Forest disaster detection method and device

Abstract

The invention provides a forest disaster detection method and device, and belongs to the field of forest disaster monitoring, and the method comprises the steps: collecting an infrared image, a stereoscopic vision image and positioning data of a target forest region in real time; an ASSA module is introduced in the feature extraction stage of the YOLOv11, and an infrared image is input into the improved YOLOv11 for flame target recognition; recognizing the stereoscopic vision image through a smoke and fire recognition algorithm, and detecting flame and smoke targets; determining a fire detection result by combining the identification results of the flame target and the smoke target; under the condition that the fire disaster is determined to occur, an HITNet algorithm is adopted to carry out ignition point positioning through a stereoscopic vision image, and the geographic coordinates of a fire disaster point are determined based on positioning data; and sending alarm information based on the fire point geographic coordinates. Therefore, the capability of extracting the flame heat radiation characteristics of the tiny fire points is effectively enhanced, and the detection result of the forest fire is improved.

Description

Technical Field

[0001] The present invention belongs to the field of forest disaster monitoring, and in particular relates to a forest disaster detection method and device. Background Art

[0002] As the demand for accurate detection of forest disaster locations continues to increase, image detection combined with sequence data is becoming increasingly important as an efficient technical means. However, traditional image detection is unable to meet the needs of real-time monitoring across vast areas. Advances in deep learning are placing higher demands on emergency response to deep forest disasters.

[0003] At present, traditional image detection and equipment have the following shortcomings: Traditional image detection methods are inefficient when processing high-resolution, large-scale surveillance images, especially when large numbers of images or videos need to be processed in real time in large-scale forest areas. They lack multi-scale target detection capabilities and find it difficult to simultaneously detect targets with obvious size differences (such as small fire spots in the distance and large flames nearby). They also have poor detection effects on small targets (such as Mars and initial smoke).

[0004] Therefore, the existing technology for monitoring forest fires has great defects, and there is an urgent need for a method and device that can improve the efficiency of real-time monitoring of forest fires. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a forest disaster detection method and device.

[0006] In order to achieve the above object, the present invention provides the following technical solutions: A forest disaster detection method, comprising: Collect infrared images, stereo vision images and positioning data of the target forest area in real time; The infrared image is input into YOLOv11, and features are extracted through deep convolution. After two deep convolutions, the ASSA module is introduced to focus the features. The C3k2 module is used to further extract high-level semantic features from the focused features. The CBAM attention mechanism is then used to enhance the high-level semantic features by combining channel and spatial attention. The flame thermal radiation features are obtained for flame target recognition. Use fireworks recognition algorithm to identify stereo vision images and detect smoke targets; Combining the recognition results of flame targets and smoke targets to determine the fire detection results; When a fire is confirmed, the HITNet algorithm is used to locate the fire point through stereo vision images, and the geographical coordinates of the fire point are determined by combining the positioning data; Issue warning information based on the geographical coordinates of the fire point.

[0007] Optionally, determining the fire detection result further includes: A large forest disaster prediction model was constructed using a spiking neural network (SNN). Historically collected infrared images and stereoscopic vision images were input into the large forest disaster prediction model, and the large forest disaster prediction model was trained based on the corresponding real alarm information. Input the real-time collected infrared images and stereo vision images into the trained forest disaster prediction model to obtain the prediction results; The prediction results are combined with the recognition results of flame targets and smoke targets to determine the fire detection results.

[0008] Optionally, the locating the fire point using a HITNet algorithm through a stereoscopic vision image includes: Calculate the disparity of the stereoscopic image to generate a depth map; A high-pass filter is used to extract the flame edge and texture features of the depth map, and a median filter is used to eliminate image noise; Identify the flame edge and texture features to eliminate noise and determine the ignition point; Combining the positioning data and the image shooting angle of the stereo vision image, the fire point position is mapped to the geographic coordinate system through spatial geometric transformation.

[0009] Optionally, the real-time collected data also includes meteorological data and soil data; When the meteorological data and soil data are within a preset standard range, the frequency of data collection is increased; the preset standard range is a range of environmental data prone to fire.

