Wheat stripe rust and gibberellic disease prediction method and system based on deep learning

Abstract

The invention discloses a wheat stripe rust and gibberellic disease prediction method and system based on deep learning, and the method comprises the steps: obtaining the feature data sets of stripe rust and gibberellic disease of different varieties of wheat in different regions, and constructing a training set and a test set; collecting knowledge related to occurrence of wheat stripe rust and gibberellic disease, and constructing a knowledge graph of wheat stripe rust and gibberellic disease; determining main factors of occurrence of wheat stripe rust and gibberellic disease based on the mapping knowledge domain of wheat stripe rust and gibberellic disease; quantifying the main factors of occurrence of the wheat stripe rust and the gibberellic disease, and establishing a wheat stripe rust and gibberellic disease prediction model; training the wheat stripe rust and gibberellic disease prediction model by adopting a training set; and predicting wheat in the target area by using the trained wheat stripe rust and gibberellic disease prediction model to obtain a wheat stripe rust and gibberellic disease prediction result. The method has the remarkable effects that the loss of characteristic information which is more sensitive to the wheat stripe rust and the gibberellic disease is prevented, the accurate and stable prediction of the wheat stripe rust and the gibberellic disease is realized, and a scientific basis can be provided for the early prevention and control of the wheat stripe rust and the gibberellic disease.

Description

Technical Field

[0001] The present invention relates to the technical field of crop disease prevention, and in particular to a method and system for predicting wheat stripe rust and head blight based on deep learning. Background Art

[0002] Wheat is the main food crop in my country, and wheat stripe rust and fusarium head blight are one of the main reasons for wheat production reduction. Wheat stripe rust and fusarium head blight are characterized by wide incidence and strong infectiousness. They have been prevalent many times in history, and in severe cases, they can reduce production by more than 60%. Therefore, accurate prediction of the occurrence of wheat stripe rust and fusarium head blight is of great guiding significance for ensuring food security and for plant protection departments to formulate effective prevention and control policies.

[0003] Commonly used meteorological prediction models for crop diseases, such as multiple linear regression, support vector regression, and stepwise regression models, have the following defects: since suitable meteorological conditions will accelerate the occurrence of crop diseases, their spectral curves will change, so the occurrence of crop diseases is closely related to meteorological factors and vegetation indices. However, the meteorological prediction model for crop diseases is not coupled with vegetation indices, and the model itself lacks support for the epidemiological mechanism of crop diseases, which leads to unclear physical meanings of the predicted values. Therefore, the root mean square error of the disease prediction model established based on multiple linear regression and support vector regression is relatively high.

[0004] However, among the various factors that affect the occurrence of stripe rust and fusarium head blight, meteorological factors can be easily observed and analyzed, so most of the existing research is based on meteorological data to predict the severity of the disease. However, the prediction method of wheat stripe rust and fusarium head blight based on meteorological data can only predict the trend of disease occurrence, and it is difficult to achieve accurate prediction over a large area, which cannot meet the actual application needs.

[0005] In addition, in actual regional field management, in addition to meteorological factors, differences in geographical environment and wheat varieties can also cause significant interference to changes in crop growth conditions caused by disease stress. However, most existing studies often ignore the impact of these differences in actual production on crop disease prediction, resulting in inaccurate predictions. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method and system for predicting wheat stripe rust and fusarium head blight based on deep learning. Based on knowledge graph and deep learning technology, it can be used to solve the problems of inaccurate prediction caused by incomplete data utilization and improper selection of disease-specific factors in the existing prediction methods of wheat stripe rust and fusarium head blight.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] In the first aspect, the present invention proposes a method for predicting wheat stripe rust and head blight based on deep learning, which mainly comprises the following steps:

[0009] Step 1: Obtain the existing stripe rust and head blight feature data sets of different varieties of wheat in different regions, and construct training sets and test sets after data preprocessing;

[0010] Step 2: Collect knowledge related to the occurrence of wheat stripe rust and fusarium head blight from the preprocessed feature data set, and construct a knowledge graph of wheat stripe rust and fusarium head blight;

[0011] Step 3: Based on the wheat stripe rust and fusarium head blight knowledge graph, determine the main factors causing the wheat stripe rust and fusarium head blight, wherein the main factors causing the wheat stripe rust and fusarium head blight include meteorological factors, geographical environment factors, and variety factors;

[0012] Step 4, quantifying the main factors causing the occurrence of wheat stripe rust and fusarium head blight, and establishing a prediction model for wheat stripe rust and fusarium head blight based on an attention neural network model;

[0013] Step 5: training the wheat stripe rust and fusarium head blight prediction model using a training set, and verifying the prediction accuracy of the trained model using a test set; if the accuracy meets the standard, the training is completed; otherwise, the training is continued after updating the parameters;

[0014] Step 6: Use the trained wheat stripe rust and fusarium head blight prediction model to predict wheat in the target area to obtain wheat stripe rust and fusarium head blight prediction results.

[0015] Furthermore, the stripe rust and head blight characteristic data set in step 1 includes hyperspectral image data, high-resolution image data, geographic environment data and meteorological data.

[0016] Furthermore, the data preprocessing includes data labeling, data enhancement, abnormal data removal, and data normalization operations.

