CN116416524A - Early-stage asymptomatic detection method for bacterial leaf blight of rice - Google Patents

Abstract

The invention discloses a method for detecting early-stage asymptomatic bacterial leaf blight of rice, which comprises the following steps: selecting spectral wavelengths with high importance scores as characteristic sensitive wavelengths in a characteristic wavelength interval by utilizing a random forest algorithm; taking the hyperspectral image at the characteristic sensitive wavelength as a sensitive image characteristic for distinguishing the blade class to which the hyperspectral image belongs; training a 3DCNN model based on sensitive image features to obtain a 3DCNN asymptomatic detection model for early asymptomatic detection of rice bacterial leaf blight; and introducing a multi-scale spectrum cavity convolution module into the 3DCNN asymptomatic detection model to perform precision optimization to obtain the MS-SDC-3DCNN model. According to the method, the dimension of the hyperspectral image is reduced by utilizing a random forest algorithm, the asymptomatic detection model is optimized by utilizing a multi-scale spectrum cavity convolution module, and the asymptomatic detection model utilizes the characteristics of a plurality of wavelength resolutions after extraction and fusion, so that important wavelength information is used more effectively, and the detection performance of the asymptomatic detection model is improved.

Description

A method for early asymptomatic detection of rice bacterial blight

technical field

The invention relates to the technical field of rice bacterial blight detection, in particular to an early asymptomatic detection method for rice bacterial blight.

Background technique

With the development of agricultural production requirements and detection technology, early asymptomatic detection of crop diseases has increasingly become a research hotspot. At present, the detection and classification of crop diseases based on computer vision mainly rely on the model constructed by the color, texture and other features displayed after the outbreak of the disease. Researchers have proposed a variety of models for detecting external changes of susceptible rice leaves based on RGB images, and achieved good results.

Among them, the crop disease detection model based on deep learning technology has attracted more and more attention of researchers. Deep learning can be understood as "feature learning". It gradually converts the initial "low-level" feature representation into a "high-level" feature representation through multi-layer processing, and uses "simple models" to complete complex tasks such as classification and regression. Since 2015, with the substantial improvement of computer data processing capabilities, deep convolutional neural networks have achieved rapid development, and their application research in crop disease detection has become more and more extensive. Various classic convolutional neural networks (such as ResNet, GoogLeNet, VGG, etc.) and custom convolutional neural networks in deep learning tools have been widely used in crop disease detection.

Although deep neural networks have achieved great success in detecting rice diseases from RGB images, it is worth noting that these networks cannot detect early asymptomatic diseases of crops based on RGB images. Without prior labeling, it is difficult to identify leaves in a healthy state or asymptomatic disease state by visual inspection due to their similar visual textures. Moreover, as far as the current development status of hyperspectral imaging technology is concerned, the massive spectral information data in hyperspectral images has not been fully processed and mined, and information processing is far from meeting real-time needs, and the current hyperspectral disease detection based on deep learning technology The model ignores the problem of the target scale (in an image of a given size, some lesions are relatively small, and some lesions are relatively large), and the use of a fixed-size convolution kernel will result in lower recognition accuracy. Therefore, in the current technology, it is impossible to detect early asymptomatic diseases of crops based on RGB images. The massive hyperspectral image data has not been fully processed and mined to meet the real-time detection requirements. The use of fixed-size convolution kernels will lead to lower recognition accuracy. rate problem.

Contents of the invention

The purpose of the present invention is to provide a method for early asymptomatic detection of rice bacterial blight, to solve the problem that in the prior art, it is impossible to detect early asymptomatic diseases of crops based on RGB images, and the massive spectral information data in hyperspectral images is not fully processed. Mining is difficult to meet the real-time requirements of detection, and the use of fixed-size convolution kernels will lead to technical problems of low recognition accuracy.

In order to solve the above technical problems, the present invention specifically provides the following technical solutions:

A method for early asymptomatic detection of rice bacterial blight, comprising the following steps:

Obtain hyperspectral images of healthy rice leaves and rice leaves infected with bacterial blight at various spectral wavelengths within the preset band to form a hyperspectral image dataset, and use the hyperspectral image dataset to construct healthy rice leaves and rice leaves infected with bacterial blight The hyperspectral curve of the rice leaf determines the characteristic wavelength interval, wherein the hyperspectral image located in the characteristic wavelength interval is characterized as an image feature that can distinguish healthy rice leaves from rice leaves infected with bacterial blight;

Using a random forest algorithm to score the importance of each spectral wavelength in the characteristic wavelength interval based on the hyperspectral image located in the characteristic wavelength interval, and select a spectral wavelength with a high importance score in the characteristic wavelength interval as the characteristic sensitive wavelength;

The hyperspectral image at the characteristic sensitive wavelength is used as a sensitive image feature for distinguishing the leaf category to which the hyperspectral image belongs, so as to achieve data dimensionality reduction of the hyperspectral image dataset;

The 3DCNN asymptomatic detection model for early asymptomatic detection of rice bacterial blight was obtained by training the 3DCNN model based on sensitive image features;

The MS-SDC-3DCNN model was obtained by introducing a multi-scale spectral atrous convolution module into the 3DCNN asymptomatic detection model for precision optimization, so as to achieve high-precision detection of early asymptomatic rice bacterial blight.

