excessively depending on certain neurons:

(4)$BN\left(x\right)=\gamma \left({\displaystyle \frac{x-\mu }{\sqrt{{\sigma }^2+ϵ}}}\right)+\beta$

(5)$Dropout\left(x\right)=x\odot M$

This combination effectively reduced the internal covariate shift and overfitting, leading to more robust learning.

4.3.4. Rectified Linear Unit (ReLU)

This method applies the rectified linear unit (ReLU) activation function to the proposed DenseNet201-based model only. The choice of ReLU for dense layers was intentional, as it introduces nonlinearity, allowing the model to learn complex patterns effectively by addressing issues such as the vanishing gradient problem, which is particularly crucial in deep networks. The ReLU is widely recognized for its computational efficiency and ability to speed up convergence, which is essential for high-accuracy, real-time fire detection. The baseline models used in our comparisons retained their default configuration ratios, including activation functions, because our primary focus was on the enhancements and customizations applied to our proposed model. By specifying the use of ReLU in the proposed model, we aim to provide readers with a clear understanding of the modifications introduced to optimize its performance in multistage fire detection tasks. Overall, ReLU activation gives the network nonlinearity, as Nair and Hinton [55] noted, allowing it to pick up intricate patterns:

(6)$ReLU\left(x\right)=max(0,x)$

It is preferred because of its ability to accelerate the training process without significantly affecting the ability of the network to model complex relationships.

4.3.5. Softmax Layer for Classification

For multiclass classification problems, the softmax layer which is the last layer, is crucial. It transforms the output logits into a probability distribution across the expected output classes:

(7)$Softmax\left(x_i\right)={\displaystyle \frac{e^{x_i}}{{\sum }_j{e}^{x_j}}}$

This layer is crucial for obtaining interpretable classification results from the network.

4.3.6. Model Overview

The model is composed of the following elements:

• DenseNet201, pretrained on the ImageNet dataset, provides a rich set of initial features.

• Three custom dense layers with 512, 256, and 8 units each, employing ReLU activation and L2 regularization (coefficient = 0.001), were used to further refine and learn from the feature maps.

• Batch normalization and dropout (rate = 0.4) after each dense layer stabilized the learning and mitigated overfitting.

• A Softmax layer was used at the end to classify the inputs into multiple categories.

4.3.7. Training Approach

To achieve model convergence and optimize efficiency, we utilized the Adam optimizer with a learning rate of $1\times {10}^{-4}$, which was determined through several trial and error attempts. Categorical cross-entropy is incorporated as a loss function in the training approach. The dataset, comprising both unaltered and improved images, was meticulously selected and arranged with a batch size of 16 to ensure stability throughout the training procedure.

To address overfitting, early stopping was implemented, halting the training process if the model’s performance did not improve for a consecutive period of three epochs. Additionally, the model checkpoints were utilized to save the best model based on its validation loss. The training parameters, outlined in Table 3, involved utilizing a range of cutting-edge models that were meticulously adjusted to achieve outstanding results on the NVIDIA Tesla T4 GPU.

A condensed training technique consisting of 100 epochs was implemented, focusing on maximizing the efficiency and allowing for the possibility of early stopping. By utilizing numerous workers and harnessing the computing power of the NVIDIA Tesla T4 GPU on Google Colab, we accelerated the process, guaranteeing fast and efficient model optimization.

4.4. Explainable AI

We used a DenseNet201-based model that was specifically trained to recognize and interpret fire and smoke to generate visual explanations for its predictions. The algorithm underwent training using data obtained from residential complexes, forests, non-fire environments, and artificially generated smoke. To provide contextual information for its predictions, we employed Grad-CAM++ and SmoothGrad techniques.

Grad-CAM++ is an enhanced version of the original Grad-CAM methodology that is employed in computer vision research to visually represent and comprehend decisions made by deep neural networks [56,57]. It provides more precise information regarding the components of an input image that impact the network’s ultimate conclusion. The following is a comprehensive equation for Grad-CAM++:

(8)$w_k^c=\sum _i\sum _j{\alpha }_{ijk}^c·\mathrm{relu}\left({\displaystyle \frac{\partial A_{ijk}}{\partial Y_c}}\right)$

(9)$L_{ij}^C=\sum _k{w}_k^C·A_{ijk}$

The notations used in the text are as follows: $w_k^C$ represents the weights of a neuron, $A_k$ represents the visualization of the kth feature map, and ${\alpha }_{ijk}^c$ are the weighting coefficients for the pixelwise gradients for class c and the convolutional feature map $A_k$, ${\displaystyle \frac{\partial A_{ijk}}{\partial Y_c}}$ is expected to have high values for feature map pixels that contribute to the presence of the object, and ReLU refers to the rectified linear unit activation function. In addition, $L_{ij}^C$ directly indicates the importance of a specific spatial location (i, j) for a particular class c.

