Edge intelligent unmanned aerial vehicle pest and disease damage real-time detection system

Abstract

The invention discloses an edge intelligent unmanned aerial vehicle disease and insect pest real-time detection system, and belongs to the technical field of intelligent agriculture, and the system comprises an edge end lightweight detection module, a self-adaptive resource scheduling module, a multi-task collaborative optimization module and an edge-cloud closed-loop learning module. Real-time reasoning of the unmanned aerial vehicle end is realized; a reasoning strategy is dynamically adjusted through self-adaptive scheduling according to the battery remaining power and the task emergency degree, and the endurance time is prolonged; detection, classification and density estimation are executed at the same time through multi-task collaborative optimization, and the calculation amount is reduced; and continuous optimization of the model is realized through edge-cloud closed-loop learning, the accuracy is improved, and efficient and reliable technical support is provided for intelligent agriculture.

Description

Technical Field

[0001] This invention relates to the fields of smart agriculture and edge computing technology, specifically to an edge intelligent drone real-time pest and disease detection system. Background Technology

[0002] Crop diseases and pests are significant factors affecting food security and agricultural production. Traditional pest and disease control relies mainly on manual inspections and experience-based judgment, which suffers from low efficiency, limited coverage, and poor timeliness. In recent years, the rapid development of drone technology and deep learning algorithms has provided new technical means for intelligent pest and disease detection.

[0003] CN120297527A discloses a method and system for delineating forest pest and disease control areas. This system establishes a database by collecting pest and disease data and environmental data for the control areas, trains a pest and disease risk prediction model using historical data, intelligently delineates target areas using GIS technology and deep learning image segmentation algorithms, and plans drone spraying routes based on path optimization algorithms. This system can improve pest and disease control efficiency and reduce repeated spraying.

[0004] However, this existing technology has the following shortcomings: First, the system mainly relies on the cloud for data processing and model inference, requiring real-time transmission of collected image data to the cloud server. In field operation environments with unstable network signals or limited bandwidth, data transmission delays can severely affect the real-time performance of detection, failing to meet the instant decision-making needs during UAV flight. Second, the deep learning models used in this system typically have a large number of parameters and high computational complexity, making them difficult to deploy directly on resource-constrained edge devices such as UAVs, thus limiting the system's practicality. Third, the system lacks an adaptive optimization mechanism tailored to the resource-constrained characteristics of UAVs. It cannot flexibly adjust detection strategies when battery power is low or computational load is too high, potentially leading to mission interruption or decreased detection quality. Fourth, the system's model updates rely on offline retraining, making it unable to utilize real-time data accumulated during operations for online learning and model optimization, thus struggling to adapt to the dynamic changes in pest and disease characteristics across different regions and time periods.

[0005] With the development of edge computing technology, it has become possible to deploy deep learning models on edge devices for local inference. However, due to the limited computing power, storage space, and battery capacity of edge devices such as drones, directly deploying traditional deep learning models faces significant challenges. Model compression techniques such as knowledge distillation, network pruning, and quantization can effectively reduce model size and computational complexity, but how to achieve extreme lightweighting while maintaining high detection accuracy, and how to dynamically adjust inference strategies according to the real-time operating environment to balance performance and energy consumption, remain key technical problems that urgently need to be solved.

[0006] Furthermore, pest and disease control requires not only accurate detection of pests and diseases but also identification of specific categories and estimation of density distribution to guide precise pesticide application. Existing systems often separate these tasks, leading to wasted computational resources and insufficient information utilization. Designing an efficient multi-task learning framework that allows multiple related tasks to share underlying feature representations and improves overall performance through inter-task collaborative optimization is a crucial approach to improving edge detection efficiency.

[0007] Meanwhile, edge devices have limited computing power and difficulty in updating models. The key to the long-term effective operation of intelligent pest and disease control systems lies in how to build an edge-cloud collaborative architecture, utilize the powerful computing resources of the cloud for model training and optimization, and promptly distribute the updated model parameters to edge devices to achieve continuous model evolution and performance improvement.

[0008] Therefore, it is necessary to provide a new edge intelligent drone real-time pest and disease detection system that can achieve efficient and accurate real-time pest and disease detection on resource-constrained edge devices such as drones, and has adaptive resource scheduling capabilities and online learning mechanisms to meet the needs of actual agricultural production. Summary of the Invention

[0009] The purpose of this invention is to provide an edge-intelligent drone real-time pest and disease detection system. By combining knowledge distillation technology with edge computing, it enables real-time inference of a lightweight deep learning model on the drone. It also dynamically balances detection performance and energy consumption through an adaptive resource scheduling mechanism, improves detection efficiency through a multi-task collaborative optimization framework, and achieves continuous model optimization through an edge-cloud closed-loop learning mechanism. This solves the problems of poor real-time performance, high resource consumption, and weak adaptability in existing technologies.

