CN109035231A - A kind of detection method and its system of the wheat scab based on deep-cycle - Google Patents

Abstract

The invention discloses a detection method for wheat head blight based on deep circulation, comprising the following steps: S1, collecting wheat ear hyperspectral image pixels, and performing sample undersampling on the hyperspectral image pixels to solve the problem of data set Unbalanced problem, so as to obtain the target data; S2, reshape the two-dimensional image of the target data as a grayscale image, and preprocess based on the grayscale image; S3, use the deep cycle neural network to process the preprocessed data Data training; S4, analyze the classification effect of the deep cycle neural network model based on the training results; the present invention uses hyperspectral imaging technology to carry out early rapid and non-destructive detection of wheat scab, which improves the area classification of the tested wheat the accuracy of the results.

Description

A detection method and system for wheat head blight based on deep circulation

technical field

The invention relates to the field of diagnostic methods for wheat scab, in particular to a detection method and system for wheat scab based on deep circulation.

Background technique

Wheat scab is an important disease of wheat. It is mainly distributed in humid and semi-humid regions, especially in temperate regions with humid and rainy climate. It has always been one of the most serious diseases in the wheat regions south of the Huaihe River and the middle and lower reaches of the Yangtze River in my country. . In recent years, the wheat head blight in Anhui has advanced from the Huainan wheat area in the frequent occurrence area to the Huaibei wheat area. Improper prevention and control of wheat scab infection will result in reduced yields, and severe crop failure, resulting in serious yield loss and quality impact on production. When wheat is infected by fungi, it will produce a variety of mycotoxins, the most serious of which is deoxynivalenol (DON), which will cause serious harm to humans and animals, and will persist in the food chain for a long time.

Hyperspectral imaging technology is an optical image acquisition and analysis technology that integrates advanced technologies such as optics, electronics, optoelectronics, computer science, and information processing. It can simultaneously acquire image information and spectral information of the research object. Because hyperspectral images have multiple bands, high resolution, simple operation, no damage to samples, narrow spectral range, no pollution, continuous bands, large data volume, and large information redundancy, they are widely researched and applied, especially in the application of in the field of agriculture. Chen Na and others used hyperspectral imaging technology to conduct rapid diagnosis research on rape sclerotinia; Zhang Hang used hyperspectral imaging technology to conduct research on the identification method of mulch in lint; Zhang Lin used hyperspectral imaging technology to conduct disease diagnosis methods. Research; Yin Lihua used hyperspectral imaging technology to classify fresh dates; Yeh et al. used different analysis methods to conduct hyperspectral diagnosis of strawberry leaf anthracnose; Gabriel et al. Correlation between data and internal hardness, soluble solids content, etc.

Most of the current detection methods for crop diseases focus on image detection and spectral detection, and hyperspectral imaging diagnosis technology is currently one of the most advanced methods for detecting crop diseases and insect pests in the world. Although hyperspectral imaging diagnostic technology has attracted the attention of experts and scholars, due to the high spatial dimension and large amount of data information of hyperspectral image data, most of the applications can only be carried out based on mature software, and it is difficult to develop from redundant The effective feature information can be accurately extracted from the data to classify the disease. Therefore, the research on large-scale, high-latitude, and large-data-level hyperspectral data modeling methods is the focus and difficulty of hyperspectral research. In order to solve the above problems of hyperspectral data, this patent is mainly based on the deep learning method, and diagnoses wheat scab by establishing a deep convolution-circular hybrid neural network model.

At this stage, there are three problems in the diagnosis of wheat scab at home and abroad: 1. Plant protection experts do not have enough time to diagnose and treat all the problems that arise in agricultural activities; 2. There are certain defects in the traditional methods of diagnosis of crop diseases. 3. The current mainstream disease detection methods usually have problems such as locality, hysteresis, destructiveness and indirectness. Therefore, it is of great significance to develop a fast and accurate detection method for wheat head blight for precise control and diagnosis of the disease.