Grad-CAM++ incorporates second-order gradients in the computation of importance weights assigned to each spatial position. The model subsequently utilizes these weights to multiply them by the activations of the feature map. This allowed the model to identify the specific areas of the input image that were most relevant to the predicted class. The ultimate visualization was achieved by combining and standardizing the weighted feature map activations.

In addition, the smooth gradient equation, which is frequently used in tandem with Grad-CAM or Grad-CAM++, reduces noise and improves the visual interpretability of the generated heatmaps [58]. During the backpropagation procedure, a regularization term was applied to the gradients. The equation used is as follows:

(10)${\widehat{M}}_c\left(x\right)={\displaystyle \frac{1}{n}}\sum M_c(x+N(0,{\sigma }^2))$

Variable n represents the number of samples, whereas $N(0,{\sigma }^2)$ denotes Gaussian noise with a standard deviation of $\sigma$.

SmoothGrad is a simple technique that randomly samples points near input x and then calculates the average sensitivity maps of these points. This technique enhances the visual clarity of sensitivity maps based on gradients.

We started by loading the pretrained DenseNet201-based model and providing class labels. Then, we specified the image repositories for each group. For each category, we selected and preprocessed a single image from the corresponding folder to construct a description. We then recorded both the actual and predicted photo labels. This provided more precise and correct explanations for the selected images. The results from using the Grad-CAM++ and SmoothGrad techniques are similarly informative, as presented in the Section 5 and Section 6.

5. Results

To validate the effectiveness of our approach, we developed a specialized DenseNet201-based model tailored for multistage fire classification and benchmarked it against several existing models, including YOLO-based architectures, as outlined by Jocher et al. (2023) [59]. The evaluation was based on key performance metrics, such as accuracy, precision, recall, and F1-score. Our proposed model demonstrated marked improvements over the baseline models across all metrics, achieving a test accuracy of 0.97, and precision and F1-scores of 0.94, which underscores its robustness and reliability in classifying different fire stages. This high level of performance not only outperforms the original DenseNet201 but also outperforms YOLO and other benchmark models, and exhibits limitations in recall and precision—critical metrics in minimizing false alarms in safety-critical applications. Our customized adjustments to DenseNet201, including the addition of batch normalization, dropout, and ReLU activations, effectively mitigated common issues such as overfitting, which was evident in the baseline models. These enhancements allowed our model to provide more stable and accurate classifications across different fire scenarios, as shown in Figure 4. This comparative analysis highlights the substantial gains in performance achieved by our approach, reinforcing its viability and superiority in high-risk fire detection environments.

The proposed DenseNet201-based solution significantly outperformed all baseline models across key evaluation metrics, as illustrated in Figure 5. A detailed analysis using confusion matrices revealed that our model achieved superior classification accuracy, with the primary diagonal of the matrix exhibiting high precision in its predictions across various fire stages. This consistency in correctly classifying the majority of images underscores the robustness of the model, particularly when compared with competing models that displayed lower precision on similar tasks.

Furthermore, Figure 6, Figure 7 and Figure 8 provide a comprehensive breakdown of our model’s performance, highlighting the accuracy, loss, precision, recall, and area under the curve (AUC) metrics. Our model maintained stable training and validation loss profiles, indicating minimal overfitting, while consistently achieving higher recall and precision scores than alternative models. The AUC values further demonstrate the model’s effective classification capabilities, reinforcing its reliability across all stages of fire detection. This comparative analysis, which is based on visual evidence from these performance metrics, demonstrates that our tailored DenseNet201 architecture yields substantial improvements over other approaches, making it particularly effective for multistage fire detection applications.

Our model demonstrated exceptional performance, outperforming competing models as evidenced by its high accuracy, consistent validation loss, and high AUC values. A key indicator of this improvement is the alignment between the training and validation accuracy curves, which significantly increases the validation accuracy and underscores the model’s strong generalization ability. Unlike many baseline models that tend to exhibit discrepancies between training and validation curves owing to overfitting, our model maintains a stable validation loss that is slightly lower than the training loss, indicating effective learning without overfitting. The precision and recall graphs, analyzed in detail, further highlight the superior performance of our model. Notably, the validation precision and recall values surpassed those of the training set, which is a clear sign that the model generalizes well to new data points. This is a substantial advantage over baseline models, which often show reduced precision and recall on the validation data, leading to more false positives and false negatives in practice. Additionally, the AUC metric, presented in Figure 8, confirms the model’s excellent classification capabilities across all fire stages, with values close to 1, reflecting its high discriminatory power.

The AUC graph indicates that our model attained a commendable AUC value, which was in close proximity to 1. The capacity to classify across numerous categories can be determined by observing a greater area under the curve (AUC) number. Consequently, our proposed framework classifies data exceptionally well.

