Tea garden insect pest early warning system based on machine learning

Abstract

The invention relates to the technical field of tea garden insect pests, and discloses a tea garden insect pest early warning system based on machine learning, which comprises a multi-mode sensing device, an edge intelligent terminal and a cloud decision platform. According to the tea garden insect pest early warning system based on machine learning, a narrow-band spectrum is dynamically selected through an adjustable spectrum imaging unit, image and sound wave features are synchronously acquired in combination with a vibration wave acquisition module, a heterogeneous calculation module of an edge intelligent terminal processes multi-source data in parallel, and a dynamic feature library management module updates a feature template in real time; the incremental learning mechanism triggers high-resolution imaging when unknown pests are detected, the meta-learning framework supports rapid migration training of a new model, the framework breaks through the limitation of a traditional single spectrum and a fixed model, and through multi-dimensional data fusion and adaptive learning, the pest recognition efficiency and adaptability are remarkably improved. The problems of low manual inspection efficiency and poor adaptability of traditional equipment are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of tea garden pests, and in particular to a tea garden pest early warning system based on machine learning. Background Art

[0002] Tea garden pests refer to biological hazards caused by various pests such as tea loopers, tea caterpillars, small green leafhoppers, and tea weevils during the growth of tea trees. These pests destroy the physiological structure of tea trees by gnawing on leaves, sucking sap, and boring into branches, resulting in a decrease in tea production and quality. In severe cases, the entire tea garden may suffer a complete crop failure.

[0003] Traditional tea garden pest monitoring relies primarily on manual inspections and single trapping devices, resulting in low monitoring efficiency and poor timeliness. Existing intelligent monitoring systems are mostly designed for single pests, using fixed spectral imaging and single model detection. These systems are unable to simultaneously identify multiple pests and lack adaptability to newly emerging pests. Furthermore, existing technologies lack predictive models for interactions between pests, making it difficult to formulate scientific, integrated prevention and control strategies. Therefore, we propose a tea garden pest early warning system based on machine learning. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of existing technologies, the present invention provides a tea garden pest early warning system based on machine learning, which has the capabilities of real-time identification of multiple pests, adaptive learning of dynamic features, and cross-species association analysis, effectively improving monitoring accuracy and the scientific nature of prevention and control decisions, and solving the problems of traditional tea garden pest monitoring that mainly relies on manual inspections and single trapping devices, resulting in low monitoring efficiency and poor timeliness. Existing intelligent monitoring systems are mostly designed for single pests, using fixed spectral imaging and single model detection, and are unable to identify multiple pests at the same time and lack adaptability to newly emerging pests. In addition, the existing technology has not established a predictive model for interactions between pests, making it difficult to formulate scientific and comprehensive prevention and control strategies.

[0006] (2) Technical solution

[0007] To achieve the aforementioned goals of real-time multi-pest identification, dynamic feature adaptive learning, and cross-species association analysis, effectively improving monitoring accuracy and the scientific nature of prevention and control decisions, the present invention provides the following technical solutions: a tea garden pest early warning system based on machine learning, comprising:

[0008] Multimodal sensing device, including:

[0009] Tunable spectral imaging unit, covering a wavelength range of 380-1000nm, with a step accuracy of 5nm, equipped with a motorized filter wheel and a telecentric lens set;

[0010] Vibration wave acquisition module, frequency response range 20-2000Hz, sensitivity ±0.5dB;

[0011] Multi-spectral fluorescence excitation unit, including 365nm, 405nm, and 470nm primary color excitation light sources;

[0012] Edge intelligent terminals, including:

[0013] Heterogeneous computing module, integrating a 64-bit ARM processor (main frequency 1.8GHz), GPU (floating-point computing capability 2.1TFLOPS) and NPU (computing power 8TOPS);

[0014] Dynamic feature library management module, supporting online feature template updates and version management;

[0015] Cloud-based decision-making platform, including:

[0016] Multi-target detection model, based on the improved YOLOv8 architecture, including feature pyramid network and attention mechanism;

[0017] Pest niche analysis module uses the MaxEnt algorithm to build a suitable habitat prediction model;

[0018] A cross-species association analysis engine that mines spatiotemporal relationships among pest communities based on graph neural networks.

[0019] Preferably, the multimodal sensing device further includes:

[0020] Modular trapping unit supports quick switching between three trapping methods: yellow board, sex attractant, and sweet and sour liquid;

[0021] Laser ranging module, with an accuracy of ±2mm, is used to obtain the three-dimensional coordinate data of the pest attachment location.