Contents of the invention

The invention provides a detection method of wheat scab based on deep circulation, which solves the problems of locality, hysteresis, destructiveness and indirectness in current mainstream disease detection methods. The model diagnoses wheat scab High accuracy. In the process of learning, the model is based on the time axis, and extracts effective time series features from it to classify images. Since the hyperspectral images in this experiment were taken continuously for 15 days, their time series features can effectively diagnose wheat scab, so the deep recurrent neural network has a high accuracy in diagnosing wheat scab.

The technical solution adopted by the present invention to solve its technical problems is: a kind of detection method of wheat scab based on deep circulation, comprises the following steps:

S1. Collect hyperspectral image pixels of wheat ears, and sample undersampling of hyperspectral image pixels to solve the problem of unbalanced data sets, so as to obtain three characteristics of background, health and disease for training depth volume The target data of the product neural network model;

S2. Reshape the two-dimensional image of the target data to obtain a grayscale image, and use the mean value removal and principal component analysis methods for preprocessing based on the gray level image. The mean value removal method is used to reduce the sampling error of different dates; the principal component Analysis methods are used to identify the main features of the data at different observation times and to improve the visualization of spectral data;

S3. Based on the preprocessed data, the deep cycle neural network model is used for training; the training of the deep cycle neural network model is based on the preprocessed data and supervised learning from top to bottom according to the structure of the model, and the error in the training process is from top to bottom. Downward transmission, and automatically fine-tune the network structure parameters, so as to achieve the effect of training the model;

S4. Analyze the classification effect of the deep recurrent neural network model based on the model accuracy rate and loss rate obtained through training.

Preferably, the sample undersampling of the hyperspectral image pixel in the step S1 specifically includes: random sampling of the original data; the data set is divided into a 70% training set and a 30% verification set.

Preferably, the step S2 specifically includes: converting the spectral two-dimensional data of the hyperspectral image pixel into a gray image; and applying standardization and mean value removal to the gray image to perform preprocessing of accelerated gradient descent.

Preferably, the deep recurrent neural network is a classic framework of time series data, which is a model framework obtained by improving the original deep recurrent neural network: LSTM framework and GRU framework; both the LSTM framework and the GRU framework have three accumulation layers;

The LSTM framework consists of three gates: a forget gate, an input gate, and an output gate;

The GRU framework has one more update gate and one input gate than the LSTM framework.

Preferably, said step S4 specifically includes: using the precision P, recall rate R and F1 score obtained after data training to evaluate the model to determine whether the wheat is infected with scab;

The accuracy P is obtained by the following formula:

The recall rate R is obtained by the following formula:

The F1 score is derived from the following formula:

Among them, TP in the formula represents the positive sample predicted by the model as positive, which is called the correct rate of judgment; TN represents the negative sample predicted by the model as negative, which is called the correct rate of false judgment; FP represents the positive rate predicted by the model The negative sample is called the false positive rate; FN represents the positive sample predicted to be negative by the model, which is called the false negative rate.

Preferably, the disease classification system of the detection method of a kind of wheat head blight based on deep circulation comprises the following modules:

The noise removal module is used to collect hyperspectral images of wheat ears, and perform denoising processing on the hyperspectral images to obtain target data;

The first processing module is used to reshape the two-dimensional image of the target data as a grayscale image, and perform preprocessing based on the grayscale image;

The second processing module is used to train the data by using the deep recurrent neural network based on the preprocessed data;

The disease classification module is used to analyze the classification effect of the deep recurrent neural network model based on the training results.

The invention provides a detection method for wheat scab based on deep circulation, and the model has high accuracy in diagnosing wheat scab. In the process of learning, the model is based on the time axis, and extracts effective time series features from it to classify images. Since the hyperspectral images in this experiment were taken continuously for 15 days, their time series features can effectively diagnose wheat scab, so the deep recurrent neural network has a high accuracy in diagnosing wheat scab.