In addition, we compared the results of our model with those of various other models, as presented in Table 4. Table 4 provides an overview of the performance evaluation results. Accuracy (0.97), precision (0.94), recall (0.94), and F1-score (0.94) were considered, and our model was superior to the competition. We focused on a set of keys, including the metrics test accuracy, precision, recall, F1-score, and loss values, to offer a clear and balanced evaluation of each model’s performance in multistage fire detection. The test accuracy, precision, recall, and F1-score were selected because they capture critical aspects of the model and classification effectiveness: accuracy indicates overall correctness, whereas precision and recall balance the model’s sensitivity and specificity, helping to avoid false positives and false negatives. The F1-score combines precision and recall to obtain a single measure of balance between the two, which is particularly relevant for a multistage classification problem. The values add further depth to the comparison by showing how well each model minimizes prediction errors. Additional metrics such as training accuracy, validation precision, validation recall, validation loss, and loss function details could add context and introduce complexity without necessarily enhancing the interpretability or practicality of the results. Training accuracy and validation metrics often fluctuate during the model training and tuning phases and may not reliably reflect the final model quality. Focusing on test accuracies and other core metrics provides a more meaningful measure of each model’s ability to generalize and perform effectively in real-world scenarios. This focused selection highlights the strengths of the proposed model, particularly its ability to generalize, and outperforms baseline models across all critical evaluation metrics. By centering on tests and loss metrics, we provide a straightforward and informative comparison that demonstrates the proposed model has robust performance and accuracy, which is crucial for the application of reliable fire detection methods.

To thoroughly evaluate our model’s interpretability and decision-making process, we utilized the Grad-CAM++ and SmoothGrad techniques, which provided visual explanations for each prediction. These methods facilitate a deeper understanding of the model’s efficacy in classifying fire images and support more confident decision-making regarding its performance. Specifically, our model aimed to detect and classify fire occurrences in images across categories such as “apartments”, “forests”, “non-fires”, and “smoke”. Figure 9 presents a selection of images, showcasing both the original input and visual outputs from Grad-CAM++ and SmoothGrad. This visualization highlights the critical areas of each image upon which the model relied to make its classification decisions, with the projected and actual labels provided for clarity. By employing these explainability techniques, we gained valuable insights into the model’s internal decision-making ability, verifying that it focuses on relevant features, such as smoke patterns and flames.

Our findings confirm that the DenseNet201-based model accurately distinguished between categories, effectively reducing misclassification rates, particularly between forest fires and apartment fires. This improvement over the baseline models demonstrates that the model’s enhanced ability to capture context-specific details is essential for multistage fire classification. However, as shown in Figure 10, misclassifications can still occur in cases where buildings are in forested settings or structures resembling forests. These instances reveal a limitation of the model, as the inherent complexity of real-world scenes can challenge its ability to accurately classify certain scenarios based solely on visual cues. The occasional difficulty of the model in differentiating between forest fires and urban settings with building-like structures highlights the need for further research into contextual learning and spatial awareness. Addressing these limitations could involve integrating spatial and contextual information, such as scene layouts or map-based characteristics, and diversifying the dataset to include more mixed scenarios. These additions may enable the model to capture subtle contextual signals that are crucial for accurately classifying complex scenes. These improvements are essential for refining the performance of the model in challenging environments, thereby increasing its reliability and versatility for real-world applications.

6. Discussion

The omnipresent threat of fires continues to jeopardize life and property, necessitating swift and accurate detection systems. In response to this pressing need, our research endeavors introduced a refined approach to fire detection through a specialized DenseNet201-based model tailored for multistage fire categorization. The reasons for choosing DenseNet201 were its high accuracy, low loss and precise performance in robust and sensitive situations. The model architecture, built upon the DenseNet201 model with augmented layers for precision classification, was developed via comprehensive methodologies encompassing data augmentations, preprocessing, and leveraging explainable AI techniques. This strategy not only increased the variety of data used for training, but also enhanced the model’s capacity to understand intricate patterns and avoid overfitting, which is essential for accessing curated fire detections in different contexts. With respect to the model architecture, the adaptation of the DenseNet201 model, which was 679, enhanced the efficiency and computational speed through depth-wise separable convolutions. Additional customized classification layers with rectified linear units (ReLUs), batch normalization, and dropout regularization facilitated the learning of complex representations, preventing overfitting and enhancing generalizability. Our modified method significantly improved upon the existing methods, especially in handling complex urban environments, and is compatible with diverse fire scenarios. The evaluation confirms the model’s effectiveness in actual fire detection scenarios by highlighting its remarkable accuracy, precision, recall, and AUC metrics. However, expanding on the practical implications can be performed in the future so that the proposed method can be enhanced not only for residential settings but also for industrial settings. In addition, multiple data sources were utilized to depict the complexity of the congested urban environment and the diversity of building information here. However, if we look to address these unique challenges further, additional fire-related data, such as street-level imagery, building information, or other geospatial data can be collected via our fire detection model [60]. Certain mapping platforms, including Mapillary, Tencent, and Baidu Maps, or any relevant mapping services, may be integrated into [61]. Using these mapping platforms, we can retrieve current or real-time imagery of the targeted areas, such as forests and apartment complexes, for indoor navigation. However, the availability and accessibility of these additional data sources may vary by country or region. In general, engineering applications necessitate customized solutions that consider the distinct problems and the requirements of fire detection systems. The suggested model effectively meets these objectives by employing customized and explainable AI techniques.

