Edge intelligent forest degradation feature unmanned aerial vehicle adaptive monitoring device and method

Abstract

The application provides an edge intelligent forest degradation feature unmanned aerial vehicle self-adaptive monitoring device and method, relates to the technical field of intelligent forest monitoring, and is composed of a hardware system and a software system. The hardware system comprises an unmanned aerial vehicle flight platform, a high-resolution visible light zoom camera, an edge computing unit, a wireless image transmission and a Beidou high-precision positioning module. The software system runs on the edge computing unit. An image enhancement module performs defogging and defuzzing processing on the video. An abnormal feature preliminary screening module uses a lightweight model to detect forest health abnormal areas in real time. A self-adaptive flight decision module automatically calculates the deviation and generates hovering, zooming or close-in flight instructions when a suspected abnormality is found. A health fine diagnosis module accurately determines the disease type and degradation level based on high-definition images. A result feedback module feeds back the real-time monitoring information with geographic coordinates to the ground end. A bottom parallel computing framework is used for data scheduling, resource scheduling and information feedback among modules.

Description

Technical Field

[0001] This invention relates to the field of smart forestry monitoring technology, and more specifically, to an adaptive monitoring device and method for edge-intelligent forest degradation characteristics using unmanned aerial vehicles (UAVs). Background Technology

[0002] Monitoring the health status and degradation level of forest and shrub-grass ecosystems is a core task of ecological restoration and forestry resource management. With climate change, ecological degradation phenomena such as the collective decline of forest populations are becoming increasingly serious. Traditional monitoring methods struggle to capture the microscopic characteristics of early-stage degradation (such as canopy thinning and leaf chlorosis), often only detecting these issues after irreversible structural degradation has occurred in the ecosystem. Furthermore, traditional monitoring methods rely heavily on manual ground inspections, involving extensive manual sample collection and harvesting, which damages the native ecosystem. This approach is not only labor-intensive and inefficient but also difficult to penetrate deep into complex terrains and dense forests, resulting in significant blind spots. While satellite remote sensing can cover large areas, its spatial and temporal resolution is limited, making it difficult to detect early, small-scale disease characteristics (such as individual tree discoloration and early insect webs). These symptoms are often only detected after large-scale outbreaks, exhibiting a significant time lag.

[0003] In recent years, unmanned aerial vehicle (UAV) technology has been widely used in forestry monitoring due to its high mobility and resolution. However, current UAV forestry monitoring technology still has significant limitations, mainly in the following two aspects:

[0004] First, it heavily relies on post-processing and network communication, lacking real-time capability. Existing drone monitoring systems mostly employ an offline mode of "flying and shooting first, then processing in the office," or rely on 4G / 5G networks to transmit images to cloud servers, where AI recognition algorithms are then used for identification. However, real-world environments are often remote, with poor transportation and limited network connectivity. The method of collecting data first and then transmitting it to a cloud server for interpretation is often impractical and fails to effectively guide fieldwork. Furthermore, manual data entry after collection is inefficient, often requiring several days to obtain the desired results. For example, current crop disease monitoring systems use drones to collect high-definition images, which are then uploaded to a cloud server via 4G / 5G networks. Deep learning models on the server analyze the images and return the disease location. However, this solution heavily relies on high-quality network connections. Poor signal coverage in remote areas leads to data transmission interruptions or delays, preventing real-time feedback from the field.

[0005] Secondly, the lack of adaptive and collaborative "perception-decision" capabilities makes it difficult to balance monitoring accuracy and efficiency. Existing drone inspections mostly employ an "open-loop" flight mode with preset routes. If the flight altitude is too high, the camera resolution is insufficient to identify minute pest and disease characteristics (such as the reddish-brown needles in the early stages of pine wilt disease); if the flight altitude is too low, although details can be seen, the coverage area of ​​a single operation is extremely limited, resulting in low efficiency. While existing edge computing devices have achieved real-time detection in some agricultural scenarios (such as wheat ear counting and weed identification), they are mostly limited to "seeing" and do not deeply couple the visual recognition results with the drone's flight control system. In other words, drones cannot automatically hover, zoom, or descend to make a diagnosis when they detect suspected anomalies, unlike human experts. This rigid "seeing but not seeing" flight mode leads to a large number of missed or false detections of diseases.

