A forest fire fireline identification method based on a unmanned aerial vehicle and a deep neural network

Abstract

The application discloses a forest fire line identification method based on a unmanned aerial vehicle and a deep neural network. In the application, compared with an existing technical scheme of combining image segmentation and Canny edge detection, the application is based on the characteristics of the fire line and focuses more on the fire line contour. When the existing scheme is directed to a fire image, when Canny edge detection is used on the non-fire line region after segmentation, some linear edges similar to the fire line are often generated, which brings challenges to subsequent forest fire line quantification and fire front tracking tasks. Unlike the existing scheme, the technical scheme of the application effectively combines the fire spread mechanism and the deep neural network technology, deeply mines the color characteristics of the fire line, highly pays attention to the information of the fire line region, and greatly improves the continuity and smoothness of the fire line contour identification result. In summary, the technical scheme of the application focuses more on the fire line itself, can obtain the fire line contour with more fine edge information and smoother fire line characteristics.

Description

Technical Field

[0001] This invention belongs to the field of forest fire prevention technology, specifically a method for identifying forest fire lines based on unmanned aerial vehicles and deep neural networks. Background Technology

[0002] With a warming and drier climate and an increase in extreme weather events, major forest fires are occurring more frequently worldwide. Real-time monitoring and early warning of forest fires is a key technological challenge in preventing and controlling such fires. Unmanned aerial vehicle (UAV)-based forest fire line identification technology is of great significance for real-time monitoring and quantification of fire lines. Frequent forest fires not only cause significant losses to people's lives and property but also affect the forest ecosystem and its balance. Major forest fires often evolve from low- to medium-intensity fires, accompanied by an accelerated fire spread. Therefore, real-time and accurate quantification of fire lines and monitoring of fire fronts are crucial technical challenges for early forest fire prevention and control.

[0003] In recent years, computer vision has become an important technology for quantifying fire lines and has attracted much attention from researchers. Early attempts by some scholars combined satellite remote sensing and computer vision methods to estimate fire line location information by identifying burning areas. However, due to the long timeframes of satellite imagery, the accuracy of identification is insufficient to meet the needs of fire prevention and control in a timely manner. Based on this, some scholars have proposed unmanned aerial vehicle (UAV)-based methods for forest fire line identification. Extensive research has revealed that, on the one hand, some methods rely too heavily on traditional empirical knowledge, resulting in low algorithm applicability; on the other hand, due to the complexity of the fire scene environment, existing deep learning methods still struggle to achieve high-precision identification.

[0004] However, in existing technologies, forest fire line identification technology based on remote sensing imagery is limited by the long cycle of satellite imagery, making it difficult to respond to fire prevention and control needs in a timely manner; UAV-based fire line identification methods based on traditional experience knowledge often exhibit weak applicability; and UAV-based fire line identification methods combined with deep learning are often limited by the complexity of the fire scene environment, making it difficult to achieve high identification accuracy. Taking the technical solution closest to this invention as an example, although it uses an image segmentation model to extract the fire line area, because this model focuses too much on the features of the area where the fire line is located and ignores the features of the fire line itself, the final identification result still cannot focus on the fire line itself, thus making it difficult to effectively improve the accuracy of fire line contour recognition.

[0005] Based on this, this invention combines the forest fire spread mechanism with deep neural network technology, constructs a dedicated image dataset of forest fire lines based on field burning experiments, and proposes a forest fire line identification method based on UAVs and deep neural networks. Summary of the Invention

[0006] The purpose of this invention is to provide a forest fire line identification method based on unmanned aerial vehicles (UAVs) and deep neural networks in order to solve the problems mentioned above.

[0007] The technical solution adopted in this invention is as follows: a forest fire line identification method based on unmanned aerial vehicles and deep neural networks, the identification method comprising the following steps:

[0008] S1: Prepare the training dataset, combining drone imagery data collected from the field burning experiment with open-source image data collected by Kaggle. 70% will be used as the training set, and 30% as the test set.

[0009] S2: Establish a network model, use the fire line color feature module to guide the network to obtain the fire line color feature extraction result image of the input image, and then use the fire line region module to constrain the input image and the extraction result. The two are cascaded to obtain the recognition result image that simultaneously contains fire line color features and fire line region structure features.

