Disease and pest recognition and prediction algorithm based on deep learning

Abstract

A disease and pest recognition and prediction algorithm based on deep learning, the algorithm comprising the following steps: step 1, selecting a material; step 2, constructing a disease and pest species data set; step 3, constructing a disease and pest recognition model; step 4, establishing a binary classification model; step 5, extracting image features from a multi-classification deep learning model, i.e., constructing a feature library; step 6, building an automatic disease and pest recognition system; step 7, constructing a disease and pest occurrence rule data set; step 8, constructing a disease and pest prediction model; step 9, correcting and updating a model; step 10, constructing a disease and pest recognition and prediction system; and step 11, constructing a disease and pest early-warning model.

Description

A disease and pest identification and prediction algorithm based on deep learning TECHNICAL FIELD

[0001] The present application relates to the technical field of disease and pest identification and prediction, in particular to a disease and pest identification and prediction algorithm based on deep learning. BACKGROUND

[0002] Pomelo is a common citrus fruit with rich nutritional and medicinal value. However, pomelo may be affected by various diseases and pests during growth [1], common pomelo diseases and pests include papilionidae pest, tea chrysomelid, anthracnose, etc., which seriously affect the quality and yield of fruits. Therefore, timely identification and prediction of pomelo diseases and pests are crucial for orchard management and yield guarantee.

[0003] Traditional pomelo disease and pest identification methods mainly include manual observation and detection, meteorological factor analysis, chemical pesticide prevention, and agricultural expert consultation. However, manual observation and detection rely on experienced agricultural practitioners, have subjectivity and limitations, and are labor-intensive and time-consuming. Meteorological factor analysis is only applicable to certain specific meteorological conditions, which has great limitations. Chemical pesticide prevention can effectively control diseases and pests, but has problems of environmental pollution and increased drug resistance. Agricultural expert consultation relies on expert experience, requires a lot of time and cost, and lacks real-time performance. Therefore, developing more intelligent and automated disease and pest identification and prediction methods is of great significance for orchard management. SUMMARY

[0004] The present application aims to provide a disease and pest identification and prediction algorithm based on deep learning, which solves the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution: a disease and pest identification and prediction algorithm based on deep learning, comprising the following steps:

[0006] Step 1: Material selection: selecting pomelo as the detection target, and collecting pomelo dataset from multiple plantations in Ji'an City, Jiangxi Province, China;

[0007] Step 2: Disease and pest species dataset construction: using multiple collection devices with different resolutions to take multi-angle photos of pomelo under different time periods of natural light conditions;

[0008] Step three, pest identification model construction: In order to build an efficient pomelo classification model, ResNet34 is selected as the pre-trained model, which has a deep network structure and excellent feature extraction capability, and a series of latest technologies are introduced to improve its performance and adaptability; Through the analysis of the actual situation of pomelo pest data set and deep technical optimization, including careful adjustment of the model to better adapt to the growth environment and pest characteristics of pomelo;

[0009] Step four, binary classification model: Through the collected data set of other categories and pomelo pictures, a binary classification model is established for classifying non-pomelo / pomelo images;

[0010] Step five, multi-classification deep learning model image feature extraction and feature library construction: using the pre-trained AlexNet, VGG16 / 19, GoogLeNet and ResNet deep models on a large-scale image library to perform feature extraction operations on the segmented image blocks, and through optimizing related parameters to fine-tune and verify the model until the expected recognition accuracy is achieved; By using different fusion methods to enhance the feature representation ability of multiple deep features, and then sending them into SVM, Softmax and Bayesian classifier, and giving the class prediction results of pests and diseases through the post-fusion method again; Related parameters include batch_size, learning rate, learning drop strategy, momentum factor, weight decay factor and Dropout; Different fusion methods include feature concatenation and element-wise sum / product, and post-fusion methods include weighted fusion and dot product;

[0011] Step six, build an automatic pest identification system: use the cascade method to combine the binary classification model with the multi-classification deep learning model of pomelo pest images to further improve the practicality of the entire pest classification system; Through the Python Flask framework, a calling interface is provided for pomelo pest identification, and the prediction results of various pests are returned in the form of probability; In this system, the binary classification model is responsible for determining whether the image contains pomelo pests, while the multi-classification deep learning model focuses on more detailed classification of pests;