[0010] A forest disaster detection device, applied to a forest disaster detection method, comprising: A sensing detection module is used to collect infrared images, stereo vision images and positioning data of the target forest area in real time; The processor is used to obtain the data collected by the sensor detection module and perform flame target recognition, flame and smoke target detection, and determine the geographical coordinates of the fire point; Communication module, used for interaction between the processor and the cloud platform and monitoring station; The cloud platform is used to build the model, update the model based on the data transmitted by the communication module, and send it to the processor; Monitoring station, which receives fire location and image data and triggers alarms.

[0011] Optionally, the sensing detection module includes: an infrared thermal imaging camera connected to the processor for transmitting infrared images; A binocular camera connected to the processor for capturing stereoscopic images; The anemometer sensor, temperature and humidity sensor, GPS positioning unit, and soil temperature and humidity sensor are all connected to the processor.

[0012] Optionally, the device further comprises: The data storage module adopts a pluggable SSD design to store raw data, algorithm models and processing results; The power module integrates solar panels and high-voltage batteries to power various components.

[0013] The forest disaster detection method provided by the present invention has the following beneficial effects: First, a multispectral perception system is constructed by real-time acquisition of infrared images and stereo vision data. The infrared images are processed using an improved YOLOv11 model. The ASSA module introduced in the model automatically focuses on more important features and suppresses irrelevant or redundant information, effectively enhancing the ability to extract flame thermal radiation features. This is particularly true for the thermal radiation features of distant, small fires (such as Mars). A multi-scale feature pyramid and channel attention weighting improve the accuracy of small target detection. The CBAM attention mechanism combines channel attention and spatial attention to focus on important areas in the image, improving the model's ability to detect flames. Furthermore, a smoke and fire recognition algorithm is used for smoke target detection. Combined with infrared detection results, this eliminates false alarms from single sensors and reduces the false alarm rate. Finally, the HITNet algorithm is used to process stereo vision images, accurately calculating the three-dimensional coordinates of the fire point, significantly reducing positioning errors. The model then automatically links the coordinates to a geographic information system to generate warning messages. Thus, the improved YOLOv11 model achieves accurate detection of small targets. Combined with stereo vision positioning, it overcomes detection blind spots caused by target size differences, ultimately achieving accurate early warning and spatial positioning of forest fires. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0015] Figure 1 The figure is a flow chart of a forest disaster detection method provided by the present invention according to an exemplary embodiment.

[0016] Figure 2 The figure is a schematic diagram of the structure of an improved YOLOv11 model provided by the present invention according to an exemplary embodiment.

[0017] Figure 3 The figure is a complete schematic diagram of the steps of a fire alarm according to an exemplary embodiment of the present invention.

[0018] Figure 4 This is a block diagram of a forest disaster detection device provided according to an exemplary embodiment of the present invention.

[0019] Figure 5 The figure is a schematic structural diagram of a forest disaster detection device provided according to an exemplary embodiment of the present invention.

[0020] Figure 6 The figure is a schematic structural diagram of a sensing detection module provided according to an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] First, the present invention provides a forest disaster detection method, specifically Figure 1 As shown, the following steps are included: S101. Collect infrared images, stereoscopic vision images, and positioning data of the target forest area in real time.

[0024] In this step, the data collection frequency is also adjusted in real time based on the environmental information. For example, if the meteorological data and soil data are within a preset standard range, the data collection frequency is increased; the preset standard range is the range of environmental data that is prone to fire.

[0025] For example, the risk of forest fires increases significantly when wind speeds exceed 9 mph, air temperatures exceed 30°C, relative humidity falls below 30%, and illuminance exceeds 1200 W / m². The risk of forest fires also increases significantly when soil relative humidity falls below 10% and soil temperatures exceed 30°C. In these situations, data collection frequency can be increased to facilitate timely monitoring and prevent deep forest fires.

[0026] S102. Input the infrared image into the improved YOLOv11, perform feature extraction through deep convolution, introduce the ASSA module for feature focusing, and use the CBAM attention mechanism for feature enhancement to obtain the flame thermal radiation features for flame target recognition.

