Unmanned aerial vehicle forest fire monitoring and analyzing method

Abstract

The invention discloses an unmanned aerial vehicle forest fire monitoring and analysis method, and relates to the technical field of artificial intelligence and unmanned aerial vehicle application. In order to solve the defects of insufficient fire detection precision and lack of a complete monitoring and analysis closed-loop system in the prior art, the technical scheme provided by the invention comprises the following steps: acquiring real-time video data of a forest area, and outputting a real-time video stream and unmanned aerial vehicle position information; in combination with the forest fire data set obtained through training, outputting a fire detection result containing a fire bounding box; in combination with context feature interaction and an attention mechanism, pixel-level segmentation is carried out on the fire disaster area, and a flame segmentation result is output; the actual flame area is calculated according to the flame pixel proportion, the flight height of the unmanned aerial vehicle and camera parameters, and flame area data and fire point positioning information are output; and the cloud platform is combined to realize fire visual display, and alarm prompt and emergency response information are generated when the flame area exceeds a threshold value. The system is suitable for real-time monitoring, positioning and fire analysis of forest fire.

Description

Technical Field

[0001] It involves the fields of artificial intelligence and drone application technology, specifically drone-based forest fire monitoring and analysis. Background Technology

[0002] Fire is inextricably linked to our lives and is an essential tool for human survival and development. However, fire is also a double-edged sword. Due to its high temperature, improper use can easily threaten our lives and property. Forest fires, as one of the classic types of fires, have become an increasingly serious problem, and their frequency is gradually increasing. When a forest fire breaks out, it spreads very rapidly, requiring a large amount of manpower, financial resources, and materials. It is characterized by its long burning time, large area, and difficulty in control, and once it occurs, it will cause significant damage to forest resources.

[0003] Forests are a vital component of the global terrestrial ecosystem, playing an irreplaceable role in maintaining ecological balance, conserving water resources, and absorbing carbon dioxide. However, forest fires are frequent and spread extremely rapidly, causing not only massive losses of timber resources but also severe casualties and economic damage. Therefore, achieving early detection and efficient monitoring of forest fires has been a key focus for research institutions both domestically and internationally.

[0004] Currently, forest fire monitoring methods mainly include manual inspections, video surveillance, lookout tower observations, ground sensor deployment, and helicopter patrols. For example, manual inspections are low-cost but have limited monitoring range and pose challenges to personal safety; lookout towers and video surveillance can expand the field of view but rely on long-term human monitoring, resulting in poor real-time performance and efficiency; sensor deployment can achieve automatic detection, but accuracy is significantly affected by environmental factors such as weather and humidity, and stability is insufficient; helicopter patrols have the advantages of wide coverage and high flexibility, but operating costs are too high, making large-scale deployment difficult. With the development of deep learning and drone technology, researchers have begun to explore using drones equipped with computer vision algorithms to monitor fires. For example, some studies have applied traditional YOLO algorithms or convolutional neural networks to the recognition of flames and smoke, while others have used segmentation networks to perform pixel-level segmentation of flame areas. However, these studies still have certain limitations in practical applications: on the one hand, target detection algorithms have limited accuracy and are easily affected by complex backgrounds; on the other hand, segmentation algorithms generally have high computational costs, making them difficult to deploy directly on embedded devices with limited computing power, such as drones. In addition, most existing systems only have single functions of detection or segmentation, and lack a closed-loop system that integrates drone inspection, fire identification, quantitative analysis of flame areas and comprehensive judgment in the background.

[0005] In summary, existing technologies suffer from shortcomings such as insufficient fire detection accuracy, difficulty in lightweight deployment of segmentation algorithms, disconnect between drone monitoring and back-end analysis, and a lack of a complete closed-loop monitoring and analysis system. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as insufficient fire detection accuracy, difficulty in lightweight deployment of segmentation algorithms, disconnect between drone monitoring and backend analysis, and lack of a complete closed-loop monitoring and analysis system, the technical solution provided by this invention is as follows: Methods for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs), including: The steps include collecting real-time video data of the forest area, performing data preprocessing, and outputting real-time video streams and drone location information; The steps involve inputting a real-time video stream into an improved YOLOv8 fire detection model for analysis, combining it with a forest fire dataset obtained from training, and outputting fire detection results containing fire bounding boxes. The steps are as follows: inputting the fire detection results into the LMAF-Net fire segmentation model, which is based on the Mobilemini-8 backbone network and MASPP structure, and combines contextual feature interaction and attention mechanism to perform pixel-level segmentation of the fire area and output flame segmentation results. The steps involve combining the flame segmentation results with the drone's location information, calculating the actual flame area based on the flame pixel ratio, drone flight altitude, and camera parameters, and outputting the flame area data and fire point location information. The steps involve inputting flame area data and fire location information into the backend analysis system, combining it with the cloud platform to achieve visual display of the fire situation, and generating alarm prompts and emergency response information when the flame area exceeds the threshold.

