Crop disease identification and adaptive decision-making method

Abstract

The invention belongs to the technical field of intelligent agriculture, and particularly discloses a crop disease identification and adaptive decision method. The method disclosed by the invention comprises the following steps: acquiring an image and an environment vector, and performing multi-scale enhancement processing on the image; then, processing the enhanced image by using a semi-supervised learning detection model based on a teacher-student architecture, and generating disease data; thirdly, fusing the disease data and the environment vector through a risk function to obtain a disease risk value; thirdly, normalizing the disease risk value, performing dynamic risk prediction based on a set threshold value, and generating a control parameter vector for prevention and control through a mapping function; and finally, performing iterative updating on the weight coefficient of the risk function by using a gradient descent method according to the calculated error and the actual error. According to the method, the dependence on annotated data is reduced, edge deployment is adapted, and the efficiency of a disease identification and prevention system is improved. The composition can be widely applied to disease control of crops.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent agriculture, and specifically relates to a crop disease identification and adaptive decision-making method. BACKGROUND

[0002] Under the background of agricultural intelligence development, the automatic detection and precise spraying system of crop diseases based on computer vision has become a key component of modern unmanned farms. The existing technical solutions are mainly based on full-supervised deep learning frameworks such as YOLO, Faster R-CNN, etc., to build a control closed loop from image acquisition, cloud reasoning to instruction generation and execution.

[0003] Specifically, the existing system periodically collects crop leaf images through camera devices deployed in the field or mobile platforms and transmits them to the server. On the server side, a detection model trained using a large amount of manually annotated data is used to identify disease areas and calculate the area ratio of disease spots. At the same time, environmental sensors independently collect data such as temperature, humidity, wind speed, and light. The decision-making link usually uses simple rules based on fixed thresholds to compare the results with fixed thresholds to generate a binary control signal, for example: when the detected disease spot area ratio exceeds a certain set threshold (such as 10%) and the environmental humidity is greater than 85%, the spraying instruction is triggered to execute spraying, otherwise it is not executed.

[0004] Although such systems can achieve certain results in a controllable environment, they have significant shortcomings:

[0005] 1. High labeling cost and poor generalization: Full-supervised models rely heavily on a large amount of high-quality manual annotation data, and each new crop or new disease requires re-labeling, which is costly; and the model has limited generalization ability in different crops, light, and background;

[0006] 2. Large computational overhead, difficult to deploy at the edge: mainstream detection models have high computational complexity (about 10 GFLOPs) and large parameter quantity (about 70 M), making it difficult to achieve real-time reasoning on limited agricultural edge devices;

[0007] 3. Rigid decision-making mechanism, lack of adaptability: the control strategy based on fixed thresholds does not consider the dynamic diffusion trend of diseases and the coupling relationship between environmental factors, and cannot flexibly adjust the prevention and control intensity according to real-time risks, which easily leads to prevention and control lag or excessive use of pesticides;

[0008] 4. Lack of self-optimization ability: system parameters are usually set and adjusted manually, and the system does not have the ability to perform self-feedback and optimization based on historical execution results, limiting the long-term intelligent level of the system.

[0009] Therefore, in order to solve the above problems, there is an urgent need for a crop disease intelligent identification and prevention method which can reduce the dependence on labeling, adapt to edge computing, and conduct dynamic risk assessment and adaptive decision-making. SUMMARY

[0010] The purpose of the present application is to provide a crop disease identification and adaptive decision-making method, which reduces the dependence on labeled data by introducing semi-supervised learning, designs a lightweight network structure to adapt to edge deployment, constructs a dynamic risk assessment model that integrates time series and environmental information, and an adaptive decision-making and optimization mechanism based on risk feedback, thereby improving the efficiency of the disease identification and prevention system.

[0011] To achieve the above purpose, the technical solution adopted by the present application is as follows:

[0012] A crop disease identification and adaptive decision-making method, the method comprising the following steps in sequence:

[0013] S1, acquiring time series images of crop leaves and corresponding environmental vectors;

[0014] S2, performing multi-scale enhancement processing on the images;

[0015] S3, using a semi-supervised learning detection model based on a teacher-student architecture to process the enhanced images to generate confidence, area ratio and morphological change rate of the disease area;

[0016] S4, fusing the confidence, area ratio, morphological change rate and part of the environmental vector data through a predefined risk function to calculate a disease risk value representing the current threat level of the disease;

[0017] S5, normalizing the disease risk value and performing dynamic risk prediction based on a set threshold;

[0018] S6, generating a control parameter vector for prevention according to the normalized disease risk value through a mapping function;

[0019] S7, acquiring actual disease state feedback in the subsequent period, defining an error function, and using gradient descent method to iteratively update the weight coefficients of the risk function with the goal of minimizing the error function.