[0027] In this step, the YOLOv11 model is improved, such as Figure 2As shown in the figure, the improved feature extraction stage of the YOLOv11 model is outlined by the dotted line. Specific improvements include replacing standard convolution with depthwise convolution, performing two consecutive depthwise convolutions to perform preliminary feature extraction on the image, then introducing an ASSA module to focus these initially extracted features to improve the detection accuracy of small objects. The C3k2 module then extracts high-level semantic features from these focused features and performs feature optimization. After another depthwise convolution, a CBAM attention mechanism is added to enhance features by combining channel attention and spatial attention.

[0028] The infrared image is input into YOLOv11, and features are extracted through two consecutive deep convolutions. After the two deep convolutions, the ASSA module is introduced for feature focusing. The C3k2 module is used to further extract high-level semantic features from the focused features. The CBAM attention mechanism is then used to enhance the high-level semantic features by combining channel and spatial attention to obtain the flame thermal radiation features for flame target recognition.

[0029] For example, after inputting an infrared image into YOLOv11, the following steps are included: In the first step, the input image first passes through two depthwise separable convolutional (DWConv) layers to improve the feature extraction ability of the model. These layers help reduce the complexity of the network while retaining important feature extraction capabilities.

[0030] In the second step, after the convolutional layer, an adaptive sparse self-attention (ASSA) module is applied, which allows the network to automatically focus on more important features when processing images or inputs, suppress irrelevant or redundant information, and enhance spatial channel features.

[0031] The third step is to enter the C3k2 module to further extract high-level semantic features. After a depth-wise separable convolution (DWConv), the model's feature extraction capability is improved to prepare for subsequent feature fusion.

[0032] In the fourth step, CBAM (Convolutional Block Attention Module) was applied, which combines channel attention (CAM) and spatial attention (SAM). This module helps the network focus on important areas in the image and improves the model's ability to detect flames and smoke.

[0033] The fifth step is to perform Conv operation on the feature map generated by the C3k2 module.

[0034] In the sixth step, during the downsampling process, the feature map generated first undergoes a Conv operation and a C3k2 module, and then undergoes a Conv operation.

[0035] In the seventh step, the feature map generated in step 6 undergoes a Conv operation, a C3k2 module, and an SPPF (Spatial Pyramid Pooling Fusion) layer to extract multi-scale information and improve feature generalization. The C2PSA (Channel-wise Spatial Attention) module further optimizes the feature map, focusing on important information at the spatial and channel levels.

[0036] In the eighth step, the feature map generated in the seventh step is concatenated with the feature map generated in the sixth step.

[0037] In the ninth step, the feature map generated in the eighth step passes through the C3k2 module and performs the Conv operation in parallel and enters the detection head.

[0038] In the tenth step, the feature map generated in the ninth step is concatenated with the feature map generated in the seventh step through the Conv operation. After passing through the C3k2 module, the Conv operation is performed in parallel and enters the detection head.

[0039] In the eleventh step, the feature map generated in the tenth step is concatenated with the feature map generated in the seventh step and enters the detection head through a C3k2 module.

[0040] In the twelfth step, after the above multi-layer convolution, upsampling and splicing processing, it finally enters the three detection heads of the ninth, tenth and eleventh steps. The three different detection output ports are responsible for detecting targets at different scales and levels, which improves the accuracy of target recognition and detects whether there is flame or smoke in the image.

[0041] S103: Recognize the stereoscopic image using a smoke recognition algorithm to detect smoke targets.

[0042] In this step, after the stereoscopic vision image is input into the fireworks recognition algorithm, the following steps are included: The first step is to process the collected raw data, including denoising, normalization, image enhancement, etc., so that the subsequent feature extraction and recognition process will be more efficient.

[0043] The second step is feature extraction, which extracts important features from the preprocessed data. These features include color information, texture, edges, and other features in the image, which are crucial for flame or smoke recognition.

[0044] The third step is fireworks recognition, which uses the extracted features to classify or identify the fireworks through the fireworks recognition algorithm to determine whether there is flame or smoke in the image.

[0045] The fourth step is to output the corresponding judgment result based on the recognition result.

[0046] S104: Determine the fire detection result by combining the recognition results of the flame target and the smoke target.

[0047] Specifically, a fire can be determined when a flame target is identified or a flame and smoke target is detected, thereby avoiding errors in the recognition results of a single algorithm that may delay the fire.