[0012] Step seven, pest occurrence regularity data set construction: In order to be able to predict the occurrence of pests in advance and take targeted measures, a pest occurrence regularity data set is made according to the historical occurrence information of pomelo pests in Jian area;

[0013] Step eight, disease and pest prediction model construction: for the disease and pest prediction model of pomelo, first, obtain the meteorological data, and divide according to the temperature for the first time, if the temperature is less than minus 7 degrees, it is easy to cause frost damage; according to the month for the second time, when the month is December, January and February, the occurrence probability of various diseases and pests is low; finally, the probability is calculated according to the comparison of meteorological data and disease and pest occurrence rule data, and the real-time disease and pest image recognition result is calibrated, and a multiple linear regression model is constructed to realize this process;

[0014] Step nine, model correction and update: a test sample set is constructed, some occurrence data of the key months from March to November with recorded local disease and pest occurrence are selected and input into the model for prediction result evaluation, the model parameter weight is corrected according to the evaluation result, and the model accuracy is gradually improved through iteration;

[0015] Step ten, disease and pest identification and prediction system construction: the real-time disease and pest identification and prediction system is constructed by using the disease and pest identification model and the prediction model of pomelo;

[0016] Step eleven, disease and pest warning model construction: including collecting the identification results of the automatic disease and pest identification system and the disease and pest warning module.

[0017] Preferably, in step two, specifically: in order to make the photographed image more in line with the real natural complex environment, in addition to the clear disease and pest leaves in the image, there is also real leaf background information, in order to ensure that the disease and pest label corresponding to the leaf in each image is real and reliable, artificial classification and identification is carried out by relevant field experts; finally, the disease and pest data set exceeds 20,000, including 13 categories.

[0018] Preferably, in step three, specifically: the method of transfer learning is adopted, ResNet34 is pre-trained on a large-scale image data set, so that the model can capture general image features; in the fine-tuning stage, the latest gradient accumulation technology is introduced, which effectively improves the robustness of the model for few-sample learning by accumulating the gradient on multiple small batches, so as to better adapt to the characteristics of pomelo disease and pest images;

[0019] In order to overcome the training problem caused by the imbalance of disease and pest categories, an improved BalanceLoader is used as a data sampler; this strategy is not only a simple weight adjustment, but also combines the latest sample rearrangement method; the sample rearrangement method dynamically adjusts the sampling order of the samples according to the performance of the model in the previous training, which can more effectively guide the model to pay attention to the difficult-to-classify samples;

[0020] In the process of model training, a self-supervised learning method is introduced; by processing the collected data, the model automatically generates labels for the data, further enriching the database and enhancing the model's learning ability of honey pomelo disease and pest characteristics.

[0021] Preferably, the step six is specifically:

[0022] The calling interface built by Python Flask framework enables related users to interact with our disease and pest classification system conveniently; related users can send image data to the interface through various channels and obtain the identification result of honey pomelo disease and pest. For the identification result, the user can click "Detailed information" to further view the introduction and control of the corresponding disease and pest, so as to take corresponding measures more effectively; if it is considered that the identification is wrong, it can also click the "Feedback" button to send information to the platform, and the platform will further optimize the training effect after collection.

[0023] Preferably, the data set in step seven is composed of an input layer, an output layer and two hidden layers, wherein the input layer is six factors affecting the type of disease and pest, the output layer is 15 types of disease and pest, and finally we obtain the disease and pest occurrence regularity data set, the total data amount is 959; the proportion of training set and test set in the data set is 80% and 20% respectively; the meteorological data is obtained by accessing the weather information from China Meteorological Network, and the obtained data includes time, place, temperature, precipitation, relative humidity, weather, wind level, which is stored according to local data structure and then processed, finally forming a standard data set.