[0022] Preferably, the multi-target detection model includes:

[0023] Adaptive feature fusion module, which dynamically selects image, sound wave, and environmental data features through gated recurrent units;

[0024] Cascade classification network, the front-stage network realizes object-level classification, and the back-stage network realizes species-level classification.

[0025] Preferably, the cloud-based decision-making platform further includes:

[0026] Pest development period prediction module, which calculates the development rate of each insect stage based on the physiological time model;

[0027] The prevention and control strategy generation module uses a multi-objective optimization algorithm to generate a comprehensive prevention and control plan.

[0028] Preferably, the edge intelligent terminal supports:

[0029] Incremental learning mechanism that automatically triggers high-resolution image acquisition when unknown pests are detected;

[0030] A meta-learning framework to enable rapid transfer training of new pest recognition models.

[0031] The tea garden pest early warning method based on machine learning, including a tea garden pest early warning system based on machine learning, further includes the following steps:

[0032] S1. Multimodal data acquisition: acquiring images, sound waves, and environmental data through the device described in claim 1;

[0033] S2. Feature extraction: Use residual attention network to extract multi-scale texture, spectrum, and sound wave features;

[0034] S3. Multi-target detection: using the model described in claim 3 to identify pest species and count their numbers;

[0035] S4. Population prediction: Combine ecological niche model and physiological time model to predict pest population dynamics;

[0036] S5. Early warning decision-making: Generate comprehensive prevention and control recommendations through cross-species association analysis.

[0037] Preferably, the feature extraction step includes:

[0038] Step 1: Local binary pattern texture analysis to extract 8-neighborhood 3-bit pattern features;

[0039] Step 2: Improve Gabor filter processing to enhance direction-selective features.

[0040] Preferably, the multi-target detection step includes:

[0041] Step 1: Dynamic anchor frame matching algorithm automatically adjusts the detection frame size according to the pest body size;

[0042] Step 2: Improve the non-maximum suppression algorithm based on the adaptive IoU threshold strategy.

[0043] (3) Beneficial effects

[0044] Compared with the existing technology, the present invention provides a tea garden pest early warning system based on machine learning, which has the following beneficial effects:

[0045] 1. This machine learning-based tea garden pest warning system dynamically selects narrowband spectra through an adjustable spectral imaging unit, and combines it with a vibration wave acquisition module to synchronously acquire image and sound wave features. The heterogeneous computing module of the edge intelligent terminal processes multi-source data in parallel, and the dynamic feature library management module updates feature templates in real time. The incremental learning mechanism triggers high-resolution imaging when unknown pests are detected, and the meta-learning framework supports the rapid migration and training of new models. This architecture breaks through the limitations of traditional single spectrum and fixed models, and through multi-dimensional data fusion and adaptive learning, it significantly improves the efficiency and adaptability of pest identification, solving the problems of low manual inspection efficiency and poor adaptability of traditional equipment.

[0046] 2. This tea garden pest early warning system based on machine learning uses a multi-target detection model on a cloud-based decision-making platform combined with a feature pyramid network and an attention mechanism to achieve cascade classification from order to species level. The pest niche analysis module predicts suitable habitats, and the cross-species association analysis engine mines the spatiotemporal relationships of communities. Combined with the developmental period prediction module and the prevention and control strategy generation module, the system can dynamically predict population dynamics and generate the optimal solution that takes into account cost, ecology, and effectiveness. By revealing the laws of interaction between pests, this system breaks through the traditional single prevention and control model and significantly improves the scientific nature and pertinence of the integrated prevention and control strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the system structure of the present invention;

[0048] Figure 2 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] See also Figure 1-2 , a tea garden pest early warning system based on machine learning, including:

[0051] Multimodal sensing device, including:

[0052] Tunable spectral imaging unit, covering a wavelength range of 380-1000nm, with a step accuracy of 5nm, equipped with a motorized filter wheel and a telecentric lens set;

[0053] Vibration wave acquisition module, frequency response range 20-2000Hz, sensitivity ±0.5dB;

[0054] Multi-spectral fluorescence excitation unit, including 365nm, 405nm, and 470nm primary color excitation light sources;

[0055] Edge intelligent terminals, including:

[0056] Heterogeneous computing module, integrating a 64-bit ARM processor (main frequency 1.8GHz), GPU (floating-point computing capability 2.1TFLOPS) and NPU (computing power 8TOPS);

[0057] Dynamic feature library management module, supporting online feature template updates and version management;

[0058] Cloud-based decision-making platform, including:

[0059] Multi-target detection model, based on the improved YOLOv8 architecture, including feature pyramid network and attention mechanism;

[0060] Pest niche analysis module uses the MaxEnt algorithm to build a suitable habitat prediction model;

[0061] A cross-species association analysis engine that mines spatiotemporal relationships among pest communities based on graph neural networks.