[0010] To achieve the above objectives, the technical solution adopted by this invention is as follows: A real-time pest and disease detection system for edge-intelligent drones is provided, comprising a lightweight edge detection module, an adaptive resource scheduling module, a multi-task collaborative optimization module, and an edge-cloud closed-loop learning module. The lightweight edge detection module constructs a lightweight neural network based on knowledge distillation technology to achieve real-time inference detection of pests and diseases on the drone. The adaptive resource scheduling module dynamically adjusts the inference accuracy and execution frequency according to battery remaining capacity, task urgency, and computational load, balancing performance and energy consumption. The multi-task collaborative optimization module simultaneously performs detection, classification, and density estimation tasks by sharing a feature extraction layer and optimizes based on a joint loss function, improving computational efficiency and detection accuracy. The edge-cloud closed-loop learning module transmits edge detection data back to the cloud for incremental training and distributes updated model parameters to the edge, achieving continuous model optimization.

[0011] The beneficial effects of this invention are as follows: By using knowledge distillation technology, knowledge from a complex teacher model is transferred to a lightweight student model, compressing the model size to less than 10% of its original size while maintaining high detection accuracy. This enables real-time inference on a single-chip microcomputer-level edge device, achieving a detection speed of over 30 frames per second, meeting the real-time requirements of UAV flight. Through an adaptive resource scheduling mechanism, the system can dynamically adjust the inference strategy based on battery level, task urgency, and computational load. When battery power is sufficient, a high-precision mode is used to ensure detection quality; when battery power is insufficient, an energy-saving mode is switched to extend flight time; and during urgent tasks, priority is given to ensuring detection accuracy in critical areas, achieving an optimal balance between performance and energy consumption, extending UAV flight time by more than 25%. Through a multi-task collaborative optimization framework, detection, classification, and density estimation tasks share underlying feature representations. Compared to executing the three tasks independently, computational load is reduced by 40%, while collaborative optimization between tasks improves overall detection accuracy by 5%. Through an edge-cloud closed-loop learning mechanism, the system can continuously learn from operational data, and model performance improves over time. After three months of deployment, accuracy improved by 8%, effectively adapting to changes in pest and disease characteristics in different regions and at different times. Overall, this invention enables efficient, accurate, and adaptive real-time detection of pests and diseases on resource-constrained drone platforms, reducing pesticide waste by more than 70% and achieving a detection accuracy rate of 93%, thus providing reliable technical support for smart agriculture. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall architecture of the edge intelligent drone pest and disease real-time detection system of the present invention; Figure 2 This is a schematic diagram of the lightweight edge detection module of the present invention; Figure 3 This is a schematic diagram of the workflow of the adaptive resource scheduling module of the present invention; Figure 4 This is a schematic diagram of the network structure of the multi-task collaborative optimization module of the present invention. Detailed Implementation

[0013] Please refer to the attached document. Figures 1-4 The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It is understood that the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1The edge intelligent drone pest and disease real-time detection system provided by the present invention includes an edge lightweight detection module 1, an adaptive resource scheduling module 2, a multi-task collaborative optimization module 3, and an edge-cloud closed-loop learning module 4.

[0015] Reference Figure 2 The lightweight edge detection module 1 includes an image acquisition unit, a knowledge distillation and compression unit, and a real-time inference unit.

[0016] The image acquisition unit acquires crop images using a high-resolution camera mounted on the drone. In one embodiment of the invention, the camera is a 5-megapixel RGB camera with a 90° field of view, achieving a ground resolution of 0.5mm at a flight altitude of 5m. The image acquisition frequency is dynamically adjusted according to flight speed and detection requirements. The acquisition frequency is reduced during high-speed flight to decrease data processing pressure, while the flight speed is reduced and the acquisition frequency is increased in critical areas to obtain more detailed image information. The acquired raw images undergo preprocessing operations, including image denoising, brightness normalization, and resizing. Denoising employs a Gaussian filtering method with a 3×3 kernel size and a standard deviation of 1.0. Brightness normalization maps image pixel values ​​to the 0-1 range, ensuring a consistent brightness distribution under different lighting conditions. Resizing scales the raw image to 416×416 pixels to match the input requirements of the subsequent neural network.