[0024] Preferably, in step eight, the specific process of building a multiple linear regression model is:

[0025] Prob = 1 - ΔM * w0-ΔT * w1-ΔH * w2-ΔS * w3-ΔR * w4-ΔW * w5+α·Identify_number

[0026] Wherein, Prob represents the probability of disease and pest occurrence, ΔM, ΔT, ΔH, ΔS, ΔR, ΔW are the differences between real-time meteorological data and baseline data month, temperature, humidity, light, rainfall and wind level respectively, w0, w1, w2, w3, w4, w5 are the weights of each feature difference, Identify_number is the real-time disease and pest image recognition result, and α is the corresponding weight.

[0027] Preferably, in step nine, if the model needs to be updated, such as adding or reducing a certain disease or pest, or there is a new description of the occurrence of a certain disease, the model is retrained by modifying the data in the training set, and the original model is replaced.

[0028] Preferably, step ten is specifically:

[0029] S1, first obtain the current time period meteorological information, and intervalize it to obtain data consistent with the disease and pest occurrence regularity data format, and then compare the two to obtain the feature difference required by the prediction model;

[0030] S2, at the same time, the disease and pest recognition model obtains real-time disease and pest images at certain time intervals, and identifies whether they contain a certain disease or pest, and if the image contains disease and pest information, it is updated;

[0031] S3, finally, the feature difference and the recognition result are jointly input into the prediction model to obtain the probability of occurrence of each disease and pest in the current time period.

[0032] Preferably, the recognition result of the disease and pest automatic recognition system in step eleven is specifically: related users send images to the server for disease and pest recognition through various channels, and also send multi-dimensional information such as location, orchard name, and time; the background identifies and feeds back the results to the user, and also analyzes and organizes the additional information, forms a three-dimensional record based on location, disease and pest species, and time, and stores it in the server for disease and pest warning.

[0033] Preferably, the disease and pest warning module in step eleven is specifically:

[0034] According to the data collected by the background, the recognition results of different areas of the demonstration orchard are statistically analyzed according to the dimensions of "seven days, one month, and three months", and a pie chart is displayed;

[0035] According to the data collected by the background, the recognition results of different areas of the demonstration orchard are statistically analyzed according to the dimensions of "seven days, one month, and three months", and a pie chart is displayed;

[0036] When the number of disease and pest recognition in a certain area or a certain disease and pest reaches a certain threshold, the warning starts; the warning rules are:

[0037] Point: the disease and pest occurrence rate of a certain area reaches 80%, and the area is warned;

[0038] Surface: multiple areas occur a disease at the same time, and the disease is warned;

[0039] Color: according to the emergency, multiple colors are displayed, red for a more serious trend, yellow for the second, and green for the weakest.

[0040] Compared with the prior art, the present application has the beneficial effects that:

[0041] The present application proposes a disease and pest identification and prediction algorithm based on deep learning, which mainly includes the construction of disease and pest identification and prediction data set, disease and pest identification and disease and pest occurrence prediction model implementation and other parts. A large number of experiments show that the disease and pest identification model precision proposed by the present application can reach 96%, and the prediction model can realize more than 80% precision. Based on the above implementation, the present application also develops a disease and pest identification and prediction system, which can automatically and real-time identify diseases and pests, and obtain meteorological information to realize efficient prediction and early warning of diseases and pests. BRIEF DESCRIPTION OF DRAWINGS

[0042] Fig. 1 is a part of the disease and pest image of the present application;

[0043] Fig. 2 is a honey pomelo disease and pest prediction flow chart of the present application;

[0044] Fig. 3 is a thrips pest identification result of the present application;

[0045] Fig. 4 is a honey pomelo disease and pest prediction identification result of the present application;

[0046] Fig. 5 is a honey pomelo disease and pest image library of the present application (taking sunburn fruit as an example);

[0047] Fig. 6 is a honey pomelo disease and pest image library of the present application (taking normal leaf as an example);

[0048] Fig. 7 is a ResNet34 residual block network structure diagram of the present application;

[0049] Fig. 8 is a honey pomelo disease and pest image deep feature library of the present application (taking sunburn fruit as an example);

[0050] Fig. 9 is a honey pomelo disease and pest image deep feature library of the present application (taking normal fruit as an example);

[0051] Fig. 10 is a disease and pest occurrence rule overview diagram of the present application;