[0010] S3: Train the network model. The Fireline dataset prepared in step S1 is uniformly batch-adjusted to a size of 224×224 pixels as the model input. Set the key parameters for network model training: initial learning rate, learning decay rate, and number of iterations. Optimize the network parameters using the designed hybrid loss function and the Adam optimizer.

[0011] S4: Perform fire line recognition. Input the above 30% of the test images into the constructed network model, load the pre-trained network parameters, perform fire line recognition, and obtain the recognized images and evaluation index results.

[0012] In a preferred embodiment, step S1 utilizes a self-constructed fireline dataset. This dataset comprises 2000 images from public resources and real wildfire experiments. Specifically, the present invention collected 600 images from the Kaggle forest fire dataset and 1400 drone images from a field burning experiment in Xichang City, Sichuan Province. In the experiment, 70% of the fireline dataset was used as the training dataset, and 30% was used as the test data.

[0013] In a preferred embodiment, in step S2, the network mainly includes a backbone network, a fire line color guidance module, and a fire line region extraction module, wherein the fire line color guidance module and the fire line region extraction module are the two core parts of the present invention. The input of the fire line color guidance module is the original image of the forest fire line, and the output is the preliminary extraction result focusing on the fire line color; the input of the fire line region extraction module is the original image and the preliminary extraction result. This module mainly uses the cross-entropy loss function to reduce the difference in fire line region size between the spatial attention map of the preliminary extraction result and the ground reference image.

[0014] In a preferred embodiment, in step S2, the color mapping function extracted by the FireWire color guiding module can be expressed as equation (1), and the equations for converting RGB to YCbCr and HSV can be described as equations (2)-(5).

[0015]

[0016] Y = 0.299R + 0.587G + 0.114B

[0017] Cb = BY

[0018] Cr=RY (2)

[0019]

[0020]

[0021]

[0022] In the formula, (R, G, B) represent the values ​​of the three color channels of the input image. The Color Mapping Guided (CMG) component applies the extracted color map to multiple multi-level features (f 2 ,f 3 ,f 4 The results are then used as a hierarchical supervised loss function to train the network, helping it focus on the flame color region and obtain a more accurate fire line region. Specifically, CMG can be represented as:

[0023] CMG i =f i ×M Color (I) (6)

[0024] Where f is i Multi-level features, where i is the layer depth in the network, and M... color This represents the extracted color map, where I is the input image.

[0025] In a preferred embodiment, in step S2, the fireline region extraction module first multiplies the input image with the preliminary extraction result and the ground reference image to extract the fireline region from the preliminary result and the ground reference image; then, a spatial attention layer is used to focus on the internal correlation between the preliminary extraction result and the ground reference image; finally, the cross-entropy loss function is used to reduce the difference in fireline region size between the spatial attention map of the preliminary extraction result and the ground reference image. In this example, following the fireline color guidance module, the proposed fireline region extraction module can help the network to focus on fireline region information on top of focusing on fireline color features. From the perspective of qualitative analysis of fireline recognition, the fireline color guidance module can effectively help the model distinguish between flames and smoke, while the fireline region extraction module can improve the continuity and smoothness of the recognized contours.

[0026] In a preferred embodiment, in step S2, the loss function uses a hybrid loss function that combines the fidelity loss function, the cross-entropy loss function, the fireline color loss function, and the fireline region regularization loss function. The hybrid function is defined as follows:

[0027]

[0028] Among them, L Fid L BCE L Color and L FLA Let λ represent the fidelity loss function, cross-entropy loss function, fireline color loss function, and fireline region regularization loss function, respectively. f , λ L , λ c and λ flr These are the weights of the four types of loss functions.

