Pest and disease detection system based on point target

Abstract

The invention discloses a pest detection system based on a point target. Relates to the technical field of spectrum detection, in particular to the technical field of pest detection based on point targets. The system comprises a spectral image data acquisition module, a spectral image data preprocessing module, a crop disease and insect pest point target detection module and a crop disease and insect pest point target verification module. The spectral image data acquisition module is used for acquiring crop spectral image data and transmitting the crop spectral image data to the spectral image preprocessing module in real time; the spectral image data preprocessing module is used for extracting a spectral band image taking a specific wavelength as a central wavelength; the crop disease and pest point target detection module is used for detecting disease abnormal point position information in a target crop; and the crop disease and insect pest point target verification module is used for constructing a crop disease and insect pest detection model, performing feature detection on the disease abnormal points output by the crop disease and insect pest point target detection module, and completing data verification of the disease abnormal points.

Description

Technical Field

[0001] The present invention relates to the technical field of spectral detection, and in particular to the technical field of pest and disease detection based on point targets. Background Art

[0002] Agricultural pests and diseases have always been a crucial part of agricultural production. As the main factor leading to crop yield reduction and quality decline, it poses a severe challenge to food security and ecological environmental sustainability. Therefore, timely and effective monitoring of crop pests and diseases is particularly important, which not only helps to detect and curb the spread of pests and diseases at an early stage, but also provides a solid guarantee for the steady development of agricultural production.

[0003] The current technology system still faces many challenges in accurately detecting the early stages of crop disease. Take red spider pests as an example. They are so small that they are difficult to detect with the naked eye (a single spider is about 0.5 mm in size). They suck the sap from crop leaves, causing symptoms such as chlorosis and yellowing in the affected parts. However, by the time these symptoms appear, the red spider pests have often become rampant, and the best time for prevention and control has been missed.

[0004] At present, the main technical means used for agricultural pest detection include digital image processing, machine learning and spectral technology. In the field of digital image processing, the patent document "A Smart Agricultural Pest Detection Method" (CN116630815B) obtains abnormal pixels in crop images according to digital image processing methods, and performs pest detection on crop images according to specific thresholds. Although the above method can screen out abnormal grayscale features on the crop surface, interference such as stains on the crop surface will cause recognition interference to the crop disease spot detection, making it difficult to accurately detect the onset of the disease in the early stage of the crop.

[0005] In the field of machine learning technology, deep neural networks are constructed to extract the characteristics of green plant lesions and the size of lesions, and identify the types of pests and diseases and the severity of the disease. For example, the patent document "A method for detecting crop pests and diseases" (CN107067043A) discloses a method for detecting crop pests and diseases. This method extracts the pixel feature distribution of different crop pests and diseases by constructing a convolutional neural network model to identify the types of different pests and diseases suffered by crops. However, in the early stages of crop disease, the effective symptom feature area of ​​the crop is often too small, and the differences between different pests and diseases in the early stages of crop disease are not significant, which makes this method have certain limitations in the early detection of crop pests and diseases.

[0006] In terms of spectral technology, the health of plants is evaluated by calculating and analyzing the vegetation index of the target area, or the disease of crops is detected by combining spectral imaging with machine learning technology solutions. For example, the patent document "Multi-spectral Pest Detection and Identification Method, System and Storage Medium" (CN114199788A) describes a method of using a multi-spectral drone to calculate the vegetation index of the target plot area and then analyze the health status of crops. However, this method is mainly applicable to the stage when the characteristics of crop diseases are obvious, and the detection effect is not ideal for the early stage of crop disease. In the patent document "Multi-spectral Pest Detection and Identification Method, System and Storage Medium" (CN114199788A), spectral data is obtained through a hyperspectral imaging system, and the machine model is trained with characteristic band spectral data, and citrus pest detection is applied. This method trains spectral feature data through a machine learning model. Although it enhances the effective feature expression of lesions to a certain extent, there is still no solution for small target pest extraction technology, and there are still certain technical difficulties and challenges in the detection of early crop diseases. Summary of the invention

[0007] In order to solve at least one of the above technical problems, the present invention provides a point target-based pest detection system, the system comprising a spectral image data acquisition module, a spectral image data preprocessing module, a crop pest point target detection module, and a crop pest point target verification module;

[0008] The spectral image data acquisition module is used to obtain crop spectral image data and transmit the crop spectral image data to the spectral image preprocessing module in real time;

[0009] The spectral image data preprocessing module is used to preprocess the input crop spectral image data and extract a spectral band image with a specific wavelength as the central wavelength;

[0010] The crop pest and disease point target detection module is used to receive the spectral band image data transmitted by the spectral image data preprocessing module and detect the location information of abnormal disease points in the target crop;

[0011] The crop pest and disease point target verification module is used to construct a crop pest and disease detection model based on the spectral data of healthy crops and diseased crops, perform feature detection on the disease abnormal points output by the crop pest and disease point target detection module, and complete data verification of the disease abnormal points.

