In the last decade, drones have generated considerable attention as a remote sensing platform with a wide application range, including traffic control, disaster response, crop protection, and satellite image analysis [1,2]. Of late, incorporating a vision system, drone applications have been developed for monitoring, perceiving, and analyzing active and passive threats at the incident sites, for example, fire detection, flood threats, car accidents, and landslide-prone areas [3,4]. Additionally, drones can be rapidly deployed because their small size permits them to participate in mission-critical decisions, enabling better resource allocation and risk reduction. Drones are expected to function in disaster-affected areas, where connection to cloud services may not be effective and high-end equipment may not be readily accessible. To ensure operational performance and real analysis, a high-level of autonomy is required. Therefore, autonomous Unmanned Aerial Vehicles (UAVs), as well as Unmanned Ground Vehicles (UGVs), rely on their onboard sensors and embedded microchips for performing tasks rather than sending the data to a central control station. Furthermore, drones can cover a larger area within a shorter time span when combined with automatic route planning techniques, with onboard visual sensors as well as autonomous navigation. With the limited computational ability and power consumption, drones, however, present their own set of challenges [5].

CNNs and Deep Learning (DL), in particular, have been widely recognized as popular solutions for a broad range of applications based on computer vision, such as activity recognition, person recognition, vehicle recognition, and classification [6–10]. In prior research, utilizing transfer learning, a pre-trained CNN was used as a feature extricate, and certain layers have been added to perform classification for the current job, learning techniques to outperform standard machine learning methods using handcrafted features. Even though CNNs have proven to be more effective in classification, their inference time is high due to their high-computational power requirements when embedded in low-power devices, such as drones, which must perform multiple vision tasks simultaneously. For certain applications, a localized, integrated approach is more desirable over cloud processing due to security and privacy concerns [11]. In addition, tiny CNNs can ascertain the accuracy needed and the performance for specialized applications where substantial information does not exist and computational resource limits are enforced. Moreover, their training process is considerably easier, and they can be conveniently updated over air owing to their computational efficiency.

Soft computing methods based on Traditional Fire Warning Systems (TFWSs) [12] and optical sensors to prevent flames from spreading have been extensively researched and developed. Various scalar sensors, including ocular sensors, inferno sensors, as well as smoke sensors closer to the blazing fire [13] have been used in a TFWS for fire detection. However, scalar sensor-based solutions do not provide additional information regarding the area coverage, level of burning, location of the fire, or size of the fire. Additionally, the above-mentioned sensor systems require human intervention, such as a visit to the fire site in the event of a disaster. To overcome these limitations, various visual sensor-based methods have been proposed [14]. In surveillance systems, for the autonomous observation of fire catastrophes, traditional, DL, and vision-driven techniques play a key role in fire detection [15,16]. These algorithms offer a number of advantages such as rapid response, low-human-intervention requirements, cost-effectiveness, and greater coverage. However, traditional fire detection is difficult and time-consuming to process because it relies on hand-crafted feature extraction, and the procedures for constructing and evaluating the features are tedious. In particular, the monitoring of early fire and alarm generation, in traditional-based approaches is difficult because of the fluctuating lighting conditions, shadows, reflections, and low-detection accuracy. This study applies CNN models motivated by their potential in several areas, such as fire detection in surveillance footage [17]. However, DL includes an End-to-End (E2E) process for identifying features, which is computationally intensive and needs a considerable volume of training data. In this article, we have fine-tuned and proposed an efficient VGG-based (E-FireNet) model that has improved detection accuracy, has fewer parameters and can be deployed in actual scenarios. E-FireNet is not only capable of classifying fire and non-fire events but also looks for Object-of-Interest (OoI) located in the image. If the input image depicts a building fire or car fire, the model classifies it as a fire event on the specific object; if no fire is detected by the model, the resultant image is classified as non-fire. Furthermore, the baseline methods use computationally complex models that are not capable of being deployed on a drone, considering that a drone is a resource constraint device with limited processing power. Therefore, a lightweight CNN model that can be deployed in drones is highly desirable. The following are the significant contributions of this study:

Early attempts for fire detection using pre-trained models with numerous parameters show limited performance when the variation in the data was low. To address these problems, this study presents a CNN-based framework for early fire detection in the images captured in diverse surveillance environments. The proposed E-FireNet model shows convincing performance with respect to the accuracy and time complexity for the Central Processing Unit (CPU) and Graphical Processing Unit (GPU), i.e., 0.98, 22.17, and 30.58 Frames Per Second (FPS), compared to renowned State-Of-The-Art (SOTA) models.

The remainder of this paper is organized as follows. Section 2 reviews the relevant literature, Section 3 describes the proposed methodology, and Section 4 presents the experimental results. Finally, Section 5 summarizes the findings and suggests future directions.

Fire is an abnormal event that leads to serious injury and death, and affects precious resources within a short duration. Several techniques were proposed to monitor and control fire events in cities for saving life and property. However, fire detection in real-time is a challenging task. For instance, a CNN-based technique was proposed in [18], which aims to improve the accuracy and reduce the false-alarm rate. For improving the performance, a pre-trained model with fine-tuning of the uppermost layers was used. Furthermore, experiments were conducted over benchmark datasets, and 94.43% accuracy was achieved. Another technique was proposed [19], to realize a false-alarm system for a fire event. Through the transfer learning technique, InceptionV3 achieved excellent performance on test data. Several researchers trained models using satellite-captured images for the classification of fire and normal scenes. The main aim was to extract the region of fire using a local binary pattern for the reduction of false detection rates, and 98% accuracy was achieved.