As a preferred solution of the present invention, the hyperspectral curves of healthy rice leaves and rice leaves infected with bacterial blight are determined by using the hyperspectral image data set to determine the characteristic wavelength range, including:

Draw all hyperspectral images belonging to healthy rice leaves in the hyperspectral image data set in the spectral coordinate system to obtain the hyperspectral curve of healthy rice leaves, wherein the ordinate in the spectral two-dimensional coordinate system is the reflection intensity, and the abscissa is the spectrum wavelength;

Draw all the hyperspectral images of the asymptomatic rice leaves in the hyperspectral image data set belonging to the rice leaves infected with bacterial blight in the spectral coordinate system to obtain the hyperspectral curve of the asymptomatic rice leaves;

Draw all hyperspectral images of rice leaves with mild symptoms in the hyperspectral image data set belonging to rice leaves infected with bacterial blight in the spectral coordinate system to obtain the hyperspectral curve of rice leaves with mild symptoms;

Draw all hyperspectral images of rice leaves with moderate symptoms in the hyperspectral image data set belonging to rice leaves infected with bacterial blight in the spectral coordinate system to obtain a hyperspectral curve of rice leaves with moderate symptoms;

Draw all hyperspectral images of rice leaves with severe symptoms in the hyperspectral image data set belonging to rice leaves infected with bacterial blight in the spectral coordinate system to obtain hyperspectral curves of rice leaves with severe symptoms;

Point-by-point in hyperspectral curves of healthy rice leaves, hyperspectral curves of asymptomatic rice leaves, hyperspectral curves of mildly symptomatic rice leaves, hyperspectral curves of moderately symptomatic rice leaves, and hyperspectral curves of severely symptomatic rice leaves Calculate the similarity, and connect all the curve points whose similarity is lower than the preset threshold point by point to obtain the characteristic distinction curve;

The spectral wavelength interval where the characteristic distinguishing curve is located is taken as the characteristic wavelength interval.

As a preferred solution of the present invention, the random forest algorithm is used to score the importance of each spectral wavelength in the characteristic wavelength interval based on the hyperspectral image located in the characteristic wavelength interval, including:

Use the random forest algorithm to randomly replace each hyperspectral image in the characteristic wavelength interval in turn, and calculate the classification error before and after the replacement to distinguish the leaf category of the hyperspectral image as the importance score of the hyperspectral image;

The importance score of the hyperspectral image is used as the importance score of the spectral wavelength corresponding to the hyperspectral image, and the spectral wavelengths are sorted from high to low according to the importance score of the spectral wavelength.

As a preferred solution of the present invention, selecting a spectral wavelength with a high importance score in the characteristic wavelength interval as a characteristic sensitive wavelength includes:

Setting the selection threshold, summing the importance scores of the sorted spectral wavelengths in the characteristic wavelength interval from high to low until the selected threshold is met, and using the corresponding spectral wavelength in the accumulation process as the characteristic sensitive wavelength;

Among them, the selected threshold is 0.9.

As a preferred solution of the present invention, the acquisition of the 3DCNN asymptomatic detection model includes:

The hyperspectral images of healthy rice leaves located at characteristic sensitive wavelengths in the hyperspectral image dataset and the hyperspectral images of rice leaves infected with bacterial blight at characteristic sensitive wavelengths in the hyperspectral image dataset are used as input to the 3DCNN model. The leaf category to which the hyperspectral image belongs is used as the output of the 3DCNN model;

Using the 3DCNN model to train based on the input of the 3DCNN model and the output of the 3DCNN model to obtain the 3DCNN asymptomatic detection model.

As a preferred solution of the present invention, the introduction of the multi-scale spectral hole convolution module in the 3DCNN asymptomatic detection model to obtain the MS-SDC-3DCNN model includes:

By selecting three-dimensional spectral hole convolution modules with different spectral hole ratios and combining them into multi-scale spectral hole convolution modules, the spectral features of the hyperspectral images are extracted respectively, wherein the three-dimensional spectral hole convolution modules with different spectral hole ratios extract The obtained spectral feature information has different characteristic scales in the spectral dimension;

The multi-scale spectral atrous convolution module fuses the obtained spectral feature information of different feature scales to improve the acquisition of spectral feature information in hyperspectral images;

The spectral feature information output by the multi-scale spectral hole convolution module is used as the input of the 3DCNN asymptomatic detection model to improve the detection performance of the 3DCNN asymptomatic detection model;

The relationship between the receptive field of the three-dimensional spectral hole convolution module and the spectral hole rate is:

R _f =2×(r _SDR −1)×(k−1)+k;

