Forest fire-fighting early warning system based on image processing

Abstract

The invention relates to the technical field of image data processing, and discloses a forest fire-fighting early warning system based on image processing, which is characterized by comprising a data acquisition module, a route optimization module, a grading early warning module and a wireless communication module, the data acquisition module is used for acquiring meteorological data, historical fire behavior records and fire recognition results; the data acquisition module comprises a meteorological data unit, a historical fire behavior recording unit and an unmanned aerial vehicle. Meteorological data, historical fire behavior records and fire recognition results are obtained at the same time, fire is judged and recognized in a multi-dimensional mode, forest fire early warning is more accurately achieved, meanwhile, the route optimization module is matched, the patrol route of the unmanned aerial vehicle is optimized, it is ensured that no dead corner exists in forest area coverage, and the forest fire early warning efficiency is improved. The problem of report failure caused by monitoring range omission is avoided, and the early warning range and reliability are enhanced under the limitation of limited resources.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and more specifically, to a forest fire early warning system based on image processing. Background Technology

[0002] Forest fires are one of the major natural disasters facing the world, characterized by long burning times, large burned areas, and high fire intensity, causing serious damage to ecosystems and the surrounding environment. Traditional forest fire prevention methods mainly rely on manual lookout and ground patrols, which are inefficient and have limited coverage. With the development of technology, aerial patrol and satellite remote sensing technologies have begun to be applied in forest fire prevention, but these technologies are limited by factors such as weather, lighting conditions, and real-time availability. Unmanned aerial vehicle (UAV) technology, with its advantages of flexibility, low cost, and real-time monitoring, is being widely used in forest fire prevention.

[0003] However, existing technologies still have many shortcomings: First, data from a single sensor is difficult to fully reflect the characteristics of a fire, which can easily lead to false alarms or missed alarms; second, drone patrol routes are mostly preset or simply dynamically adjusted, lacking comprehensive consideration of meteorological data, historical fire records and real-time fire risks.

[0004] Therefore, a forest fire early warning system is needed to improve the accuracy of fire detection during forest fire patrols. Summary of the Invention

[0005] To achieve the above objectives, this application provides a forest fire early warning system based on image processing, comprising: a data acquisition module, a route optimization module, a graded early warning module, and a wireless communication module; The data acquisition module is used to acquire meteorological data, historical fire records, and fire identification results. The data acquisition module includes: a meteorological data unit, a historical fire record unit, and a drone; The meteorological data unit is used to acquire meteorological data; The historical fire record unit is used to acquire historical fire records; The drone is used to acquire forest image data and obtain fire identification results based on the forest image data. The route optimization module is used to generate the optimal patrol route for the drone based on meteorological data, historical fire records, and fire identification results. The graded early warning module is used to calculate the fire level based on meteorological data, historical fire records and fire identification results, and to determine the early warning information based on the fire level. The drone patrols the forest according to the optimal patrol route, acquires forest image data, and obtains fire identification results based on the forest image data. The fire identification results are then transmitted to the route optimization module and the graded early warning module. The meteorological data unit transmits meteorological data to the route optimization module and the graded early warning module. The historical fire record unit transmits historical fire records to the route optimization module and the graded early warning module. The route optimization module updates the drone's optimal patrol route and feeds it back to the drone. After the graded early warning module determines the early warning information, the wireless communication module transmits the early warning information to the fire command center.

[0006] Furthermore, the drone includes: Image acquisition submodule: used to acquire forest images; the forest images include: multispectral images and thermal imaging images; Flight control submodule: Used to control the UAV to patrol the forest area according to the optimal patrol route; Image processing submodule: used to process multispectral images and thermal imaging images to obtain target image features; Fire identification submodule: Used to identify target image features and obtain fire identification results based on the target image features.

[0007] Furthermore, the image processing submodule processes multispectral images and thermal imaging images to obtain target image features, including the following steps: S101. Perform spatial alignment processing on the multispectral image and the thermal imaging image to obtain a spatially aligned image; S102. Extract image features from spatially aligned images using a convolutional neural network to obtain multi-layer spatially aligned image features; S103. Fuse multi-layer spatially aligned image features to obtain fused image features; S104. Adjust the weights of the fused image features to obtain the adjusted fused image features; S105. The adjusted fused image features are enhanced by an attention mechanism to obtain the target image features.

[0008] Furthermore, the expression for the spatial alignment process is: in, These are the horizontal coordinates of the ground points corresponding to the multispectral image. The horizontal coordinates of the ground point corresponding to the thermal imaging image; is the pixel equivalent coefficient of the multispectral camera; is the pixel equivalent coefficient of the infrared thermal imager; The focal length of the multispectral camera; The focal length of the infrared thermal imager; The baseline distance between the multispectral camera and the infrared thermal imager; For parallax; The altitude of the drone above the ground; The coordinates of the x-th pixel of the ground point in the multispectral image; The x-coordinate of the ground point pixel in the thermal image; The expression for extracting image features from a spatially aligned image is: in, This represents the feature map of the nth layer; Represents the weights of the nth convolutional layer; This represents the bias of the nth convolution layer; This represents the convolution operation; This represents the input data for the nth layer; Indicates the activation function; Indicates the network layer index; The expression for fusing multi-layer spatially aligned image features is: in, This represents the features of the nth layer after fusion; Represents the inverted index feature map; Indicates an upsampling operation; Indicates the fusion features of the previous layer; This represents the fusion features of the first layer; This represents the feature map of the 4th layer.

[0009] Furthermore, the weight adjustment of the fused image features specifically refers to adjusting the fused image features through a dynamic weight adjustment mechanism; The expression for the dynamic weight adjustment mechanism is: in, Represents the fusion feature of the nth layer; Represents the multispectral feature weights of the nth layer; This represents the feature weights of the nth thermal imaging layer; and These are the nth layer features from multispectral and thermal imaging, respectively; This represents the temperature gradient of the nth layer in thermal imaging. is the normalized vegetation index of the nth layer of the multispectral system; Sensitivity coefficient; The NDVI threshold for vegetation; The expression for enhancing the attention mechanism of the fused image features is as follows: in, For attention output features; Input features to the attention mechanism; This represents the Sigmoid activation function; These are the first-order learnable parameters; These are second-order learnable parameters; This indicates a global average pooling operation; This represents the activation function of the rectified linear unit.

[0010] Furthermore, the route optimization module includes, Fire risk index calculation unit: used to calculate the regional fire risk index based on meteorological data and historical fire records; Multi-objective reward function unit: used to acquire UAV flight parameters, perform reward calculations on the UAV flight parameters, and obtain multi-objective reward results.

[0011] Furthermore, the expression for calculating the regional fire risk index based on meteorological data and historical fire records is as follows: in, The value represents the regional fire risk index at time t; 0.6, 0.3, and 0.1 represent weighting coefficients. This indicates the fire hazard index based on temperature. This indicates the wind speed and fire risk index; This represents the normalization coefficient for historical fire point density; The density of historical fire points at time t is obtained by estimating using Gaussian kernel density. The real-time fire risk probability at time t is output by the UAV image recognition model. The expression for calculating the reward based on the UAV flight parameters is as follows: in, Indicates the target reward result; Indicates the coverage efficiency weight; Indicates the weight of fire risk avoidance; Indicates response speed weight; Indicates the energy consumption penalty weight; The area of ​​the covered region; The total patrol area; The area fire risk index at time t; To prevent division by zero for the minimum value; For response time; This is the maximum allowable response time. This represents the amount of electricity already consumed. This is the maximum battery capacity.

