Rice disease and insect pest image real-time identification and decision-making system based on deep learning

Abstract

The invention discloses a rice disease and insect pest image real-time identification and decision-making system based on deep learning, and particularly relates to the technical field of agriculture. A multi-modal fusion identification module; a resistance decision engine module; a dynamic prevention and control execution module; and a feedback optimization module. According to the invention, through a new intelligent plant protection normal form of intelligent perception-collaborative decision-precise execution-closed loop evolution, a multi-spectral imaging and microscopic feature synthesis technology is utilized, the bottleneck of traditional diagnosis is broken through, early-stage precise recognition of diseases and insect pests is realized, a medicament-pathogenic bacteria-natural enemy tripartite game model and a multi-objective optimization mechanism are innovatively constructed, and the method has a wide application prospect. The ecological regulation and control capability of the rice field is activated while the risk of drug resistance is avoided, it is ensured that prevention and control measures accurately reach a target by means of the three-dimensional profiling spraying and natural enemy synergistic release technology, residual feedback and model self-iteration are fused in the whole process, a prevention and control closed loop which is more intelligent is formed, and an evolvable digital solution is provided for rice green production.

Description

Technical Field

[0001] The present invention relates to the field of agricultural technology, and in particular to a real-time recognition and decision-making system for rice pest and disease images based on deep learning. Background Art

[0002] Rice pest and disease control is crucial for ensuring food security. Traditional methods, which rely on manual visual diagnosis, suffer from low efficiency and subjectivity, making early and accurate identification difficult. Existing image recognition technologies, mostly based on visible light images, are susceptible to interference from field lighting and shadows, and lack the ability to capture microscopic features such as mycelium, resulting in a high rate of missed detection of early lesions. At the decision-making level, overreliance on chemical pesticides and a lack of resistance management mechanisms have led to a continued increase in pathogen resistance, while biological control methods struggle to synergize with chemical measures. In the implementation phase, extensive spraying results in pesticide drift and waste, and fails to accurately cover the three-dimensional distribution of lesions. While some intelligent recognition systems currently exist, they are unable to adapt to pathogen mutations and dynamic changes in the ecological environment. Therefore, there is an urgent need to build an intelligent crop protection system that integrates multimodal perception, resistance avoidance decision-making, and precise execution to overcome diagnostic bottlenecks, curb the development of resistance, and reduce environmental pollution. Summary of the Invention

[0003] The main purpose of the present invention is to provide a real-time recognition and decision-making system for rice pests and diseases images based on deep learning, which can effectively solve the problems in the above-mentioned background technology.

[0004] To achieve the above object, the technical solution adopted by the present invention is: A real-time rice pest and disease image recognition and decision-making system based on deep learning, including: The image acquisition and enhancement module acquires multispectral images of rice fields and environmental sensor data, eliminates shadow interference through a light-adaptive normalization algorithm, and uses a generative adversarial network to synthesize microscopic features of disease spots to generate enhanced pest and disease feature maps. The multimodal fusion recognition module calls the pest and disease characteristic map, combines real-time meteorological data and the rice genotype database, extracts the spatiotemporal evolution characteristics of lesions through a spatiotemporal convolutional neural network, and outputs the pest and disease type identification result and severity level; The resistance decision engine module, based on the pest and disease type and severity level, associates the regional pesticide use history database and pathogen resistance gene library, and generates resistance avoidance agent combinations and ecological regulation strategies through a multi-agent reinforcement learning model; A dynamic control execution module integrates the resistance-avoidance agent combination and ecological regulation strategy, combines the real-time location of the drone and the tank capacity data, and generates a three-dimensional targeted spray path and natural enemy release coordinate instructions; The feedback optimization module monitors the rice field image sequence within 72 hours after the application of pesticides, quantifies the deviation of the control effect through the residual contrast network, and updates the weight parameters of the multimodal fusion recognition model.

