Paddy rice disease and insect pest intelligent monitoring system and method based on image recognition

Abstract

The invention provides a rice disease and pest intelligent monitoring system and method based on image recognition, and relates to the technical field of image processing, the rice disease and pest intelligent monitoring system based on image recognition comprises a data set construction module used for collecting field images of diseases and pests, each image is marked with a pest and disease damage type, an occurrence part and a severity level; constructing a standardized data set, segmenting a training set, a verification set and a test set according to 8: 1: 1, and laying a data foundation for model training; the model construction training module is used for training a double-branch deep learning model based on the standardized data set, a backbone network of the double-branch deep learning model adopts EfficientNet-B4 to balance efficiency and precision, and a first branch outputs disease and insect pest type classification through Softmax. According to the invention, the model is constructed to identify the disease and insect pest types and severity of rice, different treatment schemes are recommended, and the disease and insect pest treatment efficiency is improved. The problem of undefined pesticide application for rice diseases and insect pests is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a rice disease and pest intelligent monitoring system and method based on image recognition. BACKGROUND

[0002] The rice disease and pest image recognition technology is a technology that automatically identifies common diseases and pests such as rice blast, sheath blight, and rice borer by shooting field crop images through a mobile phone or professional equipment.

[0003] At present, the rice disease and pest image recognition technology has a preliminary diagnosis capability and can accurately identify the type of disease and pest, but there are still deficiencies in the severity grading judgment of disease and pest, which may lead to a mismatch between the control measures and the actual situation and needs to be improved. SUMMARY

[0004] The purpose of the present application is to provide a rice disease and pest intelligent monitoring system and method based on image recognition to solve the technical problem of the existing technology that there are still deficiencies in the severity grading judgment of disease and pest, leading to a mismatch between the control measures and the actual situation.

[0005] To solve the above technical problems, the present application provides a rice disease and pest intelligent monitoring system based on image recognition, which comprises: A data set construction module is used to collect field images of diseases and pests (such as rice blast, sheath blight, and rice borer), each image is labeled with the type of disease and pest, the occurrence site (leaf / stem / ear), and the severity level (light / medium / heavy, divided according to the proportion of disease spots), a standardized data set is constructed, and the training set, validation set, and test set are divided according to 8:1:1 to lay a data foundation for model training; A model construction and training module is used to train a double-branch deep learning model based on the standardized data set, the backbone network of the double-branch deep learning model adopts EfficientNet-B4 (a version of convolutional neural network) to balance efficiency and accuracy, branch one outputs disease and pest type classification (supports 20+ categories) through Softmax (a kind of normalization function), and branch two predicts the proportion of disease spot area through a regression network (a kind of neural network model); A treatment matching module is used to map the disease and pest type and severity (mild <10%, moderate 10%-30%, severe >30%) output by the model to a three-level treatment scheme: mild disease is preferably prevented by releasing trichogramma (a kind of wasp) to control rice borer, moderate disease adopts a combination of low-toxicity chemical agents (such as 20% tricyclazole + potassium fertilizer to enhance resistance), and severe disease starts emergency control (such as switching to thiamethoxam + unmanned aerial vehicle for unified prevention), all schemes are labeled with safety interval and ecological protection requirements to form a treatment library.

[0006] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring system for rice diseases and pests, and a model construction training module comprises: A model construction unit is configured to design a double-branch convolutional neural network, wherein a pre-trained EfficientNet-B4 is used as a backbone network to extract general features, a branch one is connected with a global pooling layer and a Softmax output layer to realize 20+ disease and pest type classification (e.g., rice blast and sheath blight), and a branch two is configured to build a regression network, wherein a hollow convolutional layer is used to capture disease spot details, a fully connected layer is connected to output a disease spot area ratio (0%-100%), and the two branches share shallow features to improve inference efficiency. A model training unit is configured to use a two-stage training model, wherein in a first stage, the backbone network is frozen, and only the head parameters of the branch one and the branch two are optimized, the branch one uses a Focal Loss (a loss function designed for class imbalance) (α=0.8-1.2) to alleviate the class imbalance problem of diseases and pests, and the branch two uses a Smooth L1 Loss (a regression loss function) (δ=0.1) to improve the noise robustness of the disease spot area ratio regression; in a second stage, the backbone network is unfrozen for joint fine-tuning, the backbone feature extractor and the double-branch parameters are simultaneously optimized (the learning rate is reduced to 5e-5), and data enhancement techniques are applied throughout the training process, including random occlusion (30% probability to simulate leaf overlap), HSV color perturbation (±20% saturation / brightness), and Gaussian noise (σ=0.05 to simulate rain and fog), and a validation set early stopping mechanism (terminated when the loss does not decrease for 5 consecutive rounds) is used to ensure model convergence. A model evaluation and compression unit is configured to use an independent test set to evaluate the model performance, wherein the type classification accuracy needs to be above 90%, and the severity regression error needs to be within ±7%; and the model is pruned in channels and quantized in INT8 (a technique for converting floating-point parameters in a deep learning model to 8-bit integers), and the model size is compressed to below 15MB.