[0006] Therefore, there is an urgent need to develop an edge-intelligence-based adaptive monitoring device for forest health and degradation characteristics that can operate independently in a network-free environment and has the ability to "fly, calculate, and make decisions while operating" in order to solve the technical bottlenecks of poor timeliness and difficulty in balancing accuracy and efficiency in current forest pest and disease monitoring. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides an edge-intelligent UAV adaptive monitoring device and method for forest degradation characteristics. From a hardware perspective, this invention integrates a high-precision, high-resolution visible light zoom camera and an embedded edge computing module. From a software perspective, it lightweights the forest disease feature extraction algorithm and flight control strategy, researches parallel computing framework technology based on ARM architecture, and adapts it to the edge computing end. By moving data acquisition and processing to the monitoring device, the timeliness of data processing is significantly improved, enabling automated and real-time detection and analysis of forest health information such as "whether it is abnormal + abnormal information". The specific technical solution is as follows:

[0008] An edge-intelligent unmanned aerial vehicle (UAV) adaptive monitoring device for forest degradation characteristics is disclosed. The device consists of a hardware system and a software system. The hardware system includes a UAV flight platform, a high-resolution visible light zoom camera, an edge computing unit, a wireless image transmission device, a BeiDou high-precision positioning module, and auxiliary equipment. The software system is deployed and runs on the edge computing unit and includes an image enhancement module, an anomaly feature screening module, an adaptive flight decision module, a health fine diagnosis module, a result feedback module, and an underlying parallel computing framework.

[0009] The system comprises a drone flight platform serving as the carrier and actuator of the device; a high-resolution visible light zoom camera as the data acquisition device carried by the drone; a wireless image transmission device for transmitting video streams and early warning data; an image enhancement module for dehazing, deblurring, and extracting regions of interest from the video stream; an anomaly feature screening module using a lightweight neural network to perform a global scan of the preprocessed video stream, identifying suspected anomalies on the forest surface in real time and outputting the coordinates and confidence levels of the anomaly areas; an adaptive flight decision module calculating the target deviation based on the anomaly location output by the screening module, generating control commands for hovering, zooming, descent, or circling flight, and sending them to the drone's flight control system; a health fine diagnosis module receiving the high-definition local images acquired after adaptive adjustment, using a high-precision classification model to determine the specific types and levels of pests and diseases, and assessing the degradation status of individual trees based on the level of damage; and a result feedback module transmitting the real-time identification and analysis results to the ground terminal device via the wireless image transmission device. An underlying parallel computing framework is used for unified scheduling and feedback of data resources, computing resources, and storage resources for each task process, including data preprocessing, anomaly screening, flight decision-making, fine diagnosis, and result feedback.

[0010] An adaptive monitoring method for edge-intelligent forest degradation characteristics using unmanned aerial vehicles (UAVs) includes:

[0011] 1) Initialization: The equipment is powered on and performs a self-test. After the self-test is passed, the abnormal feature screening module is loaded into the memory of the edge computing unit, and the UAV enters the wide-area cruise mode.

[0012] 2) Data Acquisition: The drone flies along a preset route, and a high-resolution visible light zoom camera captures the video stream in real time in wide-angle mode;

[0013] 3) Data flows into the underlying parallel computing framework: The video stream enters the underlying parallel computing framework, where it is decoded and preprocessed by the GPU task pipeline;

[0014] 4) Initial anomaly screening: The preprocessed image frames are fed into the MobileNetV3-YOLOv8n model for inference, and the output includes a list of initial screening results containing [suspected anomaly, confidence level, bounding box].

[0015] 5) Adaptive decision-making: The CPU task pipeline obtains the metadata list from the GPU, judges each target in the list, and if a suspected disease is found and the confidence level is met, the adaptive flight decision module takes over the flight control, instructs the drone to hover and controls the camera to zoom in to obtain target details.

