A method and system for monitoring diseases in the growth process of epimedium

Abstract

The present application relates to the technical field of disease monitoring, and particularly relates to a disease monitoring method and system for growth process of Epimedium brevicornum, which comprises deploying multi-source acquisition equipment in the Epimedium brevicornum planting area, acquiring plant images in real time at preset time intervals, wherein the plant images comprise RGB images and multispectral images; performing image preprocessing on the acquired RGB images to obtain standard plant images; acquiring plant growth time, performing morphological analysis on the standard plant images to extract growth characteristic parameters, inputting the growth characteristic parameters and the growth time into a pre-constructed plant growth stage prediction model to obtain the growth stage of Epimedium brevicornum. The disease monitoring method for growth process of Epimedium brevicornum provided by the present application can accurately extract key features such as disease leaves, area, color and shape by fusing RGB and multispectral images, and can realize intelligent discrimination of disease types and severity grades of Epimedium brevicornum by combining the growth stage with the disease prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disease monitoring, in particular to a disease monitoring method and system for Epimedium growth process. BACKGROUND

[0002] With the development of informationization and intelligentization of traditional Chinese medicinal material industry, Epimedium planting management is gradually transforming towards digital agriculture and intelligent disease identification. Disease spot detection, disease grade evaluation and leaf health analysis based on image recognition and AI model are becoming important technical supports for improving the planting efficiency and medicinal material quality of Epimedium industry.

[0003] The current crop disease monitoring method based on image processing has two key limitations. First, the determination of the growth stage of the plant usually relies too much on the single indicator of planting time. However, the actual growth rate is significantly affected by the differences in variety characteristics, environmental conditions and field management measures. The growth stage calculated by planting time often deviates greatly from the actual physiological state. Second, the diagnosis of the disease mainly relies on disease analysis of the collected plant image data. It is difficult to fully capture the early weak characteristics of the disease or distinguish similar phenotypes from nutritional stress and physiological disorders only by relying on visual information. Moreover, it is easily disturbed by environmental factors when shooting, resulting in insufficient sensitivity and specificity of monitoring. Therefore, the monitoring strategy relying on rough growth stage division and single image data source greatly restricts the reliability and accuracy of the results in actual application. SUMMARY

[0004] In order to solve the above problems, the present application provides a disease monitoring method and system for Epimedium growth process.

[0005] The present application adopts the following technical scheme, a disease monitoring method for Epimedium growth process, comprising:

[0006] Deploying multi-source acquisition equipment in the Epimedium planting area, acquiring plant images in real time at a preset time interval, the plant images including RGB images and multispectral images;

[0007] Performing image preprocessing on the collected RGB images to obtain standard plant images, the image preprocessing including denoising, geometric correction and image enhancement;

[0008] Acquiring the plant growth time, performing morphological analysis on the standard plant images to extract growth characteristic parameters, inputting the growth characteristic parameters and the growth time into a pre-constructed plant growth stage prediction model to obtain the growth stage of Epimedium;

[0009] Processing the RGB images and the multispectral images to extract disease leaf images and disease characteristic parameters;

[0010] Input the disease leaf image, disease characteristic parameter and epimedium growth stage into the pre-constructed disease prediction model to obtain the disease type and severity level.

[0011] As a further description of the above technical solution: the method for extracting the growth characteristic parameter by performing morphological analysis on the standard plant image comprises:

[0012] The growth characteristic parameter comprises a plant projection area and a leaf area index;

[0013] The method for obtaining the plant projection area comprises: obtaining the total number of pixels of the plant region in the image, multiplying the actual area corresponding to a single pixel block to obtain the plant projection area;

[0014] The method for obtaining the leaf area index comprises: performing gray scale processing on the collected standard plant image to extract a complete leaf pattern;

[0015] The number of pixel blocks of the complete leaf pattern is divided by the number of complete leaf patterns, and then multiplied by the actual area corresponding to a single pixel block to obtain leaf area data.

[0016] As a further description of the above technical solution: the method for extracting the complete leaf pattern comprises:

[0017] Performing gray scale processing on the collected standard plant image to obtain the gray scale value of each pixel block in the image data, obtaining standard leaf image data, obtaining the gray scale value of each pixel block in the leaf image data, marking the standard gray scale value, comparing and analyzing each real-time gray scale value in the standard plant image with the standard gray scale value, judging whether to mark the pixel block as a leaf, extracting the pixel block with the leaf mark, and mapping it to a blank background layer to obtain an extraction pattern composed of the pixel block with the leaf mark;

[0018] Obtaining the area of the pixel block of each independent region pattern in the extraction pattern, presetting a leaf area gradient threshold value, when the area of the pixel block of the independent pattern is greater than or less than the leaf area gradient threshold value, eliminating the independent region pattern, and retaining the independent region pattern that is not eliminated, and marking it as a complete leaf pattern.