[0007] Furthermore, a preferred implementation method is provided in which the forest fire dataset is obtained by fusing forest fire images taken by drones with an open-source fire dataset.

[0008] Furthermore, a preferred implementation is provided, which improves the YOLOv8 fire detection model by introducing a Hattention attention mechanism in the backbone network, using BiFPN bidirectional feature fusion in the neck structure, and introducing DYHead and deformable convolution in the detection head.

[0009] Furthermore, a preferred implementation is provided: the LMAF-Net fire segmentation model employs a Mobilemini-8 lightweight backbone network and a MASPP multi-scale dilated convolutional structure in the encoder part, combined with the CBAM attention mechanism, and recovers spatial details in the decoder part through contextual feature interaction and cross-layer feature fusion.

[0010] Furthermore, a preferred implementation method is provided, which triggers multi-level alarms when the flame area exceeds a set threshold, and generates a fire development trend map and emergency response suggestions by combining historical data records provided by the cloud platform.

[0011] It also provides drone-based forest fire monitoring and analysis equipment, including: A module that collects real-time video data of forest areas, performs data preprocessing, and outputs real-time video streams and drone location information; The module analyzes the real-time video stream input into the improved YOLOv8 fire detection model, combines it with the forest fire dataset obtained from the training, and outputs the fire detection results containing the fire bounding boxes. The fire detection results are input into the LMAF-Net fire segmentation model, which is based on the Mobilemini-8 backbone network and MASPP structure, and combines context feature interaction and attention mechanism to perform pixel-level segmentation of the fire area and output flame segmentation results. This module combines the flame segmentation results with the drone's location information, calculates the actual flame area based on the flame pixel ratio, drone flight altitude, and camera parameters, and outputs flame area data and fire point location information. The system inputs flame area data and fire location information into the backend analysis system, combines it with the cloud platform to realize the visualization of the fire situation, and generates alarm prompts and emergency response information when the flame area exceeds the threshold.

[0012] A drone-based forest fire monitoring and analysis system is also provided to implement the method described above. This system includes a drone, an image processing module, a flight control module, and a navigation and positioning module. The image processing module is used to perform fire detection and fire segmentation, and interacts with the flight control module and navigation and positioning module to transmit the detection results and location information to the background analysis system to realize forest fire monitoring and analysis.

[0013] A computer storage medium is also provided for storing a computer program, which, when read by the computer, executes the method.

[0014] A computer is also provided, including a processor and a storage medium, wherein the computer executes the method when the processor reads a computer program stored in the storage medium.

[0015] A computer program product is also provided, which, when executed, implements the method described.

[0016] Compared with the prior art, the advantages of the technical solution provided by the present invention are as follows: This approach proposes the YOLO-AttF detection algorithm based on YOLOv8, introducing a hybrid attention mechanism and a bidirectional feature pyramid network to enhance the fusion capability of features at different scales. This allows the model to more accurately distinguish flames from complex backgrounds. Compared to existing fire detection methods that only use traditional YOLO or are based on color and texture features, this approach improves detection accuracy by approximately 3 percentage points, effectively reducing the false positive rate caused by tree reflections, changes in lighting, or smoke interference, thereby improving the reliability of early fire identification.

[0017] In the fire segmentation stage, the proposed LMAF-Net employs a lightweight Mobilemini-8 backbone network and a MASPP module composed of dilated convolutions, combined with a CBAM attention mechanism to highlight flame features. Furthermore, multiple contextual feature interactions through the decoder enhance segmentation accuracy. Compared to existing common models such as UNet or Deeplabv3+, this approach achieves near- or even higher segmentation accuracy while maintaining a lower parameter count. It is suitable for deployment on drones with limited computing power, solving the problems of high computational overhead and difficulty in real-time operation associated with traditional segmentation algorithms.

[0018] In the drone platform section, this solution designs drones with path planning and positioning capabilities, and deploys improved detection and segmentation algorithms on them. Combined with 5G real-time image transmission technology, it achieves long-distance, low-latency video transmission. Compared with traditional WiFi-based or manual patrol methods, this solution ensures both the scope and flexibility of inspections while significantly reducing information transmission latency, enabling more timely detection and location of fires and providing a technological foundation for rapid response.

[0019] In the background analysis section, this solution constructs a closed loop of monitoring and analysis, capable of calculating the flame pixel ratio based on flame segmentation results and inferring the actual flame area by combining the drone's flight altitude and camera parameters. Compared to existing systems that only identify fires and issue alarms, this solution provides quantitative data support for subsequent fire situation assessment and rescue deployment, improving the scientific nature and operability of fire prevention and control.