[0052] Fig. 11 is a disease and pest analysis module (according to region + time) of the present application;

[0053] Fig. 12 is a disease and pest analysis module (according to region + time + disease and pest species) of the present application;

[0054] Fig. 13 is a warning effect diagram of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0056] Please refer to FIGS. 1-13, the present application provides a disease and pest identification and prediction algorithm based on deep learning, including the following steps:

[0057] Step one, material selection: choose honey pomelo as the detection target, and the honey pomelo dataset is collected from multiple plantations in Ji'an City, Jiangxi Province, China;

[0058] Step two, disease and pest species dataset construction: use multiple collection devices with different resolutions to take multiple-angle photos of honey pomelo under different time periods of natural light conditions;

[0059] Step three, disease and pest identification model construction: in order to construct an efficient honey pomelo classification model, ResNet34 is selected as the pre-training model, which has a deep network structure and excellent feature extraction capability, and a series of latest technologies are introduced to improve its performance and adaptability; by analyzing the actual situation of the honey pomelo disease and pest dataset and performing deep technical optimization, including carefully adjusting the model to better adapt to the growth environment and disease and pest characteristics of honey pomelo;

[0060] Step four, binary classification model establishment: a binary classification model dedicated to classifying non-honey pomelo / honey pomelo images is established by collecting the dataset of other categories and honey pomelo images;

[0061] Step five, multi-classification deep learning model image feature extraction and feature library construction: use the pre-trained AlexNet, VGG16 / 19, GoogLeNet and ResNet deep models on a large-scale image library to perform feature extraction operations on the segmented image blocks, and optimize the related parameters to fine-tune and verify the model until the expected recognition accuracy is achieved; by using different fusion methods to enhance the feature representation ability of the features, the features are then input into SVM, Softmax and Bayesian classifier, and the probability output is given again through post-fusion to give the class prediction result of the disease and pest; the related parameters include batch_size, learning rate, learning drop strategy, momentum factor, weight decay factor and Dropout; different fusion methods include feature concatenation and element-wise sum / product, and post-fusion methods include weighted fusion and dot product;

[0062] Step six, build an automatic pest identification system: a cascading method is used to combine the binary classification model with the multi-classification deep learning model of the citrus pest image to further improve the practicality of the entire pest classification system; through the Python Flask framework, an API is provided for citrus pest identification, returning the prediction results in the form of probability; in this system, the binary classification model is responsible for determining whether the image contains citrus pests, while the multi-classification deep learning model focuses on more detailed classification of pests;

[0063] Step seven, construction of pest occurrence regularity dataset: in order to predict the occurrence of pests in advance and take targeted measures, a pest occurrence regularity dataset is created based on historical pest occurrence information in Ji'an region;

[0064] Step eight, construction of pest prediction model: for the citrus pest prediction model, first, obtain the meteorological data, and perform the first division based on temperature. If the temperature is less than -7 degrees, it is prone to freeze damage. The second division is based on the month. When the month is December, January, and February, the occurrence probability of various pests is relatively low. Finally, compare the meteorological data with the pest occurrence regularity data to calculate the probability, and combine the real-time pest image recognition result for calibration to build a multiple linear regression model to realize this process;

[0065] Step nine, model correction and update: a test sample set is constructed, and some occurrence data from March to November, the key months of local pest occurrence, are selected and input into the model for prediction result evaluation. Based on the evaluation results, the model parameter weights are corrected, and the model accuracy is gradually improved through iteration;

[0066] Step ten, construction of pest identification and prediction system: use the citrus pest identification model and prediction model to build a real-time pest identification and prediction system;

[0067] Step eleven, construction of pest warning model: includes the identification results of the automatic pest identification system and the pest warning module.

[0068] Step two is: in order to make the photographed images more consistent with the real natural complex environment, in addition to clear pest leaves in the images, there are also real leaf background information. In order to ensure that the pest labels corresponding to the leaves in each image are real and reliable, artificial classification and identification are performed by experts in the relevant field; the final pest dataset exceeds 20,000, including 13 categories.