[0017] The knowledge distillation compression unit is one of the core innovations of this invention, used to build lightweight neural networks suitable for deployment on edge devices. Traditional deep learning models such as ResNet-101 and YOLOv5, while offering high detection accuracy, suffer from large parameter counts and high computational complexity, making them difficult to run in real-time on edge devices such as drones. Knowledge distillation is a model compression technique that trains a small student network to mimic the behavior of a large teacher network, thereby significantly reducing model complexity while maintaining high accuracy.

[0018] In this embodiment of the invention, the knowledge distillation and compression unit employs a teacher-student network framework. The teacher network uses a pre-trained ResNet-101 as its backbone, combined with a feature pyramid network and a detection head, and is fully trained on a large-scale pest and disease image dataset. It has 45 million parameters and achieves an average precision (mAP) of 96.5% on the test set. The student network uses the lightweight MobileNetV3 as its backbone, with only 5.5 million parameters, approximately 12% of the teacher network's parameters.

[0019] The knowledge distillation training process employs a soft-label distillation method. Specifically, the teacher network and student networks reason about the same input image, and the class probability distribution output by the teacher network is processed by a temperature parameter. After softening, the labels are used as soft labels. The student network's output needs to fit both the true labels and the soft labels simultaneously. The training loss function consists of two parts: , in, For the total loss, The cross-entropy loss between the student's network output and the real label is used. The KL divergence loss is used to determine the relationship between the student network output and the teacher network soft labels. and These are the weighting coefficients for the two parts of the loss, respectively.

[0020] Soft tag loss The calculation formula is: , in, This is the output logits vector of the teacher network. The output logits vector of the student network. This is a temperature parameter used to control the smoothness of the probability distribution. Denotes KL divergence, For category indexing, It is an exponential function. In a preferred embodiment of the present invention, The value is 4.0. The value is 0.3. The value is set to 0.7. A higher temperature parameter makes the probability distribution of the teacher network output smoother, containing more information on the similarity between categories, which helps the student network learn richer knowledge.

[0021] Through the aforementioned knowledge distillation training, the student network achieved detection capabilities approaching those of the teacher network while maintaining a lightweight design. In pest and disease detection tasks, the student network achieved an average precision (mAP) of 93.2%, only 3.3 percentage points lower than the teacher network, but the model size was compressed from 180MB to 18MB, and the inference speed increased from 15 frames per second to 35 frames per second, meeting the real-time performance and resource constraints required by edge devices.

[0022] The real-time inference unit is responsible for calling a pre-trained lightweight student network to perform forward propagation inference on the preprocessed image. The inference process includes three stages: feature extraction, region proposal generation, and object detection. In the feature extraction stage, the MobileNetV3 backbone network performs multi-scale feature extraction on the input image, generating feature maps of different resolutions. In the region proposal generation stage, the feature pyramid network fuses features from different levels to generate candidate detection boxes. In the object detection stage, the detection head classifies and regresses the location of each candidate box, outputting the pest / disease category, location coordinates, and confidence information.

[0023] The inference results include the bounding box coordinates of the detected pests and diseases. ,in The coordinates of the bounding box center in the image. and These represent the width and height of the bounding box, respectively. The system also outputs the target's category label, such as rice planthopper, rice leaf roller, and rice blast, along with a corresponding confidence score ranging from 0 to 1, indicating the model's confidence in the detection result. Detection results with a confidence score higher than 0.5 are considered valid and output to subsequent modules for processing.

[0024] Reference Figure 3 The adaptive resource scheduling module 2 includes a status monitoring unit, an urgency assessment unit, and a dynamic scheduling unit.

[0025] The status monitoring unit monitors the drone's operational status in real time, including battery level, flight speed, and computing load. Battery level is obtained through the drone's battery management system and expressed as a percentage, ranging from 0% to 100%. Flight speed is measured via the GPS module and inertial measurement unit, in m / s. Computational load is obtained by monitoring the CPU and GPU utilization of the edge computing chip, ranging from 0% to 100%. This status information is updated every second, providing real-time data support for dynamic scheduling.