[0007] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring system for rice diseases and pests, and further comprises: A severity threshold adjustment module is configured to build an environmental correction coefficient matrix, and access real-time data (temperature / humidity / rainfall) from a weather station, and when the temperature is greater than 28℃ and the humidity is greater than 90%, the severity threshold is automatically compressed by 20% (for example, the original moderate 10%-30% is adjusted to 8%-24%). A dosage adjustment module is configured to dynamically adjust the dosage, based on the corrected severity, call a pesticide response curve model (e.g., a pyraclostrobin dosage formula for rice blast), and combine the rice growth period (tiller stage with weak drug resistance, which needs to be reduced by 10%) to output a precise dose treatment plan.

[0008] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring system for rice diseases and pests, which further comprises: A mild infection severe treatment module is configured to provide a severe treatment scheme when the model detects that a specific high recurrence risk disease and pest (e.g., rice sheath blight) is in a mild infection stage, to induce rice system resistance (SAR) through a plant immune activator (chitosan, etc.), and to activate a salicylic acid signaling pathway to enhance immune memory.

[0009] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring system for rice diseases and pests, which further comprises: A model optimization module is configured to establish a dynamic optimization closed loop, to automatically screen low confidence samples (e.g., a rice blast probability of 42%+ a sheath blight probability of 38%, with a ranking of the first confidence < 50%), and to submit the low confidence samples to experts for review, and to update the model according to a period. A regional differentiation module is configured to construct regional sub-models (e.g., a southern China early rice area, a northeast japonica rice area, etc.), and to adapt to differences in disease and pest spectrum. A treatment library optimization module is configured to periodically update a treatment library based on treatment means alternation, and to ensure the timeliness of the scheme.

[0010] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring method for rice diseases and pests, which comprises the following steps: Collecting field images of diseases and pests (e.g., rice blast, sheath blight, and rice borer), marking the type of disease and pest, the occurrence site (leaf / stem / ear), and the severity level (light / medium / heavy, divided according to the proportion of disease spots), and constructing a standardized data set, and dividing the training set, the validation set, and the test set according to 8:1:1, to lay a data foundation for model training; Training a double-branch deep learning model based on the standardized data set, with the backbone network of the double-branch deep learning model adopting EfficientNet-B4 to balance efficiency and accuracy, with branch one outputting disease and pest type classification (supporting 20+ categories) through Softmax, and with branch two predicting the proportion of disease spot area through a regression network; Mapping the disease and pest type and severity (light < 10%, medium 10%-30%, and heavy > 30%) output by the model to a three-level treatment scheme: mild prevention and control (e.g., releasing trichogramma to control rice borer), medium treatment using a low-toxicity chemical agent combination (e.g., 20% tricyclazole + potassium fertilizer to enhance resistance), and heavy treatment starting emergency prevention and control (e.g., switching thiamethoxam to drug + unmanned aerial vehicle unified prevention), with all the schemes being marked with a safety interval and ecological protection requirements, to form a treatment library.

[0011] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring method for rice diseases and pests, which trains a double-branch deep learning model based on a standardized data set. The backbone network of the double-branch deep learning model balances efficiency and accuracy using EfficientNet-B4. Branch one outputs disease and pest type classification (supports 20+ categories) through Softmax. Branch two predicts the disease spot area ratio in the steps, specifically including: A double-branch convolutional neural network is designed. The backbone network uses a pre-trained EfficientNet-B4 to extract general features. Branch one is connected to a global pooling layer and a Softmax output layer to realize 20+ disease and pest type classification (such as rice blast and sheath blight). Branch two constructs a regression network to capture disease spot details through a dilated convolution layer and outputs the disease spot area ratio (0%-100%) through a fully connected layer. The double-branch shares shallow features to improve inference efficiency. The model is trained in two stages. In the first stage, the backbone network is frozen, and only the head parameters of branch one and branch two are optimized. Branch one uses Focal Loss (α=0.8-1.2) to alleviate the imbalance of disease and pest categories, and branch two uses Smooth L1 Loss (δ=0.1) to improve the noise robustness of disease spot area ratio regression. In the second stage, the backbone network is unfrozen for joint fine-tuning. The backbone feature extractor and double-branch parameters are optimized simultaneously (learning rate reduced to 5e-5). Data augmentation techniques are applied throughout the training process, including random occlusion (30% probability to simulate leaf overlap), HSV color perturbation (±20% saturation / luminance), and Gaussian noise (σ=0.05 to simulate rain and fog). The validation set early stopping mechanism (terminated if the loss does not decrease for 5 consecutive rounds) ensures model convergence. The model performance is evaluated using an independent test set. The type classification accuracy should be above 90%, and the severity regression error should be within ±7%. The channel pruning and INT8 quantization are performed on the qualified model to compress the model size to below 15MB.