[0012] Furthermore, the tiered early warning module includes: Fire rating calculation unit: used to calculate the fire rating based on the fire identification results; Fire rating correction unit: Used to correct the fire rating based on meteorological data.

[0013] Furthermore, the calculation expression for determining the fire level based on the fire identification results is as follows: in, Indicates the fire severity level; Represents the floor function; Indicates the weight of the fire area; Indicates the area of ​​the fire; Indicates the maximum fire area; Indicates the temperature gradient weights; Represents the temperature gradient; Indicates the maximum temperature gradient; Indicates the spread rate weight; Indicates the speed at which the fire spreads; Indicates the maximum spread rate; The expression for correcting the fire level based on meteorological data is as follows: in, This indicates the adjusted fire risk level. Indicates the fire severity level; Wind speed; Maximum wind speed; Humidity; Maximum humidity; Indicates the wind speed sensitivity coefficient; This is the humidity sensitivity coefficient.

[0014] Furthermore, the wireless communication module includes: Decomposition and Compression Unit: Used to perform lifting wavelet decomposition on the early warning information, obtain decomposition data, quantize and compress the high-frequency coefficients in the decomposition data, and compress the low-frequency coefficients in the decomposition data through LZW encoding. Dynamic bandwidth allocation unit: used to acquire bandwidth parameters, calculate bandwidth allocation based on bandwidth parameters, and transmit early warning information based on bandwidth allocation; The expression for calculating bandwidth allocation based on bandwidth parameters is as follows: in, This indicates the bandwidth allocation at time t; This indicates the adjusted fire risk level. Maximum bandwidth; Average bandwidth; This is the minimum bandwidth.

[0015] The beneficial effects of this invention are as follows: 1. The data acquisition module simultaneously acquires meteorological data, historical fire records, and fire identification results, making multi-dimensional judgments and identifications of fires, thereby achieving more accurate early warning of forest fires. At the same time, in conjunction with the route optimization module, it optimizes the patrol routes of drones to ensure that there are no blind spots in forest areas, avoiding the problem of missed reports due to omissions in the monitoring range, and enhancing the scope and reliability of early warning under the constraints of limited resources. 2. The image acquisition submodule in the UAV simultaneously acquires multispectral images and thermal imaging images, overcoming the limitation that single sensor data cannot fully reflect the characteristics of a fire. It can capture the spectral and temperature characteristics of a fire separately, providing a rich data foundation for accurate identification. 3. The image processing submodule in the UAV first performs spatial alignment processing on the two images, which solves the problem of spatial resolution difference of multimodal data and ensures the accuracy of subsequent feature extraction. Then, through feature fusion processing, the image features of multispectral and thermal imaging are effectively integrated to make the fire feature expression more comprehensive. The subsequent channel attention mechanism enhancement processing can highlight the features of key fire areas, suppress irrelevant background interference, and further refine the effective features. 4. The fire identification submodule in the drone identifies fire features such as smoke and flames based on target image features that have been optimized through multiple steps. Compared with the traditional identification method that relies on a single feature or has not been optimized, it can more accurately distinguish between natural landscapes and fire signs, thereby significantly improving the accuracy of fire detection during forest fire prevention patrols. 5. The graded early warning module calculates the fire level based on meteorological data, historical fire records, and fire identification results, thereby accurately predicting the fire hazard. The early warning information is then transmitted to the fire command center via the wireless communication module, enabling the fire command center to take reasonable fire-fighting measures according to the fire hazard and magnitude. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system structure provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the UAV structure provided in an embodiment of the present invention. Detailed Implementation

[0017] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Example 1: like Figure 1 As shown, a forest fire early warning system based on image processing includes: a data acquisition module, a route optimization module, a graded early warning module, and a wireless communication module; The data acquisition module is used to acquire meteorological data, historical fire records, and fire identification results. The data acquisition module is the fundamental input unit of the integrated early warning system. Its core function is to integrate multi-source heterogeneous data to provide comprehensive and accurate analytical basis for subsequent decision-making modules. It is widely used in various intelligent early warning and decision support systems. In this application, the module breaks through the limitations of traditional reliance on real-time monitoring data. On the one hand, it acquires high-precision fire identification results transmitted by UAVs, including core information such as fire location, scale, and characteristics. On the other hand, it actively collects meteorological data and historical fire records. Meteorological data, such as wind speed, humidity, and temperature, directly affect the speed and extent of fire spread, while historical fire records reflect the regional fire risk patterns and fire evolution characteristics. These three types of data form a complete data chain of real-time status, environmental impact, and historical patterns, providing multi-dimensional support for route optimization and graded early warning.

[0019] The data acquisition module includes: a meteorological data unit, a historical fire record unit, and a drone; The meteorological data unit is used to acquire meteorological data; The historical fire record unit is used to acquire historical fire records; The drone is used to acquire forest image data and obtain fire identification results based on the forest image data. The drone patrols the forest according to the optimal patrol route, acquires forest image data, and obtains fire identification results based on the forest image data. The fire identification results are then transmitted to the route optimization module and the graded early warning module. The meteorological data unit transmits meteorological data to the route optimization module and the graded early warning module. The historical fire record unit transmits historical fire records to the route optimization module and the graded early warning module. The route optimization module updates the drone's optimal patrol route and feeds it back to the drone. After the graded early warning module determines the early warning information, the wireless communication module transmits the early warning information to the fire command center.

[0020] like Figure 2 As shown, the drone includes: Image acquisition submodule: used to acquire forest images; the forest images include: multispectral images and thermal imaging images; The image acquisition submodule is a core component in various image recognition systems used to collect image data of target scenes. It can be equipped with different types of image acquisition devices according to application requirements to acquire single-modal or multi-modal image data, providing a foundation for subsequent processing and recognition.

[0021] In this embodiment, the image acquisition submodule is equipped with a multispectral camera with a wavelength range of 400-900nm and an infrared thermal imager with a wavelength range of 7500-13500nm. Together with a high-definition camera and a stabilized gimbal, it simultaneously acquires multispectral and thermal images of the forest area during drone patrols, capturing the spectral and temperature characteristics of the forest, respectively. This overcomes the limitation that single sensor data cannot fully reflect the characteristics of a fire, and provides rich multimodal data support for subsequent accurate fire identification.

[0022] Flight control submodule: Used to control the UAV to patrol the forest area according to the optimal patrol route; The flight control module is the core control unit of the UAV. It is responsible for receiving preset instructions or real-time control signals and adjusting the UAV's flight attitude, speed and direction to ensure that the UAV completes flight operations according to the predetermined task. It is widely used in various UAV patrol, surveying and mapping scenarios.

[0023] In this embodiment, the flight control module receives the optimal patrol route preset by the ground terminal according to the needs of forest fire prevention patrol, and accurately controls the drone to conduct all-round and multi-angle patrols in the forest area to ensure that there are no blind spots in the monitoring range and avoid fires being missed due to local areas not being covered, thus providing flight support for comprehensive image data collection.