[0020] It is suitable for real-time monitoring, location and fire analysis of forest fires. Attached Figure Description

[0021] Figure 1 This is a diagram of the overall system framework. Figure 2 This is a hardware system architecture diagram; Figure 3 This is a diagram of the overall software structure. Figure 4 Here is the algorithm flowchart; Figure 5 This is a schematic diagram of the algorithm simulation results. Detailed Implementation

[0022] To make the advantages and benefits of the technical solution provided by the present invention clearer, the technical solution provided by the present invention will now be described in further detail with reference to the accompanying drawings, specifically: Implementation Method 1: This implementation method provides a method for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs), including: The steps include collecting real-time video data of the forest area, performing data preprocessing, and outputting real-time video streams and drone location information; The steps involve inputting a real-time video stream into an improved YOLOv8 fire detection model for analysis, combining it with a forest fire dataset obtained from training, and outputting fire detection results containing fire bounding boxes. The steps are as follows: inputting the fire detection results into the LMAF-Net fire segmentation model, which is based on the Mobilemini-8 backbone network and MASPP structure, and combines contextual feature interaction and attention mechanism to perform pixel-level segmentation of the fire area and output flame segmentation results. The steps involve combining the flame segmentation results with the drone's location information, calculating the actual flame area based on the flame pixel ratio, drone flight altitude, and camera parameters, and outputting the flame area data and fire point location information. The steps involve inputting flame area data and fire location information into the backend analysis system, combining it with the cloud platform to achieve visual display of the fire situation, and generating alarm prompts and emergency response information when the flame area exceeds the threshold.

[0023] The forest fire dataset was obtained by fusing forest fire images taken by drones with an open-source fire dataset.

[0024] The improved YOLOv8 fire detection model introduces a Hattention mechanism in the backbone network, adopts BiFPN bidirectional feature fusion in the neck structure, and introduces DYHead and deformable convolution in the detection head.

[0025] The LMAF-Net fire segmentation model employs a lightweight Mobilemini-8 backbone network and a MASPP multi-scale dilated convolutional structure in the encoder part, combined with the CBAM attention mechanism. In the decoder part, spatial details are recovered through contextual feature interaction and cross-layer feature fusion.

[0026] When the flame area exceeds a set threshold, a multi-level alarm is triggered, and a fire development trend map and emergency response suggestions are generated by combining historical data records provided by the cloud platform.

[0027] A drone-based forest fire monitoring and analysis system is also provided to implement the method described above. This system includes a drone, an image processing module, a flight control module, and a navigation and positioning module. The image processing module is used to perform fire detection and fire segmentation, and interacts with the flight control module and navigation and positioning module to transmit the detection results and location information to the background analysis system to realize forest fire monitoring and analysis.

[0028] Implementation Method Two: This implementation method further describes the technical solution provided in Implementation Method One in more detail, specifically: A drone-based forest fire monitoring and analysis system enables real-time monitoring and quantitative assessment of forest fires through drone inspection, fire detection, flame segmentation, area estimation, and comprehensive backend analysis. The specific steps are as follows: First, a drone inspection platform was constructed. The platform used a KY-001PRO drone, equipped with a Raspberry Pi 5 as the processing core, and connected to a USB driverless AF093V1 camera for video acquisition. The Raspberry Pi 5 has a built-in quad-core Cortex-A76 processor and 8GB of memory, capable of supporting edge deployment of deep learning models. The drone is equipped with a Core-Air820 GPS module for precise positioning, and combines it with an HI3519DV500 flight control module and an N10P LiDAR navigation module for waypoint planning and attitude control. The path planning algorithm employs an A* search strategy, constructing a heuristic cost function based on the environment to generate the drone's autonomous flight path. The drone then inspects the forest area according to the planned route, collecting scene video and location information in real time. The output of this step is the drone video stream and location data.

[0029] Next, fire detection is performed on the real-time video stream. An improved YOLO-AttF detection model is deployed on a Raspberry Pi 5. This model introduces a Hattention mechanism and a BiFPN bidirectional feature fusion structure on top of YOLOv8, and employs DYHead and deformable convolution in the detection head. This design enhances flame feature extraction capabilities against complex forest backgrounds and effectively suppresses interference from smoke and lighting changes. Real-time images acquired by the drone via its camera are input into the model, which outputs detection results including the bounding box of the flame target and category labels. The output of this step is a preliminary detection image with fire area markings.

[0030] Next, the fire area is finely segmented. The backend system calls the LMAF-Net segmentation model, which uses an improved Mobilemini-8 lightweight network as its backbone. It extracts multi-scale features through layer-by-layer downsampling. The encoder part introduces a MASPP structure, combining multi-scale dilated convolutions and CBAM attention mechanisms to highlight flame region features and expand the receptive field. The decoder part employs a context interaction module and a cross-layer fusion structure to gradually restore spatial resolution and compensate for details lost during downsampling. The final output is a pixel-level flame segmentation map. Compared to traditional UNet or SegNet, this model achieves near-Deeplabv3+ segmentation accuracy with only about 20M parameters, making it more suitable for embedded deployments. The output of this step is the flame segmentation map.

[0031] Subsequently, the actual flame area is calculated based on the flame segmentation results. The backend system first calculates the percentage of flame pixels in the flame segmentation image, then combines this with the drone's real-time flight altitude and the camera's imaging parameters (e.g., the AF093V1 camera has a resolution of 2592×1944 pixels and a pixel size of 1.4 micrometers) to convert the pixel area into the actual area. Specifically, the process involves calculating the actual ground area corresponding to a single pixel in the image based on the drone's altitude and the camera's field of view, then multiplying the total number of flame pixels by the area of ​​a single pixel to obtain the actual fire coverage area. The output of this step is quantitative flame area data.