[0005] Preferably, the image acquisition and enhancement module specifically includes shadow elimination parameters, microscopic feature enhancement coefficients and multispectral fusion weights; the multimodal fusion recognition module specifically includes spatiotemporal convolution kernel size, genotype matching threshold and severity grading interval; the resistance decision engine module includes resistance probability prediction value, pesticide synergy coefficient and ecological regulation intensity parameter; the dynamic prevention and control execution module specifically includes spray elevation angle deviation tolerance, pesticide-natural enemy release time difference and three-dimensional path curvature radius; the feedback optimization module includes residual contrast threshold, weight update learning rate and model iteration trigger condition.

[0006] Preferably, the image acquisition and enhancement module includes a shadow elimination submodule, a microscopic synthesis submodule and a feature fusion submodule; The shadow removal submodule captures RGB-NIR images through a polarization filter camera, calculates the covariance matrix of saturation and brightness in the HSV space of the pixel points, uses the Mahalanobis distance threshold to segment the shadow area, and performs adaptive histogram equalization; The microscopic synthesis submodule calls the leaf stomatal structure template library, inputs the visible light image into the pre-trained Pix2Pix-HD network, generates a virtual microscopic image in the 750nm-900nm band, and enhances the morphological characteristics of pathogenic fungi hyphae; The feature fusion submodule performs channel concatenation on the shadow-removed RGB image, virtual microscopic image, and leaf temperature sensor data, uses an attention mechanism for weighted fusion, and outputs a 64-dimensional feature vector.

[0007] Preferably, the multimodal fusion recognition module includes a spatiotemporal feature extraction submodule, a resistance gene matching submodule and a severity grading submodule; The spatiotemporal feature extraction submodule inputs 5 consecutive frames of enhanced feature maps into the 3D-ResNet50 network to extract the lesion expansion rate and direction vector; The resistance gene matching submodule calls the rice variety gene database, compares the pathogen identification results with the resistance gene loci (Pi-ta / Pi-b) using the BLAST algorithm, and outputs the gene escape probability value; The severity grading submodule integrates the proportion of lesion area, chlorophyll fluorescence intensity and transpiration rate reduction ratio, and divides the severity of pests and diseases into 5 levels through a random forest model.

[0008] Preferably, the resistance decision engine module includes a drug game submodule, an ecological regulation submodule and a collaborative optimization submodule; The drug game submodule constructs a three-party game matrix of drug, pathogen and natural enemy, takes minimizing the resistance development rate as the objective function, and solves the Nash equilibrium strategy; The ecological regulation submodule calls the release amount formula of trichogrammatids and ladybugs according to the type of pests and diseases, and generates the release point coordinates based on the topological structure of the rice field; The collaborative optimization submodule adopts the NSGA-II algorithm to optimize the multi-objective pharmaceutical cost and ecological benefit and outputs the Pareto optimal solution set.

[0009] Preferably, the dynamic prevention and control execution module includes: The 3D path planning submodule reconstructs the 3D model of the rice plant based on the UAV LiDAR point cloud data, calculates the spatial coordinates and normal vector direction of the lesion, and generates a contoured spray path; The natural enemy coordination submodule plans the opening and closing sequence of the drone’s drop bay based on the release point coordinates of the ecological control strategy.

[0010] Preferably, the dynamic prevention and control execution module further includes an anti-drift control submodule, which integrates wind speed and direction sensor data and adjusts the spray pressure and droplet size in real time through a PID controller.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention significantly improves the ability to detect pests and diseases through multimodal perception and synergistic enhancement technology. At the data acquisition layer, it integrates multispectral imaging, leaf temperature sensing, and meteorological data, uses a light adaptive algorithm to eliminate environmental interference, and synthesizes the microscopic features of lesions through a generative adversarial network to visualize the hidden pathogen structure. At the recognition layer, a spatiotemporal convolutional network associates the lesion expansion dynamics with the rice gene resistance database, and combines meteorological conditions to predict epidemic risks.

[0012] 2. The present invention completely changes the extensive model of pesticide application that relies on experience through a dual-track decision-making mechanism of resistance avoidance and ecological regulation. The decision-making engine constructs a three-party game model of pesticides, pathogens and natural enemies, with the core goal of delaying pesticide resistance. It dynamically generates the optimal ratio of chemical control and biological control, automatically recommends a combination of pesticides with low resistance pressure, and simultaneously calculates the precise release amount and coordinates of trichogrammatids / ladybugs. Through a multi-objective optimization algorithm, it balances economic costs and ecological benefits, avoiding the vicious cycle of resistance caused by pesticide abuse, activating the self-regulation ability of rice field ecosystems, and realizing the sustainable management of "using insects to control insects".