Where R _f represents the receptive field of a single convolution kernel in the three-dimensional spectral hole convolution module; r _SDR represents the spectral hole rate, and k represents the size of the convolution kernel, where k is set to 3 by default;

The feature extraction operation in the multi-scale spectral atrous convolution module is implemented by 3D convolution, wherein the value at the position (x, y, z) on the jth feature map in the i-th layer of the 3D convolutional network is calculated The calculation formula is:

为第i层第j个特征图上位置(x,y,z)处的值，H_i、W_i和D_i是第i层卷积核的高度、宽度和深度，f为第i层的激活函数，/>

是连接到上一层第m个特征图的卷积核位置(h,w,d)处的权重参数，/>

是上一层第m个特征图上位置(x+h,y+w,z+d)处的值，b_ij是第i层第j个特征图的偏差；in,

is the value at position (x, y, z) on the jth feature map of the i-th layer, H _i , W _i and D _i are the height, width and depth of the convolution kernel of the i-th layer, and f is the value of the i-th layer activation function, />

is the weight parameter at the convolution kernel position (h, w, d) connected to the mth feature map of the previous layer, />

is the value at position (x+h, y+w, z+d) on the mth feature map of the previous layer, b _ij is the deviation of the jth feature map of the i-th layer;

The MS-SDC-3DCNN model is obtained by using the output of the multi-scale spectral atrous convolution module as the input of the 3DCNN asymptomatic detection model.

As a preferred solution of the present invention, said utilizing MS-SDC-3DCNN model to train based on sensitive image features, including:

Mixing hyperspectral images of healthy rice leaves at characteristic sensitive wavelengths in the hyperspectral image data set and hyperspectral images of rice leaves infected with bacterial blight at characteristic sensitive wavelengths in the hyperspectral image data set are used as training image data sets;

The training image data set is divided into a training set, a verification set and a test set in a ratio of 8:1:1, and the MS-SDC-3DCNN model is trained using the training set, verification set and test set to obtain the described detection model;

Among them, the MS-SDC-3DCNN model uses cross entropy as the loss function, and uses the stochastic gradient descent optimizer for training. The specific parameters of the MS-SDC-3DCNN model are as follows: the learning rate is set to 1×10 ^-3 , and the weight decay coefficient is set to 1×10 ^-6 , momentum is set to 0.95, ε is set to 1×10 ^-5 , epoch is set to 50, dropout is set to 0.45, and the hyperspectral image size input by MS-SDC-3DCNN is (9,9,10).

As a preferred solution of the present invention, the MS-SDC-3DCNN model is provided with a residual block to implement the residual block to avoid the problem of gradient disappearance.

As a preferred solution of the present invention, the hyperspectral image consists of 31680 pixels with a size of (132,240,10), wherein the hyperspectral image pixels are divided into three categories: healthy pixels, asymptomatic disease pixels and symptomatic disease pixels pixels.

As a preferred solution of the present invention, the characteristic wavelength range is 450-950 nm, and the characteristic wavelength range includes 232 spectral wavelengths.

Compared with the prior art, the present invention has the following beneficial effects:

The invention uses the random forest algorithm to reduce the dimension of the hyperspectral image, determines the wavelength sensitive to the disease, reduces the training time of the asymptomatic detection model of rice leaves, and constructs a multi-scale spectral hole convolution module, which can be achieved without increasing the amount of calculation. Next, expand the receptive field of the neural network to improve the feature extraction ability, and optimize the asymptomatic detection model by using the multi-scale spectral hole convolution module. Efficient use of important wavelength information to improve detection performance of asymptomatic detection models.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings that are required in the description of the embodiments or the prior art. Apparently, the drawings in the following description are only exemplary, and those skilled in the art can also obtain other implementation drawings according to the provided drawings without creative work.

Fig. 1 provides the structural representation of the structural representation of rice bacterial blight early stage asymptomatic detection method for the embodiment of the present invention;

Figure 2 is an RGB hyperspectral image of healthy rice leaves;

Figure 3 is an RGB hyperspectral image of asymptomatic rice leaves infected with bacterial blight;

Fig. 4 is hyperspectral graph;

Figure 4(A) is the importance score of 232 wavelengths;

Figure 4(B) is the ranking of the importance scores of the top 50 wavelengths among the 232 wavelengths;

Fig. 5 is a schematic diagram of a three-dimensional spectral hole convolution module with different spectral hole ratios;

Figure 5 (A) is a schematic diagram of a three-dimensional spectral void convolution module with a spectral void rate of 1;

Figure 5(B) is a schematic diagram of a three-dimensional spectral void convolution module with a spectral void rate of 2;

Figure 5(C) is a schematic diagram of a three-dimensional spectral void convolution module with a spectral void rate of 3;

Fig. 6 is a schematic diagram of a multi-scale spectral hole convolution module;

Figure 7 is a neural network framework diagram of the MS-SDC-3DCNN model.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Deep neural networks have achieved great success in detecting rice diseases from RGB images, but it is worth noting that these networks may not produce correct results for early asymptomatic BLB disease detection based on RGB images. RGB illustrations of hyperspectral images of healthy and asymptomatic leaves under rice bacterial blight stress are shown in Figures 2 and 3. Without prior labeling, it is difficult to identify by visual inspection whether leaves are healthy (Fig. 2) or asymptomatic (Fig. 3) due to their similar visual textures.