[0019] The method for comparing and analyzing each real-time gray scale value in the standard plant image with the standard gray scale value to judge whether to mark the pixel block as a leaf comprises:

[0020] Presetting a gray scale difference threshold value, when the absolute value of the difference between the real-time gray scale value and the standard gray scale value is less than or equal to the gray scale difference threshold value, marking the pixel block as a leaf, and when the absolute value of the difference between the real-time gray scale value and the standard gray scale value is greater than the gray scale difference threshold value, not marking the pixel block as a leaf.

[0021] As a further description of the above technical solution: the training method of the plant growth stage prediction model comprises:

[0022] Q sets of training data are collected in advance, Q is a positive integer greater than 0, the training data includes growth characteristic parameters and growth time, and the growth characteristic parameters and the growth time correspond to the growth stage of Epimedium;

[0023] A deep learning model is selected as the plant growth stage prediction model, the training data is used to train the plant growth stage prediction model, the growth characteristic parameters and the growth time are used as the input of the plant growth stage prediction model, the growth stage of Epimedium is used as the output of the plant growth stage prediction model, the random gradient descent method is used, the weights and biases of the plant growth stage prediction model are adjusted through the back propagation algorithm, so that the error between the prediction result of the plant growth stage prediction model and the actual result is minimized, a loss function is set, the loss function is mean square error, when the loss function value reaches convergence, the training of the plant growth stage prediction model is stopped, and the plant growth stage prediction model corresponding to the loss function value reaching convergence is used as the trained plant growth stage prediction model;

[0024] The growth stage of Epimedium includes the bud leaf stage, the young leaf elongation period stage, the functional leaf maturation stage, the vegetative reproduction stage, the reproductive differentiation stage, the flowering and fruiting stage and the wintering preparation stage.

[0025] As a further description of the above technical solution: the multi-spectral image includes a near-infrared wave band and a red light wave band.

[0026] As a further description of the above technical solution: the method for extracting the diseased leaf image comprises:

[0027] The disease coefficient of each pixel block of the multi-spectral image is calculated, a preset disease coefficient threshold is set, the pixel blocks with a disease coefficient less than the disease coefficient threshold are marked, and a primary suspicious area image is obtained;

[0028] The corresponding primary suspicious area image in the RGB image is obtained, and each independent primary suspicious area image is mapped to a blank background layer to obtain a suspicious lesion area image;

[0029] The suspicious lesion area image is input into the trained disease identification model, and whether the suspicious lesion area image is a lesion area image is output, the suspicious lesion area image with an output result of being a lesion area image is retained and marked as a lesion area image;

[0030] The green index is used to extract the leaf area from the collected lesion area image, adjacent pixels are classified into the same area, and a plurality of closed areas are obtained;

[0031] The pixel block number of all closed regions is obtained, the area of each closed region is obtained based on the pixel block number, a leaf area gradient threshold is preset, when the area of the pixel block of the closed region is greater than or less than the leaf area gradient threshold, the closed region is removed, the closed region without being removed is retained, and is marked as a diseased leaf image.

[0032] As a further description of the above technical solution: the training method of the disease identification model comprises:

[0033] Collect RGB images of Epimedium plants at different growth stages, the RGB images including images with diseases and images without diseases, and set labels for the RGB images, the labels being whether there are diseases or not, construct a group of training data by using the RGB images and the labels corresponding to the RGB images, construct a sample set by using X groups of training data, X being an integer greater than 1, divide the sample set into a training set and a test set, use the RGB images in the training set as the input of the disease identification model, use the labels in the training set as the output of the disease identification model, train the disease identification model to obtain an initial disease identification model, minimize the sum of prediction errors as a training target, use the test set to evaluate the initial disease identification model, and use the initial disease identification model when the sum of prediction errors reaches convergence as the constructed disease identification model.

[0034] The disease identification model is a deep neural network model.

[0035] As a further description of the above technical solution: the disease feature parameters include a disease spot area ratio, a shape compactness coefficient and a color difference value.

[0036] The method for obtaining the color difference value comprises: converting the diseased leaf image into an LAB color space, extracting LAB values of the disease spot and the healthy area, and calculating color differences in each channel.

[0037] The method for obtaining the disease spot area ratio comprises: obtaining the area of the disease spot region, and dividing the area by the area of the diseased image to obtain the disease spot area ratio.

[0038] The method for obtaining the shape compactness coefficient comprises: calculating the square of the contour perimeter of the diseased leaf image, and dividing the square by the product of the area of the diseased leaf image and 4π.

[0039] As a further description of the above technical solution: the training method of the disease prediction model comprises:

[0040] Pre-collect K groups of training data, K being a positive integer greater than 0, the training data including disease leaf images, disease feature parameters and growth stages of Epimedium, and disease types and severity levels corresponding to the disease leaf images, the disease feature parameters and the growth stages of Epimedium.