[0020] As a limitation, the multi-scale enhancement processing of step S2 is:

[0021] Weak enhancement is performed on the labeled samples, including random cropping, brightness adjustment, flipping and mild noise disturbance;

[0022] Strong enhancement is performed on the unlabeled samples, including Gaussian blur, color shift, random occlusion and affine transformation.

[0023] As a second limitation, the teacher network in the semi-supervised learning detection model of the teacher-student architecture in step S3 is pre-trained on a large-scale open source dataset to generate pseudo labels for unlabeled samples;

[0024] The student network adopts a lightweight Transformer-CNN hybrid structure and is trained on joint data of labeled samples and pseudo label samples using a joint loss function:

[0025] The teacher network updates parameters from the student network using an exponential moving average method

[0026] As a further limitation, the semi-supervised learning detection model of the teacher-student architecture is provided with a pseudo label error rate threshold to prevent and control pseudo label quality drift;

[0027] When the error rate of the pseudo label generated by the teacher network exceeds the set threshold, the parameters of the student network are used to update part of the parameters of the teacher network in reverse.

[0028] As a further limitation, the calculation formula of the joint loss function is:

[0029]

[0030] In the formula, is the total loss, is the supervised loss for labeled samples, is the pseudo label loss for unlabeled samples, is the consistency constraint term, and are the weight coefficients of the pseudo label loss and the consistency constraint term, respectively.

[0031] As a further limitation, the update formula of the exponential moving average method is:

[0032]

[0033] In the formula, is the updated teacher network parameter, is the teacher network parameter, is the student network parameter, is the smoothing coefficient

[0034] As a third limitation, the calculation formula of the risk function in step S4 is:

[0035]

[0036] In the formula, is the weighted disease risk value, is the disease confidence, This represents the percentage of the area affected by disease. This represents the rate of change in disease morphology. For ambient humidity, For ambient wind speed, , , , and These are the weighting coefficients, and the sum of the five weighting coefficients is 1.

[0037] As a further limitation, the normalization formula for step S5 is:

[0038]

[0039] In the formula, The normalized disease risk value, The sensitization coefficient, For risk smoothing point

[0040] As a further limitation, the mapping function mentioned in step S6 is a piecewise linear mapping function or a nonlinear mapping function, which is selected according to the control accuracy requirements of the working equipment or the preset working mode.

[0041] When in normal operation mode, a piecewise linear mapping function is used; when in fine operation mode, a non-linear mapping function is used.

[0042] The piecewise linear mapping function is:

[0043]

[0044] The nonlinear mapping function is:

[0045]

[0046] In the formula, To control the spray flow rate in the parameter vector, To minimize the spraying flow rate, For maximum spraying flow rate, The low-risk threshold The high-risk threshold It is a non-linear growth coefficient.

[0047] Preferably, the error function in step S7 is:

[0048]

[0049] In the formula, For error, This represents the actual disease risk value measured in the next cycle. For coefficients, This is the initial weight vector.

[0050] The present invention, by adopting the above-described technical solution, achieves the following technical advancements compared to existing technologies:

[0051] (1) The method of the present invention significantly reduces the dependence of traditional fully supervised models on large-scale labeled data by introducing semi-supervised learning and dynamic risk assessment mechanisms, and improves the generalization ability of the model in different crops and environments. At the same time, its lightweight design and multi-scale enhancement processing help to adapt to edge computing devices and realize low-latency and high-efficiency real-time disease detection. In addition, by integrating time-series image features and environmental data, a dynamic prediction and decision generation mechanism based on risk function is constructed, which overcomes the rigidity of fixed threshold strategy and can more accurately reflect the disease development trend and environmental coupling effect, thereby generating adaptive pesticide application control parameters. Finally, gradient descent optimization based on actual feedback is introduced, which enables the system to have the ability to continuously update itself, forming a complete intelligent closed loop from perception, decision-making to execution and optimization, effectively improving the accuracy, adaptability and long-term autonomous operation level of agricultural disease prevention and control.

[0052] (2) The method of the present invention can effectively improve the robustness of the model to common field disturbances and prevent overfitting by weakly enhancing the labeled samples while preserving the authenticity of key disease morphology and labeling information; while applying strong enhancement to a large number of unlabeled samples can greatly expand the diversity and complexity of the data, simulate the extreme imaging conditions that crops may encounter in the real environment, and thus provide rich "virtual samples" for consistency regularization training in the teacher-student framework, significantly reducing the dependence on manual labeling and enhancing the model's generalization recognition ability in different crop varieties, growth stages and complex field environments.

