1.1. Motivations

Forests play unique roles in human life and deliver various vital resources. They purify the air by drawing in CO₂ and breathing out O₂, so forests are called the “lungs of the planet”. Likewise, they are a habitat for most animals and clean drinking water by filtering out many pollutants from the water [1, 2]. Regrettably, massive forest areas have been burned and destroyed in the past years due to human activities regarded as the foremost cause of wildfires. Accordingly, elevated awareness can prevent most damages and forest losses from happening. Furthermore, early wildfire detection is crucial in diminishing associated risks and casualties and assists firefighters in fire smothering immediately in the early phases. The time spanning from wildfire detection to alerting the relevant authorities is critical and can minimize the risks and consequences. Thus, to rapidly control the fire, early detection of wildfires is vital [3].

Several techniques have been recently proposed to detect wildfires early and allocate appropriate resources to extinguish the fire. These techniques are often based on aerial and ground-based technologies, such as watchtowers with multiple sensors and satellites. However, these technologies have some constraints that lessen overall wildfire detection performance. For example, watchtowers have a limited field of view and impose high construction costs. Likewise, although satellites deliver a pervasive field of view, they come with huge costs and suffer from flexibility issues and low spatial/temporal image resolution, all of which prevent detecting the fire spot at the right time [4, 5].

1.2. Related Works

Computer vision allows machines (e.g., unmanned aerial vehicles [UAVs]) to visually perceive their surrounding environment and respond based on their predefined mission. Computer vision technologies intended for wildfire detection are classified into two classes. The first class covers machine learning (ML)-based conventional methods that work using handcrafted feature extraction techniques and identifying changes in the color [6, 7]. Choosing these features is time-consuming and requires qualified experts to pick features suitable for developing efficient algorithms. These techniques are inefficient when solving complicated problems, such as wildfire detection from images with a cluttered background in dense forests. The techniques in the second class employ deep learning (DL) algorithms to extract relevant and robust features automatically. DL algorithms allow machines to perform difficult and multifaceted tasks such as time-series analysis [8, 9], vehicle [10] and face recognition [11], self-driving cars [12], and the diagnosis of herbal diseases [13].

Recently, UAVs have been broadly utilized in several applications related to forests, like forest discovery, search and rescue operations, forest resource surveys, and wildfire smothering. Furthermore, recent technological breakthroughs have enabled drones to process visual data automatically. Fire and smoke detection in remote and hard-access areas and forests using DL-based techniques has been a subject of interest in recent years. Flame and smoke can be visual features for early and precise wildfire detection. Some studies have explored fire detection using flames [14, 15]. Likewise, some studies have inspected early fire detection using smoke as a proper signal for wildfire [16, 17]. Flames can be concealed in the early phases of wildfire, particularly in the forest [3]. For these restraints to be overcome, recent research has targeted concomitant flame and smoke detection.

Image classification-based methods classify input images (i.e., whether or not images contain fire patterns). Handcrafted pixel-based features were extracted in [18] and classified by support vector machine (SVM). Recently, several studies have worked with a convolutional neural network (CNN) to classify drone-captured wildfire images. For example, authors in [19] propose using AlexNet, a basic CNN architecture, to detect wildfire. The five CNN architectures, including AlexNet, GoogleNet, VGG, and modified versions of the last CNNs, were considered in [19] to classify UAV images into fire and non-fire events. Authors in [20] proposed a CNN-based method for early wildfire detection using a hexacopter with an optical camera. Preprocessing techniques such as histogram equalization and nonlinear filters are applied to enhance data quality and diminish noises. Similarly [16], proposes a method to classify and determine the precise location of wildfires. Other methods based on CNN were proposed in [21, 22]. Transfer learning with InceptionV3, DenseNet121, ResNet50V2, VGG19, and NASNetMobile CNNs was considered in [23] for forest wildfire detection. Reduce-VGGnet was employed in [24] for image classification with the aim of wildfire detection.

On the other side, object detection algorithms draw a bounding box around the object’s extent [25]. Here, objects can be flames or smoke. Several object detection algorithms have been reported to yield acceptable performances in recent years. The region candidates (which may contain fire events) are first generated using a selective search method in region-based algorithms. Next, these region candidates are classified depending on the occurrence or non-occurrence of the desired object. The region-based CNN (R-CNN) family is a well-established and efficient algorithm among region-based algorithms. Contrarily, single-step detectors (STDs) overlook the region candidate generation step by processing the input image in a single step, thus offering faster detection while retaining high accuracy. Yolo and RetinaNet are among the most efficient STDs [26]. A method based on YOLOv5 was proposed in [27] for domain-free fire detection. In [28], YOLOv5 and YOLOv8 were used to identify forest fires. The combination of physical and DL schemes was utilized in [29].