[0024] Image processing submodule: used to process multispectral images and thermal imaging images to obtain target image features; The image processing submodule is a key part of the forest consumption early warning system. It performs preprocessing, feature extraction, and feature optimization on the raw image data through a series of algorithms to eliminate data redundancy and interference, improve the effectiveness and recognizability of features, and provide high-quality input for the subsequent recognition module.

[0025] The image processing submodule processes multispectral images and thermal imaging images to obtain target image features, including the following steps: S101. Perform spatial alignment processing on the multispectral image and the thermal imaging image to obtain a spatially aligned image; Spatial alignment processing is based on a binocular parallax model. By using a preset formula for calculating the coordinates of ground points, it solves the problem of spatial resolution differences between multispectral images and thermal imaging images, ensuring accurate spatial matching of the two modal data.

[0026] The expression for the spatial alignment process is: in, These are the horizontal coordinates of the ground points corresponding to the multispectral image. The horizontal coordinates of the ground point corresponding to the thermal imaging image; is the pixel equivalent coefficient of the multispectral camera; is the pixel equivalent coefficient of the infrared thermal imager; The focal length of the multispectral camera; The focal length of the infrared thermal imager; The baseline distance between the multispectral camera and the infrared thermal imager; For parallax; The altitude of the drone above the ground; The coordinates of the x-th pixel of the ground point in the multispectral image; The x-coordinate of the ground point pixel in the thermal image; In multimodal image fusion processing, determining the pixel coordinates of the same target in different modal images is a prerequisite for spatial alignment. By matching pixel coordinates, the correspondence between different images can be established, providing a basis for subsequent spatial position calibration.

[0027] This technique involves simultaneously capturing images using image acquisition equipment mounted on a drone, selecting easily identifiable ground points in the forest as reference points, such as treetops, rocks, and landmarks. The pixel x-coordinates of these reference points, namely x1 and x2, are located in both multispectral and thermal imaging images, ensuring that the pixel positions for the same ground target in both types of images are traceable, thus providing accurate input for subsequent spatial coordinate calculations.

[0028] Among these parameters, the inherent parameters of the camera and imaging equipment are crucial for the transformation between image pixels and real-world spatial coordinates. These parameters are typically obtained in advance through equipment calibration and serve as the foundational data for image spatial positioning. In this application, the parameters to be obtained include the pixel equivalent coefficient k1 and focal length f1 of the multispectral camera, the pixel equivalent coefficient k2 and focal length f2 of the infrared thermal imager, and the installation baseline distance B of the two types of equipment on the UAV. These parameters, combined with the real-time altitude z of the UAV during flight, form a complete set of spatial calculation parameters to ensure the effective application of subsequent coordinate transformation formulas.

[0029] Spatial alignment is a core preprocessing step in multimodal image fusion. By establishing a mapping relationship between pixel coordinates and real spatial coordinates, it eliminates problems such as spatial offset and scale differences caused by shooting from different devices, ensuring spatial consistency of images of different modalities.

[0030] This technique is based on a binocular parallax model. First, it calculates the true horizontal coordinates of ground points in the two types of images. Then, it uses the spatial position constraint relationship X1=X2-B to calibrate the deviation. Finally, it calculates the parallax by D=x2-x1 to achieve precise spatial alignment between multispectral images and thermal imaging images, thus solving the problem of spatial resolution differences between the two types of images caused by different shooting equipment and imaging principles.

[0031] In this embodiment, spatial alignment processing is performed on multispectral images and thermal imaging images to obtain spatially aligned images. By determining the pixel coordinates of the same ground point in the two types of images, the correspondence between the multimodal images is established, providing a precise target point for spatial alignment. By acquiring the inherent parameters of the equipment and the real-time flight altitude of the UAV, the accuracy of spatial coordinate transformation is ensured. Spatial alignment is achieved through the simultaneous calculation of specific formulas, effectively eliminating spatial offsets and scale differences caused by the different imaging principles and installation positions of the multispectral camera and the infrared thermal imager. This ensures precise matching of the spatial positions of the two types of images, laying a reliable spatial foundation for subsequent feature extraction and fusion processing. It avoids the problems of mismatch and missed extraction of fire features caused by spatial misalignment, thereby ensuring the effectiveness of the input features of the fire identification model and indirectly improving the accuracy and reliability of fire detection during forest fire prevention patrols.

[0032] S102. Extract image features from spatially aligned images using a convolutional neural network to obtain multi-layer spatially aligned image features; In this embodiment, four-layer feature maps of the two spatially aligned images are extracted using a ResNet-50 network.

[0033] The expression for extracting image features from a spatially aligned image is: in, This represents the feature map of the nth layer; Represents the weights of the nth convolutional layer; This represents the bias of the nth convolution layer; This represents the convolution operation; This represents the input data for the nth layer; Indicates the activation function; Indicates the network layer index; ResNet-50 is a deep convolutional neural network based on residual connections. It has powerful deep feature extraction capabilities and can effectively solve the gradient vanishing problem in deep network training. It can extract multi-dimensional features from images layer by layer, from shallow details to deep semantics. It is a classic feature extraction model in the field of image recognition.

[0034] This technique extracts features independently from spatially aligned multispectral and thermal images. After calculation by a four-layer network, four-layer feature maps are obtained for the two types of images. These feature maps retain fire-related spectral features and temperature-related features, providing high-quality basic features for subsequent fusion.

[0035] S103. Fuse multi-layer spatially aligned image features to obtain fused image features; Multi-scale feature fusion is achieved by constructing a Feature Pyramid Network (FPN).

[0036] The expression for fusing multi-layer spatially aligned image features is: in, This represents the features of the nth layer after fusion; Represents the inverted index feature map; Indicates an upsampling operation; Indicates the fusion features of the previous layer; This represents the fusion features of the first layer; This represents the feature map of the 4th layer.

[0037] Feature Pyramid Network (FPN) is a multi-scale feature fusion architecture that integrates image features from different levels through a top-down upsampling and lateral connection mechanism. It solves the problems of disconnect between shallow detail features and deep semantic features and insufficient extraction of small target features in traditional feature fusion, and can significantly improve the comprehensiveness and effectiveness of features.

[0038] This technique is based on the four-layer feature map extracted by ResNet-50. First, the deepest layer is used as the first layer of the fusion feature. Then, the previous layer of fusion feature is upsampled and element-wise added to the corresponding feature map of the inverse index. Finally, the deep fusion of features at different scales of multispectral images and thermal imaging images is achieved, so that the fusion feature contains both detailed information and semantic information.

[0039] In this embodiment, deep feature extraction is performed on spatially aligned multispectral and thermal imaging images using a ResNet-50 network. This effectively uncovers deep, multi-dimensional fire-related features in both types of images, avoiding the problems of insufficient feature extraction and omission of key information in traditional shallow networks, thus providing a high-quality foundation for fusion. By constructing a feature pyramid network and fusing it according to a specific formula, the organic combination of shallow detail features and deep semantic features is achieved, solving the pain point of difficulty in fusing features at different scales in multimodal images. This allows the fused features to simultaneously possess spectral details, temperature correlation information, and semantic contours of the fire, significantly improving the comprehensiveness and recognizability of the features. This provides high-quality input for subsequent dynamic weight adjustment and channel attention mechanism enhancement, thereby ensuring the accurate capture of features such as smoke and flames by the fire identification module, reducing the risk of false alarms and missed alarms due to insufficient features, and improving the accuracy of fire detection during forest fire patrols.