[0029] Fidelity loss is used to ensure consistency between the output and the label (Groundtruth), and it is defined as:

[0030]

[0031] Cross-entropy loss is used to establish the relationship between ground truth and the ground significance result, and it is defined as follows:

[0032]

[0033] While cross-entropy loss can optimize the model's extraction of fire line boundaries, the model still lacks sufficient constraints on the color of the fire line, resulting in blurred boundaries between the fire line and the smoke near the flame. To address this issue, this invention constructs a color loss to explicitly learn the fire line boundaries, which can be expressed as:

[0034]

[0035] The fire region regularization loss function is used to maintain consistency between the fire region results and the labels. It is defined as follows:

[0036]

[0037] In a preferred embodiment, in step S3, the training images obtained in S1 are cropped to 224×224 pixels and subjected to data augmentation, then used as input together according to the batch size (BatchSize = 32). Regarding the hyperparameters of the proposed method, λ is empirically set... f ,λ2,λ3,λ4,λ c and λ flr Set them to 1, 0.8, 0.6, 0.4, 0.5 and 10 respectively;

[0038] During network training, the initial learning rate was set to 0.0005, and the learning rate decayed at a rate of 0.05 times every 30 iterations. The Adam optimizer was used, and the trained model weights were saved step by step.

[0039] In a preferred embodiment, in step S4, a fire detection experiment is conducted based on the 30% test dataset from S1, and MAE and S... m F max Quantitative and qualitative analysis of the evaluation indicators.

[0040] In a preferred embodiment, in step S4, the qualitative comparison results of the method of the present invention and the current mainstream methods on the dataset independently constructed by the present invention are displayed. Figure 2 Images refers to the original input images, GT refers to the labels (Groundtruth), Ours refers to the method proposed in this invention, and PoolNet, GCPANet, CTDNet, SAMNet, and HVPNet refer to current mainstream methods. Quantitative comparison results are shown in Table 1 (the best value for each comparison indicator is shown in bold). Qualitatively, compared with mainstream methods, the method proposed in this invention can obtain more complete and accurate fireline boundaries; other methods generally suffer from discontinuous firelines and uneven boundaries. Quantitatively, this invention achieves better results in mean absolute error (MAE), F-score (F...). max ) and structural similarity (S m The method achieved the best value among the three comparison indicators. The leading results in the qualitative comparison and quantitative comparison indicators demonstrate the superiority of this method.

[0041] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0042] 1. This invention constructs a dedicated image dataset for forest fire lines based on field burning experiments, proposes a forest fire line identification method based on UAVs and deep neural networks, develops a deep framework for forest fire line identification based on UAVs and deep neural networks, and designs a hybrid loss function constraint model. Real-time and accurate identification of forest fire line contours based on UAVs is achieved, providing strong support for subsequent research on fire line quantification and monitoring.

[0043] 2. Compared to existing joint image segmentation and Canny edge detection techniques, this invention focuses on the inherent characteristics of the fire line itself, giving greater emphasis to its outline. The technical solution emphasizes fire line color and region features, effectively combining fire spread mechanisms with deep neural network technology to deeply mine fire line color information and pay close attention to fire line region information, significantly improving the continuity and smoothness of fire line recognition results. Furthermore, in the model design process, the image segmentation model focusing on the fire line region faces challenges. The presence of non-fire line information within the fire line region increases the difficulty of extracting effective information and learning fire line features, resulting in the inclusion of non-fire line information in the final segmentation results. After Canny edge detection, this non-fire line information can appear as linear edges resembling fire lines, posing challenges to subsequent forest fire line quantification and fire front monitoring. In summary, this invention focuses more on the fire line itself, resulting in fire line recognition with finer edge information and smoother fire line characteristics. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the forest fire line identification system based on drones according to the present invention;

[0045] Figure 2 This is a schematic diagram illustrating the qualitative comparison results of the mainstream method of this invention on the Fireline dedicated dataset; wherein: (a) is the original input image, (b) label (Ground truth), (c) the processing result of this invention, (d) PoolNet, (e) GCPANet, (f) CTDNet, (g) SAMNet, and (h) the processing result of the HVPNet method. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] Reference Figure 1-2 ,

[0048] Example:

[0049] A forest fire line identification method based on UAVs and deep neural networks is proposed. The network extracts fire line color features and fire line area features, and uses a hybrid loss function to achieve fire line identification. The network model results are as follows: Figure 1 As shown:

[0050] A method for identifying forest fire lines based on drones and deep neural networks, the identification process of which includes the following steps:

[0051] Step S1: Prepare the training dataset by combining drone imagery data collected in the field burning experiment with open-source image data collected by Kaggle. 70% of the dataset will be used as the training set, and 30% as the test set.