The approach presented in [20] resolved the fire detection issue through classification and segmentation mechanisms. An artificial neural network was built for the binary classification and 76% accuracy was realized. Nevertheless, the segmentation method was applied to determine the fire border, whereas, for the fire mask, U-Net was used for up and down-sampling to obtain 92% precision and 84% recall. The technique introduced in [21] uses You-Look-Only-Once (YOLO) model for flame detection and extracts the visual features from video data frames. To overcome the overfitting problem and achieve efficient performance, augmentation techniques, such as rotation, flipping, and brightness adjustment were applied. Another study [22] used a Faster Region-based CNN (Fast-RCNN) to detect fire and normal scenes in an image, and later extracted the spatial features via a CNN; for temporal features, Long-Short Term Memory (LSTM) was employed to classify the target scene.

In a recent study [23], a fast and accurate algorithm was developed for extracting spatial features from surveillance video data. Three different versions of SqueezeNet were analyzed to compare their classification performances. In-depth experiments were conducted and 95.02%, 98.46%, and 98.52% accuracy were obtained with SqueezeNet1, SqueezeNet2, and SqueezeNet3, respectively. The technique presented in [24] applied different CNNs such as AlexNet, GoogLeNet, and VGG16 to recognize different events (smoke, non-fire, flame). The experimental results exhibited that the VGG16 model achieved the best performance. Similarly, in another approach [25], a lightweight CNN model was developed for flame detection in real-time, which mainly monitored and controlled the fire scenario in the early stage. A recent study [26] proposed a technique based on a deep saliency network for video-based smoke detection and compared the obtained results with those of ML and DL methods. In addition, the saliency network aimed to highlight the Region-of-Interest (RoI), i.e., the smoke area in the images. To further improve the model performance, various augmentation techniques were applied to the samples.

Several researchers with diverse background studies have applied lightweight models to detect and classify fire in real-time. For instance, [27] proposed a lightweight CNN for fire detection and classification. Furthermore, they computed the execution time to verify the model’s adaptability in real-time processing.

Similarly, [28] introduced a method that could detect fire in real-time in both indoor and outdoor surveillance videos. A multi-expert system was first employed to collect data based on color, shape, and motion analysis. The Bag-of-Words (BoW) approach was then applied for motion representation. In addition, real-time and web scraping videos of fire were utilized in experiments where the proposed model achieved better performance. Researchers in [29] introduced a real-time fire detection technique using a fusion algorithm and several sensors (smoke, flame, and temperature) in indoor and outdoor domains for fire incident detection. The included literature is summarized and then listed in Table 1.

As part of this study, a CNN architecture E-FireNet is employed, for detecting fire. Several scenarios were examined in the custom SV-Fire dataset, such as a fire in a car, a fire in a building, and non-fire. In this study, the outdoor fire conditions differ from those described in previous literature. In comparison, prior studies targeted only wildfire scenarios, including fire, smoke, non-fire scenarios, and non-smoke scenarios.

The proposed framework involves three main steps. In the first step, the collected fire images are pre-processed to increase the number of samples. In the second step, images of diverse classes are input to the proposed efficient CNN model, which effectively detects and classifies the fire into the respective class. In the final step, the model takes a decision based on the predicted label for the given input image. If the predicted label is a fire in a building or fire in a vehicle, an alert is generated to the nearest emergency response department to take early action. A detailed pictorial representation of the proposed framework is depicted in subsection 3.2; the step-wise procedure is presented in Algorithm 1. In the initialization step of Algorithm 1, the drone acquires a Video Stream (VS) and loads a pre-trained Fire Detection Model (FD_M). When the frame is read, the RoI_S is extracted and checked for the presence of fire. If it is a non-fire image, the next frame is selected; else, if a fire is detected in the frame, an alert is generated and sent to the emergency department and disaster teams. Furthermore, in the following sections, this article briefly discusses each step of the proposed framework.

This section investigates the assessment measures and evaluation metrics in detail and describes the collected dataset and the graphical outcomes. The experimental setup and performance metrics are first described, followed by a discussion on the SV-Fire dataset, and the evaluation of the results. All the models, including the proposed E-FireNet, were trained using a total of 30 epochs with a low learning-rate to ensure that the model retained most of the previously learned knowledge. The pre-trained model progressively updates the learning parameters for optimum performance on the intended dataset. In subsection 4.3 the article compares the proposed model with the SOTA models and lists the main hyper-parameters utilized in these experiments. Based on the results, each model was retrained with its default input size, with a batch size of 16, and the Stochastic Gradient Descent (SGD) optimizer was equipped with a learning-rate and momentum of 1e-4 and 0.9, respectively. The experiments were conducted on an NVIDIA RTX 2060 Super GPU with 32-GB of onboard memory, a Keras DL framework, and TensorFlow for the back-end. As shown in the following equations, the performance of the proposed model was assessed by utilizing multiple evaluation metrics, including accuracy, precision, recall, and F1-score.

To reduce social, environmental, and financial damage, CNN-based smart monitoring systems have been used to classify fire scenes in the early stages. Nevertheless, the research focuses on enhancing the accuracy, and attention to the model computation and generalization is limited. Therefore, this study presented an efficient framework (E-FireNet) that accurately classifies fire and non-fire images into their corresponding classes. The proposed E-FireNet achieves the best validation accuracy of 0.98 with limited parameters in comparison to the SOTA models. In addition, E-FireNet managed to achieve a precision of 1, a recall of 0.99, and an F1-score of 0.99. Furthermore, the SV-Fire dataset was collected since a dataset with diverse scenarios was not available for evaluating the proposed method. A set of experiments were performed using various CNN models and the proposed model, and their performances were compared in terms of accuracy, parameters, and FPS over two local systems (CPU and GPU) using the test data. Future research aims to expand the current dataset with new classes and apply vision transformers to fire detection.