7. Conclusions

Fire poses a significant danger to human life and property because of its ability to spread rapidly and cause widespread destruction. Early identification and prompt action are crucial for preventing the development of flames and minimizing their impact. The use of deep neural networks for fire detection can potentially increase the accuracy and speed of these systems. Nevertheless, we encountered difficulties in obtaining the necessary resources for constructing the deep learning model and amassing a sufficiently sized dataset. Consequently, it is imperative to merge the collected datasets. Insufficient resources also result in significant delays in conducting tests and using various methods. By employing DenseNet201, we successfully introduced a cutting-edge deep learning model that surpassed its competitors. The proposed approach can be advantageous for both commercial and residential buildings, as well as public and industrial areas. The constraints of the studies are as follows: Initially, it was necessary to conduct on-site trials because the proposed methodology has only been proven through simulations. Furthermore, this inquiry specifically concentrates on the fire phenomenon itself, rather than its size or the presence of smoke, to obtain more accurate data. It would have been wise for the proposed method to consider the building layout. Nevertheless, future requirements can be fulfilled by increasing the level of exertion. Considering the geometric layout of the apartments for future research, high-speed devices should be integrated to increase the overall efficiency of fire detection systems. Implementing this approach could lead to automation of firefighting with a significant level of accuracy.

Author Contributions

Conceptualization, M.S.C. and T.S.; methodology, M.S.C. and T.S.; software, M.S.C. and T.S.; validation, M.F.M. and N.D.; formal analysis, M.F.M. and N.D.; investigation, M.S.C. and T.S.; resources, T.S.; data curation, M.S.C. and T.S.; writing—original draft preparation, T.S. and M.S.; writing—review and editing, M.S.C., M.S. and N.D.; visualization, T.S.; supervision, M.F.M.; project administration, M.S.; funding acquisition, M.S. and N.D. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Owing to security considerations, the datasets generated and/or analyzed during this study are not publicly available. However, they can be requested by the corresponding author under reasonable conditions.

Acknowledgments

The authors extend their appreciation to King Saud University for funding this research through Researchers Supporting Project Number (RSPD2024R1027), King Saud University, Riyadh, Saudi Arabia.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:

AI	Artificial Intelligence
CNN	Convolutional Neural Network
ROC	Receiver Operating Characteristic
AUC	Area Under Curve
XAI	Explainable AI
GPU	Graphics Processing Unit
ReLU	Rectified Linear Unit
Grad-CAM	Gradient-weighted Class Activation Mapping
SVM	Support Vector Machine
ML	Machine Learning
DL	Deep Learning
BN	Batch Normalization
TPU	Tensor Processing Unit
LR	Learning Rate
TP	True Positive
FP	False Positive
TN	True Negative
FN	False Negative

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Total civilian fire deaths and home fire deaths, United States, 1977–2021 [13].

Figure 2. Sample images of each class from the merged dataset.

Figure 3. The figure illustrates the proposed approach for detecting fires in different stages including preprocessing, pretrained DenseNet201 and multiple custom layers.

Figure 4. The performance of each model is simply summarized in this bar chart.

Figure 5. Confusion matrix of different models.

Figure 6. The graphic displays the training and validation accuracies on the left side, and the training and validation losses on the right side.

Figure 7. The left side of the figure shows the graph for training and validation precision, whereas the right side shows the graph for training and validation recall.

Figure 8. This graphic illustrates the efficacy of the AUC in training and validation.

Figure 9. Explainable AI for all classes.

Figure 9. Explainable AI for all classes.

Figure 10. Explainable AI for misclassified data.

Table 1

Details of the collected dataset.

Name	Classes	Number of Images	Short Details of Dataset
1st Forest image dataset [49]	Fire, Non-fire	Fire: 1104, Non-fire: 2404	Forest images of fire and non-fire
2nd Smoke 100 k [50]	Smoke, Non-smoke	Smoke: 9762, Non-smoke: 15,000	Random images with smoke and without smoke
3rd Fire dataset [51]	Fire-images, Non-fire-images	Fire-images: 755, Non-fire-images: 244	Forest and apartment images of fire and non-fire
4th Fire detection dataset [52]	0, 1	Fire classified as 1, Non-fire classified as 0	Fire and non-fire classification
5th Fire and gun dataset [53]	Object (Fire and Gun)	7829	Images of fire and gun

Table 2

Number of images in different categories.

Category	Training (70%)	Validation (20%)	Testing (10%)	Total
Non-fire	2318	666	330	3314
Smoke	6896	1948	983	9827
Apartment (fire)	1314	389	191	1894
Forest (fire)	1957	566	279	2802
Total	12,485	3569	1783	17,837

Table 3

Training Specifications.