[0013] 3. The present invention constructs a closed loop of prevention and control based on precise spatial operation and real-time feedback optimization, which greatly improves the execution reliability. The execution module relies on the UAV LiDAR to reconstruct the three-dimensional model of the rice plant, generate a contoured spray path that fits the three-dimensional distribution of the lesions, and dynamically adjusts the droplet parameters in combination with wind speed perception to ensure the precise deposition of the agent. At the same time, it coordinates the time and space difference between the release sequence of natural enemies and the spraying of the agent to avoid accidental injury to biological natural enemies. The feedback layer quantifies the prevention and control effect through the residual comparison network, automatically identifies the recurrence of lesions or abnormal resistance, and triggers the iteration of model parameters. This "execution-monitoring-evolution" closed loop enables the system to have continuous learning capabilities, gradually adapt to local pathogen mutations and ecological environmental changes, and form a smart agricultural solution that becomes smarter with use. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0015] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0016] like Figure 1 As shown in the figure, a real-time recognition and decision-making system for rice pests and diseases based on deep learning includes: The image acquisition and enhancement module acquires multispectral images of rice fields and environmental sensor data, eliminates shadow interference through a light-adaptive normalization algorithm, and uses a generative adversarial network to synthesize microscopic features of disease spots to generate enhanced pest and disease feature maps. The multimodal fusion recognition module uses pest and disease feature maps, combines real-time meteorological data with a rice genotype database, extracts the spatiotemporal evolution characteristics of lesions through a spatiotemporal convolutional neural network, and outputs pest and disease type identification results and severity levels; The resistance decision engine module, based on the pest and disease type and severity level, links the regional pesticide use history database and pathogen resistance gene library, and generates resistance avoidance agent combinations and ecological regulation strategies through a multi-agent reinforcement learning model; The dynamic control execution module integrates resistance-avoidance pesticide combinations and ecological regulation strategies, combines the real-time location of drones and the capacity of pesticide tanks, and generates three-dimensional targeted spray paths and natural enemy release coordinate instructions; The feedback optimization module monitors the rice field image sequence within 72 hours after pesticide application, quantifies the deviation of the control effect through the residual contrast network, and updates the weight parameters of the multimodal fusion recognition model.

[0017] Specifically, the image acquisition and enhancement module uses a DJI Mavic 3 multispectral drone to obtain RGB-NIR images (resolution 3840×2160) and simultaneously collects environmental sensor data (temperature 28.5°C, humidity 75%). It then uses a light-adaptive normalization algorithm to eliminate rice plant shadow interference: the covariance matrix of saturation and brightness in the HSV space of the pixel points is calculated. When the Mahalanobis distance is greater than 3.0, it is determined to be a shadow area. CLAHE histogram equalization is performed, and then the visible light image is input into the pre-trained Pix2Pix-HD network to synthesize a virtual microscopic image in the 750-900nm band, highlighting the morphology of rice blast hyphae, and outputting a 64-dimensional feature vector to the multimodal fusion recognition module.

[0018] The image acquisition and enhancement module specifically includes shadow elimination parameters, microscopic feature enhancement coefficients and multispectral fusion weights; the multimodal fusion recognition module specifically includes spatiotemporal convolution kernel size, genotype matching threshold and severity classification interval; the resistance decision engine module includes resistance probability prediction value, pesticide synergy coefficient and ecological regulation intensity parameter; the dynamic prevention and control execution module specifically includes spray elevation angle deviation tolerance, pesticide-natural enemy release time difference and three-dimensional path curvature radius; the feedback optimization module includes residual comparison threshold, weight update learning rate and model iteration trigger condition.