[0012] Furthermore, the spectral image acquisition device in the spectral image data acquisition module is a snapshot imaging spectral camera, and the spectral band range covers the visible light to near infrared light band.

[0013] Furthermore, the spectral image data preprocessing module includes a spectral feature preprocessing module and a spectral band image output module in sequence. The spectral feature preprocessing module performs background subtraction and mean fusion processing on the crop spectral image data, and the spectral band image output module extracts a spectral band image with a specific wavelength as the center wavelength according to the characteristics of the pests and diseases to be detected.

[0014] Furthermore, the crop pest and disease point target detection module performs the following operations in sequence: performing disease feature image segmentation; detecting disease abnormal points; fusing the disease feature image segmentation result with the disease abnormal point detection result to obtain the position coordinates of the disease abnormal point after fusion processing.

[0015] Furthermore, the disease feature image segmentation is specifically as follows:

[0016] The deep residual neural network is used to train the spectral band image data, and the gradient descent strategy is used to update the weight parameters and feature gradients. The gradient of the disease abnormal point is calculated as follows: Among them, W new represents the updated weight, W old represents the weight before update, η represents the learning rate, which is used to control the step size of weight update, Represents the gradient of the loss function L with respect to the current feature weight w; the feature gradient is calculated as: Among them, T represents the gradient of the feature weight of the disease abnormal point, represents the gradient of the abnormal disease point bias, Represents the weight gradient of the disease anomaly points extracted by the neuron; constructs the Mask of the crop contour and stores the position coordinates of the crop contour.

[0017] Further, the abnormal points of the disease detection are specifically:

[0018] Based on the disease feature image segmentation prediction results, the formula Calculate the gradient amplitude between the target crop and the disease abnormal point, where G x is the horizontal gradient, G y is the vertical gradient; the abnormal disease points are screened by setting the gradient detection threshold. The calculation method of the gradient threshold is: T = G mean +α·(G max -G min )·β, where G mean is the mean gradient, G max and G min They represent the maximum and minimum values ​​of the gradient respectively, α is the adaptive factor, and β is the proportional factor.

[0019] Furthermore, when the disease feature image segmentation result is fused with the disease outlier point detection result, the crop contour area is deducted from the disease outlier point detection result according to the crop contour Mask, the disease outlier points after the crop contour deduction are constructed as a Mask matrix, and the position coordinates of the disease outlier points after feature fusion processing are saved.

[0020] Furthermore, the crop pest and disease detection model is constructed based on the spectral data of healthy crops and diseased crops as follows:

[0021] Step 1: Build a database of crop pest and disease detection models based on the characteristic spectral data of crop health and disease areas, and calibrate the characteristic spectral data to the category to which it belongs;

[0022] Step 2: Preprocess the characteristic spectrum data, including normalization of the working spectrum and smoothing of the spectrum data;

[0023] Step 3: Perform spectral data modeling based on the characteristic spectral data preprocessed in step 2 to construct a crop disease and pest detection model, wherein the crop disease and pest detection model is based on a random forest machine learning method, including randomly extracting subsamples from a training sample set and constructing a disease anomaly decision tree.

[0024] Furthermore, the data verification of the disease abnormal point is specifically as follows:

[0025] According to the position information of the abnormal disease points in the target crop detected by the crop pest and disease point target detection module, the characteristic spectrum data of the abnormal disease points are obtained; the obtained characteristic spectrum data of the abnormal disease points are subjected to the spectrum preprocessing in step 2; the preprocessed characteristic spectrum data of the abnormal disease points are input into the crop pest and disease detection model to perform abnormal disease point feature verification. The abnormal disease point feature verification calculation method is specifically as follows:

[0026] Z=model(X processed ), where model represents the detection model trained in step 3, and X processed It is the characteristic spectrum data of disease abnormal points after preprocessing.

[0027] The beneficial effects of the system of the present invention are:

[0028] The system of the present invention can accurately and efficiently capture the spectral data of crop and disease abnormalities, and covers a wide spectral band range from visible light to near-infrared light. This provides a more comprehensive and multi-dimensional spectral image data basis for detection, and enhances the details and discovery depth of detection.

[0029] The system of the present invention integrates digital image processing technology, deep neural network and spectral imaging technology to construct a point target detection scheme for early stage crop diseases. This scheme can fully capture abnormal pixel information of early crop diseases and provide an effective technical solution to the problem that the early characteristics of diseases and insect pests are not obvious and difficult to detect.