DL-based computer vision algorithms are not merely used for image classification and object recognition. They further can be employed for semantic segmentation. These algorithms are among the most effective DL techniques for wildfire detection. They mainly classify each pixel in the image according to the object class it belongs to (i.e., flame, smoke, forest, etc.). Nonetheless, semantic segmentation algorithms are highly multifaceted, entail higher computing performance, and take a more prolonged course to annotate training images. In recent years, several semantic segmentation techniques have been proposed to detect wildfires from digital images and videos captured by high-accurate UAV platforms. Examples of these techniques are DeepLab [30], U-Net [31], SegNet [32], and CTNet [33].

1.3. Contributions

Taking note of the above topics and regarding the necessity for early detection of wildfires to launch firefighting operations and avoid massive casualties, this study proposes an effective method to detect wildfires from images taken by UAVs from forests. The proposed method is based on time–frequency analysis and extracting deep features from its content. First, images are subjected to a one-level two-dimensional discrete wavelet transform (2D DWT). Wildfires exhibit distinct patterns at different scales. Previous works did not consider these patterns, and 2D DWT can extract them. Also, 2D DWT highlights specific regions in the image where changes occur. Next, the obtained sub-bands pass through CNNs to extract deep features. Since CNNs use local receptive fields to capture spatial patterns, they can be used for local analysis of 2D DWT by identifying relevant features. Then, the deep features of different sub-bands are merged and subjected to feature reduction. Ultimately, the right decision is made using suitable classifiers.

The rest of this paper is organized as follows. Section 2 describes the dataset used in this paper. The proposed method is explained in detail in Section 3. Section 4 contains the simulation results, and section 5 concludes the paper.

2.1. Dataset

As mentioned, this paper aims to present the wildfire detection algorithm for forests due to their unique roles in human life. The use of UAVs for wildfire detection in forests has been attracted due to the challenges and issues of traditional methods. The FLAME (Fire Luminosity Airborne-based Machine learning Evaluation) dataset [34] is an open-access dataset covering wildfire images. The videos were recorded by DJI Phantom 3 Professional and DJI Matrice 200 drones with Zenmuse X4S, FLIR Vue Pro thermal, and DJI Phantom 3 cameras. The first and second videos are similarly raw (16 min long) and recorded by a Zenmuse X4S camera with frames per second (FPS) of 29. The second video presents the behavior of one pile from the start of burning. The resolution of the first and second videos is 1280 × 720. The third, fourth, and fifth videos are WhiteHot, GreentHot, and Heatmap. They are 89-s, 5-min, and 25-min long, respectively, and a FLIR camera has recorded all with an FPS of 30 and a resolution of 640 × 512. The sixth video is a 17-min HD RGB video recorded by a DJI Phantom 3 camera with an FPS of 30—the resolution of this video is 3840 × 2160. The seventh and eighth repositories contain 39,375 and 8617 frames with a resolution of 254 × 254. For image classification, the ninth and tenth repositories have more than 2000 fire and ground truth mask frames with high resolution for the fire segmentation problem. The primary purpose of this dataset is to classify images into “wildfire” or “non-wildfire” and perform fire image segmentation. Furthermore, they can detect wildfires from RGB images and thermal drones. Table 1 summarizes the details of the parts of the database used in this study. Likewise, Figures 1 and 2 provide examples of images of “wildfire” and “non-wildfire” events.

2.2. Preliminaries

3.1. Preprocessing

3.2. Feature Extraction

Figure 4 demonstrates the proposed feature extraction method for wildfire detection. Images containing fire events contain considerable information on the edges that need to be extracted, although such information has not been well used in previous studies. For this, the first step is to extract the edges of the images and put the information in the high frequencies of the image. In this study, the 2D DWT is employed for this purpose.

In this study, 2D DWT was applied in one step using Daubechies 4 filters. A diagram of the single-level of 2D DWT is shown in Figure 5. As a result, four sub-bands are obtained, including the approximation band (LL), the vertical band (LH), the horizontal band (HL), and the diagonal detail band (HH), which are represented by x_a, x_h, x_v, and x_d, respectively.

For extracting deep features, each of the sub-bands obtained in the previous step is separately passed through the considered CNN, and the features obtained from the flattened layer are considered deep features for each sub-band. Notably, the dimensions of each matrix obtained for different sub-bands are altered based on the dimensions of the CNN’s input layer before being subjected to the CNN. This results in feature vectors y_a, y_h, y_v, and y_d, which correspond to the sub-bands x_a, x_h, x_v, and x_d, respectively. The size of all the feature vectors obtained is n_d × 1. These feature vectors are merged into y_f = [y_a, y_h, y_v, y_d] to obtain the final feature vector with dimensions of (4n_d × 1).

The final feature vector consists of many characteristics. There is often redundant information, such as related or duplicate factors, in this feature vector. Thus, feature dimensionality reduction (FDR) overcomes the outcomes of redundant or irrelevant information, mapping the practical information in the main features to fewer features.

3.3. Classification

In the last step, the proposed method classifies the reduced features obtained from MDS to determine if the image contains fire events. Accordingly, this study uses several classification techniques and compares their performances.