[0012] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring method for rice diseases and pests, which further includes: An environment correction coefficient matrix is constructed to access real-time data from a weather station (temperature / humidity / rainfall). When the temperature is greater than 28°C and the humidity is greater than 90%, the severity threshold is automatically compressed by 20% (for example, the original moderate 10%-30% is adjusted to 8%-24%). The dosage is dynamically adjusted based on the corrected severity. The pesticide response curve model (such as the rice blast tricyclazole dosage formula) is called to output the precise dose treatment plan in combination with the rice growth period (tiller stage with weak drug resistance, requiring a 10% reduction).

[0013] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring method for rice diseases and pests, which further comprises: When the model detects that a specific high-recurrence-risk disease and pest (such as rice sheath blight) is in a mild infection stage, a severe treatment scheme is provided to induce the rice system to obtain resistance (SAR) through a plant immune activator (chitosan, etc.) and activate the salicylic acid signaling pathway to enhance immune memory.

[0014] In one of the embodiments, the present application provides an image recognition-based intelligent monitoring method for rice diseases and pests, which further comprises: A dynamic optimization closed loop is established to automatically screen low-confidence samples (such as a rice blast probability of 42%+ a sheath blight probability of 38%, with a ranking of the first confidence <50%) for expert review, and the model is updated periodically; Regional sub-models (for example, a South China early rice area and a Northeast japonica rice area) are constructed to adapt to differences in disease and pest spectrum; a treatment library is updated periodically based on changes in treatment means to ensure the timeliness of the scheme.

[0015] By constructing a model to identify the type and severity of rice diseases and pests, the present application recommends different treatment schemes to solve the problem of unclear medication for rice diseases and pests. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the following description of the specific embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0017] Figure 1 FIG. 1 is a first part schematic diagram of an image recognition-based intelligent monitoring system for rice diseases and pests.

[0018] Figure 2 FIG. 4 is a schematic diagram of a model construction training module.

[0019] Figure 3 FIG. 7 is a second part schematic diagram of an image recognition-based intelligent monitoring system for rice diseases and pests.

[0020] Figure 4 FIG. 10 is a third part schematic diagram of an image recognition-based intelligent monitoring system for rice diseases and pests.

[0021] Figure 5 FIG. 13 is a fourth part schematic diagram of an image recognition-based intelligent monitoring system for rice diseases and pests.

[0022] Figure 6A first part schematic diagram of an intelligent monitoring method for rice diseases and insect pests based on image recognition.

[0023] Figure 7 A flowchart of a deep learning model construction and training.

[0024] Figure 8 A second part schematic diagram of an intelligent monitoring method for rice diseases and insect pests based on image recognition.

[0025] Figure 9 A third part schematic diagram of an intelligent monitoring method for rice diseases and insect pests based on image recognition.

[0026] Figure 10 A fourth part schematic diagram of an intelligent monitoring method for rice diseases and insect pests based on image recognition. DETAILED DESCRIPTION

[0027] The technical solutions of the present application will be described below in conjunction with the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the scope of the present application.

[0028] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0029] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0030] The present application will be further explained in conjunction with the specific embodiments.

[0031] In one embodiment, as shown in Figure 1 An intelligent monitoring system for rice diseases and insect pests based on image recognition, comprising: Dataset construction module 1 is used to collect field images of pests and diseases (such as rice blast, sheath blight, rice stem borer, etc.). Each image is labeled with the type of pest or disease, the affected part (leaf / stem / ear), and the severity level (light / moderate / severe, based on the proportion of lesions). A standardized dataset is constructed and the training set, validation set, and test set are divided in an 8:1:1 ratio to lay the data foundation for model training. Model building and training module 2 is used to train a dual-branch deep learning model based on a standardized dataset. The backbone network of the dual-branch deep learning model adopts EfficientNet-B4 (a version of a convolutional neural network) to balance efficiency and accuracy. Branch 1 outputs pest and disease type classification (supporting 20+ categories) through Softmax (a normalization function), and Branch 2 predicts the proportion of lesion area through a regression network (a god-level network model). The treatment matching module 3 is used to map the pest and disease types and severity (mild <10%, moderate 10%-30%, severe >30%) output by the model to three-level treatment plans: mild cases prioritize biological control (such as releasing Trichogramma wasps to control stem borers), moderate cases use a combination of low-toxicity chemical agents (such as 20% tricyclazole + potassium fertilizer to enhance resistance), and severe cases initiate emergency control (such as rotating thiamethoxam + drone-based integrated control). All plans are marked with safe intervals and ecological protection requirements, forming a treatment library.