[0052] Step S2: Establish a network model, use the fire line color feature module to guide the network to obtain the fire line color feature extraction result image of the input image, and then use the fire line region module to constrain the input image and the extraction result. The two are cascaded to obtain the recognition result image that simultaneously contains fire line color features and fire line region structure features.

[0053] Step S3: Train the network model. The Fireline dataset prepared in Step S1 is uniformly batch-adjusted to a size of 224×224 pixels as the model input. Set the key parameters for network model training: initial learning rate, learning decay rate, and number of iterations. Optimize the network parameters using the designed hybrid loss function and the Adam optimizer.

[0054] Step S4: Fireline recognition. Input the above 30% of the test images into the constructed network model, load the pre-trained network parameters, perform fireline recognition, and obtain the recognized images and evaluation index results.

[0055] According to an embodiment of the present invention, the network mainly includes a fire-line color map-guided module and a fire-line area refinement module. In the fire-line color map-guided module, the present invention utilizes a color extraction module to extract the color of the input image, making the network more focused on fire-line color features. Generally, fire-line colors are displayed as red, orange, and yellow. Therefore, we convert the image to YCbCr and HSV spaces to extract the fire-line color map. In the YCbCr space, the Y and H channels can better display the brightness of the fire, while the Cb, Cr, and S channels can better express the chromaticity. Therefore, we obtain the fire-line color map by setting the ranges of Y, H, Cb, Cr, and S. In the fire-line area refinement module, the present invention first multiplies the input image with the preliminary extraction result and the ground reference image to extract the fire-line area of ​​the preliminary result and the ground reference image; then, a spatial attention layer is used to focus on the internal correlation between the preliminary extraction result and the ground reference image; finally, the cross-entropy loss function is used to reduce the difference in fire-line area size between the spatial attention map of the preliminary extraction result and the ground reference image.

[0056] The above method specifically includes the following steps:

[0057] S1. Prepare the training dataset

[0058] Due to the scarcity of fireline datasets, this invention utilizes a self-constructed fireline dataset. This dataset comprises 2000 images from public resources and real wildfire experiments. Specifically, this invention collected 600 images from the Kaggle Forest Fire Dataset and 1400 drone images from a field burning experiment in Xichang City, Sichuan Province. In the experiment, 70% of the fireline dataset was used as the training dataset, and 30% was used as the test dataset.

[0059] S2. Building the network model

[0060] The network in this example mainly includes a backbone network, a fire line color guidance module, and a fire line region extraction module, among which the fire line color guidance module and the fire line region extraction module are the two core parts of this invention. The fire line color guidance module takes the original image of the forest fire line as input and outputs a preliminary extraction result focusing on the fire line color. The fire line region extraction module takes the original image and the preliminary extraction result as input. This module mainly uses the cross-entropy loss function to reduce the difference in fire line region size between the spatial attention map of the preliminary extraction result and the ground reference image. A detailed description follows:

[0061] (1) FireWire Color Guidance Module: In this example, the extracted color mapping function can be expressed as Equation (1), and the equations for converting RGB to YCbCr and HSV can be described as Equations (2)-(5).

[0062]

[0063] Y = 0.299R + 0.587G + 0.114B

[0064] Cb = BY

[0065] Cr=RY (2)

[0066]

[0067]

[0068]

[0069] In the formula, (R, G, B) represent the values ​​of the three color channels of the input image. The Color Mapping Guided (CMG) component applies the extracted color map to multiple multi-level features (f 2 ,f 3 ,f 4 The results are then used as a hierarchical supervised loss function to train the network, helping it focus on the flame color region and obtain a more accurate fire line region. Specifically, CMG can be represented as:

[0070] CMG i =f i ×M Color (I) (6)

[0071] Where f is i Multi-level features, where i is the layer depth in the network, and M... color This represents the extracted color map, where I is the input image.

[0072] (2) Fireline Region Extraction Module: In this example, the invention first multiplies the input image with the preliminary extraction result and the ground reference image to extract the fireline region from the preliminary result and the ground reference image; then, a spatial attention layer is used to focus on the internal correlation between the preliminary extraction result and the ground reference image; finally, the cross-entropy loss function is used to reduce the difference in fireline region size between the spatial attention map of the preliminary extraction result and the ground reference image. In this example, following the fireline color guidance module, the proposed fireline region extraction module can help the network to focus on fireline region information on the basis of focusing on fireline color features. From the perspective of qualitative analysis of fireline recognition, the fireline color guidance module can effectively help the model distinguish between flames and smoke, while the fireline region extraction module can improve the continuity and smoothness of the recognized contours.