However, the disease will change the internal structure of rice leaves, and this internal change of rice leaves will lead to changes in the reflectance of the leaves for different bands, making it possible to use hyperspectral imaging technology to detect asymptomatic infections of rice diseases. Therefore, the present invention is based on the hyperspectral image of rice leaves combined with a deep learning model for early asymptomatic detection of rice bacterial blight (a method for early asymptomatic detection of rice bacterial blight), using hyperspectral images to achieve asymptomatic detection, there is Conducive to the prevention and control of bacterial blight.

As shown in Figure 1, the present invention provides a kind of asymptomatic detection method of rice bacterial blight early stage, comprises the following steps:

Obtain hyperspectral images of healthy rice leaves and rice leaves infected with bacterial blight at various spectral wavelengths within the preset band to form a hyperspectral image dataset, and use the hyperspectral image dataset to construct healthy rice leaves and rice leaves infected with bacterial blight The hyperspectral curve of the rice leaf determines the characteristic wavelength interval, wherein the hyperspectral image located in the characteristic wavelength interval is characterized as an image feature that can distinguish healthy rice leaves from rice leaves infected with bacterial blight;

Using a random forest algorithm to score the importance of each spectral wavelength in the characteristic wavelength interval based on the hyperspectral image located in the characteristic wavelength interval, and select a spectral wavelength with a high importance score in the characteristic wavelength interval as the characteristic sensitive wavelength;

The hyperspectral image at the characteristic sensitive wavelength is used as a sensitive image feature for distinguishing the leaf category to which the hyperspectral image belongs, so as to achieve data dimensionality reduction of the hyperspectral image dataset;

The 3DCNN asymptomatic detection model for early asymptomatic detection of rice bacterial blight was obtained by training the 3DCNN model based on sensitive image features;

The MS-SDC-3DCNN model was obtained by introducing a multi-scale spectral atrous convolution module into the 3DCNN asymptomatic detection model for precision optimization, so as to achieve high-precision detection of early asymptomatic rice bacterial blight.

The disease will change the internal structure of rice leaves, and this internal change of rice leaves will lead to changes in the reflectance of the leaves for different wavelength bands. Therefore, the hyperspectral data of healthy rice leaves and rice leaves infected with bacterial blight at the same spectral wavelength The reflectance corresponding to the image is different, and the difference in reflectance can distinguish whether the hyperspectral image belongs to healthy rice leaves or rice leaves infected with bacterial blight. The spectral band of the rice leaf category that the spectral image belongs to indicates that the hyperspectral image in this spectral band contains image features that distinguish the rice leaf category. Therefore, through the drawing of the hyperspectral curve, it is possible to obtain a The image features of rice leaves with blight disease are as follows:

The hyperspectral curves of healthy rice leaves and rice leaves infected with bacterial blight are determined by using the hyperspectral image data set to determine the characteristic wavelength range, including:

Draw all hyperspectral images belonging to healthy rice leaves in the hyperspectral image data set in the spectral coordinate system to obtain the hyperspectral curve of healthy rice leaves, wherein the ordinate in the spectral two-dimensional coordinate system is the reflection intensity, and the abscissa is the spectrum wavelength;

Draw all the hyperspectral images of the asymptomatic rice leaves in the hyperspectral image data set belonging to the rice leaves infected with bacterial blight in the spectral coordinate system to obtain the hyperspectral curve of the asymptomatic rice leaves;

Draw all hyperspectral images of rice leaves with mild symptoms in the hyperspectral image data set belonging to rice leaves infected with bacterial blight in the spectral coordinate system to obtain the hyperspectral curve of rice leaves with mild symptoms;

Draw all hyperspectral images of rice leaves with moderate symptoms in the hyperspectral image data set belonging to rice leaves infected with bacterial blight in the spectral coordinate system to obtain a hyperspectral curve of rice leaves with moderate symptoms;

Draw all hyperspectral images of rice leaves with severe symptoms in the hyperspectral image data set belonging to rice leaves infected with bacterial blight in the spectral coordinate system to obtain hyperspectral curves of rice leaves with severe symptoms;

Point-by-point in hyperspectral curves of healthy rice leaves, hyperspectral curves of asymptomatic rice leaves, hyperspectral curves of mildly symptomatic rice leaves, hyperspectral curves of moderately symptomatic rice leaves, and hyperspectral curves of severely symptomatic rice leaves Calculate the similarity, and connect all the curve points whose similarity is lower than the preset threshold point by point to obtain the characteristic distinction curve;

The spectral wavelength interval where the characteristic distinguishing curve is located is taken as the characteristic wavelength interval.