[0032] Dataset Construction Module 1: Field images are labeled with pest and disease types, locations, and severity levels (based on lesion percentage). A standardized dataset is constructed and divided into training / validation / test sets (8:1:1) to provide a structured data foundation for the model. This ensures that the model learns the multidimensional features of pests and diseases (type + spatial distribution + severity), and that data segmentation conforms to standard machine learning procedures to prevent overfitting.

[0033] Model building training module 2: The EfficientNet-B4 backbone network is selected to balance computational efficiency and accuracy; branch one uses Softmax to achieve multi-type classification (>20 classes), and branch two uses a regression network to quantify the lesion area ratio. The parallel implementation of the two tasks improves the comprehensiveness of diagnosis.

[0034] Treatment Matching Module 3: Maps the model output type and severity (mild <10%, moderate 10%-30%, severe >30%) to a three-tiered treatment plan, prioritizing biological control (environmental protection), using low-toxicity agents for moderate cases (safety), and emergency prevention and control for severe cases (rapid response), and marking the safety interval and ecological requirements to ensure the scientific nature and operability of the prevention and control measures.

[0035] In one embodiment, such as Figure 2 As shown, an intelligent monitoring system for rice diseases and pests based on image recognition includes a model building and training module 2 comprising: The model construction unit 21 is used for designing a double-branch convolutional neural network: the backbone network adopts a pre-trained EfficientNet-B4 to extract general features, branch one is connected with a global pooling layer + a Softmax output layer to realize 20+ pest and disease type classification (such as rice blast, sheath blight, etc.); branch two is constructed as a regression network, the disease spot details are captured through an atrous convolution layer, and a fully connected layer is connected to output the disease spot area ratio (0%-100%), and the double branches share shallow features to improve inference efficiency. The model training unit 22 is used for training the model in two stages: in the first stage, the backbone network is frozen, and only the head parameters of branch one and branch two are optimized, wherein branch one uses Focal Loss (a loss function designed for class imbalance) (α=0.8-1.2) to alleviate the class imbalance problem of pests and diseases, and branch two uses Smooth L1 Loss (a regression loss function) (δ=0.1) to improve the noise robustness of the disease spot area ratio regression; in the second stage, the backbone network is unfrozen for joint fine-tuning, and the backbone feature extractor and the double-branch parameters are optimized synchronously (the learning rate is reduced to 5e-5), and data enhancement techniques are applied throughout the training process, including random occlusion (30% probability to simulate leaf overlap), HSV color perturbation (±20% saturation / luminance) and Gaussian noise (σ=0.05 to simulate rain and fog), and the model convergence is ensured through the validation set early stopping mechanism (terminated if the loss does not decrease for 5 consecutive rounds). The model evaluation and compression unit 23 is used for evaluating the model performance using an independent test set: the type classification accuracy needs to reach more than 90%, and the severity regression error needs to be within ±7%; the qualified model is subjected to channel pruning and INT8 quantization (a technique for converting floating-point parameters in a deep learning model to 8-bit integers), and the model size is compressed to less than 15MB.

[0036] The model construction unit 21: the backbone adopts EfficientNet-B4 to extract general features; branch one realizes efficient classification through global pooling + Softmax; branch two uses atrous convolution to capture disease spot details and improve area ratio regression accuracy; and the double branches share shallow features to reduce computational redundancy.

[0037] The training strategy of the model training unit 22: Two-stage training: first, freeze the backbone to optimize the branch head (reduce early training fluctuations), and then unfreeze the backbone for joint fine-tuning (learning rate 5e-5 to avoid damaging the pre-trained weights).

[0038] Loss function: branch one uses Focal Loss (α=0.8-1.2) to alleviate sample imbalance (such as rare disease cases); branch two uses Smooth L1 Loss (δ=0.1) to suppress regression noise.

[0039] Data augmentation: random occlusion (simulate leaf overlap), HSV perturbation (adapt to light changes), Gaussian noise (simulate rain and fog), improve model robustness.

[0040] Model evaluation compression unit 23: test set requires classification accuracy > 90%, severity error ± 7%, after reaching the standard, pass through channel pruning + INT8 quantization to compress the model to less than 15MB, meet the lightweight deployment needs of edge devices (such as field drones).

[0041] In one embodiment, as shown in Figure 3 The image recognition-based intelligent rice disease and pest monitoring system further comprises: The severity threshold adjustment module 4 is configured to construct an environmental correction coefficient matrix and access real-time data (temperature / humidity / rainfall) from a weather station. When the temperature is greater than 28°C and the humidity is greater than 90%, the severity threshold is automatically compressed by 20% (for example, the original moderate range of 10%-30% is adjusted to 8%-24%). The dosage adjustment module 5 is configured to dynamically adjust the dosage based on the corrected severity, call a pesticide response curve model (such as the triadazole dosage formula for rice blast), and output a precise dosage treatment plan in combination with the rice growth period (reduce the dosage by 10% during the tillering period due to weak resistance).