Parameter	Value
Optimizer	Adam
Learning Rate	$1\times {10}^{-4}$
Loss Function	Categorical Cross-Entropy
Batch Size	16
Epoch	100
Early Stopping Patience	3
GPU	NVIDIA Tesla T4
Libraries	Keras, TensorFlow

Table 4

Performance evaluation of the applied models with different accuracy parameters.

CNN Method Name	Accuracy	Loss	Precision	Recall	F1-Score
ConvNeXtTiny	0.887	0.286	0.812	0.806	0.803
ResNet152V2	0.942	0.151	0.897	0.899	0.897
VGG19	0.948	0.141	0.903	0.901	0.902
NASNetLarge	0.951	0.245	0.913	0.912	0.912
InceptionV3	0.955	0.108	0.916	0.918	0.917
InceptionResNetV2	0.956	0.171	0.925	0.925	0.921
Xception	0.958	0.096	0.918	0.921	0.919
Yolov8n-cls	0.959	0.076	0.921	0.920	0.919
DenseNet201	0.960	0.128	0.927	0.919	0.922
Proposed	0.970	0.164	0.943	0.943	0.943

References

1. Yeremenko, S.; Sydorenko, V.; Andrii, P.; Shevchenko, R.; Vlasenko, Y. Existing risks of forest fires in radiation contaminated areas: A critical review. Ecol. Quest.; 2021; 32, pp. 35-47. [DOI: https://dx.doi.org/10.12775/EQ.2021.022]

2. Aram, M.; Zhang, X.; Qi, D.; Ko, Y. A state-of-the-art review of fire safety of photovoltaic systems in buildings. J. Clean. Prod.; 2021; 308, 127239. [DOI: https://dx.doi.org/10.1016/j.jclepro.2021.127239]

3. de Almeida Pereira, G.H.; Fusioka, A.M.; Nassu, B.T.; Minetto, R. Active fire detection in Landsat-8 imagery: A large-scale dataset and a deep-learning study. ISPRS J. Photogramm. Remote Sens.; 2021; 178, pp. 171-186. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2021.06.002]

4. Ager, A.A.; Evers, C.R.; Day, M.A.; Alcasena, F.J.; Houtman, R. Planning for future fire: Scenario analysis of an accelerated fuel reduction plan for the western United States. Landsc. Urban Plan.; 2021; 215, 104212. [DOI: https://dx.doi.org/10.1016/j.landurbplan.2021.104212]

5. Biswas, A.; Ghosh, S.K.; Ghosh, A. Early Fire Detection and Alert System using Modified Inception-v3 under Deep Learning Framework. Procedia Comput. Sci.; 2023; 218, pp. 2243-2252. [DOI: https://dx.doi.org/10.1016/j.procs.2023.01.200]

6. Mahaveerakannan, R.; Anitha, C.; Baranitharan, K.; Rajan, S.; Muthukumar, T.; Rajulu, G.G. An IoT based forest fire detection system using integration of cat swarm with LSTM model. Comput. Commun.; 2023; 211, pp. 37-45. [DOI: https://dx.doi.org/10.1016/j.comcom.2023.08.020]

7. Avazov, K.; Hyun, A.E.; Sami S, A.A.; Khaitov, A.; Abdusalomov, A.B.; Cho, Y.I. Forest fire detection and notification method based on AI and IoT approaches. Future Internet; 2023; 15, 61. [DOI: https://dx.doi.org/10.3390/fi15020061]

8. Zheng, S.; Gao, P.; Zhou, Y.; Wu, Z.; Wan, L.; Hu, F.; Wang, W.; Zou, X.; Chen, S. An accurate forest fire recognition method based on improved BPNN and IoT. Remote Sens.; 2023; 15, 2365. [DOI: https://dx.doi.org/10.3390/rs15092365]

9. Khan, R.H.; Bhuiyan, Z.A.; Rahman, S.S.; Khondaker, S. A smart and cost-effective fire detection system for developing country: An IoT based approach. Int. J. Inf. Eng. Electron. Bus.; 2019; 12, 16.

10. Reis, H.C.; Turk, V. Detection of forest fire using deep convolutional neural networks with transfer learning approach. Appl. Soft Comput.; 2023; 143, 110362. [DOI: https://dx.doi.org/10.1016/j.asoc.2023.110362]

11. Barmpoutis, P.; Papaioannou, P.; Dimitropoulos, K.; Grammalidis, N. A review on early forest fire detection systems using optical remote sensing. Sensors; 2020; 20, 6442. [DOI: https://dx.doi.org/10.3390/s20226442] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33187292]

12. Gaudet, B.; Simeoni, A.; Gwynne, S.; Kuligowski, E.; Benichou, N. A review of post-incident studies for wildland-urban interface fires. J. Saf. Sci. Resil.; 2020; 1, pp. 59-65. [DOI: https://dx.doi.org/10.1016/j.jnlssr.2020.06.010]

13. (NSC), National Safety Council. Fire-Related Deaths & Injuries. 2022; Available online: https://injuryfacts.nsc.org/home-and-community/safety-topics/fire-related-fatalities-and-injuries/ (accessed on 26 June 2023).