[0048] In this step, a spiking neural network (SNN) is also used to construct a large forest disaster prediction model. Historically collected data is fed into the model and trained based on corresponding real-world alarm information. Real-time data is then fed into the trained model to generate predictions. The predictions are then combined with the flame and smoke target recognition results to determine the fire detection outcome. This way, if any model identifies a fire, the fire is confirmed to have occurred, further avoiding potential oversights that could result from a single algorithm.

[0049] S105. When it is determined that a fire has occurred, the HITNet algorithm is used to locate the fire point through stereoscopic vision images, and the geographic coordinates of the fire point are determined in combination with the positioning data.

[0050] In this step, the stereoscopic image is input into the HITNet algorithm, which calculates the disparity of the stereoscopic image to generate a depth map. A high-pass filter extracts the flame edge and texture features from the depth map, and a median filter removes image noise. The de-noised flame edge and texture features are then identified to determine the fire location. Combined with GPS positioning data and the camera rotation angle, the fire location is mapped to a geographic coordinate system through a spatial geometric transformation. The GPS positioning data provides the geographic location of the positioning device, while the HITNet algorithm outputs and identifies the location of the fire point relative to the positioning device. The combination of the two provides the final location of the fire point.

[0051] The specific steps include: The first step is data input, which processes the stereo image data as input data.

[0052] The second step is feature extraction. The HITNet algorithm uses a deep convolutional neural network (CNN) to extract local and global features from the image. These features include texture, color changes, and shape, which are used to identify areas of flame or smoke.

[0053] The third step is to locate the area of ​​the flame or smoke. HITNet uses a convolutional layer to process the image and extracts the features of the flame and smoke through a high-pass filter and a median filter.

[0054] In the fourth step, region matching and classification, the extracted features are fed into the matching layer, which matches them against flame or smoke templates to determine which regions may be flames or smoke. The matching process combines local image features with contextual information to improve positioning accuracy.

[0055] The fifth step is to locate the flame or smoke and output it. After feature extraction, matching and classification, the identification area where smoke or flame exists is obtained.

[0056] The sixth step is post-processing, which combines the identification area information of the flame or smoke in the image with the geographic location information and rotation angle of the edge distributed forest disaster detection device to accurately locate the geographic location where smoke and flames exist and output it.

[0057] S106. Issue an alarm based on the geographical coordinates of the fire point.

[0058] After confirming that a fire has occurred and determining the geographical location of the fire, a fire alarm signal is issued and the geographical location of the fire is sent simultaneously.

[0059] Based on the above steps, the complete steps of the fire alarm can be as follows: Figure 3 As shown in the figure, the acquired infrared image and visible light image (i.e., stereoscopic vision image) are input into the improved YOLOv11 algorithm and the fireworks recognition algorithm respectively. The above process is repeated when no fireworks target is detected. When a fireworks target is detected, the fireworks target is located using the HITNet algorithm, and the relevant information of the fireworks target is sent to the monitoring station for alarm.

[0060] Using this method, a multispectral perception system is first constructed by acquiring infrared images and stereo vision data in real time. The infrared images are processed using an improved YOLOv11 model. The ASSA module introduced in this model automatically focuses on more important features and suppresses irrelevant or redundant information, effectively enhancing the ability to extract flame thermal radiation features. This is particularly true for the thermal radiation features of distant, small fires (such as Mars). A multi-scale feature pyramid and channel attention weighting improve the accuracy of small target detection. The CBAM attention mechanism combines channel attention and spatial attention to focus on important areas in the image, improving the model's ability to detect flames. Furthermore, a smoke and fire recognition algorithm is used for smoke target detection. Combined with infrared detection results, this eliminates false alarms from a single sensor and reduces the false alarm rate. Finally, the HITNet algorithm is used to process stereo vision images, accurately calculating the three-dimensional coordinates of the fire point, significantly reducing positioning errors. The system then automatically links the coordinates to a geographic information system to generate warning messages. Thus, the improved YOLOv11 model achieves accurate detection of small targets. Combined with stereo vision positioning, it overcomes detection blind spots caused by target size differences, ultimately achieving accurate early warning and spatial positioning of forest fires.

[0061] Secondly, the present invention also provides a forest disaster detection device, such as Figure 4 As shown, including: The sensing detection module 201 is used to collect infrared images, stereoscopic vision images and positioning data of the target forest area in real time.