[0019] Furthermore, some of the above parameters are set as follows: Shadow removal parameters: Mahalanobis distance threshold 3.0, CLAHE grid size 8×8; Genotype matching threshold: when the Pi-ta gene sequence similarity is ≥85%, a resistance warning is triggered; Three-dimensional path curvature radius: minimum constraint value 1.2m; Residual contrast threshold: Model update is triggered when the lesion area reduction rate is <15%.

[0020] The image acquisition and enhancement module includes a shadow removal submodule, a microscopic synthesis submodule, and a feature fusion submodule; The shadow removal submodule captures RGB-NIR images through a polarization-filtered camera, calculates the covariance matrix of saturation and brightness in the HSV space of each pixel, segments the shadow area using a Mahalanobis distance threshold, and performs adaptive histogram equalization. The microscopy synthesis submodule uses the leaf stomatal structure template library and inputs the visible light image into the pre-trained Pix2Pix-HD network to generate a virtual microscopic image in the 750nm-900nm band to enhance the morphological characteristics of pathogenic fungi. The feature fusion submodule performs channel concatenation on the shadow-removed RGB image, virtual microscopic image, and leaf temperature sensor data, uses an attention mechanism for weighted fusion, and outputs a 64-dimensional feature vector.

[0021] Specifically, the image enhancement process is as follows: The polarization filter camera (SonyIMX586) in the shadow removal submodule captures the RGB-NIR image at 9:00, extracts the HSV values ​​of the pixels, and calculates the saturation and brightness covariance matrix Σ = [0.04, 0.02; 0.02, 0.03]. The Mahalanobis distance formula is:

[0022] Where, D: Mahalanobis distance value, x: feature vector of the current pixel, μ: mean vector of the non-shadow area, Σ: covariance matrix of the non-shadow area, Σ -1 : The inverse of the covariance matrix.

[0023] The resulting pixels are classified into the shadow area and block histogram equalization is performed; The microscopy synthesis submodule uses a stomatal template library (containing five types of rice leaf stomatal structures) to input visible light images into the Pix2Pix-HD network to generate virtual microscopic images in the 850nm band, improving the accuracy of hyphae morphology recognition to 92.3%; The feature fusion submodule uses channel concatenation (concat) to fuse RGB images, virtual microscopic images and leaf temperature data (28.5℃), and outputs a 64-dimensional feature vector through the weighted SE attention mechanism.

[0024] The multimodal fusion recognition module includes a spatiotemporal feature extraction submodule, a resistance gene matching submodule, and a severity grading submodule; The spatiotemporal feature extraction submodule inputs 5 consecutive frames of enhanced feature maps into the 3D-ResNet50 network to extract the lesion expansion rate and direction vector; The resistance gene matching submodule calls the rice variety gene database, compares the pathogen identification results with the resistance gene loci (Pi-ta / Pi-b) using the BLAST algorithm, and outputs the gene escape probability value; The severity grading submodule integrates the proportion of lesion area, chlorophyll fluorescence intensity and transpiration rate reduction ratio, and divides the severity of pests and diseases into 5 levels through the random forest model.

[0025] The specific process of multimodal recognition is as follows: First, the spatiotemporal feature extraction submodule is used to input 5 consecutive frames of feature maps (10 minutes apart) into the 3D-ResNet50 network (convolution kernel 7×7×7), and the lesion expansion vector is output; The resistance gene matching submodule was then used to call the rice gene database (containing 12,000 varieties). BLAST comparison was performed between the rice blast strain MG01 and the resistance gene Pi-ta. Three mismatches were detected (M_i=3). The escape probability was calculated by substituting the formula:

[0026] Where, P_e: resistance escape probability, exp: natural exponential function.

[0027] Finally, the severity grading submodule inputs the proportion of lesion area of ​​18%, the decrease in chlorophyll fluorescence intensity of 25%, and the decrease in transpiration rate of 30%, and the random forest model outputs a severity level of 4 (the highest level is 5).