[0053] (3) The method of the present invention uses a teacher-student architecture. The teacher network, which is pre-trained on a large-scale open-source dataset, generates high-quality pseudo-labels. The student network adopts a lightweight Transformer-CNN hybrid structure. The labeled samples and pseudo-label samples are trained in collaboration through a joint loss function. Combined with the teacher network parameter update mechanism based on exponential moving average, the student network continuously and robustly learns more generalized and robust disease features in the iteration, thereby achieving high-precision and highly adaptive detection capabilities under limited labeled data.

[0054] (4) The method of the present invention constructs a dynamic quality monitoring and correction closed loop by setting a pseudo-label error rate threshold and introducing a partial reverse update mechanism of the student network to the teacher network. It can effectively identify and intervene in the problem of "knowledge aging" of the teacher network caused by drastic environmental changes or data distribution drift, and prevent low-quality pseudo-labels from accumulating errors during training and causing continuous deterioration of model performance.

[0055] This invention belongs to the field of intelligent agriculture technology. By constructing a dynamic risk assessment model that integrates time-series and environmental information, and an adaptive decision-making and optimization mechanism based on risk feedback, the efficiency of disease identification and prevention systems is improved. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0057] In the attached diagram:

[0058] Figure 1 This is a flowchart illustrating the processing of an embodiment of the present invention;

[0059] Figure 2 This is a flowchart illustrating the training process of a semi-supervised learning detection model with a teacher-student architecture in an embodiment of the present invention.

[0060] Figure 3 This is a diagram of the lightweight Transformer-CNN hybrid structure in an embodiment of the present invention;

[0061] Figure 4 This is a comparison of the average detection accuracy of each model on the same crop disease dataset in the embodiments of the present invention;

[0062] Figure 5 This is a graph showing the average inference latency of each model in the embodiments of the present invention under the same hardware environment;

[0063] Figure 6 This is a comparison chart of the parameter values ​​of each model in the embodiments of the present invention. Detailed Implementation

[0064] The preferred embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0065] Example

[0066] like Figure 1 As shown, this embodiment is a method for crop disease identification and adaptive decision-making, which includes the following steps performed sequentially:

[0067] S1. Obtain time-series images of crop leaves and corresponding environmental vectors.

[0068] The acquired data includes crop leaf images and multidimensional environmental factors, which together constitute the appearance and conditional information of the disease state. Among them, the image sequences are acquired in RGB or multispectral form to detect the morphology, color, and texture features of lesions; the environmental vectors reflect the humidity, wind speed, temperature, and light conditions on which disease development depends.

[0069] S2. Perform multi-scale enhancement processing on the image.

[0070] To enhance the model's generalization ability, this embodiment employs a two-layer enhancement mechanism for the image:

[0071] Weak enhancements are used for labeled samples, including random cropping, brightness adjustment, flipping, and mild noise perturbation, to maintain semantic consistency;

[0072] Strong enhancements are used for unlabeled samples, including Gaussian blur, color shift, random occlusion, affine transformation, etc., to simulate changes in lighting, fog, and shooting angle.

[0073] S3. Use a semi-supervised learning detection model based on a teacher-student architecture to process the enhanced image and generate the confidence level, area ratio, and morphological change rate of the diseased area.

[0074] like Figure 2 As shown, in the semi-supervised learning detection model with a teacher-student architecture, the teacher network is pre-trained on a large-scale open-source dataset to generate pseudo-labels for unlabeled samples; the student network employs... Figure 3 The lightweight Transformer-CNN hybrid architecture shown is trained on joint data of labeled and pseudo-labeled samples using a joint loss function.

[0075] Specifically, the formula for calculating the joint loss function is as follows:

[0076]

[0077] In the formula, For the total loss, The supervised loss used for labeled samples, For pseudo-label loss used for unlabeled samples, For consistency constraints, and These are the weighting coefficients for the pseudo-label loss and the consistency constraint term, respectively.

[0078] The teacher network updates the parameters from the student network using an exponential moving average method, with the update formula as follows:

[0079]

[0080] In the formula, For the updated teacher network parameters, For teacher network parameters, For student network parameters, This is the smoothing coefficient.

[0081] The model includes a false label error rate threshold to prevent false label quality drift. When the error rate of the false labels generated by the teacher network exceeds the set threshold, the parameter knowledge of the student network is used to update some parameters of the teacher network in reverse.