[0041] The dual-branch model adopts image feature extraction and structured feature fusion, wherein the image feature extraction uses a pre-trained convolutional neural network as a backbone network, followed by a convolutional layer to extract lesion detail features and output an image feature vector; the lesion area ratio, shape compactness coefficient and color difference value are spliced into a multi-dimensional vector with the growth stage after one-hot encoding, input into a fully connected layer to output a structured feature vector; the image feature vector and the structured feature vector are spliced and input into a fully connected layer, and two parallel output layers are used: one outputs the disease type, and the other outputs the severity level.

[0042] A disease monitoring system for Epimedium growth process is used to realize the disease monitoring method for Epimedium growth process, and the system comprises:

[0043] An image data acquisition module is used to deploy multi-source acquisition equipment in an Epimedium planting area, and real-time plant images are acquired at a preset time interval, wherein the plant images comprise RGB images and multispectral images.

[0044] An image processing module is used to perform image preprocessing on the acquired RGB images to obtain standard plant images, and the image preprocessing comprises denoising, geometric correction and image enhancement.

[0045] A growth characteristic parameter acquisition module is used to acquire plant growth time, perform morphological analysis on the standard plant images to extract growth characteristic parameters, and input the growth characteristic parameters and growth time into a pre-constructed plant growth stage prediction model to obtain the Epimedium growth stage.

[0046] A disease characteristic parameter acquisition module is used to process the RGB images and multispectral images to extract disease leaf images and disease characteristic parameters.

[0047] A disease type judgment module is used to input the disease leaf images, disease characteristic parameters and Epimedium growth stage into a pre-constructed disease prediction model to obtain the disease type and severity level.

[0048] Advantages:

[0049] The disease monitoring method for Epimedium growth process provided by the application realizes the synchronous extraction of plant growth state and disease characteristics by deploying multi-source acquisition equipment in an Epimedium planting area and combining the fusion analysis of RGB and multispectral images, accurately acquires plant projection area and leaf area index and other growth parameters through image preprocessing and morphological analysis, and accurately identifies the current growth stage by using a trained growth stage prediction model in combination with growth time, thereby improving the timeliness and scientificity of planting management.

[0050] Meanwhile, the disease monitoring method fuses the RGB image and the multispectral image, accurately extracts key features such as disease leaves, areas, colors and shapes, and inputs a disease prediction model combined with a growth stage, so that intelligent discrimination of disease types and severity levels of Epimedium is realized. The disease recognition accuracy and early warning capability are enhanced, precise measures are facilitated, the prevention and treatment efficiency is improved, and the digitalization and intelligent level of the whole Epimedium planting process are improved. BRIEF DESCRIPTION OF DRAWINGS

[0051] The application will be further explained in connection with the accompanying drawings and embodiments:

[0052] Figure 1 A flowchart of a disease monitoring method for the growth process of Epimedium provided by the embodiment of the application is shown in the figure;

[0053] Figure 2 A flowchart of a method for extracting a complete leaf pattern provided by the embodiment 1 of the application is shown in the figure;

[0054] Figure 3 A flowchart of a method for extracting a disease leaf image provided by the embodiment 1 of the application is shown in the figure;

[0055] Figure 4 A module connection diagram of a disease monitoring system for the growth process of Epimedium provided by the embodiment 2 of the application is shown in the figure. DETAILED DESCRIPTION

[0056] In order to make the technical means, creative features, purposes and effects realized by the application easy to understand, the application will be further described in combination with specific figures. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0057] Embodiment 1

[0058] Please refer to Figures 1-3 The embodiment of the application provides a technical solution: a disease monitoring method for the growth process of Epimedium, comprising:

[0059] Multi-source acquisition equipment is deployed in an Epimedium planting area, and plant images are acquired in real time at a preset time interval, wherein the plant images include RGB images and multispectral images;

[0060] It should be noted that the multi-source acquisition equipment includes a high-definition RGB camera and a multispectral camera.

[0061] The acquired RGB images are subjected to image preprocessing to obtain standard plant images, wherein the image preprocessing includes denoising, geometric correction and image enhancement.

[0062] It should be noted that the image preprocessing includes: denoising, using Gaussian filtering, median filtering and other methods to remove image noise; geometric correction, correcting the perspective distortion caused by the shooting angle; image enhancement is to adjust the contrast, brightness, and highlight the key features of the plant. The above image preprocessing is prior art and will not be described in detail here.

[0063] The growth time of the plant is obtained, morphological analysis is performed on the standard plant image to extract growth characteristic parameters, and the growth characteristic parameters and the growth time are input into a pre-constructed plant growth stage prediction model to obtain the growth stage of the Epimedium.