As shown in Fig. 4, Fig. 4 (A), Fig. 4 (B) shows the hyperspectral curve graph between 378 to 1033nm wavelengths of rice leaves in different disease states under bacterial blight stress, so distinguish the described category of rice leaves Spectral characteristic information is covered between 378-1033nm spectral wavelength. The spectral resolution of the hyperspectral imaging system used in the experiment is 2.14nm, and the hyperspectral wavelength range obtained by shooting is 378.28nm-1033.05nm. The spectral wavelength interval is further intercepted to 451.276nm-949.664nm, so the present invention sets the characteristic wavelength interval to 450-950nm, which not only includes the spectral characteristic information for distinguishing the categories of rice leaves, but also reduces the spectrum at both ends due to instruments and environments. the influence of noise.

All hyperspectral images were obtained in the characteristic wavelength range of 378-1033nm, and hyperspectral images at spectral wavelengths from 450 to 950nm were intercepted to remove noise at both ends of the characteristic wavelength range and to better distinguish the spectral characteristic information of the rice leaf category. The acquired hyperspectral image data is characterized by high dimensional redundancy between adjacent wavelengths. Excessive redundant spectral information brings great challenges to detection methods and computational complexity. Therefore, it is necessary to compress the amount of data through a dimensionality reduction method and reduce the cost of subsequent processing without discarding effective characteristic spectral information. The dimensionality of the spectral data, thereby reducing the computational complexity, the specific steps are as follows:

The random forest algorithm is used to score the importance of each spectral wavelength in the characteristic wavelength interval based on the hyperspectral image located in the characteristic wavelength interval, including:

Use the random forest algorithm to randomly replace each hyperspectral image in the characteristic wavelength interval in turn, and calculate the classification error before and after the replacement to distinguish the leaf category of the hyperspectral image as the importance score of the hyperspectral image;

The importance score of the hyperspectral image is used as the importance score of the spectral wavelength corresponding to the hyperspectral image, and the spectral wavelengths are sorted from high to low according to the importance score of the spectral wavelength.

Setting the selection threshold, summing the importance scores of the sorted spectral wavelengths in the characteristic wavelength interval from high to low until the selected threshold is met, and using the corresponding spectral wavelength in the accumulation process as the characteristic sensitive wavelength;

Among them, the selected threshold is 0.9.

The importance scores of the original 232 wavelengths from 450 to 950 nm were ranked using RF (Fig. 4(A)). It can be seen that the wavelengths with the highest importance scores are mainly distributed in 500–700nm. Figure 4(B) clearly shows that the top 10 wavelengths with the highest importance scores are 547.2nm, 534.5nm, 551.4nm, 566.2nm, 697.4nm, 530.3nm, 693.0nm, 543.0nm, 538.7nm and 568.4nm, and the The top 10 wavelengths with the highest importance scores are 547.2nm, 534.5nm, 551.4nm, 566.2nm, 697.4nm, 530.3nm, 693.0nm, 543.0nm, 538.7nm and 568.4nm as the sensitive characteristic wavelengths, that is, the wavelengths for distinguishing rice leaf categories. The minimum number of features can guarantee the detection accuracy of the detection model. The hyperspectral images corresponding to the first 10 sensitive characteristic wavelengths were used as the input of the 3DCNN deep learning model, and the 3DCNN detection model was trained to detect the early asymptomatic disease of rice bacterial blight.

The acquisition of the 3DCNN asymptomatic detection model includes:

The hyperspectral images of healthy rice leaves located at characteristic sensitive wavelengths in the hyperspectral image dataset and the hyperspectral images of rice leaves infected with bacterial blight at characteristic sensitive wavelengths in the hyperspectral image dataset are used as input to the 3DCNN model. The leaf category to which the hyperspectral image belongs is used as the output of the 3DCNN model;

Using the 3DCNN model to train based on the input of the 3DCNN model and the output of the 3DCNN model to obtain the 3DCNN asymptomatic detection model.

In the process of rice leaf disease hyperspectral image analysis, due to the degree of disease invasion and the resistance of rice plants to disease, in the selected region of interest, the rice leaf hyperspectral image contains a large amount of spectral information, which makes the use of fixed It is difficult to obtain sufficient spectral information for a detection model constructed with a convolution kernel of a large size, which leads to a reduction in recognition accuracy. Therefore, the present invention introduces a three-dimensional spectral hole convolution module in the 3DCNN asymptomatic detection model to expand the receptive field of the spectral dimension. Use the three-dimensional spectral atrous convolution module to extract spectral features of different scales, and finally perform fusion calculations to obtain richer spectral information and improve model detection performance, as follows:

The MS-SDC-3DCNN model is obtained by introducing a multi-scale spectral hole convolution module in the 3DCNN asymptomatic detection model, including:

By selecting three-dimensional spectral hole convolution modules with different spectral hole ratios and combining them into multi-scale spectral hole convolution modules, the spectral features of the hyperspectral images are extracted respectively, wherein the three-dimensional spectral hole convolution modules with different spectral hole ratios extract The obtained spectral feature information has different characteristic scales in the spectral dimension;