[0042] The severity threshold adjustment module 4 implements environmental correction. The spread of diseases and pests is significantly affected by temperature and humidity (for example, rice blast spreads faster in high temperature and high humidity). When the temperature is greater than 28°C and the humidity is greater than 90%, the severity threshold is compressed by 20% (for example, the moderate range of 10%-30% is adjusted to 8%-24%). By dynamically adjusting the threshold, the risk is warned in advance, and the prevention and control opportunity is avoided.

[0043] The dosage adjustment module 5 implements precise dosing. Based on the corrected severity, a pesticide response curve (such as the triadazole dosage formula) is called, and the dosage is dynamically adjusted in combination with the rice growth period (reduce the dosage by 10% during the tillering period due to weak resistance), reducing the risk of pesticide damage and improving the efficiency of prevention and control.

[0044] In one embodiment, as shown in Figure 4 The image recognition-based intelligent rice disease and pest monitoring system further comprises: The mild infection severe treatment module 6 is configured to provide a severe treatment plan when the model detects that a specific high-recurrence-risk disease and pest (such as rice smut) is in the mild infection stage. By using a plant immune activator (chitosan, etc.), the rice system is induced to obtain resistance (SAR), and the salicylic acid signaling pathway is activated to enhance immune memory.

[0045] For specific high recurrence risk pests and diseases (such as rice sheath blight), a high-intensity biological intervention is initiated in the early stage of infection: the rice system resistance (SAR) is induced by plant immune activators (chitosan, etc.), combined with resistance variety adaptation and environmental regulation, to convert the mild stage into an "immune training window". Although the short-term investment increases, the use of plant immune memory effect significantly reduces the probability of later outbreaks and chemical pesticide dependence, achieving a decrease in the overall cycle cost.

[0046] In one embodiment, as shown in Figure 5 An intelligent rice pest and disease monitoring system based on image recognition further comprises: A model optimization module 7 is configured to establish a dynamic optimization closed loop, automatically screen low confidence samples (such as rice blast probability 42% + sheath blight probability 38%, and the first ranked confidence < 50%), and submit them to experts for review, and update the model periodically; A regional differentiation module 8 is configured to construct regional sub-models (for example, South China early rice area, Northeast glutinous rice area, etc.), and adapt to differences in pest spectrum; A treatment library optimization module 9 is configured to periodically update the treatment library based on the alternation of treatment methods, and ensure the timeliness of the solution.

[0047] The model optimization module 7: automatically screens low confidence samples (such as ambiguous cases with the highest probability < 50%) for expert review, continuously supplements high-quality training data, solves the identification bottleneck of model boundary cases, and ensures the continuous evolution of the system.

[0048] The regional differentiation module 8: for ecological differences such as the South China early rice area (high temperature and high humidity pest spectrum) and the Northeast glutinous rice area (low temperature diseases), train customized sub-models to avoid misjudgment and improve regional adaptability.

[0049] The treatment library optimization module 9: periodically synchronizes new pesticides (such as resistance replacement agents) and biological control techniques (such as new natural enemy insects) to ensure the timeliness of the solution. This design responds to changes in pests and diseases and changes in pesticide policies, and maintains the long-term effectiveness of the decision-making system.

[0050] In one embodiment, as shown in Figure 6 An intelligent rice pest and disease monitoring method based on image recognition comprises the following steps: Step S1: Collect field images of pests and diseases (such as rice blast, sheath blight, and rice borer), and mark the type of pests and diseases, the occurrence site (leaf / stem / panicle), and the severity level (light / medium / heavy, divided according to the proportion of disease spots) on each image; construct a standardized data set, and divide the training set, the validation set, and the test set in the ratio of 8:1:1 to lay a data foundation for model training; Step S2, train the dual-branch deep learning model based on the standardized dataset, the backbone network of the dual-branch deep learning model adopts EfficientNet-B4 to balance efficiency and accuracy, branch one outputs disease and pest type classification (supports 20+ categories) through Softmax, and branch two predicts the area ratio of disease spots through a regression network; Step S3, map the disease and pest type and severity (mild <10%, moderate 10%-30%, severe >30%) output by the model to a three-level treatment plan: mild prevention and control of biological agents (such as releasing trichogramma to control pests), moderate use of low-toxicity chemical agents (such as 20% tricyclazole + potassium fertilizer to enhance resistance), and severe emergency prevention and control (such as switching to thiamethoxam + unmanned aerial vehicle unified prevention), all plans are marked with safety intervals and ecological protection requirements, forming a treatment library.

[0051] The severity classification design is based on: 1. Pathology basis Studies on rice diseases (such as rice blast) have shown that: Disease spot ratio <10%: the disease is in the early stage of infection, the crop has strong self-repairing ability, and the impact on yield is limited (usually <5%); Disease spot ratio 10%-30%: the disease enters the expansion period, photosynthesis is significantly hindered, and yield loss can reach 5%-20%, requiring rapid intervention; Disease spot ratio >30%: the pathogen reproduces on a large scale, may cause regional epidemics, and yield loss exceeds 20% and is prone to absolute loss.