[0064] It should be noted that the growth time of the plant is obtained based on the number of days between the current time and the planting time;

[0065] The growth stage of the Epimedium includes the germination and leaf expansion stage, the young leaf elongation period stage, the functional leaf maturation stage, the vegetative reproduction stage, the reproductive differentiation stage, the flowering and fruiting stage, and the overwintering preparation stage.

[0066] It should be noted that the growth stage of the Epimedium is different, and the types and occurrence probabilities of the diseases are also different. Accurate acquisition of the growth stage of the Epimedium is of great help to the occurrence of the Epimedium diseases and the judgment of the disease types, because the occurrence of the Epimedium diseases has significant growth stage specificity. The physiological state, tissue tenderness, and secondary metabolite composition of the plant at different development stages are obviously different, which directly affects the infection probability and dominant species of the diseases. Accurate determination of the growth stage can significantly improve the accuracy of disease warning and provide key basis for the development of targeted control strategies.

[0067] Example: The young leaf elongation period is susceptible to disease: gray mold; the reason for high incidence: the cuticle of the new leaves is not fully formed, and the stomatal opening is large.

[0068] The core insect pests in the vegetative reproduction period are Epimedium gall midges, and the reason for high incidence is that the new branch meristems are rich in free amino acids (concentration > 12 μmol / g), which provide essential nutrients for the larvae.

[0069] The method for performing morphological analysis on the standard plant image to extract growth characteristic parameters includes:

[0070] The growth characteristic parameters include the plant projection area and the leaf area index.

[0071] The method for obtaining the plant projection area includes: obtaining the total number of pixels of the plant region in the image, multiplying the actual area corresponding to a single pixel block to obtain the plant projection area.

[0072] The method for obtaining the leaf area index includes:

[0073] The standard plant image collected is subjected to gray scale processing, and a complete leaf pattern is extracted;

[0074] The number of pixel blocks of the complete leaf pattern is divided by the number of complete leaf patterns, and then multiplied by the actual area corresponding to a single pixel block, to obtain leaf area data.

[0075] The method for extracting the complete leaf pattern comprises:

[0076] The standard plant image collected is subjected to gray scale processing, and the gray scale value of each pixel block in the image data is obtained, the standard leaf image data is obtained, the gray scale value of each pixel block in the leaf image data is obtained, and is marked as a standard gray scale value, each real-time gray scale value in the standard plant image is compared and analyzed with the standard gray scale value, it is judged whether the pixel block is marked with a leaf, the pixel block with the leaf mark is extracted, and it is mapped to a blank background layer to obtain an extraction pattern composed of the pixel block with the leaf mark;

[0077] The area of the pixel block of each independent region pattern in the extraction pattern is obtained, a preset leaf area gradient threshold value is set, when the area of the pixel block of the independent pattern is greater than or less than the leaf area gradient threshold value, the independent region pattern is removed, and the independent region pattern that is not removed is retained and is marked as a complete leaf pattern.

[0078] The method for comparing and analyzing each real-time gray scale value in the standard plant image with the standard gray scale value to judge whether the pixel block is marked with a leaf comprises:

[0079] A preset gray scale difference threshold value is set, when the absolute value of the difference between the real-time gray scale value and the standard gray scale value is less than or equal to the gray scale difference threshold value, the pixel block is marked with a leaf, and when the absolute value of the difference between the real-time gray scale value and the standard gray scale value is greater than the gray scale difference threshold value, the pixel block is not marked with a leaf.

[0080] In the embodiment, by setting a preset leaf area gradient threshold value, the independent region pattern with an excessively large or small independent region area caused by stacking or crushing of leaves is removed, and other regions are retained, so that the acquisition of the leaf area data is more accurate, and the leaf development degree is more accurately reflected, thereby facilitating identification of the growth stage of the plant based on the same.

[0081] The training method of the plant growth stage prediction model comprises:

[0082] Q sets of training data are collected in advance, Q is a positive integer greater than 0, the training data comprises growth characteristic parameters and growth time, and the growth characteristic parameters and the growth time correspond to the growth stage of the Epimedium;

[0083] The deep learning model is selected as the plant growth stage prediction model, the plant growth stage prediction model is trained using the training data, the growth characteristic parameters and the growth time are taken as the inputs of the plant growth stage prediction model, the growth stages of Epimedium are taken as the outputs of the plant growth stage prediction model, the random gradient descent method is adopted, the weight and bias of the plant growth stage prediction model are adjusted through the back propagation algorithm, so that the error between the prediction result and the actual result of the plant growth stage prediction model is minimized, the loss function is set, the loss function is mean square error, when the loss function value reaches convergence, the training of the plant growth stage prediction model is stopped, and the plant growth stage prediction model corresponding to the loss function value reaching convergence is taken as the trained plant growth stage prediction model.