[0082] After cross-modal feature fusion, the model can achieve high-confidence disease detection with low computational cost (FLOPs < 2.5G).

[0083] S4. By fusing the confidence level, area ratio, morphological change rate, and some environmental vector data through a predefined risk function, a disease risk value representing the current level of disease threat is calculated.

[0084] After obtaining the test results, the test information is integrated with environmental factors to form a model, and a quantifiable disease risk value is obtained using a risk function.

[0085] Specifically, the formula for calculating the risk function is:

[0086]

[0087] In the formula, The weighted disease risk value, For the confidence level of the disease, This represents the percentage of the area affected by disease. This represents the rate of change in disease morphology. For ambient humidity, For ambient wind speed, , , , and These are the weighting coefficients, and the sum of the five weighting coefficients is 1.

[0088] S5. Normalize the disease risk value and perform dynamic risk prediction based on the set threshold.

[0089] To enhance continuity and sensitivity, the disease risk value is normalized using the Sigmoid function:

[0090]

[0091] In the formula, The normalized disease risk value, The sensitization coefficient, This is a risk smoothing point.

[0092] After obtaining the normalized disease risk values, a low-risk threshold is set. and Conduct dynamic risk prediction.

[0093] In this embodiment, the following settings are provided: , The specific judgment logic is as follows:

[0094] when If <0.3, continue observation;

[0095] When 0.3≤ When the concentration is less than 0.7, initiate localized inspections and light spraying.

[0096] when When the value is greater than 0.7, a high-intensity prevention and control operation is triggered.

[0097] S6. Based on the normalized disease risk values, generate a vector of control parameters for prevention and control through a mapping function.

[0098] Based on the normalized disease risk values, a vector of control parameters is generated, including spray flow rate, spray angle, and operation priority.

[0099] To adapt to different operating scenarios and hardware devices, this invention presets two mapping function modes, which can be selected according to the control accuracy requirements of the operating equipment or the preset operating mode:

[0100] When the operating equipment is a standard sprinkler irrigation system or the requirement for moderate flow control accuracy is not critical, a piecewise linear mapping function is used to reduce computational complexity. The calculation formula is as follows:

[0101]

[0102] In the formula, To control the spray flow rate in the parameter vector, To minimize the spraying flow rate, For maximum spraying flow rate, The low-risk threshold This is a high-risk threshold.

[0103] When the operating equipment is a high-precision variable-rate spraying drone or requires a smoother response to avoid the impact of sudden changes in pesticide application on crops, a nonlinear mapping function is used:

[0104]

[0105] In the formula, It is a non-linear growth coefficient.

[0106] S7. In subsequent cycles, obtain actual disease status feedback, define an error function, and use the gradient descent method to iteratively update the weight coefficients of the risk function with the goal of minimizing the error function.

[0107] Specifically, the error function is defined as follows:

[0108]

[0109] In the formula, For error, This represents the actual disease risk value measured in the next cycle. For coefficients, This is the initial weight vector.

[0110] To demonstrate the effectiveness of this embodiment, it was systematically validated against existing representative models Faster-RCNN, YOLOv5s, MobileDet, YOLOv8n, and Semi-DETR in terms of detection accuracy, model lightweighting, and inference latency. The specific results are as follows: Figures 4 to 6 As shown.

[0111] from Figure 4 As can be seen, traditional detection models such as Faster-RCNN and YOLOv5s have limited accuracy under complex lighting conditions and multi-scale lesion conditions, while lightweight models MobileDet and YOLOv8n, although possessing high real-time performance, have relatively insufficient detection confidence. This embodiment, using only partially labeled samples, still achieves a detection accuracy of mAP@0.5 = 0.94, an improvement of approximately 6% compared to YOLOv5s, and outperforms Semi-DETR by approximately 1%. These results demonstrate that this embodiment can maintain high recognition performance under low-labeling conditions.

[0112] from Figure 5 As can be seen, traditional detectors like Faster R-CNN suffer from latency exceeding 100 ms due to their complex network structure, while lightweight models perform better in terms of latency. This embodiment, through a modular hybrid structure design, controls the inference time to 38 ms, achieving real-time performance comparable to YOLOv8n while maintaining higher detection accuracy. These results validate the effectiveness of this embodiment in balancing lightweight design and speed, providing a feasible software foundation for real-time agricultural monitoring and unmanned agricultural machinery control.

[0113] from Figure 6 It can be observed that the traditional fully supervised detection models (Faster-RCNN, YOLOv5s) have approximately 135 M and 70 M parameters, respectively, while the algorithm of this invention has only 7.5 M parameters, approximately 1 / 18 of the former. This result demonstrates that the Transformer-CNN fusion feature extraction in this embodiment can maintain high-performance output while significantly reducing computational cost.