[0017] Furthermore, the specific process of constructing the knowledge graph of wheat stripe rust and fusarium head blight in step 2 is as follows:

[0018] Step 2.1, knowledge expression: express the disease detection knowledge and metadata knowledge in the preprocessed feature data set in structured, semi-structured, and unstructured forms;

[0019] Step 2.2, knowledge extraction: first, entity acquisition is performed, and regular expressions and hidden Markov models are used to obtain structured disease detection knowledge and metadata knowledge; then attribute acquisition is performed, and support vector machines are used to collect specific attribute information from the feature data set; finally, relationship extraction is performed, and the LTP tool is used to perform semantic dependency syntactic analysis on the text in the feature data set to obtain the relationship between entities;

[0020] Step 2.3, knowledge fusion: Use entity disambiguation and coreference resolution to connect knowledge to the correct entity object to complete knowledge fusion;

[0021] Knowledge fusion is the process of knowledge organization and information fusion. It obtains implicit or valuable new knowledge through the acquisition, matching, integration, and mining of knowledge from numerous scattered and heterogeneous resources, while optimizing the structure and connotation of knowledge and providing knowledge services.

[0022] Step 2.4, knowledge representation: using the framework theory representation method, the data after knowledge extraction and knowledge fusion are represented in a structured form;

[0023] Knowledge representation is to transform these factors into structured data that can be understood and calculated by computers. It requires knowledge applied to data structures in order to better describe and define knowledge.

[0024] Step 2.5, knowledge processing: perform knowledge reasoning and quality assessment on the data after knowledge representation to obtain qualified knowledge;

[0025] Step 2.6: Construct the ontology layer and entity layer: Use the knowledge-processed data to construct the ontology layer and entity layer;

[0026] Step 2.7, knowledge graph construction: After extracting the entity information and the relationships between different entities in the entity layer, the extracted entity-relationship-entity triples are stored in the database and depicted as a graphical knowledge graph structure to complete the construction of the knowledge graph of wheat stripe rust and fusarium head blight.

[0027] Furthermore, the meteorological factors include the minimum humidity, maximum humidity, average humidity, average temperature, rainfall, sunshine days, average wind speed, average wind direction, average wind force, minimum evaporation, maximum evaporation, and average evaporation of the air;

[0028] The geographical environmental factors include the minimum humidity, maximum humidity, average humidity, minimum temperature, maximum temperature, average temperature, types and contents of elements required for growth, and types and abundance of microorganisms;

[0029] The variety factors include wheat moisture content, chlorophyll content, canopy structure and coverage, and wheat vegetation index.

[0030] Furthermore, the wheat stripe rust and fusarium head blight prediction model includes a feature extraction module, a spatial structure attention module, a feature fusion module, a coding and decoding module, and a classification and recognition module, wherein:

[0031] The feature extraction module is used to extract multi-scale features from the input training set data;

[0032] The spatial structure attention module is used to encode broader contextual information in the extracted multi-scale features into local features;

[0033] The feature fusion module is used to perform feature fusion on the multi-scale features output by the spatial structure attention module;

[0034] The encoding and decoding module is used to use multiple convolution kernels of different sizes to perform convolution on the fused feature matrix to extract feature information of different fields of view to obtain multiple feature maps; then strengthen important features and suppress useless features on the multiple feature maps to obtain multiple feature maps based on weight selection; splice the multiple feature maps based on weight selection in the channel dimension to obtain a spliced ​​feature map;

[0035] The classification and recognition module is used to flatten the spliced ​​feature map into a one-dimensional vector and project it to the vector space of the tail entity to obtain a hidden layer vector, where the dimension of the hidden layer vector is the same as the dimension of the tail entity vector t; perform matrix multiplication on the hidden layer vector and all the tail entity vectors in the knowledge graph to obtain the triple score of the head entity vector relationship vector and the tail entity vector; and classify and recognize stripe rust and fusarium according to the triple score.

[0036] Furthermore, the feature extraction module includes a hyperspectral image feature extraction unit, a high-resolution image feature extraction unit, a geographic environment data feature extraction unit and a meteorological data feature extraction unit, wherein:

[0037] The hyperspectral image feature extraction unit is used to extract wheat spectral features from the hyperspectral image data in the stripe rust and head blight feature data set, and upsample the obtained hyperspectral impact wheat feature map;

[0038] The high-resolution image feature extraction unit is used to extract wheat texture morphological spatial structure features from the high-resolution image data in the stripe rust and scab feature data sets;

[0039] The geographic environment data feature extraction unit is used to perform semantic level search and clustering on the geographic environment data in the stripe rust and head blight feature data set to extract the geographic environment features of the wheat planting area;

[0040] The meteorological data feature extraction unit is used to perform semantic level search and clustering on the meteorological data in the stripe rust and head blight feature data sets to achieve meteorological feature extraction in the wheat planting area.

[0041] Furthermore, the encoding module includes a channel adaptive selection unit and a receptive field adaptive selection unit. The channel adaptive selection unit is configured with a wavelet soft shrinkage subnetwork for denoising a plurality of feature maps to obtain a plurality of denoised feature maps; the receptive field adaptive selection unit is used to process the plurality of denoised feature maps separately to obtain a plurality of feature maps based on weight selection.

[0042] Furthermore, the training process of the wheat stripe rust and fusarium head blight prediction model is as follows:

[0043] Step 51, initializing the connection weights and thresholds of the model;

[0044] Step 52: Given input x i and the target output y i ;

[0045] Step 53: Determine the number of neurons in the hidden layer:

[0046] Step 54: The preceding propagation process of the signal:

[0047] Let the input of the i-th node be x i , the hidden layer input is I i , the output is limited to a certain range through the transfer function f, and the output O is obtained i :

[0048] The output of each layer is used as the input of the next layer, and the final output is obtained by iterating through the designed network.