[0084] In the embodiment, by performing gray processing and morphological analysis on the standard plant image, the key growth characteristic parameters of the plant projection area and the leaf area index are extracted, and the non-intact leaf area is effectively screened out by combining the set gray difference and the leaf area gradient threshold, the accuracy of leaf recognition and area measurement is improved, so that the growth state of Epimedium is more accurately reflected, and on this basis, the growth stage prediction model based on deep learning is constructed, and based on the growth time and the growth characteristic parameters, efficient recognition and accurate judgment of different growth stages can be realized.

[0085] And the disease resistance of Epimedium is different at different growth stages, and the manifestation forms of diseases and pests are also different, for example, the leaves are tender at the leaf unfolding stage and are easily damaged by aphids, and the disease spots spread fast; the leaves are aging at the fruiting stage, and are more susceptible to leaf spot disease, and the disease spots are hard. Accurate growth stages are obtained, the model learns the association between "stage-disease type", and the prediction pertinence is improved.

[0086] The disease leaf image and the disease characteristic parameters are extracted by processing the RGB image and the multispectral image;

[0087] The multispectral image includes a near-infrared waveband and a red light waveband;

[0088] The method for extracting the disease leaf image includes:

[0089] The disease coefficient of each pixel point of the multispectral image is calculated, a preset disease coefficient threshold is set, the pixel points with a disease coefficient less than the disease coefficient threshold are marked, and a primary suspicious area image is obtained;

[0090] The method for calculating the disease coefficient includes: ; wherein, the disease coefficient is, the value of the pixel in the near-infrared waveband is, the value of the pixel in the red light waveband is.

[0091] It should be noted that the lower the NDVI, the worse the vegetation, indicating a suspected lesion area;

[0092] Corresponding primary suspected area images in the RGB image are obtained, and each independent primary suspected area image is mapped to a blank background layer to obtain a suspected lesion area image;

[0093] The suspected lesion area image is input into the trained disease recognition model, and whether the suspected lesion area image is a lesion area image is output. The suspected lesion area image with the output result as a lesion area image is retained and marked as a lesion area image.

[0094] The training method of the disease recognition model is:

[0095] RGB images of Epimedium plants at different growth stages are collected, the RGB images include images with diseases and images without diseases, and labels are set for the RGB images, the labels being presence or absence of diseases. The RGB images and the labels corresponding to the RGB images constitute a set of training data, X sets of training data constitute a sample set, X being an integer greater than 1, and the sample set is divided into a training set and a test set. The RGB images in the training set are used as the input of the disease recognition model, and the labels in the training set are used as the output of the disease recognition model. The disease recognition model is trained to obtain an initial disease recognition model, the sum of prediction errors is minimized as the training target, the test set is used to evaluate the initial disease recognition model, and the initial disease recognition model when the sum of prediction errors reaches convergence is used as the constructed disease recognition model.

[0096] The disease recognition model is a deep neural network model.

[0097] It should be noted that in the machine learning model, the calculation formula of the prediction error is: where k is the number of feature data (i.e. extracted image), is the prediction error, is the predicted state value corresponding to the kth set of feature data, the state value being the presence or absence of disease label, and wk is the actual state value corresponding to the kth set of training data

[0098] The green index is used to extract the leaf area from the collected lesion area image, adjacent pixels are grouped into the same area, and a plurality of closed areas are obtained.

[0099] The number of pixel blocks of all closed areas is obtained, the area of each closed area is obtained based on the number of pixel blocks, a preset leaf area gradient threshold is set, when the area of the pixel block of the closed area is greater than or less than the leaf area gradient threshold, the closed area is removed, and the closed area that is not removed is retained and marked as a diseased leaf image.

[0100] The disease characteristic parameters include a disease spot area ratio, a shape compactness coefficient, and a color difference value.

[0101] The color difference value is obtained by:

[0102] The disease leaf image is converted into an LAB color space, and LAB values of the disease spot and the healthy area are extracted.

[0103] Color differences of each channel are calculated.

[0104] The calculation formula of the color differences of each channel is:

[0105] In the formula, is a luminance difference value, is a red-green channel difference value, is a yellow-blue channel difference value. is a luminance of the disease spot area, is a luminance of the healthy area, is a red-green channel of the disease spot area, is a red-green channel of the healthy area, is a yellow-blue channel of the disease spot area, is a yellow-blue channel of the healthy area.

[0106] It should be noted that the color difference value is used to reflect the difference degree of the color between the disease spot and the healthy area.

[0107] The disease spot area ratio is obtained by: obtaining an area of the disease spot area, dividing the area by an area of the disease image, and obtaining the area.

[0108] Optionally, the calculation formula of the disease spot area ratio is: In the formula, is the disease spot area ratio, is the area of the disease spot area, is the area of the disease leaf image.