[0062] The processor 202 is used to obtain data collected by the sensor detection module, and perform flame target recognition, flame and smoke target detection, and determine the geographical coordinates of the fire point.

[0063] The communication module 203 is used for interaction between the processor, the cloud platform and the monitoring station.

[0064] The monitoring station 204 receives the fire location and image data and triggers an alarm.

[0065] The cloud platform 205 is used to build the model, update the model according to the data transmitted by the communication module, and send it to the processor.

[0066] like Figure 5 As shown, the sensor detection module includes an infrared detection unit, a stereo vision detection unit, a meteorological detection unit, a soil detection unit, and a positioning unit. The infrared detection unit detects the infrared thermal spectrum of flame radiation; the stereo vision detection unit can be a binocular camera that captures stereoscopic images of the forest and simultaneously provides distance information to the fire source or smoke; the meteorological detection unit can include an anemometer sensor, a temperature and humidity sensor, a GPS positioning unit, a soil temperature and humidity sensor, etc., to collect current wind speed, wind direction, temperature, humidity, and light intensity data in real time; the soil detection unit monitors the soil moisture and temperature conditions in real time; and the positioning unit obtains the location information of the distributed forest disaster detection device at the edge.

[0067] The device also includes a data receiving module, an internal communication module, and an external communication module. The data receiving module receives data from the sensing module, organizes it, and sends it to the internal communication module. The internal communication module uses the CAN bus protocol to implement data transmission between the data receiving module and the distributed computing and processing module. The image data and sequence data from the data receiving module are transmitted to the distributed computing and processing module, which then transmits the image data and sequence data to the monitoring station and cloud platform via the external communication module.

[0068] The power module integrates solar panels and high-voltage batteries to provide power for various components.

[0069] The monitoring station receives fire location and image data and issues an alarm message.

[0070] The sensor detection module includes an infrared detection unit, a stereoscopic vision detection unit, a meteorological detection unit, a soil detection unit and a positioning unit. Figure 6 As shown, it mainly includes an infrared thermal imaging camera S001, a binocular camera S002, an antenna S003, an anemometer sensor S004, a temperature and humidity sensor S005, a high-voltage battery S006, a GPS (S007), a processor S008 and a soil temperature and humidity sensor S009.

[0071] The infrared detection unit is an infrared thermal imaging camera that can generate infrared images. These images show the spatial distribution of the surface temperature of the object being measured, usually using different colors to represent different temperature ranges.

[0072] The stereo vision detection unit includes a binocular camera, which obtains image information through the binocular camera and provides image data for training and verifying image recognition algorithm models, stereo vision algorithm models and forest disaster prediction models.

[0073] The meteorological detection unit includes an anemometer sensor and a temperature and humidity sensor, which can monitor and record wind speed, wind direction, temperature, humidity and light intensity data in real time.

[0074] The soil detection unit includes a soil temperature and humidity sensor that can detect changes in soil humidity, temperature and water content in a specific area in real time.

[0075] The positioning unit can determine the geographic coordinates of the positioning device based on the GPS positioning data. The GPS positioning data includes various information such as GPS, Beidou satellite signals and base station signals, which are used to locate the position of the edge distributed forest disaster detection device and the rotation angle of the current camera.

[0076] The data receiving module is used to receive infrared image data from the infrared detection unit, visible light three-dimensional image data from the stereo vision detection unit, and sequence data from the meteorological detection unit, positioning unit and soil detection unit.

[0077] The distributed computing and processing module is composed of several computing processing units (greater than or equal to 3). Each computing processing unit is a Sipeed M4N Dock low-power, high-computing-power edge computing board. Each computing processing unit is directly connected to the internal communication module via the network. During initialization, any computing processing unit is used as the master node, and the other computing processing units are sub-nodes. The master node and sub-nodes together constitute a distributed computing processing system. The master node receives image data and sequence data from the sensor detection module and dispatches sub-nodes to complete flame recognition using the infrared detection thermal imaging algorithm, flame and smoke recognition using the image recognition algorithm, and fire location recognition using the stereo vision algorithm.