[0052] 2. Prevention and control efficiency optimization Threshold 10%: is the critical point of economic injury level (EIL), below which biological control is more cost-effective; Threshold 30%: beyond which the disease develops exponentially and must start strong chemical control to block the transmission chain.

[0053] 3. Operational feasibility This division conforms to industry standards such as the Agricultural Department's "Rice Blast Reporting Technical Specification" (GB / T 15790-2021), facilitating integration with traditional plant protection systems and ensuring intuitive understanding of the classification by farmers and experts.

[0054] In one embodiment, as Figure 7As shown, an image recognition-based intelligent monitoring method for rice diseases and pests, the step S2, based on the standardized data set, a double-branch deep learning model is trained, the backbone network of the double-branch deep learning model adopts EfficientNet-B4 to balance efficiency and accuracy, branch one outputs disease and pest type classification (supports 20+ categories) through Softmax, and branch two predicts disease spot area ratio through a regression network. The step specifically includes: Step S21, a double-branch convolutional neural network is designed: the backbone network adopts a pre-trained EfficientNet-B4 to extract general features, branch one is connected with a global pooling layer + a Softmax output layer to realize 20+ disease and pest type classification (such as rice blast, sheath blight, etc.); branch two constructs a regression network, captures disease spot details through a dilated convolution layer, and outputs disease spot area ratio (0%-100%) through a fully connected layer. The double branches share shallow features to improve inference efficiency; Step S22, the model is trained in two stages: in the first stage, the backbone network is frozen, and only the head parameters of branch one and branch two are optimized, wherein branch one uses Focal Loss (a=0.8-1.2) to alleviate the imbalance problem of disease and pest categories, and branch two uses Smooth L1 Loss (δ=0.1) to improve the noise robustness of disease spot area ratio regression; in the second stage, the backbone network is unfrozen for joint fine-tuning, and the backbone feature extractor and double-branch parameters are optimized simultaneously (learning rate is reduced to 5e-5). At the same time, data augmentation techniques are applied throughout the training process, including random occlusion (30% probability to simulate leaf overlap), HSV color perturbation (±20% saturation / luminance), and Gaussian noise (σ=0.05 to simulate rain and fog). The validation set early stopping mechanism (terminates if the loss does not decrease for 5 consecutive rounds) ensures model convergence; Step S23, use an independent test set to evaluate model performance: type classification accuracy needs to reach more than 90%, and severity regression error needs to be within ±7%; for the qualified model, perform channel pruning and INT8 quantization to compress the model size to less than 15MB.

[0055] In Focal Loss (a=0.8-1.2), a is a class weight factor, a>1 increases the weight of rare classes (such as white leaf blight), and a<1 suppresses too many sample classes (such as rice blast). Dynamic range 0.8-1.2: through grid search (GridSearch) verification, when the proportion of rare diseases is <5%, set a=1.2 (increase the weight), when the proportion of common diseases is >30%, set a=0.8 (avoid dominant loss function), balance recall rate and accuracy.

[0056] Smooth L1 Loss (δ=0.1): Smooth L1 Loss is a regression loss function that combines the advantages of L1 loss (absolute error) and L2 loss (squared error). δ is the loss function switching threshold (use L2 loss for fine optimization when |error| < δ, and use L1 loss when > δ to resist noise). Set δ = 0.1 (i.e. 10%): due to the blurred boundaries of field image disease spots (e.g. edge diffusion caused by rain erosion), allow fine learning gradient within ±10% error, and use L1 to suppress abnormal value interference when exceeding.

[0057] In one embodiment, as shown in Figure 8 An image recognition-based intelligent monitoring method for rice diseases and pests further comprises: Step S4: Construct an environmental correction coefficient matrix and access real-time data from a weather station (temperature / humidity / rainfall). When the temperature is > 28°C and the humidity is > 90%, automatically compress the severity threshold by 20% (example: original moderate 10%-30% adjusted to 8%-24%). Step S5: Dynamically adjust the dosage based on the corrected severity, call the pesticide response curve model (such as the pyraclostrobin dosage formula for rice blast), and combine the rice growth period (tiller stage with weak resistance requiring a 10% reduction) to output a precise dosage treatment plan.

[0058] Design 28°C temperature threshold: The spore germination rate of Magnaporthe oryzae reaches a peak at 25-32°C (Japanese Plant Pathology Society, 2019), and 28°C is the critical temperature for infection force mutation.

[0059] Design 90% humidity threshold: Spore invasion of rice stomata requires continuous leaf wetness ≥ 8 hours (IRRI International Rice Research Institute data), and when the humidity is > 90%, the leaf condensation duration can meet this condition.