[0062] The tea garden pest early warning method based on machine learning, including a tea garden pest early warning system based on machine learning, further includes the following steps:

[0063] S1. Multimodal data acquisition: acquiring images, sound waves, and environmental data through the device described in claim 1;

[0064] S2. Feature extraction: Use residual attention network to extract multi-scale texture, spectrum, and sound wave features;

[0065] S3. Multi-target detection: using the model described in claim 3 to identify pest species and count their numbers;

[0066] S4. Population prediction: Combine ecological niche model and physiological time model to predict pest population dynamics;

[0067] S5. Early warning decision-making: Generate comprehensive prevention and control recommendations through cross-species association analysis.

[0068] Example 1:

[0069] The system of the present invention adopts a three-layer stacked hardware architecture, with the bottom layer being the vibration wave acquisition layer, the middle layer being the spectral imaging layer, and the top layer being the trapping execution layer. The vibration wave acquisition module integrates an 8-channel ADXL355 accelerometer and a 3-microphone array. After being processed by a PGA204 instrumentation amplifier (gain 100 times) and a second-order Butterworth low-pass filter (cut-off frequency 2000Hz), it is digitized by an ADS1278 analog-to-digital converter (24-bit, 48kHz sampling). To address the 100ms delay problem of traditional equipment, an FPGA-based clock synchronization circuit is designed to generate a 100MHz reference clock and transmit it to each module via differential signals, achieving microsecond-level alignment (<50μs) of the vibration wave and image.

[0070] The spectral imaging unit adopts a dual-optical path design: the reflection channel is equipped with a Sony IMX490 sensor (20 million pixels, global shutter), covering 400-700nm visible light; the transmission channel uses an InGaAs sensor (800×600 pixels), covering 800-1000nm near-infrared light. The electric filter wheel (model Sutter Lambda10-3) supports 5nm step switching and dynamically selects the optimal bandwidth (for example, a narrow band of 450-550nm is used for the green leafhopper) in combination with the pest size prediction algorithm. The multispectral fluorescence excitation unit consists of a 365nm / 405nm / 470nm three-primary color LED array, and uses the MAX16982 constant current driver chip to achieve 0-100% PWM dimming to ensure the effective excitation of the fluorescence characteristics of different pests.

[0071] The modular trapping unit uses a six-station turntable driven by a stepper motor, which supports rapid switching between three modes: yellow plate, sex attractant, and sweet and sour liquid. The yellow plate module integrates a laser beam counter (OPB810) with a counting accuracy of ±1%. The sex attractant module uses a polytetrafluoroethylene sustained-release membrane, and the diffusion rate follows the D=D0×exp(-k×t) model (D0=0.05mg / h, k=0.002 / h). The ratio of sweet and sour liquid is optimized through a BP neural network to sugar: vinegar: wine: water = 3:4:1:2.

[0072] Example 2:

[0073] The edge intelligent terminal is built based on the NVIDIA Jetson AGX Orin module and adopts a heterogeneous computing architecture to achieve task offload: the CPU is responsible for device control and I / O management, the GPU processes multispectral image feature extraction in parallel, and the NPU performs YOLOv8 target detection inference. The dynamic load balancing algorithm is based on the D-Waterfilling model and dynamically allocates computing resources according to the current task priority.

[0074] The dynamic feature library management module adopts a three-level storage architecture: on-chip SRAM caches the most recent 100 feature templates (32 dimensions after PCA dimensionality reduction), eMMC stores the complete feature library (supports 5000 templates), and an external SSD serves as a backup. The incremental learning trigger mechanism is implemented through the statistical process control (SPC) algorithm. When the confidence level of three consecutive detections is <0.4 and the coefficient of variation of the feature vector is >0.3, the high-resolution imaging mode (8000×6000 pixels, 16-bit depth) is automatically triggered. The meta-learning framework is based on the MAML algorithm and is pre-trained on five common pests (100 samples each). New species only require five samples to achieve 90% accuracy within 50 iterations. The feature compression algorithm combines PCA (retaining 95% variance) with Huffman coding to achieve an 8:1 compression ratio, reducing the bandwidth requirement for OTA updates.