[0028] The resistance decision engine module includes the drug game submodule, the ecological regulation submodule and the collaborative optimization submodule; The drug game submodule constructs a three-party game matrix between drugs, pathogens, and natural enemies, takes minimizing the resistance development rate as the objective function, and solves the Nash equilibrium strategy; The ecological regulation submodule uses the formula for releasing trichogrammatids and ladybugs according to the type of pests and diseases, and generates the coordinates of the release points based on the topological structure of the rice field; The collaborative optimization submodule uses the NSGA-II algorithm to optimize the cost and ecological benefits of pharmaceuticals in a multi-objective manner and outputs the Pareto optimal solution set.

[0029] Furthermore, the resistance decision process is: The game matrix is ​​constructed through the drug game submodule (drug action: tricyclazole / azoxystrobin; pathogen action: gene mutation / phenotypic adaptation; natural enemy action: trichogrammatid / ladybug) and the resistance development rate objective function is set: , solve the Nash equilibrium and get the optimal strategy "Azoxystrobin + Ladybug Alliance"; The ecological regulation submodule uses the Trichogrammatid release formula for the Chilo suppressalis:

[0030] The release point coordinates were then generated in combination with the rice field topology (elevation difference < 0.5 m).

[0031] Finally, the collaborative optimization submodule uses the NSGA-II algorithm to optimize the target and output the Pareto solution set.

[0032] The dynamic prevention and control execution module includes: The 3D path planning submodule reconstructs the 3D model of the rice plant based on the UAV LiDAR point cloud data, calculates the spatial coordinates and normal vector direction of the lesion, and generates a contoured spray path; The natural enemy coordination submodule plans the opening and closing sequence of the drone’s drop bay based on the release point coordinates of the ecological control strategy.

[0033] The dynamic prevention and control execution module also includes an anti-drift control submodule, which integrates wind speed and direction sensor data and adjusts the spray pressure and droplet size in real time through a PID controller.

[0034] Furthermore, the specific implementation methods of prevention and control are as follows: 3D path planning submodule: Uses UAV LiDAR (Velodyne VLP-16) to scan rice plants to generate a point cloud model, calculates the normal vector of the lesion, and generates a contoured spray path; Natural enemy coordination submodule: plans the launch sequence of the drop capsule at the coordinates (opens at a height of 2m and stays for 5s); Anti-drift control submodule: The wind speed sensor detects a southeast wind of 3.5 m / s, and the PID controller (proportional coefficient K_p=1.2) adjusts the spray pressure to 0.35 MPa, and the droplet size increases to 150 μm.

[0035] Example 2: Taking the rice blast outbreak demonstration area in Jiaxing, Zhejiang as an example 1. Early morning data collection and feature enhancement: System startup time: 8:30 AM In a contiguous rice demonstration area (200 mu) in Jiaxing, Zhejiang, a DJI Mavic 3 multispectral drone equipped with a polarization filter camera (Sony IMX586) was launched, capturing RGB-NIR images (resolution 3840×2160) at 9:00 a.m. Environmental data were collected simultaneously: air temperature 28.5°C, humidity 75%, and leaf temperature 28.2°C; The image acquisition and enhancement module starts working: Shadow removal: The covariance matrix of saturation and brightness in the HSV space of each pixel was calculated (Σ=[0.04, 0.02; 0.02, 0.03]). CLAHE histogram equalization (grid size 8×8) was performed on shadow areas with a Mahalanobis distance greater than 3.0 to eliminate interference from rice plant shadows.

[0036] Microscopic synthesis: The stomatal structure template library of leaves (including 5 types of stomatal morphologies) is called, and the visible light image is input into the pre-trained Pix2Pix-HD network to generate a virtual microscopic image in the 850nm band, which clarifies the morphology of rice blast hyphae (recognition accuracy is improved to 92.3%).

[0037] Multimodal fusion: The shadow-removed RGB image, virtual microscopic image, and leaf temperature data channels are concatenated and weightedly fused through the SE attention mechanism to output a 64-dimensional pest and disease feature map.