[0016] 6) Detailed diagnosis: Based on the acquired high-resolution detailed data, reason about the detailed images, output the diagnosed disease type and degradation level, and instruct the UAV to resume its original flight path.

[0017] 7) Result generation and streaming: The CPU encodes the processed structured results into data and pushes it to the ground receiving device in real time through the wireless image transmission device;

[0018] 8) Ground-based display: The ground receiving equipment connects to the wireless image transmission equipment, receives the structured result encoded data, and displays it visually.

[0019] The present invention has the following beneficial effects:

[0020] This invention proposes an integrated proactive closed-loop monitoring architecture encompassing perception, decision-making, and diagnosis. Unlike the passive mode of traditional drones that rely on "blindly shooting and then analyzing," this invention integrates an adaptive flight decision module at the edge, giving the drone the ability to "think like an expert." It can automatically approach and observe suspected diseases the moment they are detected, resolving the fundamental contradiction in traditional solutions where "flying too high makes it difficult to see disease details, while flying too low results in low coverage efficiency," significantly improving the flexibility and effectiveness of monitoring. By integrating a lightweight AI model and vegetation anomaly analysis algorithm at the data acquisition front end, it achieves real-time end-side processing from raw pixels to structured information, eliminating the dependence on networks and backend servers inherent in traditional solutions.

[0021] This invention proposes a "coarse screening-fine inspection" cascaded inference framework for ARM heterogeneous computing platforms. Addressing the limitation of edge computing power, this invention employs a lightweight model for wide-area initial screening, only invoking a high-precision model for fine diagnosis when anomalies are detected. This hierarchical processing mechanism efficiently balances real-time performance and accuracy, ensuring both cruising speed and accurate identification of early, minor lesions.

[0022] This invention boasts extremely high timeliness and disaster early warning capabilities: compared to cloud-based solutions, this invention, through end-to-end edge processing, reduces information acquisition latency from minutes or even hours to sub-second levels. Ground personnel can obtain "forest epidemic heat maps" with precise coordinates in real time and immediately conduct on-site verification and handling of high-risk areas, greatly improving the response speed of forest pest and disease control and controlling disasters in their nascent stage.

[0023] This invention enables the simultaneous interpretation of multi-dimensional health information: compared to simple applications that can only identify species, this invention is the first to achieve a leap from "species identification" to "health status quantification" at the edge. By combining high-precision visible light imagery with BeiDou positioning, the device can output structured data containing disease types, severity of damage, affected area, and precise geographical location, providing detailed data support for forest ecosystem health assessment and precision pesticide application.

[0024] This invention exhibits excellent environmental adaptability and enables low-cost deployment. It does not rely on any network connection and can operate in any remote area. Furthermore, based on a low-power ARM architecture and lightweight algorithms, the overall power consumption is low, making the device compatible with various small and medium-sized UAV platforms, easy to deploy and promote, and providing an economical and feasible technical means for the routine, grid-based monitoring of biodiversity. Attached Figure Description

[0025] Figure 1 This diagram shows the components of an edge intelligence-based adaptive monitoring device for forest health and degradation characteristics using unmanned aerial vehicles (UAVs).

[0026] Figure 2 This is a diagram illustrating the hardware system connections.

[0027] Figure 3 A data flow diagram for a software system;

[0028] Figure 4 This is a flowchart of the workflow of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0030] This invention proposes an adaptive monitoring device for edge-intelligent forest degradation characteristics using unmanned aerial vehicles (UAVs). The device consists of a hardware system and a software system. See [link to relevant documentation]. Figure 1 The hardware system includes an unmanned aerial vehicle (UAV) flight platform, a high-resolution visible light zoom camera, an edge computing unit, a wireless image transmission device, a BeiDou high-precision positioning module, and auxiliary equipment including a power supply. The software system runs on the edge computing unit and includes an image enhancement module, an anomaly feature screening module, an adaptive flight decision module, a health fine diagnosis module, a result feedback module, and a low-level parallel computing framework that provides computing power scheduling and task management for upper-layer applications.