Description of drawings

Fig. 1 is a schematic flow sheet of the method of the present invention;

Figure 2 is the hyperspectral image of wheat head blight and its region of interest;

Figure 3 is a deep recurrent neural network model.

Detailed ways

A specific embodiment of the present invention will be described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.

As shown in Fig. 1, the embodiment of the present invention provides a kind of detection method of wheat head blight based on deep cycle, comprises the following steps: S1, collects wheat ear hyperspectral image element, and carries out sample to hyperspectral image element Undersampling to solve the problem of unbalanced data sets, so as to obtain target data with three characteristics of background, health and disease for training deep convolutional neural network models;

S2. Reshape the two-dimensional image of the target data to obtain a grayscale image, and use the mean value removal and principal component analysis methods for preprocessing based on the gray level image. The mean value removal method is used to reduce the sampling error of different dates; the principal component Analysis methods are used to identify the main features of the data at different observation times and to improve the visualization of spectral data;

S3. Based on the preprocessed data, the deep cycle neural network model is used for training; the training of the deep cycle neural network model is based on the preprocessed data and supervised learning from top to bottom according to the structure of the model, and the error in the training process is from top to bottom. Downward transmission, and automatically fine-tune the network structure parameters, so as to achieve the effect of training the model;

S4. Analyze the classification effect of the deep recurrent neural network model based on the model accuracy rate and loss rate obtained through training.

Under-sampling the hyperspectral image pixel in the step S1 specifically includes: randomly sampling the original data;

The step S2 specifically includes: transforming the spectral two-dimensional data of the hyperspectral image pixel into a gray image; and applying standardization and mean value removal methods to the gray image to perform accelerated gradient descent preprocessing; the preprocessing method uses Mean-removal method; as shown in Figure 2, the left two columns in the figure are the original hyperspectral images, and the right two columns are the hyperspectral images that draw the region of interest;

Step S3 specifically includes: based on the preprocessed data in step S2, using a deep recurrent neural network to train the preprocessed data;

Deep recurrent neural network (DRNN) is a classic framework of time series data, and the output neuron can directly affect itself next time. The present invention uses the model framework improved by the original deep recurrent neural network——LSTM framework and GRU framework; LSTM The framework consists of three gates, namely a forget gate, an input gate, and an output gate; the GRU framework is the optimal form of the LSTM framework, which has an update gate and an input gate more than the LSTM framework;

The structural configuration of the deep recurrent neural network is shown in Table 1 below:

Table 1 Structural configuration of deep recurrent neural network

Described data set is divided into the training set of 70% and the verification set of 30%; As shown in Figure 3: the training experiment result of deep cycle neural network model (two models, one is LSTM, and the other is GRU), abscissa is the number of iterations (that is, the number of training cycles), the ordinate is the accuracy of the model's classification of the data, and the training set is the training set, and the validation set is the verification set.

Step S4 specifically includes: using the precision P, recall rate R and F1 score after data training to evaluate the model to determine whether the wheat is infected with scab;

The accuracy P is obtained by the following formula:

The recall rate R is obtained by the following formula:

The F1 score is derived from the following formula:

Among them, TP in the formula represents the positive sample predicted by the model as positive, which is called the correct rate of judgment; TN represents the negative sample predicted by the model as negative, which is called the correct rate of false judgment; FP represents the positive rate predicted by the model The negative sample is called the false positive rate; FN represents the positive sample predicted to be negative by the model, which is called the false positive rate.

The invention is a detection method of wheat scab based on deep circulation, which solves the problems of locality, hysteresis, destructiveness and indirectness in current mainstream disease detection methods, and the accuracy of the model in diagnosing wheat scab The rate is high. In the process of learning, the model is based on the time axis, and extracts effective time series features from it to classify images. Since the hyperspectral images in this experiment were taken continuously for 15 days, their time series features can effectively diagnose wheat scab, so the deep recurrent neural network has a high accuracy in diagnosing wheat scab. The method of the present invention is systematic, simple and fast, not only solves the problem of accurate classification of wheat scab by using hyperspectral imaging technology; but also can use hyperspectral imaging technology to carry out early rapid and non-destructive detection of wheat scab, improving The accuracy of the classification results of the tested wheat area is improved; in addition, the number of combined modules of the diagnostic system of the present invention is small, the program design is reasonable, and it can be developed and used in the later stage.