[0069] Preferably, the third step is specifically: using the method of transfer learning, pre-training ResNet34 on a large image dataset to enable the model to capture general image features; in the fine-tuning stage, introduce the latest gradient accumulation technology, effectively improve the robustness of the model for few-shot learning by accumulating gradients on multiple small batches, so as to better adapt to the characteristics of honey pomelo disease and pest images;

[0070] In order to overcome the training problem caused by the imbalance of disease and pest categories, an improved BalanceLoader is used as a data sampler; This strategy is not just a simple weight adjustment, but also combines the latest sample rearrangement method; This sample rearrangement method dynamically adjusts the sampling order of samples according to the performance of the model in the first few rounds of training, which can more effectively guide the model to focus on difficult-to-classify samples;

[0071] During the model training process, a self-supervised learning method is introduced; By processing the collected data, the model automatically generates labels for the data, thereby further enriching the database and enhancing the model's ability to learn the characteristics of honey pomelo diseases and pests.

[0072] The sixth step is specifically:

[0073] The calling interface built by Python Flask framework enables related users to easily interact with our disease and pest classification system; Related users can send image data to the interface through various channels and obtain the identification result of honey pomelo diseases and pests. For the identification result, the user can click "Detailed Information" to further view the introduction and control of the corresponding disease and pest, so as to more effectively take corresponding measures; If you think the identification is wrong, you can also click the "Feedback" button to send information to the platform, which will further optimize the training effect after collection.

[0074] The data set in step seven is composed of an input layer, an output layer and two hidden layers, where the input layer is 6 factors affecting the disease and pest species, the output layer is 15 disease and pest species, and finally we get the disease and pest occurrence regularity data set, with a total data amount of 959; The proportion of training set and test set in the data set is 80% and 20% respectively; Meteorological data is obtained by accessing weather information from China Meteorological Network, including time, place, temperature, precipitation, relative humidity, weather, wind level, and is stored in local data structure after processing to form a standard data set.

[0075] Disease and pest species (output layer)

[0076] 6 factors affecting the disease and pest species (input layer)

[0077] Month: value 1-12

[0078] Temperature: divide temperature into several grades

[0079] Weather data acquisition

[0080] Weather information is obtained from the China Weather Network through an interface. The obtained data includes time, location, temperature, precipitation, relative humidity, weather, wind level, etc. After storage according to the local data structure, the data is processed to form a standardized data set.

[0081] In step eight, the specific process of building a multiple linear regression model is as follows:

[0082] Prob = 1 - ΔM * w0- ΔT * w1- ΔH * w2- ΔS * w3- ΔR * w4- ΔW * w5+ α·Identify_number

[0083] Wherein, Prob represents the probability of occurrence of pests and diseases, ΔM, ΔT, ΔH, ΔS, ΔR, ΔW are the differences between real-time weather data and baseline data, respectively, month difference, temperature difference, humidity difference, light difference, rainfall difference, wind level difference, w0, w1, w2, w3, w4, w5 are the weights corresponding to each feature difference, Identify_number is the identification result of real-time pest and disease image, and α is the corresponding weight.

[0084] Preferably, in step nine, if the model needs to be updated, such as adding or reducing a certain pest or disease, or there is a new description of the occurrence of a certain disease, the data in the training set is modified, the model is retrained, and the original model is replaced.

[0085] Prediction effect of honey pomelo pests and diseases

[0086] The data corresponding to the time of August is selected for testing, and the test results are shown in the following table. The model can accurately predict the probability of occurrence of various pests and diseases in this time period. The prediction accuracy reaches more than 80% on the entire test set.

[0087] Step ten is as follows:

[0088] S1, first acquire the current time period weather information, and interval processing is obtained with the pest and disease occurrence regularity data format consistent data, and then the two are compared, the characteristics of the required prediction model difference;

[0089] S2, at the same time, the pest and disease identification model with a certain time interval to obtain real-time pest and disease image, and identify whether it contains a certain pest and disease, if the image contains pest and disease information, then update;

[0090] S3, finally, the feature difference and the identification result are input into the prediction model, and the probability of occurrence of each pest and disease in the current time period is obtained.