[0038] 2. Multimodal Fusion Recognition and Severity Determination Time: 9:30 The multimodal fusion recognition module starts triple analysis after receiving the feature map: Spatiotemporal evolution tracking: The feature maps of 5 consecutive frames with an interval of 10 minutes were input into the 3D-ResNet50 network (convolution kernel 7×7×7) to extract the lesion expansion vector (rate 1.2cm2 / h, direction northeast 60°); Resistance gene matching: A rice gene database containing 12,000 varieties was used to compare the rice blast strain MG01 with the resistance gene Pi-ta using the BLAST algorithm. Three mismatches were detected, and the escape probability was calculated:

[0039] (Exceeding the threshold of 0.3 triggers a resistance warning); Severity grading: The three indicators of fused lesion area ratio (18%), decreased chlorophyll fluorescence intensity (25%), and decreased transpiration rate (30%) were determined to be severity level 4 (the highest level is 5) by the random forest model.

[0040] 3. Resistance Avoidance Decision Generation Time: 10:00 The resistance decision engine module initiates multi-objective optimization based on the recognition results: Drug game: Construct a three-party game matrix between azoxystrobin, rice blast fungus, and ladybugs to solve the Nash equilibrium by minimizing the resistance development rate and determine the "azoxystrobin + ladybug combination" strategy; Ecological regulation: For the distribution of Chilo suppressalis eggs (detection quantity 1200 eggs / mu), use the formula:

[0041] The release rate of trichogrammatids (960 per mu) was calculated, and the coordinates of the release points were generated based on the topological data of the paddy field with elevation differences of <0.5 m. Collaborative optimization: The NSGA-II algorithm was used for multi-objective optimization. The Pareto solution set was output under the constraints of pesticide cost (≤¥50 / mu) and ecological benefit (≥0.8). Finally, option B (cost ¥50 / mu, benefit 0.85) was selected.

[0042] 4. Implementation of Three-Dimensional Targeted Prevention and Treatment Time: 10:30 The dynamic prevention and control execution module links the drone (with a 20L payload) and the LiDAR sensor: 3D path planning: Calculate the spatial coordinates and normal vectors of the lesions based on the rice plant point cloud model, and generate a contoured spray path with a curvature radius of ≥1.2m to ensure that the liquid precisely covers the lesions. Coordinated release of natural enemies: Plan the release capsule action at the preset coordinate point (longitude 120.76°E, latitude 30.77°N) (open at a height of 2m, stay for 5s); Anti-drift control: Wind speed data (3.5 m / s southeast wind) is integrated in real time, and the spray pressure is adjusted to 0.35 MPa and the droplet size is increased to 150 μm through a PID controller ($K_p=1.2$) to reduce the drift of the liquid medicine.

[0043] V. Feedback on Control Effects and Model Iteration Time: 72 hours after application Feedback optimization module starts closed-loop verification: The drone revisited the rice fields to collect image sequences, and the residual contrast network quantified the reduction rate of the diseased area (the measured value was 12%, which was lower than the threshold of 15%). Trigger model weight update: Adjust the 3D-ResNet50 network parameters with a learning rate of 0.001 to enhance the prediction accuracy of the lesion expansion rate; The data of this prevention and control (effects of pesticide combinations, ecological regulation responses) will be entered into the regional pesticide resistance database to provide a historical basis for subsequent decision-making.

[0044] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time recognition and decision-making system for rice pests and diseases based on deep learning, characterized by: The system comprises: The image acquisition and enhancement module acquires multispectral images of rice fields and environmental sensor data, eliminates shadow interference through a light-adaptive normalization algorithm, and uses a generative adversarial network to synthesize microscopic features of disease spots to generate enhanced pest and disease feature maps. The multimodal fusion recognition module calls the pest and disease characteristic map, combines real-time meteorological data and the rice genotype database, extracts the spatiotemporal evolution characteristics of lesions through a spatiotemporal convolutional neural network, and outputs the pest and disease type identification result and severity level; The resistance decision engine module, based on the pest and disease type and severity level, associates the regional pesticide use history database and pathogen resistance gene library, and generates resistance avoidance agent combinations and ecological regulation strategies through a multi-agent reinforcement learning model; A dynamic control execution module integrates the resistance-avoidance agent combination and ecological regulation strategy, combines the real-time location of the drone and the tank capacity data, and generates a three-dimensional targeted spray path and natural enemy release coordinate instructions; The feedback optimization module monitors the rice field image sequence within 72 hours after the application of pesticides, quantifies the deviation of the control effect through the residual contrast network, and updates the weight parameters of the multimodal fusion recognition model.