[0026] The urgency assessment unit determines the task urgency level based on the threat level of pests and diseases detected by the lightweight detection module 1 at the edge. The urgency assessment considers three factors: pest / disease type, density, and spread risk. Different types of pests and diseases cause varying degrees of damage to crops. For example, rice blast is a highly damaging type with an urgency weighting coefficient of 1.0, while a minor leaf spot disease has an urgency weighting coefficient of 0.3. Density is measured by the number of detected pests and diseases; the more targets detected per unit area, the higher the urgency. Spread risk is determined based on the spatial distribution characteristics of pests and diseases; if the detection results show a rapid spread of pests and diseases, the urgency level increases accordingly.

[0027] Urgency rating The calculation formula is: , in, This represents the hazard level coefficient corresponding to the type of pest or disease. The normalized density value, For the diffusion risk coefficient, , , The weights of the three factors are 0.5, 0.3, and 0.2, respectively, in a preferred embodiment of the invention. Urgency Score The value ranges from 0 to 1, and the urgency of the task is divided into three levels: low, medium, and high based on the rating. For low urgency, Medium urgency It is classified as a high level of urgency.

[0028] The dynamic scheduling unit is the core of the adaptive resource scheduling module 2. It is responsible for dynamically adjusting the inference accuracy level and execution frequency of the edge lightweight detection module 1 based on the information provided by the status monitoring unit and the urgency assessment unit, so as to achieve the optimal balance between performance and energy consumption.

[0029] The inference accuracy levels are divided into three modes: High Accuracy, Standard Accuracy, and Energy-Saving Accuracy. In High Accuracy mode, a complete lightweight student network is used for inference, with an input image resolution of 416×416 pixels, a detection threshold of 0.5, an average precision (mAP) of 93.2%, a single-frame inference time of 28ms, and a power consumption of 3.5W. In Standard Accuracy mode, inference speed is accelerated and power consumption is reduced by decreasing the number of network layers and lowering the input image resolution to 320×320 pixels, achieving an average precision (mAP) of 89.8%, a single-frame inference time of 18ms, and a power consumption of 2.2W. In Energy-Saving Accuracy mode, the network structure is further simplified, with the input image resolution reduced to 224×224 pixels, achieving an average precision (mAP) of 85.5%, a single-frame inference time of 12ms, and a power consumption of 1.5W.

[0030] Execution frequency refers to the number of inference operations per second, measured in Hz. High-frequency execution can detect pests and diseases promptly, but it increases computational load and energy consumption. Low-frequency execution reduces resource consumption, but may miss rapidly changing pest and disease conditions. In this invention, the execution frequency is dynamically adjusted based on flight speed and mission urgency; the faster the flight speed or the higher the mission urgency, the higher the execution frequency.

[0031] The dynamic scheduling unit makes scheduling decisions based on a multi-dimensional decision matrix. The decision matrix includes dimensions of battery remaining capacity, urgency, and load, with multiple threshold ranges set for each dimension. The battery remaining capacity dimension is divided into sufficient (… ),generally( ),insufficient( The urgency dimension is divided into three levels: low, medium, and high. The load dimension is divided into light load (…). ), mid-load ( ), heavy load ( Three intervals.

[0032] Different combinations of dimensions correspond to different scheduling strategies. For example, when battery power is sufficient, urgency is high, and load is light, the system adopts a high-precision mode with an execution frequency of 30Hz to ensure high-quality detection in critical areas. When battery power is insufficient, urgency is low, and load is heavy, the system switches to an energy-saving precision mode, reducing the execution frequency to 5Hz to prioritize drone flight time. When battery power is moderate, urgency is medium, and load is medium, the system adopts a standard precision mode with an execution frequency of 15Hz to balance detection performance and energy consumption.

[0033] To avoid system instability caused by frequent switching, the dynamic scheduling unit introduces a smoothing mechanism. When state parameters change, the system does not immediately switch modes, but waits for the parameters to remain within the new range for 3 consecutive seconds before performing the mode switch. Furthermore, a gradual transition is used during the switching process to avoid detection interruptions caused by sudden changes.

[0034] Through the aforementioned adaptive resource scheduling mechanism, the system can intelligently adjust its detection strategy based on real-time operating conditions, maximizing UAV endurance and operational efficiency while ensuring detection quality. Experimental results show that, compared to a fixed strategy, adaptive scheduling extends UAV endurance by 25%, increases the operational area by 30%, and maintains a detection accuracy rate of over 90%.

[0035] Reference Figure 4 The multi-task collaborative optimization module 3 includes a multi-branch inference unit, a joint loss calculation unit, and a drug administration parameter generation unit.

[0036] The multi-branch inference unit constructs a multi-task learning network architecture. Based on a shared feature extraction layer, it executes three tasks in parallel—pest detection, pest classification, and pest density estimation—through independent task branches. This design allows the three tasks to share the underlying image feature representation, avoiding redundant computation and significantly improving computational efficiency.