[0073] (3) Loss Function: In this example, the invention uses a hybrid loss function, which combines the fidelity loss function, cross-entropy loss function, fireline color loss function, and fireline region regularization loss function. The hybrid function is defined as follows:

[0074]

[0075] Among them, L Fid L BCE L Color and L FLA Let λ represent the fidelity loss function, cross-entropy loss function, fireline color loss function, and fireline region regularization loss function, respectively. f , λ L , λ c and λ flr These are the weights of the four types of loss functions.

[0076] Fidelity loss is used to ensure consistency between the output and the label (ground truth), and it is defined as:

[0077]

[0078] Cross-entropy loss is used to establish the relationship between the ground truth and the frontal significance result. It is defined as follows:

[0079]

[0080] While cross-entropy loss can optimize the model's extraction of fire line boundaries, the model still lacks sufficient constraints on the color of the fire line, resulting in blurred boundaries between the fire line and the smoke near the flame. To address this issue, this invention constructs a color loss to explicitly learn the fire line boundaries, which can be expressed as:

[0081]

[0082] The fire region regularization loss function is used to maintain consistency between the fire region results and the labels. It is defined as follows:

[0083]

[0084] S3, Training the network model

[0085] The proposed network is implemented using the PyTorch framework running on a GPU (NVIDIA Ti3090 GPU). The training images obtained in S1 are cropped to 224×224 pixels and subjected to data augmentation, then used as input together according to the batch size (Batch Size = 32). Regarding the hyperparameters of the proposed method, λ is empirically set... f ,λ2,λ3,λ4,λ c and λ flr Set them to 1, 0.8, 0.6, 0.4, 0.5 and 10 respectively.

[0086] During network training, the initial learning rate was set to 0.0005, and the learning rate decayed at a rate of 0.05 times every 30 iterations. The Adam optimizer was used, and the trained model weights were saved step by step.

[0087] S4. Fireline Identification Result Analysis

[0088] This example demonstrates a fire detection experiment based on 30% of the test dataset in S1, and uses MAE and S... m F max Quantitative and qualitative analyses of the evaluation index expansion results are presented. Qualitative comparisons between the method of this invention and current mainstream methods on the dataset independently constructed in this invention are shown. Figure 2 Images refers to the original input images, GT refers to the labels (Ground truth), Ours refers to the method proposed in this invention, and PoolNet, GCPANet, CTDNet, SAMNet, and HVPNet refer to current mainstream methods. Quantitative comparison results are shown in Table 1 (the best value for each comparison metric is shown in bold). Qualitatively, compared with mainstream methods, the method proposed in this invention can obtain more complete and accurate fireline boundaries; other methods generally suffer from discontinuous firelines and uneven boundaries. Quantitatively, this invention achieves better results in mean absolute error (MAE), F-score (F...). max ) and structural similarity (S m The method achieved the best value among the three comparison indicators. The leading results in the qualitative comparison and quantitative comparison indicators demonstrate the superiority of this method.

[0089] Table 1. Comparison of quantitative results with mainstream methods on the Fireline dataset.

[0090] Backbone ResNet50 ResNet50 ResNet50 ResNet50 ResNet50 ResNet50 MAE↓ 0.0202 0.0223 0.0251 0.0264 0.0283 0.0152 <![CDATA[S m ↑]]> 0.8795 0.8737 0.8261 0.8328 0.8374 0.8996 <![CDATA[F max ↑]]> 0.8798 0.8747 0.8316 0.8394 0.8390 0.9098

[0091] This invention proposes a network method for identifying forest fire lines. By mining the color features and structural features of the fire line area, and making full use of the inherent relationship between color and structural features, the method achieves effective identification of forest fire lines. The identification results provide important technical support for the quantification and monitoring of fire lines.

[0092] This invention proposes a hybrid loss function. Experiments on the test dataset show that the network of this invention can better focus on the fire line color and region information, and the identified fire line contour has better smoothness and continuity.