[0049] Step 55, calculate the mean square error: k With the expected output T k Calculate mean square error;

[0050] Step 56, error back propagation process: calculate the gradient of each parameter in the model and optimize the parameters by back propagation;

[0051] Step 57: Determine whether the model prediction accuracy meets the requirement. If yes, stop training. Otherwise, repeat steps 54 to 56 until the entire training error meets the prediction accuracy requirement.

[0052] In a second aspect, the present invention proposes a wheat stripe rust and fusarium head blight prediction system based on deep learning for implementing the method described in the first aspect, comprising:

[0053] The data acquisition module is used to obtain the existing stripe rust and head blight characteristic data sets of different varieties of wheat in different regions, and to construct training sets and test sets after data preprocessing;

[0054] The knowledge graph construction module is used to collect knowledge related to the occurrence of wheat stripe rust and fusarium head blight from the preprocessed feature data set and construct a knowledge graph of wheat stripe rust and fusarium head blight;

[0055] A disease-causing factor determination module is used to determine the main factors causing wheat stripe rust and fusarium head blight according to the wheat stripe rust and fusarium head blight knowledge graph, wherein the main factors causing wheat stripe rust and fusarium head blight include meteorological factors, geographical environment factors, and variety factors;

[0056] A prediction model building module, used to quantify the main factors causing the occurrence of wheat stripe rust and fusarium head blight, and to establish a prediction model for wheat stripe rust and fusarium head blight based on a deep learning neural network model;

[0057] A prediction model training module is used to train the wheat stripe rust and head blight prediction model using a training set, and to verify the prediction accuracy of the trained model using a test set. If the accuracy meets the standard, the training is completed; otherwise, the training continues after updating the parameters;

[0058] The prediction module is used to predict wheat in the target area using the trained wheat stripe rust and fusarium head blight prediction models to obtain wheat stripe rust and fusarium head blight prediction results.

[0059] The remarkable effects of the present invention are:

[0060] The present invention establishes a knowledge graph of wheat stripe rust and fusarium scab by coupling hyperspectral image data, high-resolution image data, geographic environment data and meteorological data to determine the main factors of wheat stripe rust and fusarium scab, and then quantifies the main factors of wheat stripe rust and fusarium scab to establish a prediction model of wheat stripe rust and fusarium scab based on an attention neural network model, and predicts wheat in the target area. The above method steps can be effectively and intuitively applied to the detection of changes in wheat physiological states, and by analyzing a variety of relevant data of wheat stripe rust and fusarium scab, disease-specific factors for identifying wheat stripe rust and fusarium scab are obtained, and characteristic information that is more sensitive to wheat stripe rust and fusarium scab is prevented from being lost, thereby achieving accurate and stable prediction and forecasting of wheat stripe rust and fusarium scab, and providing a scientific basis for the early prevention and control of wheat stripe rust and fusarium scab. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a flow chart of the method of the present invention;

[0062] Figure 2 It is a flowchart for building a knowledge graph;

[0063] Figure 3 It is a schematic diagram of the structure of the prediction model for wheat stripe rust and fusarium head blight;

[0064] Figure 4 This is the training flow chart of the prediction model for wheat stripe rust and fusarium head blight;

[0065] Figure 5 It is a principle block diagram of the system of the present invention. DETAILED DESCRIPTION

[0066] The specific implementation manner and working principle of the present invention are further described in detail below with reference to the accompanying drawings.

[0067] Embodiment 1:

[0068] like Figure 1 As shown, this embodiment proposes a method for predicting wheat stripe rust and fusarium head blight based on deep learning, and the specific steps are as follows:

[0069] Step 1: Obtain the existing stripe rust and head blight feature data sets of different varieties of wheat in different regions, and construct training sets and test sets after data preprocessing;

[0070] In this example, the stripe rust and head blight feature data sets include hyperspectral image data, high-resolution image data, geographic environment data, and meteorological data.

[0071] Furthermore, the data preprocessing includes data labeling, data enhancement, abnormal data removal, and data normalization operations.

[0072] In this embodiment, the wheat vegetation index includes: Normalized Difference Vegetation Index NDVI, Ratio Vegetation Index SR, Structure Enhanced Pigment Vegetation Index SIPI, Optimized Soil Conditioning Vegetation Index OSAVI, and Crop Nitrogen Response Index NRI.

[0073] Step 2: Collect knowledge related to the occurrence of wheat stripe rust and fusarium head blight from the preprocessed feature data set, and construct a knowledge graph of wheat stripe rust and fusarium head blight;

[0074] The wheat disease knowledge graph is a network structure formed by connecting wheat disease entities and the relationships between entities. Triples are a common representation form of knowledge graphs, and each piece of wheat disease knowledge in the knowledge graph can be intuitively represented as <head entity, relationship, tail entity>. As the most basic element in the knowledge graph, wheat disease entities (head entities or tail entities) are mainly composed of diseases, environmental information, symptoms, prevention and control methods, parts, etc.; relationships exist between different head entities or tail entities, mainly including categories, disease manifestations, causes, pathogenesis, preventive measures, pesticide effects, etc. The construction process of the wheat stripe rust and fusarium knowledge graph: First, express disease detection knowledge and metadata knowledge, and effectively organize and manage disease detection knowledge and metadata knowledge in the form of entity connections; then use disease knowledge and metadata to construct the ontology layer and entity layer, including data acquisition and processing (cleaning, fusion, etc.), ontology layer construction, entity layer construction, etc.; finally, construct the knowledge graph.