[0037] The shared feature extraction layer employs a lightweight MobileNetV3 network derived from knowledge distillation to extract features from the input image at multiple scales. The network outputs three feature maps at different resolutions, corresponding to high, medium, and low-level semantic information, respectively. The high-level feature maps are smaller in size but richer in semantic information, making them suitable for object classification; the low-level feature maps are larger in size but contain more detailed information, making them suitable for object localization.

[0038] The detection branch is responsible for determining whether pests or diseases exist in the image and providing their location information. The detection branch employs a feature pyramid network structure, fusing feature maps from different levels to generate multi-scale candidate detection boxes. For each candidate box, the detection branch outputs a binary classification probability, indicating whether pests or diseases exist at that location, along with the bounding box's coordinate correction value.

[0039] The classification branch is responsible for identifying the specific category of detected pests and diseases. Based on the candidate boxes output by the detection branch, the classification branch performs fine-grained classification on each region containing pests or diseases. The network structure of the classification branch contains multiple convolutional layers and fully connected layers, ultimately outputting a multi-class probability vector representing the probability that the target belongs to each pest or disease category. In embodiments of this invention, the system supports the identification of 20 common crop pests and diseases, including rice planthoppers, rice leaf rollers, rice blast, sheath blight, bacterial blight, and rice false smut.

[0040] The density estimation branch is responsible for estimating the density distribution of pests and diseases within the detection area. Density estimation is crucial for guiding precise pesticide application; high-density areas require increased pesticide dosage, while low-density areas can require reduced dosage, thus achieving efficient use of the pesticide solution. The density estimation branch uses a regression method to output a continuous density value for each detection area, representing the number of pests and diseases per unit area.

[0041] Density estimation is performed under supervised point labeling. During training, each pest or disease target is labeled, and the ground truth density is used to generate a density map using a Gaussian kernel function. During inference, the density estimation branch outputs the predicted density map. The total number of pests or diseases in the region is obtained by integrating the density map, and then the density value is calculated.

[0042] The joint loss calculation unit is responsible for calculating the loss values ​​of the three tasks and performing weighted fusion for parameter optimization during the model training phase. The design of the joint loss function takes into account the correlation and importance differences between tasks, and achieves collaborative optimization among tasks by dynamically adjusting the weights of the losses for each task.

[0043] Joint loss function The form is: , in, The loss for the detection task includes classification loss and bounding box regression loss. For the cross-entropy loss of the classification task, The mean squared error loss for the density estimation task. , , These are the weighting coefficients for the losses of the three tasks.

[0044] In this innovative embodiment of the invention, the weight coefficients are not fixed constants, but are dynamically adjusted according to the training status of the task. Specifically, the weights of the detection and classification tasks are set based on the model's prediction confidence for the current sample. When the model's prediction confidence for a sample is low, it indicates that the sample is a difficult sample, and the weight of the corresponding task needs to be increased to strengthen learning. The weight adjustment formula is: , , in, and Based on weights, and These represent the confidence scores for the detection task and the classification task, respectively. For adjustment coefficients. In a preferred embodiment of the present invention, , , .

[0045] The weight of the density estimation task is dynamically adjusted based on the degree of pest and disease aggregation. When the pest and disease density in the detection area is high, accurate density estimation is more important for guiding pesticide application; therefore, the weight of the density estimation task is increased. The weight adjustment formula is as follows: , in, Based on weights, This represents the density value for the current region. The maximum density value in the dataset. For adjustment coefficients. In a preferred embodiment of the present invention, , .

[0046] Through the aforementioned dynamic weight adjustment mechanism, the model can adaptively focus on difficult samples and high-density regions, improving overall detection performance. Experimental results show that, compared with fixed-weight multi-task learning, dynamic weight adjustment improves detection accuracy by 2.3% and reduces density estimation error by 15%.

[0047] The pesticide application parameter generation unit generates differentiated application parameters based on the type, density distribution, and location information of pests and diseases, providing guidance for precise pesticide application by drones. These parameters include pesticide type, spray concentration, spray volume, and spray speed.

[0048] The type of pesticide is determined based on the type of pest or disease. Different pests and diseases have different sensitivities to pesticides, requiring the selection of targeted pesticides. The system has a built-in pest-pesticide mapping table that records the optimal pesticide type for each pest or disease. For example, imidacloprid is used for rice planthoppers, tricyclazole for rice blast, and jinggangmycin for rice sheath blight.