The multi-scale spectral atrous convolution module fuses the obtained spectral feature information of different feature scales to improve the acquisition of spectral feature information in hyperspectral images;

The spectral feature information output by the multi-scale spectral hole convolution module is used as the input of the 3DCNN asymptomatic detection model to improve the detection performance of the 3DCNN asymptomatic detection model;

The relationship between the receptive field of the three-dimensional spectral hole convolution module and the spectral hole rate is:

R _f =2×(r _SDR −1)×(k−1)+k;

Where R _f represents the receptive field of a single convolution kernel in the three-dimensional spectral hole convolution module; r _SDR represents the spectral hole rate, and k represents the size of the convolution kernel, where k is set to 3 by default;

In Figure 5(A), the black cube is a convolution kernel with a size of (3,3,3), and the spectral hole ratio SDR=1 indicates that no hole is added, and the size of the receptive field is also (3,3,3). In Figure 5(B), the spectral hole ratio SDR=2, the black cube is the position of the convolution kernel after the hole is added, and the white cube is the size of the convolution kernel receptive field after the hole is added (7,7,7). In Figure 5(C), the spectral hole ratio SDR=3, the black cube is the position of the convolution kernel after the hole is added, and the white cube is the size of the convolution kernel receptive field after the hole is added (11, 11, 11). Black cubes represent convolution kernels, and white cubes cover receptive fields. That is, when the SDR is set to 1, 2, and 3, the receptive fields of the convolution kernels in the three-dimensional spectral atrous convolution module are 3×3×3, 7×7×7, and 11×11×11, respectively. The three-dimensional spectral hole convolution modules with different spectral hole ratios are combined into a multi-scale spectral hole convolution module, which realizes the spectral feature extraction of the hyperspectral image respectively.

Specifically, the feature extraction operation in the multi-scale spectral atrous convolution module is implemented by 3D convolution, wherein the position (x, y, z) on the jth feature map in the i-th layer of the 3D convolutional network is calculated The formula for calculating the value at is:

为第i层第j个特征图上位置(x,y,z)处的值，H_i、W_i和D_i是第i层卷积核的高度、宽度和深度，f为第i层的激活函数，/>

是连接到上一层第m个特征图的卷积核位置(h,w,d)处的权重参数，/>

是上一层第m个特征图上位置(x+h,y+w,z+d)处的值，b_ij是第i层第j个特征图的偏差。in,

is the value at position (x, y, z) on the jth feature map of the i-th layer, H _i , W _i and D _i are the height, width and depth of the convolution kernel of the i-th layer, and f is the value of the i-th layer activation function, />

is the weight parameter at the convolution kernel position (h, w, d) connected to the mth feature map of the previous layer, />

is the value at position (x+h, y+w, z+d) on the mth feature map of the previous layer, b _ij is the deviation of the jth feature map of the i-th layer.

The MS-SDC-3DCNN model is obtained by using the output of the multi-scale spectral atrous convolution module as the input of the 3DCNN asymptomatic detection model.

Described utilizing MS-SDC-3DCNN model to carry out training based on sensitive image features, including:

Mixing hyperspectral images of healthy rice leaves at characteristic sensitive wavelengths in the hyperspectral image data set and hyperspectral images of rice leaves infected with bacterial blight at characteristic sensitive wavelengths in the hyperspectral image data set are used as training image data sets;

The training image data set is divided into a training set, a verification set and a test set in a ratio of 8:1:1, and the MS-SDC-3DCNN model is trained using the training set, verification set and test set to obtain the described detection model;

Among them, the MS-SDC-3DCNN model uses cross entropy as the loss function, and uses the stochastic gradient descent optimizer for training. The specific parameters of the MS-SDC-3DCNN model are as follows: the learning rate is set to 1×10 ^-3 , and the weight decay coefficient is set to 1×10 ^-6 , momentum is set to 0.95, ε is set to 1×10 ^-5 , epoch is set to 50, dropout is set to 0.45, and the hyperspectral image size input by MS-SDC-3DCNN is (9,9,10).

The MS-SDC-3DCNN model is provided with a residual block to implement the residual block to avoid the problem of gradient disappearance.

The hyperspectral image consists of 31680 pixels, and the size is (132, 240, 10). Among them, the hyperspectral image pixels are divided into three categories: healthy pixels, asymptomatic disease pixels and symptomatic disease pixels. Crop the hyperspectral image into image blocks of size (9,9,10) as the hyperspectral image size input by MS-SDC-3DCNN, and assign the classification label corresponding to the central pixel of each image block to the image piece.

The characteristic wavelength range is 450-950nm, and the characteristic wavelength range includes 232 spectral wavelengths. The spectral resolution of the hyperspectral imaging system used in the experiment is 2.14nm, and the selected characteristic wavelength range is 450-950nm. Dividing the wavelength range by the spectral resolution can be calculated to include 232 wavelengths in the wavelength range.