[0075] Example 3:

[0076] The multi-target detection model is improved based on YOLOv8 and adopts mixed precision training (FP16+FP32). ResNet50-SE (with SE attention module added) is used for image feature extraction. The acoustic wave features are converted to 130 dimensions through MFCC parameters (40Mel filter, 512FFT). The environmental data (temperature, humidity, light intensity) are encoded into time series features through LSTM network (128 hidden units). The adaptive feature fusion module uses GRU network to fuse image (2048 dimensions), acoustic wave (130 dimensions), and environmental (64 dimensions) features into a 2048-dimensional vector. The front stage of the cascade classification network uses MobileNetV3-Large (224×224 input) to achieve object-level classification, and the back stage uses EfficientNet-B4 (380×380 input) to achieve species-level classification. Focal Loss (α=0.25,γ=2) and Center Loss (λ=0.5) are jointly optimized.

[0077] The pest niche analysis module constructs a suitable habitat prediction model based on the MaxEnt algorithm, inputting 19 environmental variables such as annual average temperature and annual precipitation and the geographic coordinates of 5,000 pest samples. The model's generalization ability is improved through the regularization parameter C=10 and the feature combination LQH. The cross-species association analysis engine uses a graph neural network. Node features include pest population density, developmental stage, and environmental factors, and edge attributes are based on predator-prey relationships (ecological survey data) and competition index. The spatiotemporal convolution module combines GCN and LSTM, and uses the Apriori algorithm to mine frequent item sets (support > 0.2, confidence > 0.8) to reveal the spatiotemporal relationship of pest communities.

[0078] Example 4:

[0079] The pest developmental period prediction module is based on a physiological time model, with the developmental rate formula being V = V0 × exp(-Ea(R × T)). After parameter calibration, the average absolute error in tea geometrid egg stage prediction was less than 1.2 days. The control strategy generation module employed an improved NSGA-III algorithm. The optimization objectives included minimizing control costs, minimizing ecological risks, and maximizing control effectiveness. Constraints included the pesticide safety interval (7-14 days) and a natural enemy protection threshold (parasitic wasp survival rate >60%).

[0080] The system is deployed in an equilateral triangle layout with three sensing terminals per hectare. Terminal-to-edge communication is achieved through LoRaWAN (SF=12, BW=125kHz), and edge-to-cloud communication is achieved through a 5G private network (latency <20ms). The solar power supply system (150W solar panel + 200Ah lithium battery) supports 7 days of rainy and cloudy conditions. Test results show that the average mAP@0.5 for 15 pest species is 0.927, the response time is <30s, the false alarm rate is <0.8 times / week, and the average daily data transmission volume is <20MB, verifying the system's efficiency and reliability.

[0081] In summary, the tea garden pest early warning system based on machine learning dynamically selects narrowband spectra through the adjustable spectral imaging unit, combines the vibration wave acquisition module to synchronously obtain image and sound wave features, and the heterogeneous computing module of the edge intelligent terminal processes multi-source data in parallel. The dynamic feature library management module updates the feature template in real time. The incremental learning mechanism triggers high-resolution imaging when unknown pests are detected, and the meta-learning framework supports the rapid migration and training of new models. This architecture breaks through the limitations of traditional single spectrum and fixed models, and significantly improves the efficiency and adaptability of pest identification through multi-dimensional data fusion and adaptive learning, solving the problems of low efficiency of manual inspections and poor adaptability of traditional equipment.

[0082] In addition, the tea garden pest early warning system based on machine learning, the multi-target detection model of the cloud decision-making platform combines the feature pyramid network and the attention mechanism to realize cascade classification from order level to species level, the pest niche analysis module predicts the suitable habitat, and the cross-species association analysis engine mines the spatiotemporal relationship of the community. Combined with the developmental period prediction module and the prevention and control strategy generation module, the system can dynamically predict the population dynamics and generate the optimal solution that takes into account cost, ecology and effect. This system breaks through the traditional single prevention and control model by revealing the laws of interaction between pests, significantly improves the scientificity and pertinence of the integrated prevention and control strategy, and solves the problem that traditional tea garden pest monitoring mainly relies on manual inspections and single trapping devices, with low monitoring efficiency and poor timeliness. Existing intelligent monitoring systems are mostly designed for single pests, using fixed spectral imaging and single model detection. They cannot identify multiple pests at the same time and lack adaptability to newly emerging pests. In addition, the existing technology has not established a predictive model for the interaction between pests, making it difficult to formulate a scientific integrated prevention and control strategy.