[0075] like Figure 2 As shown, the specific process of constructing the knowledge graph of wheat stripe rust and fusarium head blight in this embodiment is as follows:

[0076] Step 2.1, knowledge expression: express the disease detection knowledge and metadata knowledge in the preprocessed feature data set in structured, semi-structured, and unstructured forms;

[0077] Step 2.2, knowledge extraction: first, entity acquisition is performed, and regular expressions and hidden Markov models are used to obtain structured disease detection knowledge and metadata knowledge; then attribute acquisition is performed, and support vector machines are used to collect specific attribute information from the feature data set; finally, relationship extraction is performed, and the LTP tool is used to perform semantic dependency syntactic analysis on the text in the feature data set to obtain the relationship between entities;

[0078] Step 2.3, knowledge fusion: Use entity disambiguation and coreference resolution to connect knowledge to the correct entity object to complete knowledge fusion;

[0079] Knowledge fusion is the process of knowledge organization and information fusion. It obtains implicit or valuable new knowledge through the acquisition, matching, integration, and mining of knowledge from numerous scattered and heterogeneous resources, while optimizing the structure and connotation of knowledge and providing knowledge services.

[0080] Step 2.4, knowledge representation: using the framework theory representation method, the data after knowledge extraction and knowledge fusion are represented in a structured form;

[0081] Knowledge representation is to transform these factors into structured data that can be understood and calculated by computers. It requires knowledge applied to data structures in order to better describe and define knowledge.

[0082] Step 2.5, knowledge processing: perform knowledge reasoning and quality assessment on the data after knowledge representation to obtain qualified knowledge;

[0083] Step 2.6: Construct the ontology layer and entity layer: Use the knowledge-processed data to construct the ontology layer and entity layer;

[0084] Step 2.7, obtain the knowledge graph: After extracting the entity information and the relationships between different entities in the entity layer, store the extracted entity-relationship-entity triples in the database, depict them as a graphical knowledge graph structure, and complete the construction of the knowledge graph of wheat stripe rust and fusarium head blight.

[0085] Step 3: Based on the wheat stripe rust and fusarium head blight knowledge graph, determine the main factors causing the wheat stripe rust and fusarium head blight, wherein the main factors causing the wheat stripe rust and fusarium head blight include meteorological factors, geographical environment factors, and variety factors;

[0086] In this example, the meteorological factors include the minimum humidity, maximum humidity, average humidity, average temperature, rainfall, sunshine days, average wind speed, average wind direction, average wind force, minimum evaporation, maximum evaporation, and average evaporation;

[0087] The geographical environmental factors include the minimum humidity, maximum humidity, average humidity, minimum temperature, maximum temperature, average temperature, types and contents of elements required for growth, and types and abundance of microorganisms;

[0088] The variety factors include wheat moisture content, chlorophyll content, canopy structure and coverage, and wheat vegetation index.

[0089] Step 4, quantifying the main factors causing the occurrence of wheat stripe rust and fusarium head blight, and establishing a prediction model for wheat stripe rust and fusarium head blight based on an attention neural network model;

[0090] like Figure 3 As shown, the wheat stripe rust and fusarium head blight prediction model includes a feature extraction module, a spatial structure attention module, a feature fusion module, a coding and decoding module, and a classification and recognition module, wherein:

[0091] The feature extraction module is used to extract multi-scale features from the input training set data;

[0092] The spatial structure attention module is used to encode broader contextual information in the extracted multi-scale features into local features;

[0093] The feature fusion module is used to perform feature fusion on the multi-scale features output by the spatial structure attention module;

[0094] The encoding and decoding module is used to use multiple convolution kernels of different sizes to perform convolution on the fused feature matrix to extract feature information of different fields of view to obtain multiple feature maps; then strengthen important features and suppress useless features on the multiple feature maps to obtain multiple feature maps based on weight selection; splice the multiple feature maps based on weight selection in the channel dimension to obtain a spliced ​​feature map;

[0095] The classification and recognition module is used to flatten the spliced ​​feature map into a one-dimensional vector and project it to the vector space of the tail entity to obtain a hidden layer vector, where the dimension of the hidden layer vector is the same as the dimension of the tail entity vector t; perform matrix multiplication on the hidden layer vector and all the tail entity vectors in the knowledge graph to obtain the triple score of the head entity vector relationship vector and the tail entity vector; and classify and recognize stripe rust and fusarium according to the triple score.

[0096] Furthermore, the feature extraction module includes a hyperspectral image feature extraction unit, a high-resolution image feature extraction unit, a geographic environment data feature extraction unit and a meteorological data feature extraction unit, wherein:

[0097] The hyperspectral image feature extraction unit is used to extract wheat spectral features from the hyperspectral image data in the stripe rust and head blight feature data set, and upsample the obtained hyperspectral impact wheat feature map;

[0098] The high-resolution image feature extraction unit is used to extract wheat texture morphological spatial structure features from the high-resolution image data in the stripe rust and scab feature data sets;

[0099] The geographic environment data feature extraction unit is used to perform semantic level search and clustering on the geographic environment data in the stripe rust and head blight feature data set to extract the geographic environment features of the wheat planting area;

[0100] The meteorological data feature extraction unit is used to perform semantic level search and clustering on the meteorological data in the stripe rust and head blight feature data sets to achieve meteorological feature extraction in the wheat planting area.