[0031] The UAV flight platform, serving as the carrier and actuator of the device, includes the fuselage, power system, and flight control system. The flight control system connects to the edge computing unit via serial port or bus, receiving and executing yaw, pitch, hover, and zoom commands from the adaptive flight decision module.

[0032] A high-resolution visible light zoom camera serves as the data acquisition device, operating in red, green, and blue bands, supporting a resolution of at least 1920×1080 pixels, and equipped with a continuous optical zoom lens. In a preferred embodiment, the camera is a global shutter CMOS sensor to reduce image distortion and motion blur caused by the UAV's high-speed flight. The camera connects to the flight control system via a GIFbal interface, supporting PWM or serial port control of the zoom magnification.

[0033] Edge computing units serve as computing and storage centers, including computing devices such as GPUs or NPUs, with a computing power of at least 10 TOPS and a low-power ARM CPU architecture. For example, the Jetson Orin Nano development kit can be used, whose built-in GPU can provide up to 20 TOPS (INT8) of AI computing power, meeting the real-time inference requirements of lightweight models such as MobileNetV3-YOLOv8n, while its multi-core ARM CPU handles other non-intensive computing tasks.

[0034] The wireless image transmission equipment is responsible for transmitting the real-time recognition results from the drone to the ground terminal, which can then access the display device for viewing at any time.

[0035] The Beidou high-precision positioning module supports RTK differential positioning technology to obtain three-dimensional geographic coordinates with centimeter-level accuracy, ensuring the accuracy of disease location marking.

[0036] Image enhancement module: This module is used to eliminate fog and motion blur common in forest areas, and improve the signal-to-noise ratio of the video stream.

[0037] To achieve low-latency enhancement at the edge, this embodiment employs an improved algorithm based on dark channel prior. Assuming the original input image is I(x), the dehazed radiance map is J(x), the transmittance is t(x), and the global atmospheric illumination is A, the enhancement model is expressed as:

[0038] ;

[0039] In the formula, J(x) is the pixel value of the output image after dehazing; I(x) is the pixel value of the original foggy image captured by the camera; x is the pixel coordinate in the image; A is the global atmospheric illumination value, usually taken as the pixel value of the brightest pixel in the dark channel corresponding to the original image; t(x) is the transmittance of the medium, which characterizes the ability of light to penetrate fog. The lower limit of transmittance is set (0.1 in this embodiment). To further improve the contrast of anomalous features (such as withered yellow leaves), this module also introduces adaptive histogram equalization, whose pixel remapping function... Defined as:

[0040] ;

[0041] In the formula, This is the grayscale transformation function, i.e., the grayscale value after transformation; Let k be the gray level value of the input image; k is the index of the current gray level, ranging from 0 to L-1; L is the total number of gray levels in the image. Let be the probability density function of the j-th gray level in the histogram, that is, the proportion of pixels at that gray level to the total number of pixels; this formula enhances the texture details of suspected lesion areas in the image.

[0042] Anomaly Feature Preliminary Screening Module: This module, based on the lightweight MobileNetV3-YOLOv8n framework, performs a global scan of the wide-angle cruise video stream to quickly identify suspected anomaly areas. The specific process is as follows:

[0043] (a) Model lightweight optimization:

[0044] Operator fusion optimization was performed for the ARM architecture, transferring the feature representation of larger models such as YOLOv8l to a lightweight model to improve recognition accuracy; TensorRT was used to perform INT8 quantization on the trained model, dynamically converting some weights and activations from FP32 to INT8 to reduce memory usage and improve inference speed; continuous operators such as convolution, normalization, and activation were fused to reduce the number of kernel calls and memory accesses, improving running efficiency on edge GPUs.