[0032] Finally, the flame area data is combined with the drone's location information to generate a complete fire analysis report. The backend system receives the video stream and analysis results uploaded by the drone using a user interface developed with QT Designer. The interface displays real-time fire footage, flame segmentation areas, flame area values, and corresponding geographic coordinates. The backend further generates a fire situation map and development trend prediction. If the fire area exceeds a preset threshold, multi-level early warning prompts are automatically triggered. The output of this step is a visualized fire analysis report, providing a basis for fire prevention command and emergency rescue.

[0033] Through the above steps, a closed-loop system is formed, encompassing drone inspection, fire detection, flame segmentation, and back-end quantitative analysis. This system not only improves the accuracy of fire detection and the real-time performance of segmentation but also enables automatic calculation of fire area and early warning linkage. It solves the technical problems of insufficient fire detection accuracy, difficulty in lightweight deployment of segmentation algorithms, and disconnect between drone monitoring and back-end analysis in existing technologies.

[0034] A drone-based forest fire monitoring and analysis system comprises a video acquisition circuit, a video processing circuit, a drone flight control circuit, a navigation and positioning circuit, a motor drive circuit, and a backend analysis system, forming a complete closed loop from data acquisition, processing, transmission to backend analysis. The specific structure and construction are as follows: First, the video acquisition circuit includes an AF093V1 camera and an interface module. The camera connects to the Raspberry Pi processing unit via a USB interface to acquire real-time images of the forest area. The camera has a built-in CMOS sensor and autofocus mechanism, enabling clear imaging at distances from 5 cm to 10 meters, with a resolution of 2592×1944 pixels. The video acquisition circuit is stably powered by a power module, and the output video signal is transmitted to the video processing circuit.

[0035] Secondly, the video processing circuit consists of a Raspberry Pi 5 core board, a storage module, and peripheral interfaces. The Raspberry Pi 5 features a built-in quad-core Cortex-A76 processor, high-speed PCIe expansion, and dual MIPI interfaces, supporting parallel video data input and processing. The storage module in the video processing circuit stores the weight files of the trained YOLO-AttF detection model and LMAF-Net segmentation model. The processing unit inputs the acquired images into the algorithm model and outputs fire target recognition results and segmentation data. The video processing circuit also includes a wireless communication module, enabling real-time transmission of images and data via 5G or WiFi networks.

[0036] Furthermore, the UAV flight control circuit consists of the HI3519DV500 flight control module, the MPU6050 inertial detection unit, and the attitude calculation circuit. The flight control module is the core control unit of the system, responsible for receiving navigation and positioning signals, performing real-time calculations of the UAV's attitude, and generating control commands to output to the motor drive circuit. The inertial detection unit is responsible for collecting acceleration and angular velocity data, which the attitude calculation circuit combines with the flight control commands to achieve stable hovering, trajectory flight, and autonomous return of the UAV.

[0037] Furthermore, the navigation and positioning circuit consists of an N10P LiDAR module and a Core-Air820 GPS module. The N10P LiDAR uses Time-of-Flight (TOF) ranging to output point cloud information with centimeter-level accuracy for environmental perception and obstacle avoidance. The GPS module supports joint positioning using multiple systems such as BeiDou, GPS, and GLONASS, providing stable dual-frequency signals in complex forest areas. The navigation and positioning circuit combines the LiDAR and GPS information through a data fusion processor and employs filtering algorithms to achieve precise UAV positioning and path planning.

[0038] The motor drive circuit consists of an electronic speed controller (ESC) and a brushless motor. The ESC receives control signals from the flight control module, converting DC power into three-phase AC power to drive the T-Motor MN3110 brushless motor. The motor uses a silicon steel core and star windings, featuring high efficiency, long lifespan, and low noise, providing lift and maneuverability for the UAV. The motor drive circuit achieves independent control of the four motors through a power distribution module to meet the multi-degree-of-freedom flight requirements of the UAV.

[0039] The backend analysis system consists of a data receiving module, an image analysis module, and a user interaction module. The data receiving module receives images and recognition results uploaded by the drone via a wireless network; the image analysis module combines the LMAF-Net segmentation results with the drone's flight altitude and camera parameters to calculate the actual area of ​​the flames; the user interaction module displays the drone inspection footage, fire detection markers, segmented areas, and fire area values ​​on a QT interface, while simultaneously locating the fire source on a map using GPS coordinates and triggering an alarm based on a set threshold.

[0040] In summary, the UAV forest fire monitoring and analysis system of the present invention achieves real-time detection, segmentation, and area quantification analysis of fires through the organic integration of video acquisition circuits, video processing circuits, flight control circuits, navigation and positioning circuits, motor drive circuits, and a background analysis system, ensuring the integrity and feasibility of the system.