Support vector machine (SVM): SVM has been proposed as a robust classifier. SVM is broadly used in various studies with binary scenarios because of its lower computational complexity and easy processing. The optimal hyperplane in SVM maximizes the marginal distance between classes. The linear SVM is utilized in this paper.

k-Nearest neighbor (kNN): kNN is a well-established method in ML-based classification algorithms. kNN is a simple classifier used for wildfire detection. The distance between the test data and the nearest samples in training data determines the test data class. Thus, the value of k plays a unique role in the performance of kNN.

Decision tree: This classifier is a supervised ML technique where a dataset is constantly divided into subsets based on a specific parameter. This classifier uses a tree-like structure consisting of a root, internal decision, and end nodes. The root node is regarded as the whole dataset classified into branches. The internal subsets are called decision nodes; the final node represents the predicted class.

4.1. Simulation Setup

The simulations were carried out using MATLAB R2023a on the computer with the following hardware specifications: CPU: Core i7 13700 H, RAM: 32 GB, GeForce RTX3050, HDD: 1TB. We consider the 10-fold cross-validation to partition the dataset into the train and test data. This scheme randomly divides the dataset into 10 equally sized parts, and the training and test procedure is repeated 10 times. In each cross-validation, one part is considered test data, and the remaining train the proposed method. This procedure was repeated for all parts as test data, and the results were averaged.

4.2. Classification Accuracy

Table 3 shows the classification accuracy for various mother wavelet, CNN, and classifier combinations. Mother wavelets are Daubechies 4 (db4), Symlet 8 (sym8), biorthogonal 1.5 (bior 1.5), and reverse biorthogonal 3.5 (rbior 3.5). CNNs, including AlexNet, MobileNet, and InceptionV3, were used to extract deep features. Likewise, SVM, kNN, and decision tree classifiers were utilized to classify reduced deep features. From the results, we observe: (1) the db4 wavelet consistently yields better classification outcomes than sym8, bior1.5, and rbior3.5, likely because it balances edge characterization and noise suppression effectively; (2) InceptionV3 outperforms AlexNet and MobileNet in nearly every wavelet–classifier pairing, indicating that its deeper multi-scale filters are beneficial for wildfire vs. nonwildfire discrimination; and (3) SVM tends to produce higher accuracy compared to kNN and decision tree under the same feature sets, likely thanks to its robust margin-maximizing properties. Overall, the db4 + InceptionV3 + SVM combination achieves the highest accuracy, about 96.8%.

4.3. Confusion Matrix

Table 4 gives the confusion matrix for the proposed method. Confusion matrix is a square matrix with dimensions n_c × n_c, where n_c denotes the number of classes that are equal to two in this study. The confusion matrix represents the prediction summary in the matrix form and indicates how many predictions are correct and incorrect per class. Such a representation allows us to understand the classes that are being confused by the model as other classes. As shown, from 180 images with wildfire events, only three are wrongly detected as non-wildfire events. Likewise, from 200 images without wildfire events, nine are wrongly recognized as images with wildfire events. The confusion matrix further represents the sensitivity of each class, which is equal to the correct classification accuracy for each class. The sensitivity and precision are 0.9833 and 0.9516, respectively. Regarding the importance of early wildfire detection and the fast action by firefighters, these results confirm the efficiency of the proposed wildfire detection method. Also, the value of 0.9672 for the F₁ score indicates the balanced ability of the proposed model to handle positive cases while maintaining accuracy.

4.4. Ablation Study

Table 5 presents the results of investigating the feature reduction algorithm’s effect on the proposed method’s accuracy. As discussed earlier, CNNs extract many deep features from 2D DWT sub-bands, which are highly correlated and thus prolong classification, increase complexity, and reduce efficiency. Thus, using feature reduction methods is crucial. Table 5 compares the accuracy of MDS, PCA, and linear discriminant analysis (LDA). Furthermore, the accuracy of the proposed method is reported while using no feature reduction method. As shown, classification accuracy is improved after using any feature reduction method. Furthermore, supervised LDA outperforms unsupervised PCA, while MDS achieves the highest classification accuracy.

Table 6 compares the efficacy of various optimization algorithms in training CNNs. According to the results, the optimizer stochastic gradient descent with momentum (SGDM) achieves the highest accuracy than the adaptive moment estimation (ADAM) and root mean squared propagation (RMSProp) ().

4.5. Convergence Speed

Figure 6 presents the convergence speed of the proposed method for wildfire detection. As observed, the proposed method has an appreciated convergence speed. Also, the difference between the training and test data accuracy indicates that the model is converged and there is no overfitting or underfitting.

4.6. Computational Time Analysis

In that section, we present a temporal-cost table that quantifies the average processing time for each major step of our approach—namely, (1) wavelet decomposition, (2) CNN-based deep feature extraction, (3) dimensionality reduction (MDS), and (4) final classification. Table 7 illustrates the proposed method’s overall computational. These results confirm that the CNN’s forward pass is the primary driver of runtime. However, given modern GPUs or on-board accelerators (like the NVIDIA Jetson family), our entire pipeline remains practical for near real-time applications, mainly if further optimizations (such as model pruning or quantization) are applied.