[0045] (b) Network structure optimization:

[0046] Backbone Network: The backbone network replaces CSPDarknet with MobileNetV3, leveraging its depthwise separable convolutions and lightweight activation function h-swish to significantly reduce the number of parameters and computational cost; while retaining the C2f module design of YOLOv8n to enhance multi-level feature extraction capabilities. To address the issue of missed detection of small targets against a forest background, this embodiment introduces a Wise-IoU (WIoU) mechanism into the loss function to focus on difficult samples. Total Loss Function Defined as:

[0047] ;

[0048] In the formula, This represents the total loss value during model training. The bounding box regression loss is used to measure the positional deviation between the predicted box and the ground truth box. This is the binary cross-entropy loss, used to measure the accuracy of category classification; This is a distributed focus loss used to optimize the fine-grained localization of the bounding box; , , These are the weighting coefficients corresponding to the three losses mentioned above. The dynamic non-monotonic focusing bounding box loss is calculated using the following formula:

[0049] ;

[0050] In the formula, This is the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. and These are the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. These are the width and height of the smallest bounding box containing the predicted box and the ground truth box, respectively; The superscript indicates that this term is used as a constant or normalization baseline and does not participate in gradient backpropagation to ensure the scale invariance of the distance metric. Through this loss function, the model can more accurately capture subtle early color-changing features.

[0051] (c) Model deployment and inference process:

[0052] The optimized MobileNetV3-YOLOv8n model is loaded onto an edge GPU to acquire video stream images frame by frame, and batch reading and parallel inference are performed on N consecutive frames.

[0053] (d) Anomaly detection and confidence output:

[0054] Based on the inference results, the detected abnormal regions and their corresponding confidence levels are output.

[0055] Adaptive Flight Decision Module: This module is the control center for "active monitoring". When the anomaly feature screening module detects a target confidence level > 0.6, this module takes over the flight control, calculates the target deviation, and generates control commands.

[0056] (a) Target deviation calculation:

[0057] Let the coordinates of the image center be... The center coordinates of the detected suspected abnormal area are: Then the pixel deviation e(t) is:

[0058] ;

[0059] In the formula, e(t) is the Euclidean distance (pixel value) of the target's deviation from the image center at time t; , respectively, represent the x and y coordinates of the abnormal target detected at time t in the image coordinate system; These are the x and y coordinates of the center point of the image resolution, respectively.

[0060] (b) PID flight control law:

[0061] To control the drone to hover and align with the target, a visual feedback-based discrete PID control algorithm is used to generate yaw rate commands. :

[0062] ;

[0063] In the formula, The yaw rate control command received by the UAV flight control system at time t; , and These are the proportional coefficient, integral coefficient, and differential coefficient, respectively; e(t) is the Euclidean distance (pixel value) of the target's deviation from the image center at time t; t is the time variable; This is the cumulative integral term of the deviation over time, used to eliminate steady-state error; This is the differential term (rate of change) of the deviation with respect to time.

[0064] (c) Adaptive zoom strategy:

[0065] To see the details of abnormal areas more clearly, the module determines the current pixel proportion of the target. Compared with the preset optimal diagnostic ratio Calculate the required optical zoom magnification :

[0066] ;

[0067] In the formula, The calculated target optical zoom factor; This represents the camera's current optical zoom level.

[0068] The command is sent to the zoom camera via serial port to "zoom in" on the target of the disease.

[0069] Health Precision Diagnosis Module: This module is activated after the drone hovers and zooms to acquire high-resolution local images. It utilizes the DeepLabV3+ semantic segmentation network to perform pixel-level analysis of the affected areas. The specific workflow of this module is as follows:

[0070] (a) Quantification and grading of the degree of degradation:

[0071] Let the total number of pixels in the tree canopy projection be... The semantic segmentation model outputs the number of pixels for "slightly discolored areas (such as chlorosis in leaves)". The number of pixels for "severely dead areas (such as dead branches, fallen leaves, and exposed xylem)" is [number missing]. The formula for calculating VCDI is:

[0072] ;

[0073] in, and This is a severity weighting coefficient. In a preferred embodiment, it is set with reference to the standards for the occurrence and severity of forest pests. (Indicating a decline in physiological function) (Characteristic of structural necrosis).