[0078] The distributed computing and processing module uses infrared detection thermal imaging technology to detect whether there is a flame; and uses image recognition algorithms to further detect whether there is smoke and flame. The distributed computing and processing module uses the cloud platform large model combined with current data to evaluate the probability of fire. The distributed computing and processing module will publish the original image data and the processed JSON format data to the cloud platform and the monitoring station, and store these data and the recognition results in the data storage module. The specific process of the stereoscopic vision image detection method is to perform denoising, normalization, and pre-processing of the stereoscopic vision image into a tensor, perform model reasoning on the pre-processed data, decode the bounding box, screen the candidate box, label the target, and finally output the target detection result. The infrared image detection method is characterized in that the specific process is to perform denoising, contrast enhancement, image normalization, and conversion into a tensor on the infrared image, perform model reasoning, bounding box decoding, non-large extreme value suppression, threshold screening, target labeling, and finally output the target detection result.

[0079] The distributed computing and processing module includes a plurality of (greater than or equal to 3) distributed computing and processing units, which are used to process the image data (from the infrared detection unit and the stereo detection unit) and sequence data (from the meteorological detection unit, the soil detection unit and the positioning unit) received from the sensor detection module, and use the infrared detection thermal imaging algorithm to detect whether there is a flame; use the image recognition algorithm to further detect whether there is smoke and flame; use the stereo vision algorithm to estimate the distance between the fire location and the positioning unit; use the forest disaster prediction model issued by the cloud platform in combination with the current data to evaluate the probability of fire; store the current data and the recognition results in the data storage module; all the original data are sent to the cloud platform to train the forest disaster prediction model, and the fire location and image data are sent to the monitoring station to display the alarm information.

[0080] The external communication module realizes data interaction between the edge-end distributed forest disaster detection device and the cloud platform, including uploading all data and issuing all updated models, and sending the recognition results to the monitoring station.

[0081] In this device, the external communication module enables data exchange between the detection device and the cloud platform, including uploading all data and distributing all updated models, while also sending recognition results to the monitoring station. The cloud platform receives all raw data from distributed forest disaster detection devices at the edge and stores it in a historical database, building a continuously updated knowledge base. The cloud platform uses newly acquired incremental infrared image data to train and optimize infrared detection and thermal imaging algorithm models, and newly acquired incremental three-dimensional image information to train and optimize image recognition and stereo vision algorithm models. Data from multiple devices is used to train a large forest disaster prediction model, deeply exploring the connection between data and disasters and predicting the probability of future disasters. The updated infrared detection and thermal imaging algorithm model, image recognition algorithm model, and stereo vision algorithm model are distributed to the distributed forest disaster detection devices at the edge. Furthermore, the cloud platform utilizes a spiking neural network (SNN) to construct a large forest disaster prediction model, predicting forest disasters based on real-time data to achieve disaster prevention.

[0082] In addition, the data storage module adopts a pluggable SSD design to store raw data, algorithm models and processing results. It is used to save the data collected by the sensor detection module, the recognition results of the distributed computing and processing module, as well as the infrared detection thermal imaging algorithm model, image recognition algorithm model, and stereo vision algorithm model.

[0083] Based on the above device, the following steps can be implemented: In the first step, the infrared detection unit collects infrared images and detects flame and smoke targets through the improved YOLOv11 algorithm.

[0084] In the second step, the stereo vision detection unit collects visible light images and uses the fireworks recognition algorithm to perform target detection.

[0085] In the third step, when no smoke or fire is detected in the first and second steps, the process returns to the first step.

[0086] In the fourth step, when fireworks are detected in the first or second step, the fire point is located using the existing HITNet algorithm on the three-dimensional image collected by the stereo vision detection unit.

[0087] The fifth step is to determine the final geographic location information (geographic coordinates) of the fire point through the positioning information of the fire point (position in the visible light image), the geographic location information (geographic coordinates) of the device positioned by the positioning unit, and the rotation angle of the device, and transmit the final geographic location information to the monitoring station.

[0088] Step 6. Return to step 1.