[0101] As a preferred embodiment, the spatial structure attention module includes a first part, a second part and a third part, wherein:

[0102] The first part includes a first convolution unit, a second convolution unit, a first activation function unit, a third convolution unit, a second activation function unit, and a first multiplication processing unit. The feature maps of the scales other than the bottom-level feature map in the multi-scale feature map output by the feature extraction module are respectively input into the corresponding first convolution unit. The feature maps of different scales obtained after the multi-scale feature map output by the feature extraction module and the up-convolution are all input into the corresponding second convolution unit. The feature map output by the first convolution unit is added to the feature map of the same scale output by the second convolution unit and then input into the first activation function unit. The first activation function unit strengthens the features in the feature map and then inputs the feature map into the second activation function unit after convolution processing by the third convolution unit. The second activation function unit suppresses the noise in the feature map and then inputs the feature map into the first multiplication processing unit. The first multiplication processing unit multiplies it with the feature map of the same scale input into the first convolution unit and then outputs it.

[0103] The second part includes a second multiplication processing unit and a third activation function unit. The feature maps of the remaining scales except the bottom-level feature map in the multi-scale feature map are respectively input into the corresponding second multiplication processing unit, and the second multiplication processing unit multiplies the feature maps with the first half of the channels in the feature map of the same scale after upsampling. The feature map output by the second multiplication processing unit is added to the second half of the channels in the feature map of the same scale after upsampling, and then input into the third activation function unit for processing and output;

[0104] The third part is used to perform addition processing on the feature map output by the first part and the feature map output by the second part.

[0105] Furthermore, the first activation function unit and the third activation function unit both use ReLU function, and the second activation function unit uses Sigmoid function.

[0106] In this example, the encoding module includes a channel adaptive selection unit and a receptive field adaptive selection unit, and the channel adaptive selection unit is configured with a wavelet soft shrinkage subnetwork, which is used to denoise multiple feature maps using the following formula to obtain multiple denoised feature maps;

[0107]

[0108] Feature Map X 21 is the feature map X 11 Feature map after denoising; |X11| avgis the channel-adaptive global average pooling feature, σ represents the Sigmoi function, and δ represents the ReLU function. W1 and W2 are the network parameters of the two fully connected layers in the channel-adaptive fully connected neural network. According to the channel-adaptive global average pooling feature |X11| avg Calculated.

[0109] The receptive field adaptive selection unit is used to process the multiple denoised feature maps respectively to obtain multiple feature maps based on weight selection, specifically:

[0110] The denoised feature map X is added element by element 21 The fused feature U is obtained;

[0111] The fused features are shrunk to the spatial dimension H×W to obtain the receptive field adaptive global average pooling feature S:

[0112]

[0113]

[0114] Among them U k is the k-th element of weight U;

[0115] The receptive field adaptive global average pooling feature S is used to create a compact feature through the receptive field adaptive fully connected neural network:

[0116]

[0117]

[0118] Where σ2 is the softmax function, W3 and W4 are the weight matrices of the two fully connected layers in the receptive field adaptive fully connected network;

[0119] Divide Z into two parts along the channel dimension, and obtain the receptive field adaptive attention weights b1, b2 and b3 corresponding to the multi-scale convolution;

[0120]

[0121] Finally, the feature map based on weight selection is obtained according to the formula:

[0122] X 31 =F scale (X 21 , b1).

[0123] Step 5: training the wheat stripe rust and fusarium head blight prediction model using a training set, and verifying the prediction accuracy of the trained model using a test set; if the accuracy meets the standard, the training is completed; otherwise, the training is continued after updating the parameters;

[0124] The training process of the wheat stripe rust and fusarium head blight prediction model is as follows:

[0125] Step 51, initializing the connection weights and thresholds of the model;

[0126] At this stage, the system automatically and randomly initializes all weights and thresholds. In order to prevent the network from entering a saturated state too early, they are generally set to smaller random numbers.

[0127] Step 52: Given input x i and the target output y i ;

[0128] Step 53: Determine the number of neurons in the hidden layer:

[0129] Step 54: The preceding propagation process of the signal:

[0130] Let the input of the i-th node be x i , the hidden layer input is I i , the output is limited to a certain range through the transfer function f, and the output O is obtained i :

[0131] The output of each layer is used as the input of the next layer, and the final output is obtained by iterating through the designed network.

[0132] Step 55, calculate the mean square error: k With the expected output T k Calculate mean square error;

[0133] Step 56, error back propagation process: calculate the gradient of each parameter in the model and optimize the parameters by back propagation;

[0134] The public announcement of parameter optimization is as follows:

[0135]

[0136]

[0137] Where η is the learning rate; Δw ij is the weight of the connection between nodes i and j after correction; Δθ i is the threshold of neuron i after correction;

[0138] Step 57: Determine whether the model prediction accuracy meets the requirement. If yes, stop training. Otherwise, repeat steps 54 to 56 until the entire training error meets the prediction accuracy requirement.

[0139] Prediction accuracy ACC refers to the ratio of the number of samples correctly predicted by the model to the total number of samples for a given validation data set:

[0140]

[0141] Where TP is the positive sample predicted by the model as positive, TN is the negative sample predicted by the model as negative, FP is the negative sample predicted by the model as positive, and FN is the positive sample predicted by the model as negative.

[0142] Step 6: Use the trained wheat stripe rust and fusarium head blight prediction model to predict wheat in the target area to obtain wheat stripe rust and fusarium head blight prediction results.