[0041] The workflow of the above methods and systems is as follows: First, the airborne system is established. The drone is equipped with dual-light cameras and a Raspberry Pi 5 processing unit. The cameras are used to collect real-time video data of the forest area, the Raspberry Pi 5 preprocesses the collected video, and the data is transmitted back via a millimeter-wave communication module MOD6320-TX and a 5G wireless link. The drone has autonomous flight path planning capabilities, enabling it to perform inspection tasks according to a preset path, and hovering and reporting its location information after detecting a fire. The output of this step is a real-time video stream and location information.

[0042] Next, the video stream is input into the algorithm core for fire detection. The algorithm core is based on a dataset trained on 2467 drone fire images and deploys an improved YOLOv8 detection model. This model effectively improves the recognition accuracy of flames against complex forest backgrounds by enhancing multi-scale feature fusion capabilities and attention mechanisms. The detection results output the bounding box information of the fire area. The output of this step is the fire detection result.

[0043] Next, the detection results are input into the segmentation algorithm module for flame region extraction. The segmentation algorithm uses LMAF-Net with Mobilemini-8 as the backbone network and combined with the MASPP structure, which has the characteristics of being lightweight and highly accurate. Through the synergistic effect of multi-scale convolution, contextual feature interaction, and attention mechanism, pixel-level segmentation of the flame region is achieved, resulting in a clear flame mask map. The output of this step is the flame segmentation result.

[0044] Subsequently, the backend processing module receives the flame segmentation results and the drone's location information to perform quantitative fire analysis. The backend system calculates the actual flame area based on the flame pixel ratio, drone flight altitude, and camera parameters, and uses GPS coordinates to determine the specific location of the fire. The output of this step is flame area data and fire location information.

[0045] Finally, the backend system performs fire analysis and situation display via a cloud platform. The backend interface visualizes flame detection images, segmented areas, actual area, and fire location, and generates a fire situation report. When the detection results exceed a set threshold, the system automatically triggers an alarm and pushes emergency response suggestions to the user. The output of this step is the fire situation assessment result and alarm information.

[0046] Implementation Method 3: Combination Figure 1-5 This embodiment describes the technical solution provided above in further detail through specific examples. Specifically: Based on system requirements and functions, the overall system architecture is as follows: Figure 1 As shown, this forest fire monitoring system adopts a three-layer architecture: airborne terminal, communication link, and cloud platform. The airborne terminal is equipped with dual-light cameras and a Raspberry Pi 5 processor, running the YOLO-AttF algorithm detection model and the lightweight segmentation algorithm LMAF-Net, improving detection accuracy by 3%. Data is transmitted back in real time via a MOD6320-TX millimeter-wave module. The backend system performs real-time analysis based on two core algorithms, automatically generating dynamic fire situation maps and triggering multi-level alarms. The cloud platform provides data storage, multi-terminal synchronization, and historical disaster analysis functions.

[0047] like Figure 2 As shown, the hardware devices used in this system mainly include a USB driverless AF093V1 camera, a Raspberry Pi 5, a KY-001PRO drone, a HI3519DV500 flight control module, an N10P intelligent navigation module, a Core-Air820 GPS positioning module, and a T-Motor MN3110 motor and drive module.

[0048] Image acquisition is the fundamental step in achieving precise flame positioning, detection, and segmentation in this implementation. The system uses a Raspberry Pi paired with an external camera. The camera acquires scene images in real time and transmits them to the Raspberry Pi, where the built-in algorithms process them instantly, forming a complete data processing chain from image acquisition to intelligent analysis.

[0049] This article uses the USB driverless AF093V1 camera, which is plug-and-play and compatible with 99.99% of systems. It features a 5MP 2K sensor with a resolution of 2592*1944, a built-in noise-canceling microphone, an OV industrial main control chip on a precision motherboard, an imported CMOS sensor, and is integrally molded using a high-precision nanometer lithography machine. It also features autofocus and is currently used in industrial applications such as video recording, aerial photography, face recognition, and intelligent document scanning.

[0050] The video transmission module uses a Raspberry Pi 5, which offers significant hardware and performance improvements compared to the Raspberry Pi 4, making it suitable for developers and enthusiasts to develop various projects. The Raspberry Pi 5 features the new Broadcom BCM2712 processor, a quad-core Cortex-A76 architecture with a clock speed of up to 2.4GHz, offering 2-3 times the performance of its predecessor. GPU performance has also been greatly enhanced, supporting more efficient graphics processing. It supports 4GB and 8GB memory options, and adds a PCIe 2.0 interface for high-speed peripheral connections. The USB ports have been upgraded to two USB 3.0 ports supporting 5Gbps transfer rates, while two USB 2.0 ports are retained. The camera and display interfaces have been upgraded to two 4-lane MIPI DSI / CSI ports, compatible with more peripherals. This system adopts a server-side training and edge-side deployment architecture. First, the improved YOLOv8 flame detection model is trained and optimized on the server platform, and then the generated model weight file is deployed to the Raspberry Pi 5 embedded device. The system receives real-time video stream data from a camera via a USB interface and uses a ported YOLOv8 model to achieve efficient fire detection and segmentation functions.