[0074] (b) Grading determination:

[0075] Based on the calculated VCDI value, the health / degradation status of individual trees is divided into four levels:

[0076] Healthy (no degenerative characteristics): VCDI < 5%;

[0077] Mild degradation (early stress): 5%≤VCDI<25%, indicating that the trees are under drought or early pest and disease stress, and the canopy shows a small amount of discoloration;

[0078] Moderate degradation (structural damage): 25%≤VCDI<60%, characterized by obvious canopy dieback or partial canopy death, and a significant decline in growth vigor;

[0079] Severe degradation (dying / dead standing trees): VCDI≥60%, indicating that the main crown of the tree is dead or the whole tree is dead, losing its ecological function.

[0080] Result feedback module: pushes the recognition results via wireless image transmission.

[0081] Underlying parallel computing framework: This framework is designed for the specific task flow of this equipment, and its core idea is to create a heterogeneous computing task pipeline.

[0082] (a) Establish a real-time identification workflow model: The video stream collected by the sensor serves as the data source for workflow input.

[0083] (b) GPU task pipeline: The framework offloads computationally intensive tasks such as video frame decoding, image enhancement, and abnormal region recognition model inference to the GPU for execution.

[0084] (c) CPU task pipeline: The framework retrieves the results of the anomaly detection model inference from the GPU memory and then executes logic-intensive and serial tasks on the CPU, such as drone control, data packaging and result streaming.

[0085] (d) Synchronization and scheduling: The thread synchronization mechanism ensures the consistency of data transfer between the GPU and CPU, thereby achieving parallelization and pipelined processing of the entire process and maximizing the use of the heterogeneous computing resources of the ARM SoC.

[0086] The interaction relationships between hardware modules are shown in the appendix. Figure 2With an edge computing unit at its core, it connects a high-precision, high-resolution visible light zoom camera, a BeiDou high-precision positioning module, wireless image transmission equipment, and a drone flight platform, and is powered by a mobile power supply.

[0087] For details on the interaction relationships between software modules, please refer to the appendix. Figure 3 The underlying parallel computing framework serves as the central control unit for equipment operation. It interfaces with the image enhancement module, anomaly feature screening module, adaptive flight decision-making module, health detail diagnosis module, and result feedback module, enabling unified management of task allocation, data transmission, computing and storage resource scheduling, and feedback information. The image enhancement module provides video frame data to the anomaly feature screening module, which infers and identifies the video frames, outputting structured results including whether anomalies exist, the confidence level, and the location of the anomaly. The adaptive flight decision-making module receives these structured results and controls the UAV's flight and zoom lens according to the confidence level. The health detail diagnosis module receives high-resolution camera data and analyzes anomaly areas to obtain parameters such as pest type and degradation degree. The result feedback module receives the outputs from the anomaly feature screening module and the health detail diagnosis module and transmits the real-time monitoring information to ground equipment via wireless image transmission.

[0088] The workflow diagram of an edge intelligence-based adaptive monitoring device for forest health and degradation characteristics using unmanned aerial vehicles (UAVs) is shown below. Figure 4 As shown.

[0089] 1) Initialization: Power on the equipment, perform system self-check, and load the initial screening model of abnormal features into the edge computing unit.

[0090] 2) Wide-area cruise data acquisition: The UAV flies along a preset route, and the high-resolution visible light zoom camera is in wide-angle mode to capture video streams in real time.

[0091] 3) Data flows into the underlying parallel computing framework: The video stream enters the underlying parallel computing framework, where it is decoded and preprocessed by the GPU task pipeline.

[0092] 4) Initial Anomaly Screening: After image enhancement, the video stream is sent to the initial anomaly feature screening module. If no anomalies are found, cruise continues. If a suspected anomaly region is detected (confidence > threshold), an interrupt signal is triggered.

[0093] 5) Active attitude adjustment: The adaptive flight decision module calculates the PID control quantity, directs the UAV to hover and controls the camera's optical zoom until the target clarity meets the diagnostic requirements.