14. Mukherjee, A.; Rakshit, S.; Panja, A.K.; De, D.; Dey, N. A 6G-Enabled Edge-Assisted Internet of Drone Things Ecosystem for Fire Detection. Proceedings of the International Conference on Frontiers in Computing and Systems; Mandi, India, 16–17 October 2023; Springer: Berlin/Heidelberg, Germany, 2023; pp. 13-25.

15. Wang, Y.; Marsden, J.; Kelly, M. Challenges of fire fighting in fire engineered built environment. Procedia Eng.; 2011; 11, pp. 583-592. [DOI: https://dx.doi.org/10.1016/j.proeng.2011.04.699]

16. de Aquino Aragão, M.; Fiedler, N.C.; Ramalho, A.H.C.; Menezes, R.A.S.; da Silva, E.C.G.; Juvanhol, R.S.; Biazatti, L.D.; da Silva, E.F.; Pereira, R.S.; da Silva Guanaes, G.M. et al. Risk of forest fires occurrence on a transition island Amazon-Cerrado: Where to act?. For. Ecol. Manag.; 2023; 536, 120858. [DOI: https://dx.doi.org/10.1016/j.foreco.2023.120858]

17. Laschi, A.; Foderi, C.; Fabiano, F.; Neri, F.; Cambi, M.; Mariotti, B.; Marchi, E. Forest road planning, construction and maintenance to improve forest fire fighting: A review. Croat. J. For. Eng. J. Theory Appl. For. Eng.; 2019; 40, pp. 207-219.

18. Shees, A.; Ansari, M.S.; Varshney, A.; Asghar, M.N.; Kanwal, N. FireNet-v2: Improved Lightweight Fire Detection Model for Real-Time IoT Applications. Procedia Comput. Sci.; 2023; 218, pp. 2233-2242. [DOI: https://dx.doi.org/10.1016/j.procs.2023.01.199]

19. García-Aguilar, I.; García-González, J.; Luque-Baena, R.M.; López-Rubio, E. Automated labeling of training data for improved object detection in traffic videos by fine-tuned deep convolutional neural networks. Pattern Recognit. Lett.; 2023; 167, pp. 45-52. [DOI: https://dx.doi.org/10.1016/j.patrec.2023.01.015]

20. Vasconcelos, R.N.; Franca Rocha, W.J.; Costa, D.P.; Duverger, S.G.; Santana, M.M.d.; Cambui, E.C.; Ferreira-Ferreira, J.; Oliveira, M.; Barbosa, L.d.S.; Cordeiro, C.L. Fire Detection with Deep Learning: A Comprehensive Review. Land; 2024; 13, 1696. [DOI: https://dx.doi.org/10.3390/land13101696]

21. Jin, C.; Wang, T.; Alhusaini, N.; Zhao, S.; Liu, H.; Xu, K.; Zhang, J. Video Fire Detection Methods Based on Deep Learning: Datasets, Methods, and Future Directions. Fire; 2023; 6, 315. [DOI: https://dx.doi.org/10.3390/fire6080315]

22. Diaconu, B.M. Recent advances and emerging directions in fire detection systems based on machine learning algorithms. Fire; 2023; 6, 441. [DOI: https://dx.doi.org/10.3390/fire6110441]

23. Xanthopoulos, G.; Delogu, G.M.; Leone, V.; Correia, F.J.; Magalhães, C.G. Firefighting approaches and extreme wildfires. Extreme Wildfire Events and Disasters; Elsevier: Amsterdam, The Netherlands, 2020; pp. 117-132.

24. Yang, S.; Huang, Q.; Yu, M. Advancements in remote sensing for active fire detection: A review of datasets and methods. Sci. Total. Environ.; 2024; 943, 173273. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2024.173273] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38823698]

25. Sharma, A.; Kumar, H.; Mittal, K.; Kauhsal, S.; Kaushal, M.; Gupta, D.; Narula, A. IoT and deep learning-inspired multi-model framework for monitoring Active Fire Locations in Agricultural Activities. Comput. Electr. Eng.; 2021; 93, 107216. [DOI: https://dx.doi.org/10.1016/j.compeleceng.2021.107216]

26. Shen, X.; Cao, Z.; Liu, H.; Cong, B.; Zhou, F.; Ma, Y.; Zou, X.; Wei, S. Inverse tracing of fire source in a single room based on CFD simulation and deep learning. J. Build. Eng.; 2023; 76, 107069. [DOI: https://dx.doi.org/10.1016/j.jobe.2023.107069]

27. Peng, T.; Ke, W. Urban fire emergency management based on big data intelligent processing system and Internet of Things. Optik; 2023; 273, 170433. [DOI: https://dx.doi.org/10.1016/j.ijleo.2022.170433]