[0143] The prediction model of wheat stripe rust and fusarium head blight described in this embodiment is compared with other 5 existing disease prediction models, which are: based on adaptive discriminant deep belief network (Deep Belief Networks, DBN), based on time series neural network prediction model (Neural Networks, NN) and plant disease prediction model based on time series and RBF network (Time Series and RBF Networks, TSRBF) and crop disease prediction method based on long short-term memory network prediction model (Long Short-Term Memory, LSTM) and bidirectional LSTM (Bi-directional Long Short-Term Memory, Bi-LSTM). The 5 comparison methods DBN, NN, TSRBF, LSTM and Bi-LSTM methods are all vector data converted from knowledge graphs like the prediction model described in this embodiment. Under the same experimental conditions, when the number of iterations is 400, the accuracy of the prediction model described in this embodiment on the training set is 90.27%, and the accuracy on the validation set is 94.38%, which are much higher than the other 5 existing models.

[0144] Embodiment 2:

[0145] like Figure 5 As shown, this embodiment proposes a prediction system for implementing the method described in Embodiment 1, including:

[0146] The data acquisition module is used to obtain the existing stripe rust and head blight characteristic data sets of different varieties of wheat in different regions, and to construct training sets and test sets after data preprocessing;

[0147] The knowledge graph construction module is used to collect knowledge related to the occurrence of wheat stripe rust and fusarium head blight from the preprocessed feature data set and construct a knowledge graph of wheat stripe rust and fusarium head blight;

[0148] A disease-causing factor determination module is used to determine the main factors causing wheat stripe rust and fusarium head blight according to the wheat stripe rust and fusarium head blight knowledge graph, wherein the main factors causing wheat stripe rust and fusarium head blight include meteorological factors, geographical environment factors, and variety factors;

[0149] A prediction model building module, used to quantify the main factors causing the occurrence of wheat stripe rust and fusarium head blight, and to establish a prediction model for wheat stripe rust and fusarium head blight based on a deep learning neural network model;

[0150] A prediction model training module is used to train the wheat stripe rust and head blight prediction model using a training set, and to verify the prediction accuracy of the trained model using a test set. If the accuracy meets the standard, the training is completed; otherwise, the training continues after updating the parameters;

[0151] The prediction module is used to predict wheat in the target area using the trained wheat stripe rust and fusarium head blight prediction models to obtain wheat stripe rust and fusarium head blight prediction results.

[0152] Finally, this embodiment carried out the following experiments:

[0153] Wheat stripe rust and fusarium scab were taken as the detection objects. The relevant data of wheat stripe rust and fusarium scab came from the Henan Provincial Plant Protection Station, and the environmental information of wheat stripe rust and fusarium scab was collected from various cities in Henan Province from 2015 to 2020; the meteorological data came from the Shaanxi Provincial Meteorological Information Center, and the temperature, precipitation, relative humidity and other data from October of the previous year to May of the following year were collected from meteorological stations in 8 cities in Henan Province from 2015 to 2020. Each single data item is the average value of the data collected for 5 days as the meteorological data sequence in the wheat-growing area. In order to enable accurate prediction, factors that have the greatest impact on the occurrence of wheat stripe rust and corn blight were selected for testing, including the minimum humidity, maximum humidity, average humidity, average temperature, rainfall, number of sunshine days, average wind speed, average wind direction, average wind force, minimum evaporation, maximum evaporation, and average evaporation of the air; the minimum humidity, maximum humidity, average humidity, minimum temperature, maximum temperature, average temperature of the soil, types and contents of elements required for growth, types and abundance of microorganisms; wheat moisture content, chlorophyll content, canopy structure and coverage, wheat vegetation index, etc.

[0154] In addition to the above actual data, the corpus collection tool written in Python programming language was used to capture 200 entries about wheat stripe rust and fusarium from multiple corpora such as China Agricultural Information Network, Xingnong Network, Vegetable Network and Baidu Encyclopedia as corpus, and these corpora were stored in text files in the form of GBK encoding. Since the directly obtained entries are unstructured texts, it is difficult to directly use their data. Therefore, the corpus was preprocessed using the three modules of word segmentation, part-of-speech tagging and dependency syntax in the open source tool LTP for natural language processing, and the results were stored in the format of the annotation file XML.

[0155] 36,600 entity relationship graphs were extracted from 30,000 entities to form 36,600 triples for disease prediction. A 10-fold cross-validation method was used for wheat stripe rust prediction test, and the validation set size was 3,660 triples. The method of this embodiment was compared with five other existing disease prediction methods:

[0156] The five existing disease prediction methods are: Deep Belief Networks (DBN) based on adaptive discriminant, Neural Networks (NN) based on time series, Plant disease prediction method based on time series and RBF networks (TSRBF), and Crop disease prediction method based on long short-term memory networks (LSTM) and bidirectional LSTM (Bi-directional LongShort-TermMemory, Bi-LSTM).

[0157] All six methods use meteorological data to predict stripe rust and head blight. Experimental equipment configuration and environment: 32G memory, Intel Core i5-4200U CPU @ 2.30GHz; GPU GEFORCE GTX 1080ti; Ubuntu 14.0.

[0158] Table 1 shows the average prediction accuracy of 6 methods after 30 tests using the 10-fold cross validation method.