[0091] The identification result of the pest and disease automatic identification system collected in step eleven is: by using the calling interface, the relevant users send images to the server for pest and disease identification through various channels, and also send multi-dimensional information such as location, orchard name, time, etc.; After the background identification, the feedback result is fed back to the user, and the additional information is also analyzed and sorted, forming a three-dimensional record based on location, pest and disease type, and time, and stored in the server for pest and disease warning.

[0092] The pest and disease warning module in step eleven is:

[0093] According to the data collected by the background, the identification results of different areas of the demonstration orchard are statistically analyzed according to the dimensions of "seven days, one month, three months", and a pie chart is given, as shown in Figure 11;

[0094] According to the data collected by the background, the identification results of different areas of the demonstration orchard are statistically analyzed according to the dimensions of "seven days, one month, three months", and a pie chart is given, as shown in Figure 11;

[0095] When the number of pest and disease identification in a certain area or a certain pest and disease reaches a certain threshold, the warning begins; The warning rule is:

[0096] Point: the occurrence rate of pest and disease in a certain area reaches 80%, and the area is warned;

[0097] Face: multiple areas occur a disease at the same time, and the disease is warned;

[0098] Color: according to the emergency, multiple colors are displayed, red for serious trend, yellow for second, and green for weakest, as shown in Figure 13.

[0099] The pest and disease identification data of the present application comes from the constructed honey pomelo pest and disease species data set, including more than 20,000 images of 13 species. The pest and disease prediction data comes from the pest and disease occurrence regularity data set we constructed, including 959 data of 14 species. The pest and disease prediction model test data comes from the additional constructed pest and disease occurrence regularity data set, which is realized by randomly collecting meteorological information in each month of the year, including 40 data of 14 species. The pest and disease identification pre-training model is ResNet34, which is trained using the Pytorch framework and on the NVIDIA Tesla T4 server. In the pest and disease prediction model, w0, w1, w2, w3, w4, w5 are 0.1, 0.15, 0.05, 0.04, 0.06, 0.04 respectively, and a is 0.1. The constructed model is used to develop a honey pomelo pest and disease identification website, and the accurate and efficient identification of pests and diseases is realized through the double-end interaction of the applet and the website.

[0100] Honey pomelo pest and disease identification effect:

[0101] The image of the fruit containing the thrips disease is identified, as shown in Figure 3.

[0102] The results show that the model not only can correctly identify the honey pomelo pest and disease, but also can give the probability of identifying other pests and diseases. In addition, in the whole data set test, the accuracy of the model is 96%.

[0103] Honey pomelo pest and disease prediction effect:

[0104] The present application selects the data of frequent occurrence of pests and diseases for testing, and the corresponding time is August, and the test result is shown in Figure 4. The model can accurately predict the probability of occurrence of various pests and diseases in this time period, and the prediction accuracy reaches more than 80% on the whole test set.

[0105] The present application proposes a pest and disease identification and prediction algorithm based on deep learning, which mainly includes the construction of pest and disease identification and prediction data set, the realization of pest and disease identification and pest and disease occurrence prediction model, etc. A large number of experiments show that the pest and disease identification model proposed in the present application can achieve an accuracy of 96%, and the prediction model can achieve an accuracy of more than 80%. Based on the above realization, the present application also develops a pest and disease identification and prediction system, which can automatically and real-time identify pests and diseases, and obtain meteorological information to realize efficient prediction and early warning of pests and diseases.