[0094] 6) Detailed Health Diagnosis: Obtain high-definition detailed images, calculate the VCDI index, and determine the level of degradation.

[0095] 7) Result generation and streaming: The CPU encodes and packages the processed structured results (such as [Elm black spot disease, mild degradation (early stress), Beidou coordinates (E116.3, N41.9)]) and pushes them to the ground terminal in real time through wireless image transmission equipment.

[0096] 8) Ground-based display: Ground receiving devices (such as tablets or laptops) are connected to wireless image transmission devices to receive and visualize the identified structured data, allowing users to view the identification results in real time.

[0097] There are multiple alternative solutions, all of which should fall within the protection scope of this invention:

[0098] 1. Sensor alternatives: High-resolution visible light zoom cameras can replace or add multispectral cameras or lightweight hyperspectral cameras to help identify species or assess health status by analyzing vegetation indices.

[0099] 2. Alternatives to the anomaly feature screening model: The anomaly feature screening model is not limited to MobileNetV3-YOLOv8n, but can also be other lightweight object detection models suitable for edge deployment, such as EfficientDet, NanoDet, etc.

[0100] 3. Edge computing platform alternatives: ARM architecture devices are not limited to the Jetson series; other brands of edge computing chips can also be used. In specific cases, FPGAs (Field-Programmable Gate Arrays) with integrated vision processing units can also be used.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0106] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An adaptive monitoring device for edge-intelligent forest degradation characteristics using unmanned aerial vehicles (UAVs), characterized in that, The device comprises a hardware system and a software system. The hardware system includes a UAV flight platform, a high-resolution visible light zoom camera, an edge computing unit, wireless image transmission equipment, and a BeiDou high-precision positioning module. The software system is deployed and runs on the edge computing unit and includes an image enhancement module, an anomaly feature initial screening module, an adaptive flight decision module, a health fine-tuning module, a result feedback module, and an underlying parallel computing framework. The UAV flight platform serves as the carrier and actuator of the device. A high-resolution visible light zoom camera is the data acquisition device carried by the UAV, used to acquire wide-angle or telephoto images of the trees. A wireless image transmission device is the data transmission link, used for the transmission of video streams and monitoring and early warning data. A BeiDou high-precision positioning module provides centimeter-level spatial geographic coordinates for the monitoring device. An image enhancement module performs dehazing and deblurring on the video stream to improve the image signal-to-noise ratio. An anomaly feature screening module uses a lightweight neural network model to perform real-time detection on the video stream output by the image enhancement module, identifying suspected health abnormalities or deterioration features on the tree surface, including canopy discoloration, dead branches, and exposed branches, and outputs the location and confidence level of this area in the image coordinate system. An adaptive flight decision module... The module receives abnormal area information output by the anomaly feature screening module, calculates target deviation, generates hovering, zooming, or close-range flight control commands, and sends them to the UAV's flight control system to obtain high-resolution local images of suspected disease areas. The health fine diagnosis module analyzes the acquired high-resolution local images, uses a high-precision classification model to determine the specific pest and disease types and damage levels, and assesses the degradation status of trees based on the damage levels. The result feedback module transmits the real-time judgment and analysis results to the ground terminal equipment through the wireless image transmission device. The underlying parallel computing framework is used to uniformly schedule and provide feedback on data resources, computing resources, and storage resources for each task process, including image enhancement, anomaly feature screening, adaptive flight decision-making, health fine diagnosis, and result feedback.

2. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, The high-resolution visible light zoom camera covers the red, green, and blue bands, and has an output resolution of no less than 1920×1080 pixels. It also supports continuous optical zoom with a zoom factor of no less than 10x.

3. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, Edge computing units include GPUs or NPUs, and the FP16 computing power is no less than 10 TOPS.

4. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, The abnormal feature screening module acquires video stream images frame by frame, sequentially reads N images into the GPU and uses the MobileNetV3-YOLOv8n model for parallel inference, outputting structured information of suspected disease areas in each image, including bounding boxes, category labels including discoloration, necrosis, and initial screening confidence.

5. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, The adaptive flight decision module is used to implement closed-loop control of "detection-adjustment". When the anomaly feature screening module detects that the confidence of the target exceeds the preset threshold, this module calculates the pixel deviation value between the target center point and the current image center point. If the deviation value is greater than the set range, yaw and pitch adjustment commands are generated to center the target. If the target's proportion in the image is less than the set threshold, zoom or descent commands are generated until the target texture clarity meets the requirements of fine diagnosis.

6. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, The specific workflow of the health precision diagnosis module is as follows: (a) High-precision image acquisition: Receive high-resolution local images acquired after adaptive flight adjustment; (b) Feature segmentation and classification: The image is input into a high-precision semantic segmentation model to perform pixel-level segmentation of the forest degradation feature region, distinguishing physiological degradation features including leaf chlorosis and discoloration and structural degradation features including dieback and exposed xylem; (c) Quantification of degradation degree: The Canopy Visual Degradation Index (VCDI) is used as a quantitative indicator to characterize the degradation degree of individual trees; Considering the differences in the threat to tree survival posed by different disease characteristics, a weighted cumulative method was used for calculation, and the tree status was divided into four levels according to the VCDI value: healthy, slightly degraded, moderately degraded, and severely degraded. (d) Geocoding: Read the coordinate data of the Beidou high-precision positioning module at this time, bind the disease type, damage level and geographic coordinates to generate a tree health record.

7. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, The underlying parallel computing framework interfaces with the image enhancement module, the anomaly feature screening module, the adaptive flight decision module, the health fine diagnosis module, and the result feedback module, respectively, to realize the scheduling and feedback of data resources, computing resources, and storage resources in each task. An interrupt-resumption scheduling mechanism is introduced: in the wide-area cruise state, resources are prioritized for the anomaly feature screening module; when the adaptive flight command is triggered, the underlying framework suspends the screening task, dynamically allocates the main computing resources to the health fine diagnosis module, releases the resources after the diagnosis is completed, and resumes the screening task.

8. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 4, characterized in that, The backbone network of the abnormal feature screening module adopts a lightweight structure to reduce the number of parameters and computational load. The neck network is used for multi-scale feature fusion. The detection head is optimized for early discolored leaves and dead tree crowns in a forest background. The output includes two categories: suspected abnormalities and normal background, in order to reduce the false negative rate.

9. The edge intelligent forest degradation characteristic UAV adaptive monitoring device according to claim 1, characterized in that, High-resolution visible light zoom cameras can also be used for multispectral or hyperspectral cameras.

10. An adaptive monitoring method for edge intelligent forest degradation characteristics using a UAV according to any one of claims 1-9, characterized in that, include: 1) Initialization: The equipment is powered on and performs a self-test. After the self-test is passed, the edge computing unit loads the abnormal feature screening module, the UAV takes off and enters the wide-area cruise mode. 2) Wide-area initial screening: The UAV flies along the preset route, the high-resolution visible light zoom camera is in wide-angle mode, and the anomaly feature initial screening module performs a global scan of the real-time video stream; 3) Adaptive triggering: When the initial screening module detects a suspected abnormal area and the confidence level meets the standard, an interrupt signal is triggered to enter the fine monitoring mode; 4) Flight attitude adjustment: The adaptive flight decision module takes over the flight control, controls the camera to zoom or the drone to hover, locks onto and magnifies suspected defective targets; 5) Detailed diagnosis: Acquire high-definition detailed images of the target, and the health detailed diagnosis module performs disease classification and degradation level assessment; 6) Result generation and transmission: The result transmission module packages the diagnostic results, including disease type, degradation level, photos and BeiDou coordinates, and pushes them to the ground receiving device in real time through wireless image transmission equipment; 7) Mission recovery: After the diagnosis is completed, control the camera to return to wide-angle mode, the drone returns to the original flight path and speed, and continues to perform wide-area patrol missions until the flight path ends.

Images