[0159] Table 1 Average prediction accuracy and running time of six methods

[0160] method BN NN TSRBF LSTM Bi-LSTM The present invention Prediction accuracy 86.25 81.62 83.66 88.52 89.93 92.43

[0161] As can be seen from Table 1, the prediction accuracy of the method described in this embodiment is the highest, and is much higher than other methods. The reason is that the knowledge graph of the present invention makes full use of the interaction and influence between meteorological data and wheat stripe rust and scab, as well as between different meteorological data. Other prediction methods are completely based on input meteorological data, which may introduce irrelevant environmental factors and noise, resulting in low prediction accuracy. At the same time, the prediction model of the present invention can obtain disease-specific factors for identifying wheat stripe rust and scab, and prevent the loss of characteristic information that is more sensitive to wheat stripe rust and scab, thus achieving accurate and stable prediction of wheat stripe rust and scab.

[0162] The method of the present invention uses a stripe rust and head blight feature data set including hyperspectral image data, high-resolution image data, geographic environment data and meteorological data to predict stripe rust and head blight. Under the same experimental conditions, the average prediction accuracy obtained by performing 30 tests of the 10-fold cross validation method is shown in Table 2:

[0163] Table 2 Average prediction accuracy of different data used in the present invention

[0164] data Average Precision Weather data 92.43 Feature Dataset 96.48

[0165] As can be seen from Table 2, the method described in this embodiment uses a feature data set that combines hyperspectral image data, high-resolution image data, geographic environment data, and meteorological data for pre-testing, and the prediction accuracy is significantly higher than the accuracy of prediction using traditional meteorological data. It can be seen that the feature data set can effectively prevent the loss of feature information that is more sensitive to wheat stripe rust and fusarium scab, thereby improving the prediction accuracy of wheat stripe rust and fusarium scab.

[0166] In summary, the embodiment of the present invention couples hyperspectral image data, high-resolution image data, geographic environment data and meteorological data to establish a knowledge graph of wheat stripe rust and fusarium to determine the main factors of wheat stripe rust and fusarium, and then quantifies the main factors of wheat stripe rust and fusarium to establish a prediction model of wheat stripe rust and fusarium based on the attention neural network model to predict wheat in the target area. The above method steps can be effectively and intuitively applied to the detection of changes in the physiological state of wheat. By analyzing a variety of relevant data of wheat stripe rust and fusarium, disease-specific factors for identifying wheat stripe rust and fusarium are obtained, and the loss of characteristic information that is more sensitive to wheat stripe rust and fusarium is prevented, and accurate and stable prediction and forecasting of wheat stripe rust and fusarium are achieved, which can provide a scientific basis for the early prevention and control of wheat stripe rust and fusarium.

[0167] The technical solution provided by the present invention is described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting wheat stripe rust and fusarium head blight based on deep learning, characterized in that: The steps include: Step 1: Obtain the existing stripe rust and head blight feature data sets of different varieties of wheat in different regions, and construct training sets and test sets after data preprocessing; Step 2: Collect knowledge related to the occurrence of wheat stripe rust and fusarium head blight from the preprocessed feature data set, and construct a knowledge graph of wheat stripe rust and fusarium head blight; Step 3: Based on the wheat stripe rust and fusarium head blight knowledge graph, determine the main factors causing the wheat stripe rust and fusarium head blight, wherein the main factors causing the wheat stripe rust and fusarium head blight include meteorological factors, geographical environment factors, and variety factors; Step 4, quantifying the main factors causing the occurrence of wheat stripe rust and fusarium head blight, and establishing a prediction model for wheat stripe rust and fusarium head blight based on an attention neural network model; Step 5: training the wheat stripe rust and fusarium head blight prediction model using a training set, and verifying the prediction accuracy of the trained model using a test set; if the accuracy meets the standard, the training is completed; otherwise, the training is continued after updating the parameters; Step 6: Use the trained wheat stripe rust and fusarium head blight prediction model to predict wheat in the target area to obtain wheat stripe rust and fusarium head blight prediction results.

2. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 1, characterized in that: The stripe rust and head blight characteristic data sets described in step 1 include hyperspectral image data, high-resolution image data, geographic environment data and meteorological data.

3. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 2, characterized in that: The data preprocessing includes data labeling, data enhancement, abnormal data removal, and data normalization operations.

4. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 1, characterized in that: The specific process of constructing the knowledge graph of wheat stripe rust and fusarium head blight in step 2 is as follows: Step 2.1, knowledge expression: express the disease detection knowledge and metadata knowledge in the preprocessed feature data set in structured, semi-structured, and unstructured forms; Step 2.2, knowledge extraction: first, entity acquisition is performed, and regular expressions and hidden Markov models are used to obtain structured disease detection knowledge and metadata knowledge; then attribute acquisition is performed, and support vector machines are used to collect specific attribute information from the feature data set; finally, relationship extraction is performed, and the LTP tool is used to perform semantic dependency syntactic analysis on the text in the feature data set to obtain the relationship between entities; Step 2.3, knowledge fusion: Use entity disambiguation and coreference resolution to connect knowledge to the correct entity object to complete knowledge fusion; Step 2.4, knowledge representation: using the framework theory representation method, the data after knowledge extraction and knowledge fusion are represented in a structured form; Step 2.5, knowledge processing: perform knowledge reasoning and quality assessment on the data after knowledge representation to obtain qualified knowledge; Step 2.6: Construct the ontology layer and entity layer: Use the knowledge-processed data to construct the ontology layer and entity layer; Step 2.7, knowledge graph construction: After extracting the entity information and the relationships between different entities in the entity layer, the extracted entity-relationship-entity triples are stored in the database and depicted as a graphical knowledge graph structure to complete the construction of the knowledge graph of wheat stripe rust and fusarium head blight.

5. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 1, characterized in that: The meteorological factors include the minimum humidity, maximum humidity, average humidity, average temperature, rainfall, sunshine days, average wind speed, average wind direction, average wind force, minimum evaporation, maximum evaporation, and average evaporation; The geographical environmental factors include the minimum humidity, maximum humidity, average humidity, minimum temperature, maximum temperature, average temperature, types and contents of elements required for growth, and types and abundance of microorganisms; The variety factors include wheat moisture content, chlorophyll content, canopy structure and coverage, and wheat vegetation index.

6. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 1, characterized in that: The wheat stripe rust and fusarium head blight prediction model includes a feature extraction module, a spatial structure attention module, a feature fusion module, a coding and decoding module, and a classification and recognition module, wherein: The feature extraction module is used to extract multi-scale features from the input training set data; The spatial structure attention module is used to encode broader contextual information in the extracted multi-scale features into local features; The feature fusion module is used to perform feature fusion on the multi-scale features output by the spatial structure attention module; The encoding and decoding module is used to use multiple convolution kernels of different sizes to perform convolution on the fused feature matrix to extract feature information of different fields of view to obtain multiple feature maps; then strengthen important features and suppress useless features on the multiple feature maps to obtain multiple feature maps based on weight selection; splice the multiple feature maps based on weight selection in the channel dimension to obtain a spliced ​​feature map; The classification and recognition module is used to flatten the spliced ​​feature map into a one-dimensional vector and project it to the vector space of the tail entity to obtain a hidden layer vector, where the dimension of the hidden layer vector is the same as the dimension of the tail entity vector t; perform matrix multiplication on the hidden layer vector and all the tail entity vectors in the knowledge graph to obtain the triple score of the head entity vector relationship vector and the tail entity vector; and classify and recognize stripe rust and fusarium according to the triple score.

7. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 6, characterized in that: The feature extraction module includes a hyperspectral image feature extraction unit, a high-resolution image feature extraction unit, a geographic environment data feature extraction unit and a meteorological data feature extraction unit, wherein: The hyperspectral image feature extraction unit is used to extract wheat spectral features from the hyperspectral image data in the stripe rust and head blight feature data set, and upsample the obtained hyperspectral impact wheat feature map; The high-resolution image feature extraction unit is used to extract wheat texture morphological spatial structure features from the high-resolution image data in the stripe rust and scab feature data sets; The geographic environment data feature extraction unit is used to perform semantic level search and clustering on the geographic environment data in the stripe rust and head blight feature data set to extract the geographic environment features of the wheat planting area; The meteorological data feature extraction unit is used to perform semantic level search and clustering on the meteorological data in the stripe rust and head blight feature data sets to achieve meteorological feature extraction in the wheat planting area.

8. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 6, characterized in that: The encoding module includes a channel adaptive selection unit and a receptive field adaptive selection unit, wherein the channel adaptive selection unit is configured with a wavelet soft shrinkage subnetwork for denoising a plurality of feature maps to obtain a plurality of denoised feature maps; The receptive field adaptive selection unit is used to process multiple denoised feature maps respectively to obtain multiple feature maps based on weight selection.

9. The method for predicting wheat stripe rust and head blight based on deep learning according to claim 1, characterized in that: The training process of the wheat stripe rust and fusarium head blight prediction model is as follows: Step 51, initializing the connection weights and thresholds of the model; Step 52: Given input x i and the target output y i ; Step 53: Determine the number of neurons in the hidden layer: Step 54: The preceding propagation process of the signal: Let the input of the i-th node be x i , the hidden layer input is I i , the output is limited to a certain range through the transfer function f, and the output O is obtained i : The output of each layer is used as the input of the next layer, and the final output is obtained by iterating through the designed network. Step 55, calculate the mean square error: k With the expected output T k Calculate mean square error; Step 56, error back propagation process: calculate the gradient of each parameter in the model and optimize the parameters by back propagation; Step 57: Determine whether the model prediction accuracy meets the requirement. If yes, stop training. Otherwise, repeat steps 54 to 56 until the entire training error meets the prediction accuracy requirement.

10. A wheat stripe rust and head blight prediction system based on deep learning for implementing the method according to any one of claims 1 to 9, characterized in that: include: The data acquisition module is used to obtain the existing stripe rust and head blight characteristic data sets of different varieties of wheat in different regions, and to construct training sets and test sets after data preprocessing; The knowledge graph construction module is used to collect knowledge related to the occurrence of wheat stripe rust and fusarium head blight from the preprocessed feature data set and construct a knowledge graph of wheat stripe rust and fusarium head blight; A disease-causing factor determination module is used to determine the main factors causing wheat stripe rust and fusarium head blight according to the wheat stripe rust and fusarium head blight knowledge graph, wherein the main factors causing wheat stripe rust and fusarium head blight include meteorological factors, geographical environment factors, and variety factors; A prediction model building module, used to quantify the main factors causing the occurrence of wheat stripe rust and fusarium head blight, and to establish a prediction model for wheat stripe rust and fusarium head blight based on a deep learning neural network model; A prediction model training module is used to train the wheat stripe rust and head blight prediction model using a training set, and to verify the prediction accuracy of the trained model using a test set. If the accuracy meets the standard, the training is completed; otherwise, the training continues after updating the parameters; The prediction module is used to predict wheat in the target area using the trained wheat stripe rust and fusarium head blight prediction models to obtain wheat stripe rust and fusarium head blight prediction results.

Images