[0060] Double-factor superposition: When the temperature is > 28°C and the humidity is > 90%, the latent period of rice blast is shortened from 7 days to 3 days (experimental data from the Chinese Academy of Agricultural Sciences), and the disease spread rate is increased by 2.1 times.

[0061] In one embodiment, as shown in Figure 9 An image recognition-based intelligent monitoring method for rice diseases and pests further comprises: Step S6: When the model detects a specific high-recurrence-risk disease or pest (such as rice smut) in the mild infection stage, provide a severe treatment plan, induce rice system resistance (SAR) through plant immune activators (chitosan, etc.), and activate the salicylic acid signaling pathway to enhance immune memory.

[0062] For example, in the case of rice sheath blight, chitosan (COS) can be compounded with methyl jasmonate (MeJA) or β-aminobutyric acid (BABA) to activate the salicylic acid (SA) and jasmonic acid (JA) dual signaling pathways. This avoids the rice sheath blight fungus escaping the single salicylic acid pathway defense, and the jasmonic acid pathway can enhance cell wall lignification (such as increased phenylalanine ammonia lyase activity).

[0063] In one embodiment, as shown in Figure 10 A rice disease and pest intelligent monitoring method based on image recognition further comprises: Step S7, a dynamic optimization closed loop is established to automatically screen low confidence samples (such as rice blast probability 42%+ sheath blight probability 38%, ranking first confidence <50%) for expert review, and the model is updated periodically; Step S8, a regional sub-model (for example, South China early rice area, Northeast japonica rice area, etc.) is constructed to adapt to the differences in disease and pest spectrum; Step S9, based on the alternation of treatment means, the treatment library is updated periodically to ensure the timeliness of the scheme.

[0064] When the treatment library is updated, the new scheme (such as the proportion of pesticide combination) can be written into the smart contract of the blockchain, and the spraying record, environmental data and control effect are associated. If the new scheme data is not effective, the treatment library can be removed.

[0065] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A smart monitoring system for rice diseases and pests based on image recognition, characterized in that, This image recognition-based intelligent monitoring system for rice diseases and pests includes: The dataset construction module is used to collect field images of pests and diseases, and each image is labeled with the type of pest or disease, the location of occurrence, and the severity level; a standardized dataset is constructed, and the training set, validation set, and test set are split in an 8:1:1 ratio to lay the data foundation for model training; The model building and training module is used to train a dual-branch deep learning model based on a standardized dataset. The backbone network of the dual-branch deep learning model adopts EfficientNet-B4 to balance efficiency and accuracy. Branch 1 outputs the classification of pest and disease types through Softmax, and Branch 2 predicts the proportion of lesion area through a regression network. The treatment matching module maps the pest and disease types and severity output by the model to three-tiered treatment plans: mild cases prioritize biological control, moderate cases use a combination of low-toxicity chemical agents, and severe cases initiate emergency control measures. All plans are marked with safe intervals and ecological protection requirements, forming a treatment library.

2. The intelligent monitoring system for rice diseases and pests based on image recognition according to claim 1, characterized in that, The model building and training module includes: The model building unit is used to design a dual-branch convolutional neural network: the backbone network uses a pre-trained EfficientNet-B4 to extract general features, and branch one connects a global pooling layer + a Softmax output layer to achieve classification of 20+ types of pests and diseases; branch two builds a regression network, captures lesion details through dilated convolutional layers, and outputs the lesion area ratio through a fully connected layer. The two branches share shallow features to improve inference efficiency. The model training unit employs a two-stage training approach: In the first stage, the backbone network is frozen, and only the head parameters of branch one and branch two are optimized. Branch one uses Focal Loss to mitigate the imbalance of pest and disease categories, while branch two uses Smooth L1 Loss to improve the noise robustness of lesion area proportion regression. In the second stage, the backbone network is unfrozen for joint fine-tuning, simultaneously optimizing the backbone feature extractor and dual-branch parameters. Data augmentation techniques are applied throughout the training process, including random occlusion, HSV color perturbation, and Gaussian noise. An early stopping mechanism on the validation set ensures model convergence. The model evaluation compression unit is used to evaluate model performance using an independent test set: the type classification accuracy should be above 90%, and the severity regression error should be within ±7%; for qualified models, channel pruning and INT8 quantization are performed to compress the model size to below 15MB.

3. The intelligent monitoring system for rice diseases and pests based on image recognition according to claim 1, characterized in that, The image recognition-based intelligent monitoring system for rice diseases and pests also includes: The severity threshold adjustment module is used to construct an environmental correction coefficient matrix and input real-time weather station data. When the temperature is >28℃ and the humidity is >90%, the severity threshold is automatically compressed by 20%. The dosage adjustment module is used to dynamically adjust the dosage. Based on the corrected severity, it calls the pesticide response curve model and outputs a precise dosage treatment plan in combination with the rice growth stage.