[0040] S104. Adjust the weights of the fused image features to obtain the adjusted fused image features; Based on the temperature gradient ∆T of the thermal imaging n and multispectral normalized vegetation index (NDVI) n The fusion weights of the two types of features are automatically adjusted to make the fused features better suited to the needs of fire identification.

[0041] The aforementioned weight adjustment of the fused image features specifically refers to adjusting the fused image features through a dynamic weight adjustment mechanism; The expression for the dynamic weight adjustment mechanism is: in, Represents the fusion feature of the nth layer; Represents the multispectral feature weights of the nth layer; This represents the feature weights of the nth thermal imaging layer; and These are the nth layer features from multispectral and thermal imaging, respectively; This represents the temperature gradient of the nth layer in thermal imaging. is the normalized vegetation index of the nth layer of the multispectral system; Sensitivity coefficient; The NDVI threshold for vegetation; Dynamic weight adjustment mechanism is the core optimization technology of multimodal feature fusion. By dynamically allocating the weights of different modal features according to the saliency of the data itself and the adaptability of the scene, it breaks through the limitation that fixed weights cannot adapt to complex scenes, making the fused features more in line with the specific task requirements, thereby improving the accuracy of subsequent recognition models.

[0042] This mechanism achieves precise control through three sets of correlation formulas: the first set of fusion formulas clarifies the weighted fusion logic of multispectral features and thermal imaging features; the second set of weight formulas takes the temperature gradient of thermal imaging as the core basis, the larger the temperature gradient, the more likely it is to be a high-temperature area of ​​fire, and the weight of multispectral features is adaptively reduced to avoid interference from irrelevant spectral information; the third set of weight formulas takes the deviation between the normalized vegetation index (NDVI) of multispectral features and the vegetation threshold as the basis, the further the NDVI value deviates from the normal vegetation range, the more likely the fire has caused vegetation damage or smoke obstruction, and the weight of thermal imaging features is adaptively increased, ultimately achieving the optimal fusion ratio of the two types of features under different scenarios.

[0043] Based on a dynamic weight adjustment mechanism and three specific formulas, the fused features can adaptively allocate weights according to the temperature gradient of thermal imaging and the NDVI value of multispectral imaging. In suspected fire areas, when the temperature gradient is large and the NDVI value deviates from the normal vegetation range, the weight of thermal imaging features is reasonably increased, strengthening the expression of fire-related temperature core features. In normal forest areas, the two types of features are balanced and integrated to avoid information loss caused by an excessive proportion of a single feature. This dynamically adaptable fusion method makes the adjusted fused features more targeted and recognizable, effectively solving the problem that fixed-weight fusion cannot adapt to complex fire scenarios. It provides higher-quality feature input for subsequent channel attention mechanism enhancement and fire identification module, thereby significantly improving the accuracy of fire identification and reducing false alarms and missed alarms caused by unreasonable feature fusion. It provides key technical support for efficient monitoring of forest fire prevention patrols.

[0044] S105. The adjusted fused image features are enhanced by an attention mechanism to obtain the target image features.

[0045] Attention mechanism enhancement processing is introduced by introducing ECA, or efficient channel attention mechanism, which enhances the model's attention to key features of the fire area, such as smoke and flames, through a specific formula, while suppressing interference from irrelevant background features, and finally outputting purified target image features.

[0046] The expression for enhancing the attention mechanism of the fused image features is as follows: in, For attention output features; Input features to the attention mechanism; This represents the Sigmoid activation function; These are the first-order learnable parameters; These are second-order learnable parameters; This indicates a global average pooling operation; This represents the activation function of the rectified linear unit.

[0047] Channel attention mechanism is a key technology in the field of deep learning for optimizing feature representation. By assigning adaptive weights to different channels of image features, it strengthens the key channel features that are relevant to the task and suppresses redundant information in irrelevant channels, thereby improving the model's ability to recognize targets. ECA, or efficient channel attention mechanism, is a lightweight channel attention architecture that can efficiently capture the dependencies between channels without dimensionality reduction operations. It reduces computational complexity while ensuring performance and is suitable for scenarios with limited computing power, such as drones.

[0048] First, Global Average Pooling is used to process the fused image features, compressing the spatial dimension while preserving channel-level global information. Then, a linear transformation is performed using first-order learnable parameters, followed by ReLU (Rectified Linear Unit) activation to introduce non-linearity and enhance the model's ability to capture non-linear features. Next, second-order learnable parameters are used to further adjust the feature dimension. Finally, the output is mapped to the 0-1 range using the σ (Sigmoid) activation function to obtain the attention weights for each channel. These weights are then multiplied element-wise with the original fused features to enhance key fire channel features (e.g., smoke and flame) and suppress redundant background features (e.g., normal vegetation and terrain), outputting purified target image features.

[0049] The enhanced processing of fused features is achieved through a specific formula of the ECA channel attention mechanism, which can efficiently capture the dependencies between channels without dimensionality reduction, achieving high efficiency in feature optimization under the limited computing power of UAVs. Global average pooling preserves channel-level global information, ensuring that the overall expression of key fire features is not missed. The combination of ReLU and Sigmoid activation functions effectively filters out effective features and suppresses invalid information. The learnable parameters are trained on massive fire samples to adapt to forest fire prevention scenarios, accurately identifying fire-related feature channels and assigning them high weights. This processing further refines the fused features, making the target image features more focused on core fire features such as smoke and flames, significantly reducing the interference of background redundancy information on subsequent recognition, directly improving the recognition accuracy and anti-interference ability of the fire recognition module, effectively reducing false alarms caused by feature mixing, and providing key feature optimization support for accurate fire detection in forest fire prevention patrols.

[0050] Fire identification submodule: Used to identify target image features and obtain fire identification results based on the target image features.

[0051] The fire identification submodule is a functional unit that judges fires based on image features. Through a trained deep learning model, it filters out fire-specific features from the input image features, and then judges whether a fire exists and outputs relevant results. It is the core functional module of fire monitoring equipment.

[0052] In this embodiment, the fire identification submodule relies on the AI ​​computing unit carried by the drone. Based on the target image features optimized by the image processing module, it accurately identifies fire features such as smoke and flames in the forest through a trained multimodal data fusion deep learning model. It effectively distinguishes natural landscapes such as clouds and shadows from fire signs, while marking key information such as the location and scale of the fire, and outputs accurate fire identification results to provide a basis for subsequent early warning.

[0053] In this embodiment, during a forest fire prevention patrol mission in a temperate deciduous forest covering an area of ​​80 square kilometers, a route optimization module was used to plan a preset route for the UAV with a grid pattern and key area densification, covering key areas such as forest edges, ridges, and historical fire spots.

[0054] After the drone takes off, the flight control submodule precisely controls the drone to fly at a speed of 60km / h according to the preset route. During the flight, the multispectral camera of the image acquisition module collects multispectral images of forest vegetation in real time, and the infrared thermal imager simultaneously collects thermal images of the forest area. The stabilized gimbal always keeps the lens facing the ground to ensure that the collected images are clear and without blur.