[0114] In summary, this embodiment achieves a balance between accuracy, latency, and model size. It significantly outperforms most fully supervised models in terms of mAP; and is comparable to or better than advanced lightweight models in terms of latency and parameter count, thus improving the efficiency of disease identification and control systems.

Claims

1. A method for identifying and adaptively deciding on crop diseases, characterized in that, The method includes the following steps performed sequentially: S1. Obtain time-series images of crop leaves and corresponding environmental vectors; S2. Perform multi-scale enhancement processing on the image; S3. Use a semi-supervised learning detection model based on a teacher-student architecture to process the enhanced image and generate the confidence score, area ratio, and morphological change rate of the diseased area; S4. By fusing the confidence level, area ratio, morphological change rate and some environmental vector data through a predefined risk function, a disease risk value representing the current level of disease threat is calculated. S5. Normalize the disease risk value and perform dynamic risk prediction based on the set threshold. S6. Based on the normalized disease risk values, generate a vector of control parameters for prevention and control through a mapping function; S7. In subsequent cycles, obtain actual disease status feedback, define an error function, and use the gradient descent method to iteratively update the weight coefficients of the risk function with the goal of minimizing the error function.

2. The crop disease identification and adaptive decision-making method according to claim 1, characterized in that, The multi-scale enhancement process described in step S2 is as follows: Weak enhancements are applied to the labeled samples, including random cropping, brightness adjustment, flipping, and mild noise perturbation. Strong enhancements are applied to unlabeled samples, including Gaussian blur, color shift, random occlusion, and affine transformation.

3. The crop disease identification and adaptive decision-making method according to claim 1, characterized in that, In the semi-supervised learning detection model with a teacher-student architecture described in step S3, the teacher network is pre-trained on a large-scale open-source dataset to generate pseudo-labels for unlabeled samples; The student network employs a lightweight Transformer-CNN hybrid architecture, trained on joint data of labeled and pseudo-labeled samples using a joint loss function: The teacher network updates the parameters from the student network using an exponential moving average method.

4. The crop disease identification and adaptive decision-making method according to claim 3, characterized in that, The semi-supervised learning detection model with a teacher-student architecture includes a false label error rate threshold to prevent false label quality drift. When the error rate of the pseudo-labels generated by the teacher network exceeds a set threshold, the parameter knowledge of the student network is used to update some parameters of the teacher network in reverse.

5. A method for identifying and adaptively deciding on crop diseases according to claim 3 or 4, characterized in that, The formula for calculating the joint loss function is as follows: In the formula, For the total loss, The supervised loss used for labeled samples, For pseudo-label loss used for unlabeled samples, For consistency constraints, and These are the weighting coefficients for the pseudo-label loss and the consistency constraint term, respectively.

6. A method for crop disease identification and adaptive decision-making according to claim 3 or 4, characterized in that, The update formula for the exponential moving average method is as follows: In the formula, For the updated teacher network parameters, For teacher network parameters, For student network parameters, This is the smoothing coefficient.

7. The crop disease identification and adaptive decision-making method according to claim 1, characterized in that, The formula for calculating the risk function in step S4 is as follows: In the formula, The weighted disease risk value, For the confidence level of the disease, This represents the percentage of the area affected by disease. This represents the rate of change in disease morphology. For ambient humidity, For ambient wind speed, , , , and These are the weighting coefficients, and the sum of the five weighting coefficients is 1.

8. The crop disease identification and adaptive decision-making method according to claim 7, characterized in that, The normalization formula for step S5 is: In the formula, The normalized disease risk value, The sensitization coefficient, This is a risk smoothing point.

9. The crop disease identification and adaptive decision-making method according to claim 8, characterized in that, The mapping function mentioned in step S6 is a piecewise linear mapping function or a nonlinear mapping function, which is selected according to the control accuracy requirements of the working equipment or the preset working mode. When in normal operation mode, a piecewise linear mapping function is used; when in fine operation mode, a non-linear mapping function is used. The piecewise linear mapping function is: The nonlinear mapping function is: In the formula, To control the spray flow rate in the parameter vector, To minimize the spraying flow rate, For maximum spraying flow rate, The low-risk threshold The high-risk threshold It is a non-linear growth coefficient.

10. The crop disease identification and adaptive decision-making method according to claim 9, characterized in that, The error function mentioned in step S7 is: In the formula, For error, This represents the actual disease risk value measured in the next cycle. For coefficients, This is the initial weight vector.

Images