[0109] It should be noted that the area of the disease spot area can be obtained by multiplying the number of pixel blocks of the disease spot area by the actual area corresponding to a single pixel block, and the area of the whole leaf is obtained in the same way as the area of the disease spot area, which is not described herein again.

[0110] The shape compactness coefficient is obtained by:

[0111] In the formula, is the shape compactness coefficient, is a contour perimeter of the disease leaf image, is the area of the disease leaf image.

[0112] It should be noted that the contour perimeter of the diseased leaf image is directly obtained by ImageJ image software provided by the image processing library.

[0113] The specific method for extracting the leaf region based on the green index is as follows:

[0114] The green index is calculated by the formula , ; wherein R, G and B are respectively the red, green and blue components (value range 0-255) of each pixel, and the green index G is

[0115] The green index threshold T is set, and when T (such as T = 20), the leaf region is marked, otherwise it is not marked.

[0116] The disease type and severity level are obtained by inputting the diseased leaf image, disease characteristic parameters and growth stage of Epimedium into the pre-constructed disease prediction model.

[0117] The training method of the disease prediction model comprises:

[0118] K sets of training data are collected in advance, K is a positive integer greater than 0, and the training data includes the diseased leaf image, the disease characteristic parameters and the growth stage of Epimedium, and the disease type and severity level corresponding to the diseased leaf image, the disease characteristic parameters and the growth stage of Epimedium.

[0119] It should be noted that the disease type is determined by plant protection experts according to the leaf image, disease spot characteristics and field observation, such as the leaf performance of powdery mildew, aphid disease, anthracnose and root rot, and recorded as discrete categories (such as “powdery mildew = 01, aphid disease = 02”).

[0120] The severity level is divided in combination with the lesion area ratio: such as light (lesion area ratio < 15%), moderate (15%≤ lesion area ratio < 30%) and severe (lesion area ratio ≥ 30%).

[0121] The dual-branch model of "image feature extraction + structured feature fusion" is adopted, wherein the image feature extraction uses a pre-trained convolutional neural network as a backbone, the parameters of the first 10 layers are frozen, the general image features learned on ImageNet are used, and then 2 convolutional layers (32, 16 channels) are connected to extract disease spot detail features such as texture and edge, and output an image feature vector (such as 256 dimensions). The disease spot area ratio, shape compactness coefficient, and color difference value (3 numerical features) are spliced with the growth stage after one-hot encoding (such as 4 dimensions) into a 7-dimensional vector, which is input into a 3-layer fully connected layer (64, 32, 16 neurons) to output a structured feature vector. The image feature vector and the structured feature vector are spliced and input into a fully connected layer (128 neurons), and two parallel output layers are used: one outputs the disease type, such as 8 categories, using Softmax activation, and the other outputs the severity level.

[0122] The joint cross-entropy loss is used as the loss function for model training, and the loss function is ; The cross-entropy loss for type prediction is The cross-entropy loss for level prediction is The weight coefficient is = 0.5 to balance the weights of the two.

[0123] The Adam optimizer is selected, the initial learning rate is 0.001, and every 50 rounds are attenuated by 1 / 10. The collected training data set is divided into a training set and a test set. The disease prediction model is trained by the training set, the sum of prediction errors is minimized as the training target, and the test set is used to evaluate the initial disease prediction model. The disease prediction model with the converged prediction error sum is used as the constructed disease prediction model.

[0124] In this embodiment, by deploying multi-source acquisition equipment in the Epimedium plantation area, combining RGB and multispectral image fusion analysis, the growth state and disease characteristics of the plant are synchronously extracted. Through image preprocessing and morphological analysis, the growth parameters such as plant projection area and leaf area index are accurately obtained, and combined with the growth time, the trained growth stage prediction model is used to accurately identify the current growth stage, improving the timeliness and scientificity of planting management;

[0125] At the same time, by fusing RGB and multispectral images, the key features such as disease leaf area, color and shape are accurately extracted, and combined with the growth stage, the disease prediction model is input, which can realize intelligent discrimination of Epimedium disease type and severity level. This method enhances the accuracy and early warning ability of disease identification, helps to take precise measures, improves the prevention and control efficiency, and improves the digitalization and intelligentization level of the whole process of Epimedium planting.

[0126] Embodiment 2

[0127] Referring to Figure 4 The embodiment of the present application provides a disease monitoring system for growth process of Epimedium, which is used for realizing the disease monitoring method for growth process of Epimedium, and the system comprises:

[0128] An image data acquisition module is configured to deploy multi-source acquisition equipment in an Epimedium planting area, and to acquire plant images in real time at a preset time interval, wherein the plant images comprise RGB images and multispectral images.

[0129] An image processing module is configured to perform image preprocessing on the acquired RGB images to obtain standard plant images, wherein the image preprocessing comprises denoising, geometric correction and image enhancement.