[0083] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the existing technology. The computer software programs or protocols involved in the functional modules are themselves technologies that are well known to those skilled in the art and are not improvements to this system. The improvements to this system are the interaction or connection relationships between the modules, that is, improvements to the overall structure of the system to solve the corresponding technical problems to be solved by this system.

[0084] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The tea garden pest early warning system based on machine learning is characterized by: include: Multimodal sensing device, including: Tunable spectral imaging unit, covering a wavelength range of 380-1000nm, with a step accuracy of 5nm, equipped with a motorized filter wheel and a telecentric lens set; Vibration wave acquisition module, frequency response range 20-2000Hz, sensitivity ±0.5dB; Multi-spectral fluorescence excitation unit, including 365nm, 405nm, and 470nm primary color excitation light sources; Edge intelligent terminals, including: Heterogeneous computing module, integrating a 64-bit ARM processor (main frequency 1.8GHz), GPU (floating-point computing capability 2.1TFLOPS) and NPU (computing power 8TOPS); Dynamic feature library management module, supporting online feature template updates and version management; Cloud-based decision-making platform, including: Multi-target detection model, based on the improved YOLOv8 architecture, including feature pyramid network and attention mechanism; Pest niche analysis module uses the MaxEnt algorithm to build a suitable habitat prediction model; A cross-species association analysis engine that mines spatiotemporal relationships among pest communities based on graph neural networks.

2. The tea garden pest early warning system based on machine learning according to claim 1 is characterized in that: The multimodal sensing device further includes: Modular trapping unit supports quick switching between three trapping methods: yellow board, sex attractant, and sweet and sour liquid; Laser ranging module, with an accuracy of ±2mm, is used to obtain the three-dimensional coordinate data of the pest attachment location.

3. The tea garden pest early warning system based on machine learning according to claim 1 is characterized in that: The multi-target detection model includes: Adaptive feature fusion module, which dynamically selects image, sound wave, and environmental data features through gated recurrent units; Cascade classification network, the front-stage network realizes object-level classification, and the back-stage network realizes species-level classification.

4. The tea garden pest early warning system based on machine learning according to claim 1 is characterized in that: The cloud-based decision-making platform also includes: Pest development period prediction module, which calculates the development rate of each insect stage based on the physiological time model; The prevention and control strategy generation module uses a multi-objective optimization algorithm to generate a comprehensive prevention and control plan.

5. The tea garden pest early warning system based on machine learning according to claim 1 is characterized in that: The edge intelligent terminal supports: Incremental learning mechanism that automatically triggers high-resolution image acquisition when unknown pests are detected; A meta-learning framework to enable rapid transfer training of new pest recognition models.

6. A tea garden pest early warning method based on machine learning, characterized in that: The tea garden pest early warning system based on machine learning according to claims 1 to 5 further comprises the following steps: S1. Multimodal data acquisition: acquiring images, sound waves, and environmental data through the device described in claim 1; S2. Feature extraction: Use residual attention network to extract multi-scale texture, spectrum, and sound wave features; S3. Multi-target detection: using the model described in claim 3 to identify pest species and count their numbers; S4. Population prediction: Combine ecological niche model and physiological time model to predict pest population dynamics; S5. Early warning decision-making: Generate comprehensive prevention and control recommendations through cross-species association analysis.

7. The tea garden pest early warning method based on machine learning according to claim 6, characterized in that: The step S2, feature extraction, comprises: Step 1: Local binary pattern texture analysis to extract 8-neighborhood 3-bit pattern features; Step 2: Improve Gabor filter processing to enhance direction-selective features.

8. The tea garden pest early warning method based on machine learning according to claim 6, characterized in that: The multi-target detection step in S3 includes: Step 1: Dynamic anchor frame matching algorithm automatically adjusts the detection frame size according to the pest body size; Step 2: Improve the non-maximum suppression algorithm based on the adaptive IoU threshold strategy.

Images