[0055] When the drone flies to the vicinity of a historical fire site, the image processing submodule first performs spatial alignment processing on the acquired multispectral and thermal images. The ground point coordinates are calculated using the binocular parallax model formula to accurately match the spatial positions of the two images. Then, a ResNet-50 network is used to extract four-layer feature maps from the two types of images. Based on the feature pyramid network, features at different scales are fused. Combined with a dynamic weight adjustment mechanism, the fusion weights are automatically adjusted according to the temperature gradient of the region in the thermal image and the NDVI value of the multispectral image to enhance the proportion of thermal imaging features. Then, through the ECA attention mechanism, the focus is on areas in the image where the temperature rises abnormally and the NDVI value deviates from the normal vegetation range, thus refining the target image features.

[0056] Based on the characteristics of the target image, the fire identification submodule quickly identifies the presence of diffuse smoke in the area, marks its geographical coordinates (latitude and longitude) and the smoke coverage area (approximately 0.2 square kilometers), outputs that there are signs of fire in the area, and transmits the data to the graded early warning module.

[0057] The image acquisition submodule in the UAV simultaneously acquires multispectral and thermal images, overcoming the limitation that single-sensor data cannot comprehensively reflect fire characteristics. It can capture the spectral and temperature characteristics of fires separately, providing a rich data foundation for accurate identification. The image processing submodule in the UAV first performs spatial alignment processing on the two types of images, solving the problem of spatial resolution differences in multimodal data and ensuring the accuracy of subsequent feature extraction. Then, through feature fusion processing, it effectively integrates the image features of multispectral and thermal imaging, making the fire characteristics more comprehensive. Subsequent channel attention mechanism enhancement processing can highlight the characteristics of key fire areas and suppress irrelevant background interference, further refining effective features. The fire identification submodule in the UAV identifies fire features such as smoke and flames based on target image features optimized in multiple steps. Compared with traditional identification methods that rely on single features or have not undergone optimization processing, it can more accurately distinguish between natural landscapes and fire signs, thereby significantly improving the accuracy of fire detection during forest fire prevention patrols.

[0058] The route optimization module is used to generate the optimal patrol route for the drone based on meteorological data, historical fire records, and fire identification results. The route optimization module is the core decision-making unit of the dynamic patrol system. By introducing an artificial intelligence model and combining real-time environmental parameters and task objectives, it dynamically adjusts the route to solve the problems of insufficient flexibility of the preset route and its inability to adapt to complex scenarios.

[0059] This module uses a reinforcement learning algorithm, namely the SAC algorithm and memory enhancement mechanism, as the core to build a patrol route optimization model. It takes meteorological data, such as the need to prioritize patrolling high-wind-speed areas to prevent the rapid spread of fire, and fire identification results, such as the need to increase patrol density in identified fire areas and prioritize coverage of uncovered areas, as input to the model. Through a multi-objective reward function, it comprehensively weighs coverage efficiency, fire risk avoidance, response speed, and energy consumption, and finally generates the optimal patrol route that adapts to the current fire risk situation, replacing the traditional fixed route or simple dynamic adjustment method.

[0060] The route optimization module includes, Fire risk index calculation unit: used to calculate the regional fire risk index based on meteorological data and historical fire records; The expression for calculating the regional fire risk index based on meteorological data and historical fire records is as follows: in, The value represents the regional fire risk index at time t; 0.6, 0.3, and 0.1 represent weighting coefficients. This indicates the fire hazard index based on temperature. This indicates the wind speed and fire risk index; This represents the normalization coefficient for historical fire point density; The density of historical fire points at time t is obtained by estimating using Gaussian kernel density. The real-time fire risk probability at time t is output by the UAV image recognition model. The fire risk index is a core indicator for quantifying regional fire risk. Its calculation needs to comprehensively reflect environmental impact, historical patterns, and real-time situation. Traditional fire risk indices often lack accuracy due to their single dimension, making it difficult to support dynamic route optimization.

[0061] This unit achieves multi-dimensional integration through a weighted summation formula: the weight coefficients of 0.6, 0.3, and 0.1 clearly define the core logic of prioritizing real-time environmental fire risk, supplementing it with historical fire risk patterns, and further supplementing it with real-time monitoring results. Specifically, Gaussian kernel density estimation is used to process historical fire records to accurately reflect the high probability of regional fires, while real-time fire risk probabilities identified by drones are directly adopted to ensure that the index responds instantly to the current fire situation. The final calculated regional fire risk index can comprehensively and objectively characterize the regional fire risk level at time t, providing accurate risk quantification basis for the subsequent reward function.

[0062] Multi-objective reward function unit: used to acquire UAV flight parameters, perform reward calculations on the UAV flight parameters, and obtain multi-objective reward results.

[0063] The expression for calculating the reward based on the UAV flight parameters is as follows: in, Indicates the target reward result; Indicates the coverage efficiency weight; Indicates the weight of fire risk avoidance; Indicates response speed weight; Indicates the energy consumption penalty weight; The area of ​​the covered region; The total patrol area; The area fire risk index at time t; To prevent division by zero for the minimum value; For response time; This is the maximum allowable response time. This represents the amount of electricity already consumed. This is the maximum battery capacity.

[0064] In reinforcement learning, the reward function is key to guiding the model to learn the optimal policy. Multi-objective reward functions need to balance multiple conflicting or complementary optimization objectives to avoid the one-sidedness of the path caused by traditional single-objective functions.

[0065] The fire risk index calculation unit, through multi-dimensional parameter weighted fusion, breaks through the limitations of traditional single-factor fire risk assessment, accurately quantifies the regional fire risk level, and provides an objective and comprehensive risk basis for route optimization. The multi-objective reward function unit, through reasonable design of reward and penalty terms, achieves a dynamic balance between coverage efficiency, fire risk avoidance, response speed, and energy consumption, avoiding the problem of route optimization being biased towards a single objective. The synergistic effect of the two units enables the reinforcement learning model to learn the optimal patrol strategy adapted to complex fire situations and environments. Compared with traditional preset routes or simple dynamic adjustment methods, this not only improves the integrity of patrol coverage and the timeliness of fire risk response, but also reduces the energy consumption of drones and the flight risk in high-fire-risk areas, providing core decision support for the efficient operation of the fire early warning system and further ensuring the intelligence and efficiency of forest fire prevention patrols.

[0066] The graded early warning module is used to calculate the fire level based on meteorological data, historical fire records and fire identification results, and to determine the early warning information based on the fire level. The graded early warning module is the core output unit of the early warning system. By establishing scientific grading standards, it transforms monitoring results into precise and actionable early warning signals, avoiding the problems of over- or under-response caused by traditional single early warning modes. In this application, the module abandons the simple early warning logic based on the presence or absence of fire. Instead, it uses fire identification results, such as fire area, temperature gradient, and spread rate, as a basis, combined with meteorological data such as wind speed accelerating spread and humidity inhibiting spread. It calculates the fire level through a dynamic formula and further corrects the fire hazard level, ultimately classifying it into four levels: blue, yellow, orange, and red. Different levels correspond to differentiated response measures, achieving a precise match between fire hazard level, early warning level, and response measures.