2. The real-time rice pest and disease image recognition and decision-making system based on deep learning according to claim 1, characterized in that: The image acquisition and enhancement module specifically includes shadow elimination parameters, microscopic feature enhancement coefficients and multispectral fusion weights; the multimodal fusion recognition module specifically includes spatiotemporal convolution kernel size, genotype matching threshold and severity classification interval; the resistance decision engine module includes resistance probability prediction value, pesticide synergy coefficient and ecological regulation intensity parameter; the dynamic prevention and control execution module specifically includes spray elevation angle deviation tolerance, pesticide-natural enemy release time difference and three-dimensional path curvature radius; the feedback optimization module includes residual comparison threshold, weight update learning rate and model iteration trigger condition.

3. The real-time rice pest and disease image recognition and decision-making system based on deep learning according to claim 1, characterized in that: The image acquisition and enhancement module includes a shadow elimination submodule, a microscopic synthesis submodule and a feature fusion submodule; The shadow removal submodule captures RGB-NIR images through a polarization filter camera, calculates the covariance matrix of saturation and brightness in the HSV space of the pixel points, uses the Mahalanobis distance threshold to segment the shadow area, and performs adaptive histogram equalization; The microscopic synthesis submodule calls the leaf stomatal structure template library, inputs the visible light image into the pre-trained Pix2Pix-HD network, generates a virtual microscopic image in the 750nm-900nm band, and enhances the morphological characteristics of pathogenic fungi hyphae; The feature fusion submodule performs channel concatenation on the shadow-removed RGB image, virtual microscopic image, and leaf temperature sensor data, uses an attention mechanism for weighted fusion, and outputs a 64-dimensional feature vector.

4. The real-time rice pest and disease image recognition and decision-making system based on deep learning according to claim 1, characterized in that: The multimodal fusion recognition module includes a spatiotemporal feature extraction submodule, a resistance gene matching submodule and a severity grading submodule; The spatiotemporal feature extraction submodule inputs 5 consecutive frames of enhanced feature maps into the 3D-ResNet50 network to extract the lesion expansion rate and direction vector; The resistance gene matching submodule calls the rice variety gene database, compares the pathogen identification results with the resistance gene loci (Pi-ta / Pi-b) using the BLAST algorithm, and outputs the gene escape probability value; The severity grading submodule integrates the proportion of lesion area, chlorophyll fluorescence intensity and transpiration rate reduction ratio, and divides the severity of pests and diseases into 5 levels through a random forest model.

5. The real-time rice pest and disease image recognition and decision-making system based on deep learning according to claim 1, characterized in that: The resistance decision engine module includes a drug game submodule, an ecological regulation submodule and a collaborative optimization submodule; The drug game submodule constructs a three-party game matrix of drug, pathogen and natural enemy, takes minimizing the resistance development rate as the objective function, and solves the Nash equilibrium strategy; The ecological regulation submodule calls the release amount formula of trichogrammatids and ladybugs according to the type of pests and diseases, and generates the release point coordinates based on the topological structure of the rice field; The collaborative optimization submodule adopts the NSGA-II algorithm to optimize the multi-objective pharmaceutical cost and ecological benefit and outputs the Pareto optimal solution set.

6. The deep learning-based real-time rice pest and disease image recognition and decision-making system according to claim 1, characterized in that: The dynamic prevention and control execution module includes: The 3D path planning submodule reconstructs the 3D model of the rice plant based on the UAV LiDAR point cloud data, calculates the spatial coordinates and normal vector direction of the lesion, and generates a contoured spray path; The natural enemy coordination submodule plans the opening and closing sequence of the drone’s drop bay based on the release point coordinates of the ecological control strategy.

7. The deep learning-based real-time rice pest and disease image recognition and decision-making system according to claim 1, characterized in that: The dynamic prevention and control execution module also includes an anti-drift control submodule, which integrates wind speed and direction sensor data and adjusts the spray pressure and droplet size in real time through a PID controller.

Images