[0051] The drone's flight controller (FC) is the "brain" of the drone, responsible for processing sensor data, executing flight control algorithms, and coordinating the work of various hardware modules to ensure stable flight. In this implementation, the MCU main controller is the HI3519DV500, and the inertial detection unit is the MPU6050, including functions such as optical flow altitude hold, multi-functional return-to-home, 360° surround flight, trajectory-based fixed-point flight, and one-click takeoff.

[0052] The control module is responsible for calculating the current attitude of the UAV, optimizing the control, and generating corresponding control quantities for the drive module; the detection module is responsible for measuring the current attitude of the UAV and providing data to the control module; the drive module is responsible for driving the four motors to rotate, thereby enabling the UAV to fly normally.

[0053] The core of the radar navigation module is the N10P lidar sensor. The N10P lidar is a key component of the navigation system, employing Time-of-Flight (TOF) ranging technology for high-precision environmental perception. This sensor emits a laser beam and receives reflected signals, enabling the construction of a 3D point cloud map with centimeter-level accuracy, effectively covering a detection range of 100-300 meters. In terms of hardware architecture, the system adopts a modular design. The radio frequency front-end uses phased array technology for electronic scanning, avoiding mechanical rotating parts. The signal processing unit is equipped with a high-performance DSP processor, which extracts key parameters such as target distance, azimuth, and velocity in real time through Fast Fourier Transform. The data fusion processor integrates radar data with information from multiple sensor sources such as IMU and GPS, employing algorithms such as Kalman filtering to achieve precise positioning.

[0054] The satellite positioning module tested and used in this embodiment is the Core-Air820. The Core-Air820 is a dual-frequency, high-performance, highly integrated multi-mode satellite positioning and navigation module. It is small in size and low in power consumption, and can be used in GNSS positioning applications such as vehicle navigation, smart wearables, and drones. It supports satellite positioning systems such as BeiDou-3, GPS, GLONASS, GALILEO, QZSS, and SBAS, supports simultaneous positioning on both L1 and L5 frequency bands, and supports multi-system joint positioning and single-system independent positioning.

[0055] A brushless DC motor (BLDC) mainly consists of a permanent magnet rotor, a multi-phase winding stator, and a Hall effect position sensor (optional). Its working principle is based on the interaction of electromagnetic induction and magnetic fields: the stator contains multiple sets of coils (usually wound with insulated wire) fixed to the motor housing; the rotor is composed of permanent magnets arranged according to a specific number of pole pairs. Brushless motors must rely on an electronic speed controller (ESC) for control. The ESC converts the DC power from the battery into three-phase AC power and precisely adjusts the motor speed according to flight control commands.

[0056] This embodiment uses a T-Motor MN3110 brushless motor. The core material is made of laminated silicon steel sheets to reduce eddy current losses. It uses enameled wire windings with a star (Y) connection. The wire diameter is selected according to the current requirement, ranging from 0.3mm to 1.2mm. Brushless DC motors do not have commutators or brushes, have a relatively simple structure, and feature long service life, low noise, and high speed.

[0057] The UAV path planning in this embodiment adopts the A* algorithm. The A* algorithm is the most efficient direct search algorithm for solving the optimal path in a static environment. It is an improved algorithm of Dijkstra's algorithm, inheriting and retaining the advantages of Dijkstra's algorithm. While ensuring that the optimal path is obtained, it adopts heuristic search.

[0058] The A* algorithm uses two lists during path planning: an Open list to store nodes to be searched and a Closed list to store points that have been explored and whose minimum cost has been calculated using an evaluation function. The A* algorithm determines the search direction and next-hop node using the evaluation function V(x). Generally, it assumes the agent can move in four directions (up, down, left, right) or eight directions (up, down, left, right, upper left, lower left, upper right, lower right). The next-hop node for the current node X is determined by the minimum cost of its surrounding nodes calculated by the evaluation function V(x). This process is repeated from the starting point S to the target point D until a planned path is generated. The points in the Closed list are then returned, representing the desired path.

[0059] This implementation uses 5G image transmission technology for real-time high-definition video transmission with a latency as low as 0.5 seconds. Compared to traditional image transmission (such as WiFi), it offers faster transmission speeds (supporting 4K or even 8K high-definition images) and lower latency (end-to-end latency down to milliseconds), ensuring real-time image transmission and preventing misjudgments of fire situations or loss of target tracking due to latency. The specific process is as follows: The video stream is first placed into a sub-thread buffer queue for temporary storage and to regulate the data flow rate. Next, the video data stream in the queue is transferred to the FFmpeg memory buffer. Leveraging FFmpeg's powerful audio and video processing capabilities, the data is separated into single frames at specific time intervals (00 00 00 01), then encapsulated in FLV format, and sent to the server-side streaming device via the RTMP protocol. After receiving the data, the server transmits it to the server for processing. Finally, the playback device obtains the data from the server, decodes it, and then plays the video.