28. Solórzano, A.; Eichmann, J.; Fernández, L.; Ziems, B.; Jiménez-Soto, J.; Marco, S.; Fonollosa, J. Early fire detection based on gas sensor arrays: Multivariate calibration and validation. Sens. Actuators B Chem.; 2020; 352, 130961. [DOI: https://dx.doi.org/10.1016/j.snb.2021.130961]

29. Quttineh, N.; Olsson, P.; Larsson, T.; Lindell, H. An optimization approach to the design of outdoor thermal fire detection systems. Fire Saf. J.; 2020; 129, 103548. [DOI: https://dx.doi.org/10.1016/j.firesaf.2022.103548]

30. Kizilkaya, B.; Ever, E.; Yatbaz, H.; Yazici, A. An effective forest fire detection framework using heterogeneous wireless multimedia sensor networks. ACM Trans. Multimed. Comput. Commun. Appl. (TOMM); 2020; 18, pp. 1-21. [DOI: https://dx.doi.org/10.1145/3473037]

31. Baba, M.; Grado, J.; Solis, D.; Roma, I.; Dellosa, J. A Multisensory Arduino-Based Fire Detection and Alarm System using GSM Communications and RF Module with an Android Application for Fire Monitoring. Int. J. Innov. Sci. Res. Technol. (IJISRT); 2020; 5, pp. 964-968. [DOI: https://dx.doi.org/10.5281/zenodo.6433836]

32. Dampage, U.; Bandaranayake, L.; Wanasinghe, R.; Kottahachchi, K.; Jayasanka, B. Forest fire detection system using wireless sensor networks and machine learning. Sci. Rep.; 2022; 12, 46. [DOI: https://dx.doi.org/10.1038/s41598-021-03882-9]

33. Khan, F.; Xu, Z.; Sun, J.; Khan, F.; Ahmed, A.; Zhao, Y. Recent advances in sensors for fire detection. Sensors; 2022; 22, 3310. [DOI: https://dx.doi.org/10.3390/s22093310]

34. He, X.; Feng, Y.; Xu, F.; Chen, F.F.; Yu, Y. Smart fire alarm systems for rapid early fire warning: Advances and challenges. Chem. Eng. J.; 2022; 450, 137927. [DOI: https://dx.doi.org/10.1016/j.cej.2022.137927]

35. Jia, Y.; Chen, W.; Yang, M.; Wang, L.; Liu, D.; Zhang, Q. Video smoke detection with domain knowledge and transfer learning from deep convolutional neural networks. Optik; 2021; 240, 166947. [DOI: https://dx.doi.org/10.1016/j.ijleo.2021.166947]

36. Liu, Z.; Zhang, K.; Wang, C.; Huang, S. Research on the identification method for the forest fire based on deep learning. Optik; 2020; 223, 165491. [DOI: https://dx.doi.org/10.1016/j.ijleo.2020.165491]

37. Chen, G.; Zhou, H.; Li, Z.; Gao, Y.; Bai, D.; Xu, R.; Lin, H. Multi-Scale Forest Fire Recognition Model Based on Improved YOLOv5s. Forests; 2023; 14, 315. [DOI: https://dx.doi.org/10.3390/f14020315]

38. Jagatheesaperumal, S.; Saeidi, V.; Kalantar, B.; Ueda, N.; Halin, A. Fire-Net: A deep learning framework for active forest fire detection. J. Sens.; 2022; 2022, 8044390. [DOI: https://dx.doi.org/10.1155/2022/8044390]

39. Mukhiddinov, M.; Abdusalomov, A.B.; Cho, J. Automatic Fire Detection and Notification System Based on Improved YOLOv4 for the Blind and Visually Impaired. Sensors; 2022; 22, 3307. [DOI: https://dx.doi.org/10.3390/s22093307]

40. Abdusalomov, A.; Islam, B.; Nasimov, R.; Mukhiddinov, M.; Whangbo, T. An improved forest fire detection method based on the detectron2 model and a deep learning approach. Sensors; 2023; 23, 1512. [DOI: https://dx.doi.org/10.3390/s23031512]

41. Ahn, Y.; Choi, H.; Kim, B. Development of early fire detection model for buildings using computer vision-based CCTV. J. Build. Eng.; 2023; 65, 105647. [DOI: https://dx.doi.org/10.1016/j.jobe.2022.105647]

42. Khan, A.; Hassan, B.; Khan, S.; Ahmed, R.; Abuassba, A. DeepFire: A Novel Dataset and Deep Transfer Learning Benchmark for Forest Fire Detection. Mob. Inf. Syst.; 2022; 2022, 5358359. [DOI: https://dx.doi.org/10.1155/2022/5358359]

43. de Venâncio, P.; Lisboa, A.; Barbosa, A. An automatic fire detection system based on deep convolutional neural networks for low-power, resource-constrained devices. Neural Comput. Appl.; 2022; 34, pp. 15349-15368. [DOI: https://dx.doi.org/10.1007/s00521-022-07467-z]