[0106] Although the present application has been described in detail with reference to the foregoing embodiments, the technical solutions recorded in the foregoing embodiments can be modified, or some of the technical features can be replaced by equivalent features, by those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based pest and disease identification and prediction algorithm, characterized by: The following steps are involved: Step 1: Material selection: Honey pomelo was selected as the detection target. The honey pomelo dataset was collected from multiple plantations in Ji'an City, Jiangxi Province, China. Step 2: Constructing a dataset of pests and diseases: Using a variety of acquisition devices with different resolutions, photograph the honey pomelo from multiple angles under natural lighting conditions at different time periods. Step 3: Pest and disease identification model construction: To build an efficient honey pomelo classification model, we selected ResNet34 as the pre-trained model. This model has a deep network structure and excellent feature extraction capabilities. We also introduced a series of cutting-edge technologies to improve its performance and adaptability. We analyzed the actual honey pomelo pest and disease dataset and conducted in-depth technical optimization, including careful adjustments to the model to better adapt to the honey pomelo's growth environment and pest and disease characteristics. Step 4: Build a binary classification model: Using the collected datasets of other categories and pomelo images, build a binary classification model specifically for classifying non-pomelo / pomelo images; Step 5: Image feature extraction and feature library construction for a multi-classification deep learning model: AlexNet, VGG16 / 19, GoogLeNet, and ResNet deep models, pre-trained on a large-scale image library, are used to perform feature extraction on the segmented image blocks. The models are fine-tuned, trained, and verified by optimizing relevant parameters until the expected recognition accuracy is achieved. The representational capabilities of the features are enhanced by using different fusion methods for multiple deep features. These features are then fed into SVM, Softmax, and Bayesian classifiers, and their probability outputs are post-fused to give the pest and disease category prediction results. Relevant parameters include batch size, learning rate, learning descent strategy, momentum factor, weight decay factor, and Dropout. Different fusion methods include feature concatenation and element-wise sum / product, while post-fusion methods include weighted fusion and dot product. Step 6: Build an automatic pest and disease identification system. This involves combining a binary classification model with a multi-classification deep learning model for pomelo pest and disease images using a cascade approach to further enhance the practicality of the entire pest and disease classification system. Using the Python Flask framework, a calling interface is provided for pomelo pest and disease identification, returning predictions for various pests and diseases in the form of probabilities. In this system, the binary classification model determines whether an image contains pomelo pests and diseases, while the multi-classification deep learning model provides a more detailed classification of pests and diseases. Step 7: Constructing a dataset on pest and disease occurrence patterns: To predict pest and disease occurrences in advance and take targeted measures, a dataset on pest and disease occurrence patterns was created based on historical information on pomelo pest and disease occurrences in Ji'an. Step 8: Pest and disease prediction model construction: For the honey pomelo pest and disease prediction model, meteorological data is first obtained and divided according to temperature. If the temperature is less than -7 degrees Celsius, frost damage is likely to occur. A second division is made according to the month. When the months are December, January, and February, the probability of occurrence of various pests and diseases is low. Finally, the probability is calculated by comparing the meteorological data with the pest and disease occurrence pattern data. The results are combined with real-time pest and disease image recognition to calibrate the results and construct a multivariate linear regression model to achieve this process. Step 9: Model revision and update: Construct a test sample set. Select some local pest and disease occurrence data from March to November, the key months with recorded local pests and diseases. Input these data into the model to evaluate the prediction results. Based on the evaluation results, adjust the model parameter weights and gradually improve the model accuracy through iteration. Step 10: Construction of pest and disease identification and prediction system: Utilize the honey pomelo pest and disease identification model and prediction model to build a real-time pest and disease identification and prediction system; Step 11: Construction of pest and disease early warning model: including collecting the identification results of the pest and disease automatic identification system and the pest and disease early warning module.

2. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: Specifically, step 2 includes the following steps: to make the captured images more consistent with the real natural and complex environment, in addition to clear diseased and insect-infested leaves, the images also include real leaf background information. To ensure that the disease and insect pest labels corresponding to the leaves in each image are authentic and reliable, manual classification and identification are performed by experts in related fields. The final disease and insect pest dataset exceeds 20,000 images, including 13 species.

3. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: Specifically, step three involves: using transfer learning to pre-train ResNet34 on a large-scale image dataset, enabling the model to capture common image features; during the fine-tuning phase, introducing the latest gradient accumulation technology to effectively improve the model's robustness to few-shot learning by accumulating gradients over multiple small batches, thereby better adapting to the characteristics of honey pomelo pest and disease images; To overcome training issues caused by imbalanced pest and disease classifications, an improved BalanceLoader was used as the data sampler. This strategy goes beyond simple weight adjustment and incorporates a state-of-the-art sample reordering method. This method dynamically adjusts the sample order based on the model's performance in previous rounds of training, effectively guiding the model to focus on difficult-to-classify samples. During the model training process, a self-supervised learning method was introduced; by processing the collected data, the model automatically generated labels for the data, thereby further enriching the database and enhancing the model's ability to learn the characteristics of pomelo pests and diseases.

4. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: The specific steps of step six are as follows: A Python Flask-based API allows users to easily interact with our pest and disease classification system. Users can send image data to the API through various channels to obtain identification results for pomelo pests and diseases. Users can click "Details" to view detailed information about the identified pests and diseases, enabling them to more effectively implement appropriate measures. If they believe the identification is incorrect, they can click "Feedback" to submit a response to the platform, which will collect the information and use it to further optimize the training results.

5. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: The data set in step seven consists of an input layer, an output layer, and two hidden layers. The input layer is composed of 6 factors that affect the types of pests and diseases, and the output layer is composed of 15 types of pests and diseases. Finally, we obtain a data set of the occurrence patterns of pests and diseases, with a total data volume of 959. The training set and test set account for 80% and 20% of the data set respectively. The meteorological data is obtained by obtaining weather information from the China Meteorological Network through an interface. The obtained data includes time, location, temperature, precipitation, relative humidity, weather, and wind level. After being stored in the local data structure, it is processed to form a standardized data set.

6. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: In step eight, the specific process of constructing the multiple linear regression model is: Prob = 1-ΔM * w0-ΔT * w1-ΔH * w2-ΔS * w3-ΔR * w4-ΔW * w5+α·Identify_number Among them, Prob represents the probability of occurrence of pests and diseases, ΔM, ΔT, ΔH, ΔS, ΔR, and ΔW are the monthly difference, temperature difference, humidity difference, light difference, rainfall difference, and wind level difference between the real-time meteorological data and the benchmark data, respectively, w0, w1, w2, w3, w4, and w5 are the weights corresponding to each feature difference, Identify_number is the real-time pest and disease image recognition result, and α is its corresponding weight.

7. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: In step nine, if the model needs to be updated, such as adding or reducing a certain pest or disease, or a new description of the occurrence pattern of a certain disease, the model is retrained by modifying the data in the training set to replace the original model.

8. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: The step ten is specifically as follows: S1. First, obtain the meteorological information of the current time period and perform interval processing on it to obtain data in the same format as the data on the occurrence pattern of pests and diseases. Then, compare the two to obtain the characteristic difference required by the prediction model; S2. At the same time, the pest and disease recognition model obtains real-time pest and disease images at a certain time interval and identifies whether they contain certain pests and diseases. If the image contains pest and disease information, it is updated; S3. Finally, the feature difference and the recognition result are input into the prediction model to obtain the probability of occurrence of each disease and insect pest in the current time period.

9. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: The specific steps of collecting the identification results of the automatic pest and disease identification system in step 11 are as follows: using the calling interface, the relevant users send images to the server through multiple channels for pest and disease identification, and also send multi-dimensional information such as location, orchard name, and time; while the background identifies and feeds back the results to the user, it also analyzes and organizes the additional information to form a three-dimensional record based on location, pest and disease type, and time, which is stored in the server for use in pest and disease early warning.

10. The deep learning-based pest identification and prediction algorithm according to claim 1, characterized in that: The pest warning module in step 11 is specifically: Based on the data collected by the backend, a statistical analysis of the identification results of demonstration orchards in different regions was conducted based on the dimensions of "seven days, one month, and three months", and a pie chart was presented. Based on the data collected by the backend, statistical analysis of different types of pests and diseases is conducted according to the dimensions of "seven days, one month, and three months", and a radar chart is provided; When the number of identified pests or diseases in a certain area reaches a certain threshold, an early warning will be initiated; the early warning rules are as follows: Point: If the incidence of pests and diseases in a certain area reaches 80%, an early warning will be issued for the area; Surface: If a disease occurs in multiple regions at the same time, an early warning will be issued for that disease; Color: Displayed in multiple colors according to the urgency of the situation. Red means the trend is more serious, yellow is the second most serious, and green is the weakest.

Images