[0067] The tiered early warning module includes: Fire rating calculation unit: used to calculate the fire rating based on the fire identification results; The calculation expression for determining the fire level based on the fire identification results is as follows: in, Indicates the fire severity level; Represents the floor function; Indicates the weight of the fire area; Indicates the area of ​​the fire; Indicates the maximum fire area; Indicates the temperature gradient weights; Represents the temperature gradient; Indicates the maximum temperature gradient; Indicates the spread rate weight; Indicates the speed at which the fire spreads; Indicates the maximum spread rate; Accurate quantification of fire severity is the core foundation of graded early warning. Traditional fire severity calculation often relies on a single parameter, resulting in a one-sided assessment of severity and failing to fully reflect the degree of fire hazard.

[0068] The fire level calculation unit achieves scientific calculation through a formula that weights and fuses multiple key parameters: three core parameters are selected, namely fire area A, temperature gradient ΔT, and spread rate v, which correspond to the fire's impact range, energy intensity, and spread trend, respectively. They are assigned reasonable weights and normalized by dividing by their respective maximum values ​​to eliminate the dimensional differences between different parameters. Finally, the integer fire level Fire_Level is obtained through the Round function, ensuring that the level classification is clear and quantifiable and avoiding fuzzy judgments.

[0069] Fire rating correction unit: Used to correct the fire rating based on meteorological data.

[0070] The expression for correcting the fire level based on meteorological data is as follows: in, This indicates the adjusted fire risk level. Indicates the fire severity level; Wind speed; Maximum wind speed; Humidity; Maximum humidity; Indicates the wind speed sensitivity coefficient; This is the humidity sensitivity coefficient.

[0071] The actual severity of a fire depends not only on its own characteristics but also on environmental parameters. Traditional fire rating systems often ignore environmental factors, leading to a disconnect between early warnings and actual fire risks.

[0072] The fire risk level correction unit uses environmental parameters as the core basis for correction and achieves dynamic correction through formulas: wind speed, as a contributing factor to fire spread, needs to be positively corrected, for example, the higher the wind speed, the higher the fire risk level; humidity, as an inhibiting factor to fire spread, needs to be negatively corrected, for example, the higher the humidity, the lower the fire risk level; k and λ are sensitivity coefficients used to adjust the degree of influence of environmental parameters on the fire risk level, and finally obtain the corrected fire risk level Adjusted_Risk_Level. The corrected fire risk level can accurately match the comprehensive situation of the fire's own characteristics and environmental influences.

[0073] The fire severity calculation unit, by integrating three core parameters—fire area, temperature gradient, and spread rate—and performing normalization and rounding, overcomes the limitations of traditional single-parameter fire severity determination. This allows the initial fire severity rating to comprehensively reflect the fire's scope, intensity, and spread trend, ensuring the objectivity and comprehensiveness of the severity rating assessment. The fire hazard rating correction unit introduces two key environmental parameters—wind speed and humidity—and dynamically adjusts the degree of environmental influence on fire development through sensitivity coefficients. This effectively compensates for the initial fire severity rating's failure to consider environmental factors, making the adjusted fire hazard rating more closely reflect the actual fire hazard evolution. The combined effect of these two units achieves a dual consideration of fire characteristics and environmental influencing factors, making the classification of early warning levels more accurate and targeted. This avoids insufficient or excessive response due to level determination bias, providing reliable early warning basis for fire command centers to formulate scientific rescue strategies and significantly improving the practicality and effectiveness of the fire early warning system.

[0074] The wireless communication module is used to transmit early warning information to the fire command center.

[0075] The wireless communication module is the information interaction unit of the early warning system, responsible for transmitting front-end processing results and decision-making information to the back-end command platform in real time and stably. Its core requirements are high transmission efficiency and strong reliability, ensuring information transmission without delay or loss. In this application, this module, as a key link in the early warning closed loop, not only realizes the real-time transmission of early warning information such as warning level, fire location, and fire risk evolution trend to the fire command center, but also provides communication support for subsequent rescue command issuance and resource scheduling. Simultaneously, in conjunction with subsequently optimized compression strategies and bandwidth allocation mechanisms, it ensures transmission stability in complex forest environments, ensuring that the command center obtains critical information in a timely manner. The wireless communication module includes: Decomposition and Compression Unit: Used to perform lifting wavelet decomposition on the early warning information, obtain decomposition data, quantize and compress the high-frequency coefficients in the decomposition data, and compress the low-frequency coefficients in the decomposition data through LZW encoding. Lifting wavelet decomposition is an efficient data preprocessing technique that can decompose raw data into low-frequency coefficients containing core information and high-frequency coefficients containing detailed information, providing a basis for differentiated compression; quantization compression is a common lossy compression method that achieves efficient compression by reducing data precision and is suitable for non-core detailed data; LZW coding is a lossless compression algorithm that can compress data without losing information and is suitable for core and critical data.

[0076] Fire area data is the core basis for early warning decisions, and it is necessary to balance compression efficiency and data accuracy: after wavelet decomposition, high-frequency coefficients correspond to the detailed noise of fire images, such as slight pixel fluctuations. Redundancy is removed by quantization compression without affecting the core judgment; low-frequency coefficients correspond to the core features of the fire, such as flame outline, smoke range, and temperature distribution. The core data uses LZW encoding to retain key information without loss, ensuring the accuracy of subsequent analysis and decision-making.

[0077] In this embodiment, the data from non-fire areas is compressed into a static background using run-length encoding and then sent to the fire command center. This provides further reference for the fire command center to make fire-fighting decisions, making the fire-fighting measures arranged by the fire command center more reasonable.

[0078] Run-length encoding is a lossless compression algorithm designed for data with high repetition and large redundancy. It simplifies storage and transmission by recording the length and value of repeated data. It is especially suitable for scenarios with high data repetition rates, such as static backgrounds and uniform regions. It has high compression efficiency and low computational complexity.

[0079] Data from non-fire areas consists mainly of static backgrounds, such as large areas of normal vegetation and terrain. Pixel values ​​vary little and have extremely high repetition, so there is no need to retain too much detailed information. Run-length encoding can quickly compress the data volume, significantly reducing the bandwidth consumption of non-critical data, while avoiding excessive computing power consumption by complex compression algorithms, thus adapting to resource-constrained scenarios in UAV wireless communication.

[0080] A layered strategy of lifting wavelet decomposition, high-frequency quantization compression, and low-frequency LZW lossless compression is adopted for fire area data. This strategy not only removes non-critical details and redundancy through quantization compression, improving transmission efficiency, but also preserves the core fire feature data completely through LZW encoding, ensuring that the fire command center can formulate rescue strategies based on accurate data.

[0081] In this embodiment, run-length encoding is used for static background data in non-fire areas. Leveraging its high compression efficiency and low computational cost, it significantly reduces the bandwidth usage of redundant data and prevents non-critical data from crowding out communication resources. The two compression methods are specifically adapted to the characteristics of different types of data, and together they significantly improve the efficiency of wireless communication transmission. This solves the problem that traditional single compression algorithms either sacrifice the accuracy of fire data or have insufficient compression efficiency, ensuring that critical data in high-risk areas can be transmitted to the command center with priority and speed, providing communication support for the timeliness and reliability of forest fire early warning.

[0082] Dynamic bandwidth allocation unit: used to acquire bandwidth parameters, calculate bandwidth allocation based on bandwidth parameters, and transmit early warning information based on bandwidth allocation; The expression for calculating bandwidth allocation based on bandwidth parameters is as follows: in, This indicates the bandwidth allocation at time t; This indicates the adjusted fire risk level. Maximum bandwidth; Average bandwidth; This is the minimum bandwidth.