[0130] A growth characteristic parameter acquisition module is configured to acquire plant growth time, perform morphological analysis on the standard plant images to extract growth characteristic parameters, and input the growth characteristic parameters and the growth time into a pre-constructed plant growth stage prediction model to obtain the growth stage of Epimedium.

[0131] A disease characteristic parameter acquisition module is configured to process the RGB images and the multispectral images to extract disease leaf images and disease characteristic parameters.

[0132] A disease type judgment module is configured to input the disease leaf images, the disease characteristic parameters and the growth stage of Epimedium into a pre-constructed disease prediction model to obtain a disease type and a severity level.

[0133] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method of disease monitoring of the growth process of Epimedium, characterized by, The method comprises the following steps: deploying multi-source acquisition equipment in the Epimedium plantation area, acquiring plant images in real time at preset time intervals, the plant images including RGB images and multispectral images; performing image preprocessing on the collected RGB images to obtain standard plant images, the image preprocessing including denoising, geometric rectification and image enhancement; acquiring plant growth time, performing morphological analysis on the standard plant images to extract growth characteristic parameters, inputting the growth characteristic parameters and the growth time into a pre-constructed plant growth stage prediction model to obtain the growth stage of Epimedium; processing the RGB images and the multispectral images to extract diseased leaf images and disease characteristic parameters; the method for extracting the diseased leaf images comprises the following steps: calculating a lesion coefficient for each pixel block of the multispectral image, presetting a lesion coefficient threshold, marking the pixel blocks with a lesion coefficient less than the lesion coefficient threshold to obtain a primary suspicious area image; acquiring the corresponding primary suspicious area image in the RGB image, and mapping each independent primary suspicious area image to a blank background layer to obtain a suspicious lesion area image; inputting the suspicious lesion area image into a trained disease identification model to output whether the suspicious lesion area image is a lesion area image, retaining the suspicious lesion area image with the output result being a lesion area image and marking it as a lesion area image; using a green index to extract leaf regions from the collected lesion area image, grouping adjacent pixels into the same region to obtain a plurality of closed regions; acquiring the number of pixel blocks of all the closed regions, acquiring the area of each closed region based on the number of pixel blocks, presetting a leaf area gradient threshold, removing the closed region when the area of the pixel block of the closed region is greater than or less than the leaf area gradient threshold, retaining the closed region that is not removed and marking it as a diseased leaf image; the disease characteristic parameters include a lesion area ratio, a shape compactness coefficient and a color difference value; the method for acquiring the color difference value comprises the following steps: converting the diseased leaf image into an LAB color space, extracting the LAB values of the lesion and the healthy region, and calculating the color difference of each channel; the method for acquiring the lesion area ratio is to acquire the lesion area, divide it by the disease image area, and obtain the lesion area ratio; the method for acquiring the shape compactness coefficient is to calculate the square of the contour perimeter of the diseased leaf image, divide it by the product of the area of the diseased leaf image and 4π, and obtain the shape compactness coefficient; inputting the diseased leaf image, the disease characteristic parameters and the growth stage of Epimedium into a pre-constructed disease prediction model to obtain the disease type and the severity level.

2. The method of claim 1, wherein the disease monitoring method is used for the growth process of Epimedium brevicornum. the method for performing morphological analysis on the standard plant images to extract growth characteristic parameters comprises the following steps: the growth characteristic parameters include a plant projection area and a leaf area index; the method for acquiring the plant projection area comprises the following steps: acquiring the total number of pixels of the plant region in the image, multiplying it by the actual area corresponding to a single pixel block, and obtaining the plant projection area; the method for acquiring the leaf area index comprises the following steps: performing gray processing on the collected standard plant images to extract a complete leaf pattern; acquiring the number of pixel blocks of the complete leaf pattern, dividing it by the number of complete leaf patterns, multiplying it by the actual area corresponding to a single pixel block, and obtaining leaf area data.

3. The method of claim 2, wherein the disease monitoring method is used for the growth process of Epimedium brevicornum. The method for extracting the complete leaf pattern comprises the following steps: The standard plant image is subjected to gray processing to obtain the gray value of each pixel block in the image data; The standard leaf image data is obtained to obtain the gray value of each pixel block in the leaf image data, which is marked as a standard gray value; The real-time gray value of each block in the standard plant image is compared and analyzed with the standard gray value to determine whether the pixel block is marked with a leaf, the pixel block marked with a leaf is extracted and mapped to a blank background layer to obtain an extraction pattern composed of the pixel block marked with a leaf; The area of the pixel block of each independent region pattern in the extraction pattern is obtained, a preset leaf area gradient threshold value is set, when the area of the pixel block of the independent pattern is greater than or less than the leaf area gradient threshold value, the independent region pattern is removed, and the independent region pattern that is not removed is retained and marked as a complete leaf pattern; The method for comparing and analyzing the real-time gray value of each block in the standard plant image with the standard gray value to determine whether the pixel block is marked with a leaf comprises the following steps: A preset gray difference threshold value is set, when the absolute value of the difference between the real-time gray value and the standard gray value is less than or equal to the gray difference threshold value, the pixel block is marked with a leaf, and when the absolute value of the difference between the real-time gray value and the standard gray value is greater than the gray difference threshold value, the pixel block is not marked with a leaf.