[0093] The network proposed in this invention has achieved an advanced level in the task of forest fire line identification based on UAVs. Through quantitative and qualitative comparative experimental analysis, it can be seen that the index values ​​of the network of this invention are all ranked first.

[0094] Compared to existing joint image segmentation and Canny edge detection techniques, this invention focuses on the inherent characteristics of the fire line, giving it greater emphasis on the fire line outline. The technical solution of this invention focuses on the color and regional features of the fire line, effectively combining fire spread mechanisms with deep neural network technology to deeply mine fire line color information and pay close attention to fire line region information, significantly improving the continuity and smoothness of fire line recognition results.

[0095] Furthermore, during the model design process, the image segmentation model focusing on the fire line area faces challenges. The presence of non-fire line information within the fire line area increases the difficulty for the model to extract valid information and learn fire line features, resulting in the inclusion of non-fire line information in the final segmentation results. After Canny edge detection, this non-fire line information appears as linear edges resembling fire lines, posing challenges to subsequent forest fire line quantification and fire front monitoring.

[0096] In summary, when it comes to identifying forest fire lines, the technical solution of this invention focuses more on the fire line itself, and the identified fire line results have more refined edge information and smoother fire line characteristics.

[0097] This invention proposes a network method for identifying forest fire lines. By mining the color features and structural features of the fire line area, and fully utilizing the inherent relationship between color and structural features, it achieves effective identification of forest fire lines. The identification results provide important technical support for the quantification and monitoring of fire lines. The proposed fire line color guidance module and fire line area extraction module are the key points and areas to be protected in this invention.

[0098] This invention proposes a hybrid loss function. Experiments on the test dataset show that the network of this invention can better focus on the color and region information of the fire line, and the identified fire line contours have better smoothness and continuity. The hybrid loss function designed specifically based on the proposed network model serves as the key point and protection point of this invention.

[0099] This invention constructs a dedicated dataset of forest fire line images based on unmanned aerial vehicles (UAVs) by combining field burning experiments. This dataset provides crucial data support for forest fire line identification and has significant engineering application value. The constructed dedicated dataset is the key point and the area to be protected in this invention.

[0100] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying forest fire lines based on unmanned aerial vehicles (UAVs) and deep neural networks, characterized in that: The identification method includes the following steps: S1: Prepare the training dataset by combining drone imagery data collected from the field burning experiment with open-source image data collected by Kaggle; 70% of the dataset is used as the training set and 30% as the test set. S2: Establish a network model, use the fire line color feature module to guide the network to obtain the fire line color feature extraction result image of the input image, and then use the fire line region module to constrain the input image and the extraction result. The two are cascaded to obtain the recognition result image that simultaneously contains fire line color features and fire line region structure features. S3: Train the network model. The Fireline dataset prepared in step S1 is uniformly batch-adjusted to a size of 224×224 pixels as the model input. Set the key parameters for network model training: initial learning rate, learning decay rate, and number of iterations. Optimize the network parameters using the designed hybrid loss function and the Adam optimizer. S4: Perform fire line recognition. Input the above 30% of the test images into the constructed network model, load the pre-trained network parameters, perform fire line recognition, and obtain the recognition images and evaluation index results. In step S2, the network mainly includes a backbone network, a fire line color guidance module, and a fire line region extraction module, among which the fire line color guidance module and the fire line region extraction module are two core parts. The input of the fire line color guidance module is the original image of the forest fire line, and the output is the preliminary extraction result of the fire line color. The input of the fire line region extraction module is the original image and the preliminary extraction result. This module mainly uses the cross-entropy loss function to reduce the difference in the size of the fire line region between the spatial attention map of the preliminary extraction result and the ground reference image. In step S2, the color mapping function extracted by the FireWire color guide module can be expressed as equation (1), and the equations for converting RGB to YCbCr and HSV can be described as equations (2)-(5). (1) (2) (3) (4) (5) In the formula, (R, G, B) represent the values ​​of the three color channels of the input image; the Color Mapping Guided (CMG) component applies the extracted color map to multiple multi-level features. The results are then used as a hierarchical supervised loss function to train the network, helping it focus on the flame color region and obtain a more accurate fire line region; specifically, CMG can be represented as: (6) Among them Multi-level features The layer depth in the network. This represents the extracted color image. The input image.