[0083] Dynamic bandwidth allocation is a resource optimization technology in the field of wireless communication. Its core is to dynamically adjust the bandwidth allocation scheme according to data priority and service urgency, so as to solve the problem of insufficient bandwidth for high-priority data and low-priority data occupying redundant resources in fixed bandwidth allocation. It is especially suitable for mobile monitoring scenarios with limited communication resources.

[0084] The bandwidth dynamic allocation unit uses fire risk level as the core logic to determine data priority. It uses the adjusted fire risk level as the sole criterion for bandwidth allocation, presets three types of bandwidth parameters, and achieves differentiated allocation through a piecewise function: when the fire risk level is ≥4 (orange or red alert), corresponding to highly urgent data such as high-definition fire images and real-time spread data, bandwidth allocation is... Ensure high-speed transmission of critical data; when 2 ≤ fire hazard level < 4, i.e., a yellow alert, corresponding to moderately urgent data such as routine warning information and patrol status data, allocate... Balancing transmission efficiency and resource consumption; when the fire risk level is <2, i.e., a blue alert, corresponding to low urgency data, such as when there is no patrol feedback from the fire area, data is allocated... Save bandwidth resources and ensure that bandwidth resources are allocated to high-priority data.

[0085] The dynamic bandwidth allocation unit uses fire risk level as the core criterion and achieves differentiated and precise bandwidth allocation through a piecewise function. High fire risk areas are allocated the maximum bandwidth to ensure priority and rapid transmission of critical fire risk data, avoiding information delays or loss due to insufficient bandwidth, and providing timely support for the fire command center to quickly formulate rescue strategies. Medium and low fire risk areas are allocated corresponding bandwidth as needed, effectively saving communication resources and preventing low-priority data from crowding out the transmission channels of high-priority data, thus solving the resource waste problem of traditional fixed bandwidth allocation. This technology, in conjunction with the hierarchical compression strategy of the wireless communication module, further enhances the targeting and efficiency of wireless communication transmission, ensuring optimal bandwidth resource allocation in the complex communication environment of forests. This guarantees the transmission quality of critical information in high-risk areas while also considering the stability and resource utilization of the overall communication system, providing reliable communication resource support for the efficient operation of the fire early warning system.

[0086] The data acquisition module provided by this invention simultaneously acquires meteorological data, historical fire records, and fire identification results, enabling multi-dimensional fire assessment and identification. This allows for more accurate forest fire early warning. Simultaneously, the route optimization module optimizes the drone's patrol route, ensuring comprehensive forest coverage and preventing missed reports due to monitoring gaps. This enhances the scope and reliability of early warnings under limited resources. The image acquisition submodule in the drone simultaneously acquires multispectral and thermal images, overcoming the limitation of single-sensor data in comprehensively reflecting fire characteristics. It can capture the spectral and temperature features of fires separately, providing a rich data foundation for accurate identification. The image processing submodule in the drone first performs spatial alignment processing on the two types of images, resolving the spatial resolution difference problem of multimodal data and ensuring the accuracy of subsequent feature extraction. Feature fusion processing effectively integrates multispectral and thermal imaging image features, resulting in a more comprehensive representation of fire characteristics. Subsequent channel attention enhancement processing highlights key fire area features, suppresses irrelevant background interference, and further refines effective features. The fire identification submodule in the UAV identifies fire features such as smoke and flames based on target image features optimized through multiple steps. Compared with traditional identification methods that rely on single features or have not undergone optimization processing, it can more accurately distinguish between natural landscapes and fire signs, thereby significantly improving the accuracy of fire detection during forest fire prevention patrols. The graded early warning module calculates the fire level based on meteorological data, historical fire records, and fire identification results, thereby accurately predicting the fire hazard. The early warning information is then transmitted to the fire command center via a wireless communication module, enabling the fire command center to take appropriate fire-fighting measures based on the fire hazard and magnitude.

[0087] Example 2: In a mountainous forest area of ​​120 square kilometers, including 3 historically high-risk fire zones and 2 sensitive forest edge areas, the fire early warning system described in this application will be deployed: The data acquisition module collects meteorological data in real time, including: wind speed 2.8 m / s, humidity 55%, and temperature 28℃; The historical fire records of the area over the past 5 years were retrieved, showing that the western valley is a high-incidence area, with an average of 1.2 fires per year, mostly caused by dryness and strong winds; Meanwhile, the fire identification results transmitted by the forest fire patrol drone were received, including: smoke features were detected in the western valley area, covering an area of ​​0.3 square kilometers, with a temperature gradient of 10°C, which was initially judged to be an initial fire. Subsequently, the route optimization module inputs the above data into a patrol route optimization model based on the SAC algorithm and memory enhancement mechanism. The model calculates the optimal route through a multi-objective reward function: the fire risk index of the western valley is currently high, so it needs to be prioritized for intensive patrols; the uncovered area on the east side needs to take into account coverage efficiency; and the high-altitude, strong-wind area on the north side is avoided to reduce energy consumption. Finally, the optimal route is generated: intensive patrol of the western valley → regular patrol of the uncovered area on the east side → key patrol of the historical fire risk area on the south side. Compared with the original preset route, the route is shortened by 18%, and the response time is expected to be shortened by 25%. The graded early warning module is based on fire identification results, including: area of ​​0.3 square kilometers and temperature gradient of 10℃; meteorological data, including: wind speed of 2.8 m / s accelerating the spread and humidity of 55% having no significant inhibitory effect. The fire level calculation formula yields Fire_Level=3, and after adjustment for fire risk level, Adjusted_Risk_Level=3.2, it is determined to be a level two, i.e., a yellow warning. The corresponding response measures are to send warning information and adjust the patrol route of drones to increase monitoring density. Finally, the wireless communication module transmits early warning information, including early warning level, fire location, longitude 119°15′E, latitude 35°40′N, fire risk evolution prediction, and optimal patrol route, to the fire command center in real time through the wireless image and data transmission integrated network. The command center then pushes the information to the mobile terminals of the patrol personnel in the area.

[0088] This invention integrates meteorological data, historical fire records, and UAV fire identification results through a data acquisition module, constructing a multi-dimensional, full-chain data support system to avoid the biased decision-making caused by single data sources. The route optimization module dynamically generates optimal patrol routes through an intelligent model, significantly improving patrol coverage efficiency and fire response speed compared to traditional preset routes, while reducing UAV energy consumption. The graded early warning module achieves accurate fire level classification and differentiated early warning based on multi-source data, solving the problem of inaccurate response in traditional simple early warning mechanisms. The wireless communication module ensures real-time and stable transmission of early warning information and decision-making instructions, forming a complete closed loop of data acquisition, route optimization, graded early warning, and instruction issuance. This not only significantly improves the intelligence level of forest fire prevention but also buys valuable time for fire rescue through accurate early warning and efficient patrols, effectively reducing the risk and severity of forest fire spread. It provides an efficient and reliable system solution for forest fire prevention in large-area, complex terrain.