[0060] This implementation method trains on two datasets to improve the accuracy of fire detection. The first dataset consists of images taken by drones by faculty and students of Northeast Forestry University simulating forest fires. The second dataset is the publicly available FireFly dataset, which covers various types of ash, from burning leaves to sparks. This implementation method merges the self-made dataset with the open-source dataset to form a dataset of 2467 forest fire images from the perspective of drones.

[0061] Fire detection algorithm Overall Algorithm Structure: The fire detection algorithm YOLO-AttF proposed in this implementation consists of a backbone extraction network, a neck fusion network, and a detection head network. It sequentially completes image feature extraction, multi-scale feature fusion, and final target detection and classification tasks. The input image is processed by the backbone network to extract basic features, and then the neck network fuses and enhances features at different levels. Finally, the detection head outputs the target location and category information.

[0062] In the backbone extraction network, a Hattention mechanism is introduced, enabling the model to focus on more critical regions in the image and suppress irrelevant background interference. In the neck fusion network, a 4-input, 3-output BiFPN bidirectional fusion method combined with upsampling and other operations is used to support top-down and bottom-up bidirectional feature fusion, efficiently integrating features at different scales and enhancing multi-scale feature interaction. In the detection head network, a dynamic detection head, DYHead, is used with deformable convolutions to adapt to the irregular shapes of fire targets, improving target detection accuracy and robustness in complex scenes.

[0063] Hattention Attention Mechanism: This attention mechanism architecture focuses on deep image feature extraction and enhancement. First, it obtains basic features through deep feature extraction, constructing a residual hybrid attention group (RHAG) consisting of a hybrid attention module (HAB) and an overlapping cross attention module (OCAB). Layer normalization is used to stabilize the training process, combined with channel attention (CAB) to enhance key features, self-attention is used to capture long-range spatial dependencies, and overlapping cross attention expands the receptive field to uncover multi-scale associations. The architecture introduces a multilayer perceptron for nonlinear transformation and extensively employs residual connections to prevent information loss. After multiple rounds of multi-dimensional attention enhancement and progressive feature fusion, a more discriminative feature representation is finally output.

[0064] Bi-FPN (Bidirectional Feature Pyramid Network): Bi-FPN is an improvement and optimization of PANet. Both are used for feature pyramid construction to handle multi-scale features. PANet presents a relatively simple hierarchical transmission, with features mainly transmitted along unidirectional or basic cross-layer connections, resulting in relatively limited feature interaction between layers. BiFPN, on the other hand, adds more cross-layer connections, constructing bidirectional feature flow paths. This allows for more thorough bidirectional fusion and iterative optimization of high-level semantic features and low-level detailed features, strengthening information interaction between features at different scales and improving feature reuse and fusion effects.

[0065] DYHead's dynamic detection head employs a multi-dimensional attention perception and deformable convolution, constructing a cascaded structure of three attention modules: scale, space, and task perception. Input features first pass through the scale-aware attention module, which uses average pooling, 1×1 convolution, ReLU, and Hard Sigmoid activation to extract and weight features at different scales. Next, the spatial-aware attention module utilizes 3×3 convolution, offset calculation, and deformable convolution v3 to adapt to target spatial deformations and enhance features. Finally, the task-aware attention module precisely adjusts features according to task requirements through average pooling, fully connected layers, ReLU, and normalization. These modules fuse attention weights using multiplication operations, achieving multi-dimensional fine-grained optimization of input features. This facilitates efficient extraction of key information for detection tasks and improves the detection accuracy and adaptability for targets in complex scenes.

[0066] The fire segmentation algorithm, LMAF-Net, consists of two core modules: an encoder and a decoder. It achieves the task of forest fire detection through feature extraction and feature restoration. The specific process is as follows: input image → encoder extracts multi-scale features → decoder fuses features and outputs detection results.

[0067] The encoder consists of a Mobilemini-8 backbone network and a MASPP module. Mobilemini-8 extracts multi-scale features through layer-by-layer downsampling, with low-level features preserving texture details and high-level features containing semantic information. The MASPP module is an improvement on ASPP, employing multi-scale convolution and pooling for feature fusion, combined with the CBAM attention mechanism to highlight flame features, while adjusting the dilation coefficient to expand the receptive field, and adding pooling layers to enhance high-level semantic features, thereby improving the model's generalization ability.

[0068] The decoder processes the encoder output through an SFF feature fusion module. This module multiplies the encoder features with the original features to compensate for the spatial details lost during downsampling, then restores the resolution through upsampling, and gradually restores the image scale by combining cross-layer connections. Finally, after 3×3 convolution thinning and 1×1 convolution, a single-channel prediction map is output, achieving accurate segmentation of the flame region.