[0089] Using the above device, a multispectral perception system is first constructed by real-time acquisition of infrared images and stereo vision data. The infrared images are processed using an improved YOLOv11 model. The ASSA module introduced in the model automatically focuses on more important features and suppresses irrelevant or redundant information, effectively enhancing the ability to extract flame thermal radiation characteristics. This is especially true for the thermal radiation characteristics of distant small fire points (such as Mars). The multi-scale feature pyramid and channel attention weighting improve the accuracy of small target detection. The CBAM attention mechanism combines channel attention and spatial attention to focus on important areas in the image, improving the model's ability to detect flames. A smoke recognition algorithm is also used for smoke target detection. Combined with infrared detection results, it eliminates false alarms from single sensors and reduces the false alarm rate. Finally, the HITNet algorithm is used to process stereo vision images to accurately calculate the three-dimensional coordinates of the fire point, significantly reducing positioning errors. The system automatically links the coordinates to a geographic information system to generate warning information. In this way, the improved YOLOv11 model achieves accurate detection of small targets. Combined with stereo vision positioning, it overcomes detection blind spots caused by target size differences, ultimately achieving early and accurate warning and spatial positioning of forest fires.

[0090] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0092] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0093] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0094] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification has described the present invention in detail, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are included in the scope of protection of the patent for the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

Claims

1. A forest disaster detection method, characterized in that: The method comprises: Collect infrared images, stereo vision images and positioning data of the target forest area in real time; The infrared image is input into YOLOv11, and features are extracted through deep convolution. After two deep convolutions, the ASSA module is introduced to focus the features. The C3k2 module is used to further extract high-level semantic features from the focused features. The CBAM attention mechanism is then used to enhance the high-level semantic features by combining channel and spatial attention. The flame thermal radiation features are obtained for flame target recognition. Use fireworks recognition algorithm to identify stereo vision images and detect smoke targets; Combining the recognition results of flame targets and smoke targets to determine the fire detection results; When a fire is confirmed, the HITNet algorithm is used to locate the fire point through stereo vision images, and the geographical coordinates of the fire point are determined by combining the positioning data; Issue warning information based on the geographical coordinates of the fire point.

2. The method according to claim 1, characterized in that Determining the fire detection result further includes: A large forest disaster prediction model was constructed using a spiking neural network (SNN). Historically collected infrared images and stereoscopic vision images were input into the large forest disaster prediction model, and the large forest disaster prediction model was trained based on the corresponding real alarm information. Input the real-time collected infrared images and stereo vision images into the trained forest disaster prediction model to obtain the prediction results; The prediction results are combined with the recognition results of flame targets and smoke targets to determine the fire detection results.

3. The method according to claim 1, characterized in that The method of using the HITNet algorithm to locate the fire point through stereo vision images includes: Calculate the disparity of the stereoscopic image to generate a depth map; A high-pass filter is used to extract the flame edge and texture features of the depth map, and a median filter is used to eliminate image noise; Identify the flame edge and texture features to eliminate noise and determine the ignition point; Combining the positioning data and the image shooting angle of the stereo vision image, the fire point position is mapped to the geographic coordinate system through spatial geometric transformation.

4. The method according to claim 1, wherein The data collected in real time also includes meteorological data and soil data; When the meteorological data and soil data are within a preset standard range, the frequency of data collection is increased; the preset standard range is a range of environmental data prone to fire.

5. A forest disaster detection device, characterized in that: The forest disaster detection method according to claim 1, wherein the device comprises: A sensing detection module is used to collect infrared images, stereo vision images and positioning data of the target forest area in real time; The processor is used to obtain the data collected by the sensor detection module and perform flame target recognition, flame and smoke target detection, and determine the geographical coordinates of the fire point; Communication module, used for interaction between the processor and the cloud platform and monitoring station; The cloud platform is used to build the model, update the model based on the data transmitted by the communication module, and send it to the processor; Monitoring station, which receives fire location and image data and triggers alarms.

6. The device according to claim 5, characterized in that The sensing detection module includes: an infrared thermal imaging camera connected to the processor for transmitting infrared images; A binocular camera connected to the processor for capturing stereoscopic vision images; The anemometer sensor, temperature and humidity sensor, GPS positioning unit, and soil temperature and humidity sensor are all connected to the processor.

7. The device according to claim 6, characterized in that The device further comprises: The data storage module adopts a pluggable SSD design to store raw data, algorithm models and processing results; The power module integrates solar panels and high-voltage batteries to power various components.

Images