[0049] The spraying concentration is determined based on the severity of the pest or disease. Severity is determined by a combination of factors, including density and spread risk; higher density and faster spread necessitate a higher spraying concentration. The formula for calculating the spraying concentration is: , in, For the target spray concentration, As the baseline concentration, The density value of the detection area. For reference density value, For the diffusion risk coefficient, and For adjustment coefficients. In a preferred embodiment of the invention, for rice planthoppers, mg / L, Units / square meter , .

[0050] The spraying volume is determined based on the area of ​​the monitored region and the density of pests and diseases. The larger the area and the higher the density, the greater the spraying volume should be. The formula for calculating the spraying volume is: , in, The spray volume is expressed in mL. The area to be tested is (unit: square meters). The baseline spraying rate is (unit: mL / m²). This is a density adjustment factor. In a preferred embodiment of the invention, mL / m² .

[0051] The spraying speed is determined based on the uniformity of the pest and disease distribution. If the pests and diseases are evenly distributed, a faster spraying speed can be used to improve operational efficiency; if the distribution is uneven, the spraying speed needs to be reduced to achieve precise application. The spraying speed range is from 2 m / s to 8 m / s.

[0052] The generated application parameters are sent to the drone's spraying control system in real time via a wireless communication module. The spraying control system adjusts the nozzle's on / off status, flow rate, and flight speed according to the parameters to achieve precise and efficient pest and disease control.

[0053] The edge-cloud closed-loop learning module 4 includes a data filtering and uploading unit, a cloud-based incremental training unit, and a parameter distribution and update unit.

[0054] The data filtering and uploading unit is responsible for selecting high-value samples from the large amount of data generated during edge detection and uploading them to the cloud platform for model training. Due to the massive amount of image data generated during drone operations, network bandwidth and cloud storage costs limit the uploading of all data. Therefore, an effective data filtering strategy needs to be designed to select the samples most valuable for model improvement.

[0055] This invention employs a data filtering strategy based on uncertainty sampling. For unfamiliar or difficult samples, the model's prediction results often have high uncertainty; these samples are most valuable for model training. Specifically, the data filtering upload unit calculates the prediction uncertainty of each detected sample and selects samples with uncertainty exceeding a set threshold as high-value samples.

[0056] Prediction uncertainty is measured by the variance of the results from multiple forward propagations. During inference, a slight random perturbation is added to the same input image, and multiple forward propagations (typically 5) are performed, resulting in multiple predictions. Uncertainty The calculation formula is: , in, Forward propagation count, For the first The predicted probability obtained from the first forward propagation. for The average of the predicted probabilities. Higher uncertainty indicates that the model's predictions for that sample are less stable, and therefore the sample is more valuable.

[0057] In a preferred embodiment of the invention, the uncertainty threshold is set to 0.15. When At that time, the sample was marked as a high-value sample and uploaded to the cloud platform along with its corresponding detection results and metadata (such as time, location, weather conditions, etc.). To avoid uploading too much data, the system also sets an upload frequency limit, with a maximum of 1,000 samples uploaded per task.

[0058] In addition to uncertainty sampling, the data filtering and uploading unit also considers sample diversity. If the uploaded samples are too similar, the gain for model training is limited. Therefore, among the samples that meet the uncertainty condition, representative samples are further selected through clustering. Specifically, K-means clustering is performed on the feature vectors of the candidate samples, and the sample closest to each cluster center is selected for uploading, ensuring that the uploaded samples cover different data distributions.

[0059] The cloud-based incremental training unit receives high-value samples uploaded from the edge and combines them with existing training data in the cloud to incrementally train the model to improve performance. Incremental training refers to continuing to train the model using new data based on an existing model, allowing the model to adapt to changes in data distribution, rather than training a new model from scratch.

[0060] The main challenge of incremental training is catastrophic forgetting, where the model may forget previously learned knowledge when learning new data, leading to performance degradation on older data. To address this issue, this invention employs an experience replay mechanism. A historical sample buffer is maintained in the cloud, storing representative samples used in previous training. During incremental training, newly uploaded samples are mixed with historical samples, ensuring that the model retains its memory of old knowledge while learning new knowledge.

[0061] Specifically, the historical sample buffer pool has a capacity of 50,000 samples and adopts a first-in, first-out (FIFO) update strategy. During each incremental training iteration, historical samples are randomly selected from the buffer pool and mixed with new samples at a 1:1 ratio to form training batches. Training uses a stochastic gradient descent optimizer with a learning rate of 0.001, a batch size of 32, and the number of training iterations is dynamically determined based on the number of new samples, typically ranging from 500 to 2000 iterations.