2. The forest fire line identification method based on UAVs and deep neural networks as described in claim 1, characterized in that: In step S1, a self-constructed fireline dataset was used; this dataset consists of 2,000 images composed of public resources and real wildfire experiments; specifically, 600 images were collected from the Kaggle forest fire dataset, and 1,400 drone images were collected from a field burning experiment in Xichang City, Sichuan Province; in the experiment, 70% of the fireline dataset was used as the training dataset, and 30% of the fireline dataset was used as the test dataset.

3. The forest fire line identification method based on UAVs and deep neural networks as described in claim 1, characterized in that: In step S2, the fire region extraction module first multiplies the input image with the preliminary extraction result and the ground reference image to extract the fire region of the preliminary result and the ground reference image. Then, a spatial attention layer is used to focus on the internal correlation between the preliminary extraction result and the ground reference image. Finally, the cross-entropy loss function is used to reduce the difference in the size of the fire region between the spatial attention map of the preliminary extraction result and the ground reference image. In this example, following the fire color guidance module, the proposed fire region extraction module can help the network to focus on the fire region information on the basis of focusing on the fire color features. From the perspective of qualitative analysis of fire recognition, the fire color guidance module can effectively help the model distinguish between flames and smoke, while the fire region extraction module can improve the continuity and smoothness of the recognized contours.

4. The forest fire line identification method based on UAVs and deep neural networks as described in claim 1, characterized in that: In step S2, a hybrid loss function is used, which combines the fidelity loss function, cross-entropy loss function, fireline color loss function, and fireline region regularization loss function; the hybrid function is defined as: (7) in, , , and Let represent the fidelity loss function, cross-entropy loss function, fireline color loss function, and fireline region regularization loss function, respectively. , , and These are the weights of the four types of loss functions; Fidelity loss is used to ensure consistency between the output and the label (Groundtruth), and it is defined as: (8) Cross-entropy loss is used to establish the relationship between ground truth and the frontier significance result. It is defined as follows: (9) While cross-entropy loss can optimize the model's extraction of fire line boundaries, the model still lacks sufficient constraints on the color of the fire line, resulting in a blurred boundary between the fire line and the smoke near the flame. To address this issue, a color loss is constructed to explicitly learn the fire line boundaries, which can be expressed as: (10) The fire region regularization loss function is used to maintain consistency between the fire region results and the labels. It is defined as follows: (11)。 5. The forest fire line identification method based on UAVs and deep neural networks as described in claim 1, characterized in that: In step S3, the training images obtained in S1 are cropped to 224×224 pixels and subjected to data augmentation. Then, they are used as input along with the batch size (BatchSize=32). Regarding the hyperparameters of the proposed method, based on experience... , , , , and Set them to 1, 0.8, 0.6, 0.4, 0.5 and 10 respectively; When training the network, the initial learning rate is set to 0.0005, and the learning rate decays at a rate of 0.05 times every 30 iterations; the Adam optimizer is used, and the trained model weights are saved step by step.

6. The forest fire line identification method based on UAVs and deep neural networks as described in claim 1, characterized in that: In step S4, a fire detection experiment is conducted based on the 30% test dataset from S1, and MAE is used. , Quantitative and qualitative analysis of the evaluation indicators.

7. The forest fire line identification method based on UAVs and deep neural networks as described in claim 1, characterized in that: In step S4, the qualitative comparison results of this method and current mainstream methods on the self-constructed dataset are shown in Figure 2. Images refer to the original input images, GT refers to the labels (Groundtruth), Ours refers to this method, and PoolNet, GCPANet, CTDNet, SAMNet, and HVPNet refer to current mainstream methods. The quantitative comparison results are shown in Table 1 (the best value of each comparison index is shown in bold). From the qualitative comparison, compared with mainstream methods, the proposed method can obtain more complete and accurate fire line boundaries. The fire line contours extracted by other methods generally have problems of discontinuous fire lines and unsmooth boundaries. From the quantitative comparison, in terms of mean absolute error (MAE), F score ( ) and structural similarity ( The method achieved the best value among the three comparison indicators. The leading results in the qualitative comparison and quantitative comparison indicators demonstrate the superiority of this method.

Images