4. The intelligent monitoring system for rice diseases and pests based on image recognition according to claim 1 or 3, characterized in that, The image recognition-based intelligent monitoring system for rice diseases and pests also includes: The mild infection and severe treatment module is used to provide severe treatment when the model detects that a specific high-recurrence-risk pest or disease is in the mild infection stage. It induces resistance in the rice system through plant immune activators and activates the salicylic acid signaling pathway to enhance immune memory.

5. The intelligent monitoring system for rice diseases and pests based on image recognition according to claim 1, characterized in that, The image recognition-based intelligent monitoring system for rice diseases and pests also includes: The model optimization module is used to establish a dynamic optimization closed loop, automatically screen low-confidence samples for expert review, and update the model periodically. The regional differentiation module is used to build regional sub-models to adapt to differences in the spectrum of diseases and pests; The treatment library optimization module is used to periodically update the treatment library based on changes in treatment methods to ensure the timeliness of the solutions.

6. A method for intelligent monitoring of rice diseases and pests based on image recognition, characterized in that, The image recognition-based intelligent monitoring method for rice diseases and pests includes the following steps: Collect field images of pests and diseases, and label each image with the type of pest or disease, the location of occurrence, and the severity level; construct a standardized dataset, and split the training set, validation set, and test set in an 8:1:1 ratio to lay the data foundation for model training; A dual-branch deep learning model is trained based on a standardized dataset. The backbone network of the dual-branch deep learning model adopts EfficientNet-B4 to balance efficiency and accuracy. Branch 1 outputs the classification of pest and disease types through Softmax, and Branch 2 predicts the proportion of lesion area through a regression network. The model outputs pest and disease types and severity levels are mapped to three levels of treatment plans: mild cases prioritize biological control, moderate cases use a combination of low-toxicity chemical agents, and severe cases initiate emergency control measures. All plans are marked with safety intervals and ecological protection requirements, forming a treatment library.

7. The intelligent monitoring method for rice diseases and pests based on image recognition according to claim 6, characterized in that, The step of training a dual-branch deep learning model based on a standardized dataset, where the backbone network of the dual-branch deep learning model uses EfficientNet-B4 to balance efficiency and accuracy, and branch one outputs pest and disease type classification through Softmax, while branch two predicts the proportion of lesion area through a regression network, specifically includes: Design a dual-branch convolutional neural network: The backbone network uses pre-trained EfficientNet-B4 to extract general features. Branch 1 connects a global pooling layer and a Softmax output layer to achieve classification of 20+ types of pests and diseases. Branch 2 constructs a regression network, which captures lesion details through dilated convolutional layers and outputs the lesion area ratio through fully connected layers. The two branches share shallow features to improve inference efficiency. A two-stage training model is adopted: In the first stage, the backbone network is frozen, and only the head parameters of branch 1 and branch 2 are optimized. Branch 1 uses Focal Loss to alleviate the imbalance of pest and disease categories, while branch 2 uses Smooth L1 Loss to improve the noise robustness of the regression of lesion area proportion. In the second stage, the backbone network is unfrozen and jointly fine-tuned, and the backbone feature extractor and dual-branch parameters are optimized in a synchronous manner. At the same time, data augmentation techniques are applied throughout the training process, including random occlusion, HSV color perturbation, and Gaussian noise. The model convergence is ensured by the early stopping mechanism on the validation set. Evaluate model performance using an independent test set: the type classification accuracy should be above 90%, and the severity regression error should be within ±7%; perform channel pruning and INT8 quantization on qualified models to compress the model size to below 15MB.

8. The intelligent monitoring method for rice diseases and pests based on image recognition according to claim 6, characterized in that, The image recognition-based intelligent monitoring method for rice diseases and pests also includes: An environmental correction coefficient matrix is ​​constructed, and real-time weather station data is input. When the temperature is greater than 28℃ and the humidity is greater than 90%, the severity threshold is automatically compressed by 20%. The dosage is dynamically adjusted, and based on the corrected severity, the pesticide response curve model is invoked to output a precise dosage treatment plan in combination with the rice growth stage.

9. The intelligent monitoring method for rice diseases and pests based on image recognition according to claim 6 or 8, characterized in that, The image recognition-based intelligent monitoring method for rice diseases and pests also includes: When the model detects that a specific high-recurrence-risk pest or disease is in a mild infection stage, it provides a severe treatment plan, which induces the rice system to acquire resistance through plant immune activators and activates the salicylic acid signaling pathway to enhance immune memory.

10. The intelligent monitoring method for rice diseases and pests based on image recognition according to claim 6, characterized in that, The image recognition-based intelligent monitoring method for rice diseases and pests also includes: Establish a dynamic optimization closed loop, automatically screen low-confidence samples for expert review, and update the model periodically; Construct regional sub-models to adapt to differences in the spectrum of diseases and pests; Based on the evolution of treatment methods, the treatment database is updated periodically to ensure the timeliness of the treatment plans.

Images