[0062] The cloud-based incremental training unit is also responsible for validating and testing the trained model. A separate validation set is maintained in the cloud, containing images of pests and diseases under various scenarios and conditions, to evaluate model performance. After training, metrics such as mean AP and recall are calculated on the validation set. The new model's parameters are only distributed to the edge if its performance surpasses that of the currently deployed model; otherwise, the current model is retained to avoid performance degradation.

[0063] The parameter delivery and update unit is responsible for delivering the new model parameters trained in the cloud to the edge devices, enabling online updates of the edge models. The parameter delivery employs a differential update strategy, transmitting only the changed model parameters instead of the entire model, thereby reducing data transmission volume and lowering network bandwidth requirements.

[0064] Differential updates are implemented based on incremental changes in model parameters. The cloud compares the parameters of the new model with those of the old model, calculates the parameter difference, compresses and encodes the difference, and then sends it down. The edge device receives the parameter difference, adds it to the local model parameters, and obtains the updated model. Parameter compression uses quantization and Huffman coding methods, quantizing floating-point parameters into 8-bit integers and then performing entropy coding, achieving a compression ratio of 10:1.

[0065] The parameters are sent during breaks in drone operations or while charging to avoid disrupting normal operations. After the parameters are sent, the lightweight edge detection module 1 loads the new model parameters, performs functional tests, and officially puts the system into use once the model is confirmed to be working properly. If loading fails or the test performance is substandard, the edge device automatically rolls back to the old model to ensure system stability.

[0066] Through the aforementioned edge-cloud closed-loop learning mechanism, the system achieves continuous model optimization and performance improvement. Experimental results show that after three months of online learning, the model's detection accuracy in new scenarios increased from 86.5% to 94.3%, effectively adapting to changes in pest and disease characteristics across different regions and time periods. Simultaneously, the closed-loop learning mechanism enables the system to self-evolve; with usage time and data accumulation, model performance continues to improve, ensuring long-term stable operation.

[0067] In one specific embodiment of the present invention, the edge-intelligent drone pest and disease real-time detection system is applied to rice pest and disease control operations. The drone, equipped with the system, flies at a height of 5 meters above the rice paddy at a speed of 3 m / s, covering an area of ​​100 mu (approximately 6.7 hectares). During flight, the system acquires real-time images of the rice paddy and performs real-time inference through the edge-lightweight detection module 1, detecting three pests and diseases: rice planthopper, rice leaf roller, and rice blast. The adaptive resource scheduling module 2 dynamically adjusts the inference strategy based on battery remaining capacity and the threat level of the pests and diseases. A high-precision mode is used in high-threat areas, while an energy-saving mode is switched to in low-threat areas, extending the drone's flight time to 45 minutes, 12 minutes longer than a fixed strategy. The multi-task collaborative optimization module 3 simultaneously outputs pest and disease type, location, and density information, generating differentiated application parameters. In high-density areas, the pesticide concentration is 30 mg / L, and the spraying rate is 80 mL / m²; in low-density areas, the pesticide concentration is 15 mg / L, and the spraying rate is 40 mL / m², achieving precise application and reducing pesticide waste by 75%. The edge-cloud closed-loop learning module 4 uploads high-value samples collected during the operation to the cloud. After incremental training, the model's detection accuracy in the region increased from 90.1% to 93.8%. The overall operation effect is significant, with pest and disease control efficiency improved by 60% and crop yield loss reduced by 80%.

[0068] The edge intelligent drone pest and disease real-time detection system provided by this invention achieves model lightweighting through knowledge distillation technology, balances performance and energy consumption through adaptive resource scheduling, improves detection efficiency through multi-task collaborative optimization, and achieves continuous model evolution through edge-cloud closed-loop learning. It effectively solves the problems of poor real-time performance, high resource consumption, and weak adaptability in existing technologies, and provides efficient and reliable technical support for smart agriculture and precision plant protection.