4. The method of claim 1, wherein the disease monitoring method is used for the growth of Epimedium. The training method of the plant growth stage prediction model comprises the following steps: Q sets of training data are collected in advance, Q is a positive integer greater than 0, the training data comprises growth characteristic parameters and growth time, and the growth characteristic parameters and the growth time correspond to the growth stage of Epimedium; A deep learning model is selected as the plant growth stage prediction model, the training data is used to train the plant growth stage prediction model, the growth characteristic parameters and the growth time are used as the input of the plant growth stage prediction model, the growth stage of Epimedium is used as the output of the plant growth stage prediction model, the random gradient descent method is adopted, the weight and bias of the plant growth stage prediction model are adjusted through the back propagation algorithm, so that the error between the prediction result of the plant growth stage prediction model and the actual result is minimized, a loss function is set, the loss function is mean square error, when the loss function value reaches convergence, the training of the plant growth stage prediction model is stopped, and the plant growth stage prediction model corresponding to the loss function value reaching convergence is used as the trained plant growth stage prediction model; The growth stage of Epimedium comprises a germination and leaf expansion stage, a young leaf elongation period stage, a functional leaf maturation stage, a vegetative reproduction stage, a reproductive differentiation stage, a flowering and fruiting stage and a wintering preparation stage.

5. The method of claim 1, wherein the method is used for monitoring the growth of Epimedium. The multi-spectral image comprises a near-infrared wave band and a red light wave band.

6. The method of claim 1, wherein the method is used for monitoring the growth of Epimedium. The training method of the disease identification model comprises the following steps: The RGB images of Epimedium plants at different growth stages are collected, the RGB images include images with diseases and images without diseases, and labels are set for the RGB images, the labels are whether there are diseases, a group of training data is formed by the RGB images and the labels corresponding to the RGB images, X groups of training data form a sample set, X is an integer greater than 1, and the sample set is divided into a training set and a test set; the RGB images in the training set are taken as the input of the disease identification model, the labels in the training set are taken as the output of the disease identification model, the disease identification model is trained, an initial disease identification model is obtained, the sum of prediction errors is minimized as a training target, the initial disease identification model is evaluated by using the test set, and the initial disease identification model when the sum of prediction errors reaches convergence is taken as the constructed disease identification model. The disease identification model is a deep neural network model.

7. The method of claim 1, wherein the method is used for monitoring the growth of Epimedium. The training method of the disease prediction model comprises: K groups of training data are collected in advance, K is a positive integer greater than 0, the training data include disease leaf images, disease characteristic parameters and growth stages of Epimedium, and disease types and severity levels corresponding to the disease leaf images, the disease characteristic parameters and the growth stages of Epimedium; A double-branch model of image feature extraction + structured feature fusion is used, wherein a pre-trained convolutional neural network is used as a backbone network backbone, followed by a convolutional layer to extract disease spot detail features and output an image feature vector; a disease spot area ratio, a shape compactness coefficient and a color difference value are spliced into a multi-dimensional vector with a growth stage after one-hot encoding, input into a fully connected layer and output a structured feature vector; the image feature vector and the structured feature vector are spliced and input into a fully connected layer, and two parallel output layers are used: one outputs a disease type and the other outputs a severity level.

8. A system for monitoring diseases in the growth process of Epimedium according to any one of claims 1 to 7, characterized in that The system comprises: An image data acquisition module is configured to deploy multi-source acquisition devices in an Epimedium planting area, acquire plant images in real time at preset time intervals, and the plant images include RGB images and multispectral images; An image processing module is configured to perform image preprocessing on the collected RGB images to obtain standard plant images, and the image preprocessing includes denoising, geometric correction and image enhancement; A growth characteristic parameter acquisition module is configured to acquire plant growth time, perform morphological analysis on the standard plant images to extract growth characteristic parameters, and input the growth characteristic parameters and the growth time into a pre-constructed plant growth stage prediction model to obtain the growth stage of Epimedium; A disease characteristic parameter acquisition module is configured to process the RGB images and the multispectral images to extract disease leaf images and disease characteristic parameters; A disease type judgment module is configured to input the disease leaf images, the disease characteristic parameters and the growth stage of Epimedium into a pre-constructed disease prediction model to obtain a disease type and a severity level.

Images