[0089] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A forest fire early warning system based on image processing, characterized in that, include: Data acquisition module, route optimization module, graded early warning module, and wireless communication module; The data acquisition module is used to acquire meteorological data, historical fire records, and fire identification results. The data acquisition module includes: a meteorological data unit, a historical fire record unit, and a drone; The meteorological data unit is used to acquire meteorological data; The historical fire record unit is used to acquire historical fire records; The drone is used to acquire forest image data and obtain fire identification results based on the forest image data. The route optimization module is used to generate the optimal patrol route for the drone based on meteorological data, historical fire records, and fire identification results. The graded early warning module is used to calculate the fire level based on meteorological data, historical fire records and fire identification results, and to determine the early warning information based on the fire level. The drone patrols the forest according to the optimal patrol route, acquires forest image data, and obtains fire identification results based on the forest image data. The fire identification results are then transmitted to the route optimization module and the graded early warning module. The meteorological data unit transmits meteorological data to the route optimization module and the graded early warning module. The historical fire record unit transmits historical fire records to the route optimization module and the graded early warning module. The route optimization module updates the drone's optimal patrol route and feeds it back to the drone. After the graded early warning module determines the early warning information, the wireless communication module transmits the early warning information to the fire command center.

2. The forest fire early warning system according to claim 1, characterized in that, The drone includes: Image acquisition submodule: used to acquire forest images; the forest images include: multispectral images and thermal imaging images; Flight control submodule: Used to control the UAV to patrol the forest area according to the optimal patrol route; Image processing submodule: used to process multispectral images and thermal imaging images to obtain target image features; Fire identification submodule: Used to identify target image features and obtain fire identification results based on the target image features.

3. The forest fire early warning system according to claim 2, characterized in that, The image processing submodule processes multispectral images and thermal imaging images to obtain target image features, including the following steps: S101. Perform spatial alignment processing on the multispectral image and the thermal imaging image to obtain a spatially aligned image; S102. Extract image features from spatially aligned images using a convolutional neural network to obtain multi-layer spatially aligned image features; S103. Fuse multi-layer spatially aligned image features to obtain fused image features; S104. Adjust the weights of the fused image features to obtain the adjusted fused image features; S105. The adjusted fused image features are enhanced by an attention mechanism to obtain the target image features.

4. The forest fire early warning system according to claim 3, characterized in that, The expression for the spatial alignment process is: in, The horizontal coordinates of the ground points corresponding to the multispectral image; The horizontal coordinates of the ground point corresponding to the thermal imaging image; is the pixel equivalent coefficient of the multispectral camera; is the pixel equivalent coefficient of the infrared thermal imager; The focal length of the multispectral camera; The focal length of the infrared thermal imager; The baseline distance between the multispectral camera and the infrared thermal imager; For parallax; The altitude of the drone above the ground; The coordinates of the x-th pixel of the ground point in the multispectral image; The x-coordinate of the ground point pixel in the thermal image; The expression for extracting image features from a spatially aligned image is: in, This represents the feature map of the nth layer; Represents the weights of the nth convolutional layer; This represents the bias of the nth convolution layer; This represents the convolution operation; This represents the input data for the nth layer; Indicates the activation function; Indicates the network layer index; The expression for fusing multi-layer spatially aligned image features is: in, This represents the features of the nth layer after fusion; Represents the inverted index feature map; Indicates an upsampling operation; Indicates the fusion features of the previous layer; This represents the fusion features of the first layer; This represents the feature map of the 4th layer.

5. The forest fire early warning system according to claim 3, characterized in that, The aforementioned weight adjustment of the fused image features specifically refers to adjusting the fused image features through a dynamic weight adjustment mechanism; The expression for the dynamic weight adjustment mechanism is: in, Represents the fusion feature of the nth layer; Represents the multispectral feature weights of the nth layer; This represents the feature weights of the nth thermal imaging layer; and These are the nth layer features from multispectral and thermal imaging, respectively; This represents the temperature gradient of the nth layer in thermal imaging. is the normalized vegetation index of the nth layer of the multispectral system; Sensitivity coefficient; The NDVI threshold for vegetation; The expression for enhancing the attention mechanism of the fused image features is as follows: in, For attention output features; Input features to the attention mechanism; This represents the Sigmoid activation function; These are the first-order learnable parameters; These are second-order learnable parameters; This indicates a global average pooling operation; This represents the activation function of the rectified linear unit.

6. The forest fire early warning system according to claim 1, characterized in that, The route optimization module includes, Fire risk index calculation unit: used to calculate the regional fire risk index based on meteorological data and historical fire records; Multi-objective reward function unit: used to acquire UAV flight parameters, perform reward calculations on the UAV flight parameters, and obtain multi-objective reward results.

7. The forest fire early warning system according to claim 6, characterized in that, The expression for calculating the regional fire risk index based on meteorological data and historical fire records is as follows: in, The value represents the regional fire risk index at time t; 0.6, 0.3, and 0.1 represent weighting coefficients. This indicates the fire hazard index based on temperature. This indicates the wind speed and fire risk index; This represents the normalization coefficient for historical fire point density; The density of historical fire points at time t is obtained by estimating using Gaussian kernel density. The real-time fire risk probability at time t is output by the UAV image recognition model. The expression for calculating the reward based on the UAV flight parameters is as follows: in, Indicates the target reward result; Indicates the coverage efficiency weight; Indicates the weight of fire risk avoidance; Indicates response speed weight; Indicates the energy consumption penalty weight; The area of ​​the covered region; The total patrol area; The area fire risk index at time t; To prevent division by zero for the minimum value; For response time; Maximum allowable response time; This represents the amount of electricity already consumed. This is the maximum battery capacity.

8. The forest fire early warning system according to claim 1, characterized in that, The tiered early warning module includes: Fire rating calculation unit: used to calculate the fire rating based on the fire identification results; Fire rating correction unit: Used to correct the fire rating based on meteorological data.

9. The forest fire early warning system according to claim 8, characterized in that, The calculation expression for determining the fire level based on the fire identification results is as follows: in, Indicates the fire severity level; Represents the floor function; Indicates the weight of the fire area; Indicates the area of ​​the fire; Indicates the maximum fire area; Indicates the temperature gradient weights; Represents the temperature gradient; Indicates the maximum temperature gradient; Indicates the spread rate weight; Indicates the speed at which the fire spreads; Indicates the maximum spread rate; The expression for correcting the fire level based on meteorological data is as follows: in, This indicates the adjusted fire risk level. Indicates the fire severity level; Wind speed; Maximum wind speed; Humidity; Maximum humidity; Indicates the wind speed sensitivity coefficient; This is the humidity sensitivity coefficient.

10. The forest fire early warning system according to claim 1, characterized in that, The wireless communication module includes: Decomposition and Compression Unit: Used to perform lifting wavelet decomposition on the early warning information, obtain decomposition data, quantize and compress the high-frequency coefficients in the decomposition data, and compress the low-frequency coefficients in the decomposition data through LZW encoding. Dynamic bandwidth allocation unit: used to acquire bandwidth parameters, calculate bandwidth allocation based on bandwidth parameters, and transmit early warning information based on bandwidth allocation; The expression for calculating bandwidth allocation based on bandwidth parameters is as follows: in, This indicates the bandwidth allocation at time t; This indicates the adjusted fire risk level. Maximum bandwidth; Average bandwidth; This is the minimum bandwidth.

Images