[0069] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0070] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. An edge-intelligent unmanned aerial vehicle (UAV) real-time pest and disease detection system, characterized in that, include: The edge-end lightweight detection module is used to perform real-time reasoning on the collected crop images based on a lightweight neural network of knowledge distillation during the flight of the drone, and output pest and disease categories, location coordinates and confidence information. An adaptive resource scheduling module, connected to the edge lightweight detection module, is used to dynamically determine the inference accuracy level and execution frequency based on the current battery level, the urgency of the detection task, and the real-time inference load, and to generate resource scheduling instructions. The multi-task collaborative optimization module is connected to the edge lightweight detection module and the adaptive resource scheduling module. It is used to synchronously execute pest detection, pest classification and pest density estimation according to the resource scheduling instructions, and to perform weighted fusion of multiple task outputs based on the joint loss function to generate accurate pesticide application parameters. The edge-cloud closed-loop learning module is connected to the edge lightweight detection module and the multi-task collaborative optimization module. It is used to transmit key detection data back to the cloud platform in real time via wireless network for incremental model training, and obtain updated model parameters and prevention and control strategies from the cloud and feed them back to the edge lightweight detection module to form a closed-loop optimization system.

2. The edge intelligent drone pest and disease real-time detection system according to claim 1, characterized in that, The lightweight edge detection module includes: The image acquisition unit is used to acquire crop images and perform preprocessing. The knowledge distillation and compression unit is used to transfer knowledge from complex deep models to lightweight student networks based on a teacher-student network framework. The real-time inference unit is used to call the lightweight student network to perform forward propagation inference on the preprocessed image and output the pest and disease detection results.

3. The edge intelligent drone pest and disease real-time detection system according to claim 1, characterized in that, The adaptive resource scheduling module includes: The status monitoring unit is used to monitor the drone's remaining battery level, flight speed, and computing load in real time. An urgency assessment unit, connected to the edge-end lightweight detection module, is used to determine the urgency level of the task based on the detected threat level of pests and diseases. The dynamic scheduling unit, connected to the status monitoring unit and the urgency assessment unit, is used to dynamically adjust the inference accuracy and execution frequency based on the accuracy-energy consumption balance strategy.

4. The edge intelligent drone pest and disease real-time detection system according to claim 1, characterized in that, The multi-task collaborative optimization module includes: The multi-branch inference unit is used to perform multiple tasks in parallel through independent detection branches, classification branches, and density estimation branches, based on a shared feature extraction layer; A joint loss calculation unit, connected to the multi-branch inference unit, is used to calculate the loss values ​​of multiple tasks and perform weighted fusion. The pesticide application parameter generation unit, connected to the joint loss calculation unit, is used to generate differentiated pesticide application parameters based on pest and disease type, density distribution, and location information.

5. The edge intelligent drone pest and disease real-time detection system according to claim 1, characterized in that, The edge-cloud closed-loop learning module includes: The data filtering and uploading unit is used to filter high-value test samples and upload them to the cloud platform via wireless network. The cloud-based incremental training unit is used for incremental training of the model based on uploaded edge detection data. The parameter distribution and update unit is connected to the cloud-based incremental training unit and the edge-end lightweight detection module, and is used to distribute the updated model parameters to the edge end to realize online model updates.

6. The edge intelligent unmanned aerial vehicle (UAV) real-time pest and disease detection system according to claim 2, characterized in that, The knowledge distillation and compression unit employs a soft-label distillation method with temperature parameter adjustment to achieve knowledge transfer by minimizing the difference between the student network output and the teacher network soft labels.

7. The edge intelligent unmanned aerial vehicle (UAV) real-time pest and disease detection system according to claim 3, characterized in that, The dynamic scheduling unit performs scheduling based on a multi-dimensional decision matrix, which includes battery remaining capacity dimension, urgency dimension and load dimension. Each dimension has multiple threshold intervals, and different interval combinations correspond to different inference accuracy levels.

8. The edge intelligent unmanned aerial vehicle (UAV) real-time pest and disease detection system according to claim 4, characterized in that, The joint loss calculation unit dynamically adjusts the loss weights of each task based on the correlation between tasks. The weights of the detection and classification tasks are set according to the confidence level, and the weights of the density estimation task are set according to the degree of pest and disease aggregation.

9. The edge intelligent unmanned aerial vehicle (UAV) real-time pest and disease detection system according to claim 5, characterized in that, The data filtering and uploading unit adopts an uncertainty sampling strategy, selecting samples whose model prediction confidence is lower than a set threshold as high-value samples for uploading.

10. The edge intelligent unmanned aerial vehicle (UAV) real-time pest and disease detection system according to claim 5, characterized in that, The cloud-based incremental training unit employs an experience replay mechanism to maintain a historical sample buffer pool. During incremental training, new samples and historical samples are mixed to prevent catastrophic forgetting.

Images