The above disclosures are only specific embodiments of the present invention, but the embodiments of the present invention are not limited thereto, and any changes conceivable by those skilled in the art shall fall within the protection scope of the present invention.

Claims (6)

1. a kind of detection method of the wheat scab based on deep-cycle, which comprises the following steps:

S1, acquisition wheat wheat head high spectrum image pixel, and sample lack sampling is carried out to high spectrum image pixel, to solve data Collect unbalanced problem, to obtain having three kinds of background, health, disease features to be used to train depth convolutional neural networks mould The target data of type；

S2, the remodeling for carrying out two dimensional image to target data obtain gray level image, and based on gray level image with mean value removal and Two methods of principal component analysis are pre-processed, and mean value removal method is used to reduce the sampling error of not same date；Principal component analysis Method is used to identify the main feature of data, and the visualization for improving spectroscopic data in different observing times；

S3, it is trained based on pretreated data application deep-cycle neural network model；Deep-cycle neural network model Training be the supervised learning carried out from top to bottom based on the obtained data of pretreatment according to the structure of model, in training process Error is transmitted from up to down, and is finely adjusted automatically to network architecture parameters, to achieve the effect that training pattern；

S4, based on the obtained model accuracy rate of training and loss late come the classifying quality of analysis depth Recognition with Recurrent Neural Network model.

2. a kind of detection method of wheat scab based on deep-cycle according to claim 1, which is characterized in that institute It states in step S1 and high spectrum image pixel progress sample lack sampling is specifically included: random sampling is carried out to initial data；It is described Data set be divided into 70% training set and 30% verifying collection.

3. a kind of detection method of wheat scab based on deep-cycle according to claim 1, which is characterized in that institute It states step S2 to specifically include: converting gray image for the spectrum two-dimensional data of high spectrum image pixel；And gray level image is transported The pretreatment of accelerating gradient decline is carried out with the method for standardization, mean value removal.

4. a kind of detection method of wheat scab based on deep-cycle according to claim 1, which is characterized in that institute The canonical frame that deep-cycle neural network is time series data is stated, improves to obtain by original deep-cycle neural network Model framework: LSTM frame and GRU frame；There are three stack layers for the LSTM frame and GRU frame；

The LSTM frame is made of three doors: one is forgotten door, an input gate, an out gate；

A update door and an input gate more than the GRU frame ratio LSTM frame.

5. a kind of detection method of wheat scab based on deep-cycle according to claim 1, which is characterized in that institute State step S4 to specifically include: the accuracy P that is obtained after training using data, recall rate R and F1 score are come assessment models with true Determine whether wheat infects head blight；

The accuracy P is obtained by following formula:

The recall rate R is obtained by following formula:

The F1 score is obtained by following formula:

Wherein, the TP representative model in formula predicts the positive sample being positive, and is referred to as judged as genuine accuracy；TN representative model is pre- The negative sample being negative is surveyed, is referred to as judged as false accuracy；FP representative model predicts the negative sample being positive, referred to as rate of false alarm；FN Representative model predicts the positive sample being negative, referred to as rate of failing to report.

The system 6. a kind of disease of the detection method of wheat scab based on deep-cycle according to claim 1 is classified System, which is characterized in that comprise the following modules:

Noise remove module for acquiring wheat wheat head high spectrum image, and is removed noise processed to high spectrum image, obtains To target data；

First processing module for the remodeling to target data progress two dimensional image as gray level image, and is based on grayscale image As being pre-processed；

Second processing module, for being trained based on pretreated data application deep-cycle neural network to data；

Disease categorization module, for based on trained result come the classifying quality of analysis depth Recognition with Recurrent Neural Network model.