The invention uses the random forest algorithm to reduce the dimension of the hyperspectral image, determines the wavelength sensitive to the disease, reduces the training time of the asymptomatic detection model of rice leaves, and constructs a multi-scale spectral hole convolution module, which can be achieved without increasing the amount of calculation. Next, expand the receptive field of the neural network to improve the feature extraction ability, and optimize the asymptomatic detection model by using the multi-scale spectral hole convolution module. Efficient use of important wavelength information to improve detection performance of asymptomatic detection models.

The above embodiments are only exemplary embodiments of the present application, and are not intended to limit the present application, and the protection scope of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements shall also be deemed to fall within the protection scope of the present application.

Claims (10)

1. The early asymptomatic detection method for the bacterial leaf blight of rice is characterized by comprising the following steps of:

acquiring hyperspectral images of healthy rice leaves and rice leaves infected with bacterial leaf blight at each spectral wavelength in a preset wave band to form a hyperspectral image dataset, and constructing hyperspectral curves of the healthy rice leaves and the rice leaves infected with bacterial leaf blight by using the hyperspectral image dataset to determine a characteristic wavelength interval, wherein the hyperspectral images in the characteristic wavelength interval are characterized as image features of the healthy rice leaves and the rice leaves infected with bacterial leaf blight;

carrying out importance scoring on each spectrum wavelength in a characteristic wavelength interval based on the hyperspectral image in the characteristic wavelength interval by utilizing a random forest algorithm, and selecting the spectrum wavelength with high importance score in the characteristic wavelength interval as a characteristic sensitive wavelength;

taking the hyperspectral image at the characteristic sensitive wavelength as a sensitive image characteristic for distinguishing the blade class to which the hyperspectral image belongs, so as to realize the data dimension reduction of a hyperspectral image dataset;

training a 3DCNN model based on sensitive image features to obtain a 3DCNN asymptomatic detection model for early asymptomatic detection of rice bacterial leaf blight;

and introducing a multi-scale spectrum cavity convolution module into the 3DCNN asymptomatic detection model to perform precision optimization to obtain an MS-SDC-3DCNN model so as to realize early asymptomatic high-precision detection of the bacterial leaf blight of the rice.

2. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 1, wherein the method comprises the steps of: the method for determining the characteristic wavelength interval by utilizing the hyperspectral curve of the healthy rice leaves and the rice leaves infected with bacterial leaf blight disease constructed by the hyperspectral image dataset comprises the following steps:

drawing all hyperspectral images belonging to healthy rice leaves in a hyperspectral image dataset in a spectrum coordinate system to obtain a hyperspectral curve of the healthy rice leaves, wherein the ordinate in the spectrum two-dimensional coordinate system is reflection intensity, and the abscissa is spectrum wavelength;

drawing all hyperspectral images of asymptomatic rice leaves in the hyperspectral image data set belonging to rice leaves infected with bacterial leaf blight to obtain hyperspectral curves of asymptomatic rice leaves in a spectrum coordinate system;

drawing all hyperspectral images of the mild symptom rice leaves in the hyperspectral image data set, which belong to the rice leaves infected with bacterial leaf blight, in a spectrum coordinate system to obtain a hyperspectral curve of the mild symptom rice leaves;

drawing all hyperspectral images of moderate symptom rice leaves in the hyperspectral image data set belonging to the rice leaves infected with bacterial leaf blight to obtain hyperspectral curves of the moderate symptom rice leaves in a spectrum coordinate system;

drawing all hyperspectral images of the rice leaves with severe symptoms in the hyperspectral image data set, which belong to the rice leaves infected with bacterial leaf blight, in a spectrum coordinate system to obtain hyperspectral curves of the rice leaves with severe symptoms;

the method comprises the steps of performing point-by-point similarity measurement in a hyperspectral curve of healthy rice leaves, a hyperspectral curve of asymptomatic rice leaves, a hyperspectral curve of mild symptomatic rice leaves, a hyperspectral curve of moderate symptomatic rice leaves and a hyperspectral curve of severe symptomatic rice leaves, and performing point-by-point connection on all curve points with similarity lower than a preset threshold value to obtain a characteristic distinguishing curve;

and taking the spectral wavelength interval in which the characteristic distinguishing curve is located as the characteristic wavelength interval.

3. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 2, wherein the method comprises the steps of: the scoring the importance of each spectral wavelength in the characteristic wavelength interval based on the hyperspectral image in the characteristic wavelength interval by using a random forest algorithm comprises the following steps:

carrying out random replacement on each hyperspectral image in the characteristic wavelength interval in sequence by utilizing a random forest algorithm, and calculating classification errors for distinguishing the classes of the blades to which the hyperspectral images belong before and after replacement as importance scores of the hyperspectral images;

and taking the importance scores of the hyperspectral images as the importance scores of the spectral wavelengths corresponding to the hyperspectral images, and sorting the spectral wavelengths from high to low according to the importance scores of the spectral wavelengths.

4. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 3, wherein the method comprises the steps of: the selecting the spectral wavelength with high importance score in the characteristic wavelength interval as the characteristic sensitive wavelength comprises the following steps:

setting a selected threshold value, summing importance scores of the spectrum wavelengths sequenced in the characteristic wavelength interval from high to low until the selected threshold value is met, and taking the spectrum wavelengths corresponding to the summation process as the characteristic sensitive wavelengths;

wherein the selected threshold is 0.9.

5. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 3 or 4, wherein the method comprises the steps of: the obtaining of the 3DCNN asymptomatic detection model comprises the following steps:

taking hyperspectral images of healthy rice leaves positioned at the characteristic sensitive wavelength in the hyperspectral image data set and hyperspectral images of rice leaves infected with bacterial blight and positioned at the characteristic sensitive wavelength in the hyperspectral image data set as inputs of a 3DCNN model, and taking the leaf category to which the hyperspectral images belong as outputs of the 3DCNN model;

and training by using a 3DCNN model based on the input of the 3DCNN model and the output of the 3DCNN model to obtain the 3DCNN asymptomatic detection model.

6. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 3, wherein the method comprises the steps of: introducing a multi-scale spectrum cavity convolution module into the 3DCNN asymptomatic detection model to obtain an MS-SDC-3DCNN model, wherein the method comprises the following steps:

the three-dimensional spectrum cavity convolution modules with different spectrum cavity rates are selected to be combined into a multi-scale spectrum cavity convolution module, spectrum characteristic extraction is respectively carried out on the hyperspectral image, wherein the spectrum characteristic information extracted by the three-dimensional spectrum cavity convolution modules with different spectrum cavity rates has different characteristic scales in spectrum dimension;

the multi-scale spectrum cavity convolution module fuses the obtained spectrum characteristic information with different characteristic scales so as to improve the acquisition of the spectrum characteristic information in the hyperspectral image;

taking the spectral characteristic information output by the multi-scale spectral cavity convolution module as the input of a 3DCNN asymptomatic detection model to improve the detection performance of the 3DCNN asymptomatic detection model;

the relation between the receptive field and the spectrum void ratio of the three-dimensional spectrum void convolution module is as follows:

R _f ＝2×(r _SDR -1)×(k-1)+k；

wherein R is _f Representing the receptive field of a single convolution kernel in the three-dimensional spectrum cavity convolution module; r is (r) _SDR Representing spectral void fraction, k represents the size of the convolution kernel, where k is set to 3 by default;

the feature extraction operation in the multi-scale spectrum hole convolution module is realized by 3D convolution, wherein a calculation formula for calculating a value at a position (x, y, z) on a j-th feature map in an i-th layer of the 3D convolution network is as follows:

wherein,,

is the value at the position (x, y, z) on the j-th feature map of the i-th layer, H _i 、W _i And D _i Is the height, width and depth of the convolution kernel of the ith layer, f is the activation function of the ith layer,/>

Is the weight parameter at the convolution kernel position (h, w, d) connected to the mth feature map of the upper layer, +.>

Is the value at the position (x+h, y+w, z+d) on the mth feature map of the upper layer, b _ij Is the deviation of the j-th feature map of the i-th layer;

and taking the output of the multi-scale spectrum cavity convolution module as the input of the 3DCNN asymptomatic detection model to obtain the MS-SDC-3DCNN model.

7. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 4, wherein the training based on sensitive image features using an MS-SDC-3DCNN model comprises:

mixing hyperspectral images of healthy rice leaves positioned at the characteristic sensitive wavelength in the hyperspectral image data set and hyperspectral images of rice leaves infected with bacterial blight positioned at the characteristic sensitive wavelength in the hyperspectral image data set to serve as a training image data set;

dividing a training image data set into a training set, a verification set and a test set according to the proportion of 8:1:1, and training the MS-SDC-3DCNN model by utilizing the training set, the verification set and the test set to obtain the detection model;

the MS-SDC-3DCNN model uses cross entropy as a loss function, and is trained by using a random gradient descent optimizer, and specific parameters of the MS-SDC-3DCNN model are as follows: the learning rate is set to 1×10 ^-3 The weight attenuation coefficient is set to 1×10 ^-6 Momentum is set to 0.95 and ε is set to 1×10 ^-5 The epoch was set to 50, dropout was set to 0.45, and the hyperspectral image size of the MS-SDC-3DCNN input was (9,9,10).

8. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 2, wherein a residual block is arranged in the MS-SDC-3DCNN model to realize the residual block to avoid the problem of gradient disappearance.

9. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 2, wherein the hyperspectral image consists of 31680 pixels and has a size of (132,240,10), wherein the hyperspectral image pixels are classified into three types: healthy pixels, asymptomatic disease pixels, and symptomatic disease pixels.

10. The method for early asymptomatic detection of bacterial leaf blight of rice according to claim 2, wherein the characteristic wavelength interval is 450-950nm and the characteristic wavelength interval comprises 232 spectral wavelengths.

Images