[0069] The input image undergoes background and other redundant information removal, retaining only the flame region for more accurate segmentation. Specifically, polygons are drawn and labeled for the flame region during dataset processing, and the labeled flame targets are cropped and randomly pasted onto a completely black image. Mobilemini-8: The input image first undergoes preliminary feature extraction through the "CBA-HS" module, outputting a 16×112×112 feature map. The core of the network consists of eight cascaded "bneck" modules, which extract multi-scale features through downsampling and channel number adjustment. Specifically, bneck1 reduces the feature map size from 112×112 to 56×56 to obtain global information, while subsequent bneck modules capture different levels of flame features by adjusting the number of channels (e.g., 24, 40, etc.).

[0070] Each bneck module employs channel expansion and a two-stage CBA structure: the first stage uses ordinary convolution to quickly extract global contour features, while the second stage uses separable depthwise convolution to extract spatial features with fine precision while reducing computational cost. An SE attention mechanism is introduced to automatically focus on key channel information, enhancing flame-related features.

[0071] CBA, as a fundamental building block, combines the advantages of ordinary convolution and separable depthwise convolution, along with the H-Swish / ReLU activation function, to effectively extract flame features while maintaining lightweight architecture. The entire Mobilemini-8 network achieves a balance between lightweight architecture and feature representation capabilities through the collaborative work of multiple modules, providing strong support for accurate forest flame segmentation.

[0072] CBAM Attention Module: CRAM adopts a "channel-spatial" cascaded structure. First, the channel attention module extracts channel features using dual pooling branches, generates channel weights after convolution and ReLU activation, highlighting flame-related channels. Subsequently, the spatial attention module also uses dual pooling to compress features, generates spatial weights through convolution and Sigmoid activation, enhances the flame region, and suppresses the background.

[0073] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs), characterized in that, include: The steps include: collecting real-time video data of the forest area, performing data preprocessing, and outputting real-time video streams and drone location information; The steps involve inputting a real-time video stream into an improved YOLOv8 fire detection model for analysis, combining it with a forest fire dataset obtained from training, and outputting fire detection results containing fire bounding boxes. The steps are as follows: inputting the fire detection results into the LMAF-Net fire segmentation model, which is based on the Mobilemini-8 backbone network and MASPP structure, and combines contextual feature interaction and attention mechanism to perform pixel-level segmentation of the fire area and output flame segmentation results. The steps involve combining the flame segmentation results with the drone's location information, calculating the actual flame area based on the flame pixel ratio, drone flight altitude, and camera parameters, and outputting the flame area data and fire point location information. The steps involve inputting flame area data and fire location information into the backend analysis system, combining it with the cloud platform to achieve visual display of the fire situation, and generating alarm prompts and emergency response information when the flame area exceeds the threshold.

2. The method for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The forest fire dataset was obtained by fusing forest fire images taken by drones with an open-source fire dataset.

3. The method for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The improved YOLOv8 fire detection model introduces a Hattention mechanism in the backbone network, adopts BiFPN bidirectional feature fusion in the neck structure, and introduces DYHead and deformable convolution in the detection head.

4. The method for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The LMAF-Net fire segmentation model employs a lightweight Mobilemini-8 backbone network and a MASPP multi-scale dilated convolutional structure in the encoder part, combined with the CBAM attention mechanism. In the decoder part, spatial details are recovered through contextual feature interaction and cross-layer feature fusion.

5. The method for monitoring and analyzing forest fires using unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, When the flame area exceeds a set threshold, a multi-level alarm is triggered, and a fire development trend map and emergency response suggestions are generated by combining historical data records provided by the cloud platform.

6. A drone-based forest fire monitoring and analysis device, characterized in that, include: A module that collects real-time video data of forest areas, performs data preprocessing, and outputs real-time video streams and drone location information; The module analyzes the real-time video stream input into the improved YOLOv8 fire detection model, combines it with the forest fire dataset obtained from the training, and outputs the fire detection results containing the fire bounding boxes. The fire detection results are input into the LMAF-Net fire segmentation model, which is based on the Mobilemini-8 backbone network and MASPP structure, and combines context feature interaction and attention mechanism to perform pixel-level segmentation of the fire area and output flame segmentation results. This module combines the flame segmentation results with the drone's location information, calculates the actual flame area based on the flame pixel ratio, drone flight altitude, and camera parameters, and outputs flame area data and fire point location information. The system inputs flame area data and fire location information into the backend analysis system, combines it with the cloud platform to realize the visualization of the fire situation, and generates alarm prompts and emergency response information when the flame area exceeds the threshold.

7. A drone-based forest fire monitoring and analysis system, characterized in that, The system for implementing the method of claim 1 includes a drone, an image processing module, a flight control module, and a navigation and positioning module. The image processing module is used to perform fire detection and fire segmentation, and interacts with the flight control module and navigation and positioning module to transmit the detection results and location information to the background analysis system to realize forest fire monitoring and analysis.

8. A computer storage medium for storing computer programs, characterized in that, When the computer program is read by the computer, the computer executes the method of claim 1.

9. A computer, comprising a processor and a storage medium, characterized in that, When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1.

10. A computer program product, as a computer program, is characterized by: When the computer program is executed, it implements the method of claim 1.

Images