44. Dogan, S.; Barua, P.; Kutlu, H.; Baygin, M.; Fujita, H.; Tuncer, T.; Acharya, U. Automated accurate fire detection system using ensemble pretrained residual network. Expert Syst. Appl.; 2022; 203, 117407. [DOI: https://dx.doi.org/10.1016/j.eswa.2022.117407]

45. Bahhar, C.; Ksibi, A.; Ayadi, M.; Jamjoom, M.; Ullah, Z.; Soufiene, B.; Sakli, H. Wildfire and Smoke Detection Using Staged YOLO Model and Ensemble CNN. Electronics; 2023; 12, 228. [DOI: https://dx.doi.org/10.3390/electronics12010228]

46. Kodur, V.; Kumar, P.; Rafi, M.M. Fire hazard in buildings: Review, assessment and strategies for improving fire safety. PSU Res. Rev.; 2020; 4, pp. 1-23. [DOI: https://dx.doi.org/10.1108/PRR-12-2018-0033]

47. Chou, J.; Cheng, M.; Hsieh, Y.; Yang, I.; Hsu, H. Optimal path planning in real time for dynamic building fire rescue operations using wireless sensors and visual guidance. Autom. Constr.; 2019; 99, pp. 1-7. [DOI: https://dx.doi.org/10.1016/j.autcon.2018.11.020]

48. Wu, D.; Zhang, C.; Ji, L.; Ran, R.; Wu, H.; Xu, Y. Forest Fire Recognition Based on Feature Extraction from Multi-View Images. Trait. Signal; 2021; 38, pp. 775-783. [DOI: https://dx.doi.org/10.18280/ts.380324]

49. Ever, E.; Yatbaz, H.Y.; Kizilkaya, B.; Yazici, A. Effective Forest Fire Detection Data-Set for Heterogeneous Wireless Multimedia Sensor Networks. 2022; Available online: https://data.mendeley.com/datasets/g5nzp6j3bt/3 (accessed on 1 May 2023). [DOI: https://dx.doi.org/10.17632/g5nzp6j3bt.3]

50. Cheng, H.; Yin, J.; Chen, B.; Yu, Z. Smoke 100k: A Database for Smoke Detection. 2019; Available online: https://bigmms.github.io/cheng_gcce19_smoke100k/ (accessed on 5 May 2023).

51. Saied, A. FIRE Dataset. 2018; Available online: https://www.kaggle.com/phylake1337/fire-dataset (accessed on 5 May 2023).

52. Kumar, A. Fire Detection Dataset. 2019; Available online: https://www.kaggle.com/atulyakumar98/test-dataset (accessed on 5 May 2023).

53. Kumar, A. Fire and Gun Dataset. 2020; Available online: https://www.kaggle.com/atulyakumar98/fire-and-gun-dataset (accessed on 5 May 2023).

54. Huang, G.; Liu, Z.; van der Maaten, L.; Weinberger, K.Q. Densely Connected Convolutional Networks. Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); Honolulu, HI, USA, 21–26 July 2017; [DOI: https://dx.doi.org/10.1109/CVPR.2017.243]

55. Nair, V.; Hinton, G.E. Rectified Linear Units Improve Restricted Boltzmann Machines. Proceedings of the 27th International Conference on Machine Learning ICML; Haifa, Israel, 21–24 June 2010.

56. Zhang, Y.; Hong, D.; McClement, D.; Oladosu, O.; Pridham, G.; Slaney, G. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging. J. Neurosci. Methods; 2021; 353, 109098. [DOI: https://dx.doi.org/10.1016/j.jneumeth.2021.109098]

57. Chattopadhay, A.; Sarkar, A.; Howlader, P.; Balasubramanian, V.N. Grad-cam++: Generalized gradient-based visual explanations for deepconvolutional networks. Proceedings of the 2018 IEEE Winter Conference on Applications of Computer Vision (WACV); Lake Tahoe, NV, USA, 12–15 March 2018; pp. 839-847.

58. Omeiza, D.; Speakman, S.; Cintas, C.; Weldermariam, K. Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv; 2019; arXiv: 1908.01224

59. Jocher, G.; Chaurasia, A.; Qiu, J. YOLO by Ultralytics. 2023; Available online: https://github.com/ultralytics/ultralytics (accessed on 1 June 2023).

60. Xia, Z.; Li, H.; Chen, Y.; Yu, W. Detecting urban fire high-risk regions using colocation pattern measures. Sustain. Cities Soc.; 2019; 49, 101607. [DOI: https://dx.doi.org/10.1016/j.scs.2019.101607]

61. Biljecki, F.; Ito, K. Street view imagery in urban analytics and GIS: A review. Landsc. Urban Plan.; 2021; 215, 104217. [DOI: https://dx.doi.org/10.1016/j.landurbplan.2021.104217]