[0030] The system of the present invention constructs a crop disease point target verification module through spectral technology, which can effectively filter out interfering abnormal points and accurately identify disease characteristics, thereby improving detection accuracy and reducing false alarm rate.

[0031] The system of the present invention provides strong data support and algorithm guarantee for the accurate detection of pests and diseases in farmland and greenhouses, has high detection efficiency and accuracy, and can be widely used in pest and disease monitoring and early warning in agricultural production. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The following is a flowchart of the system operation of the present invention. DETAILED DESCRIPTION

[0033] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0034] Embodiment 1,

[0035] This embodiment provides a point target-based pest detection system, such as Figure 1 As shown, the system includes a spectral image data acquisition module, a spectral image data preprocessing module, a crop disease and insect pest point target detection module, and a crop disease and insect pest point target verification module;

[0036] The spectral image data acquisition module is used to obtain crop spectral image data and transmit the crop spectral image data to the spectral image preprocessing module in real time;

[0037] The spectral image data preprocessing module is used to preprocess the input crop spectral image data and extract a spectral band image with a specific wavelength as the central wavelength;

[0038] The crop pest and disease point target detection module is used to receive the spectral band image data transmitted by the spectral image data preprocessing module and detect the location information of abnormal disease points in the target crop;

[0039] The crop pest and disease point target verification module is used to construct a crop pest and disease detection model based on the spectral data of healthy crops and diseased crops, perform feature detection on the disease abnormal points output by the crop pest and disease point target detection module, and complete data verification of the disease abnormal points.

[0040] Embodiment 2,

[0041] This embodiment is a further limitation of Embodiment 1. The spectral image acquisition device in the spectral image data acquisition module is a snapshot imaging spectral camera. The spectral band range covers 350nm to 950nm and can support a variety of light source detection environments, including outdoor sunlight, D65, etc.

[0042] The spectral image data preprocessing module includes a spectral feature preprocessing module and a spectral band image output module in sequence. The spectral feature preprocessing module performs background subtraction and mean fusion processing on the crop spectral image data, wherein the mean fusion can be performed in the entire spectral range or in a specific band or feature area. The spectral band image output module extracts a spectral band image with a specific wavelength as the central wavelength according to the characteristics of the pests and diseases to be detected. The specific band can be selected according to the characteristics of different pests and diseases. Furthermore, this example extracts spectral band images with 550nm, 690nm, and 770nm as central wavelengths according to the differences in the spectral characteristics of healthy crops and red spider diseases.

[0043] Embodiment 3,

[0044] This embodiment is a further limitation of Embodiment 1, and the following operations are performed in the crop pest and disease point target detection module in sequence: performing disease feature image segmentation; detecting disease abnormal points; fusing the disease feature image segmentation result with the disease abnormal point detection result to obtain the position coordinates of the disease abnormal point after fusion processing.

[0045] The disease feature image segmentation is specifically as follows:

[0046] The deep residual neural network is used to train the spectral band image data, and the gradient descent strategy is used to update the weight parameters and feature gradients. The gradient of the disease abnormal point is calculated as follows: Among them, W new represents the updated weight, W old represents the weight before update, η represents the learning rate, which is used to control the step size of weight update, Represents the gradient of the loss function L with respect to the current feature weight w; the feature gradient is calculated as: Among them, T represents the feature weight gradient of the disease abnormal point, represents the gradient of the abnormal disease point bias, Represents the weight gradient of the disease anomaly points extracted by the neuron; constructs the Mask of the crop contour and stores the position coordinates of the crop contour.

[0047] The specific detection disease abnormal points are:

[0048] Based on the disease feature image segmentation prediction results, the formula Calculate the gradient amplitude between the target crop and the disease abnormal point, where G x is the horizontal gradient, G y is the vertical gradient; the abnormal disease points are screened by setting the gradient detection threshold. The calculation method of the gradient threshold is: T = G mean +α·(G max -G min )·β, where G mean is the mean gradient, G max and G min They represent the maximum and minimum values ​​of the gradient respectively, α is the adaptive factor, and β is the proportional factor.

[0049] When the disease feature image segmentation results are fused with the disease anomaly point detection results, the crop contour area is deducted from the disease anomaly point detection results according to the crop contour Mask, the disease anomaly points after the crop contour deduction are constructed as a Mask matrix, and the position coordinates of the disease anomaly points after feature fusion processing are saved.

[0050] Embodiment 4,

[0051] This embodiment is a further limitation of the first embodiment. According to the spectral data of healthy crops and diseased crops, a crop disease and insect pest detection model is constructed as follows:

[0052] Step 1: Construct a database of crop pest and disease detection model based on the characteristic spectral data of crop health and disease areas, and calibrate the characteristic spectral data to the category to which it belongs; further, the characteristic spectral data calibration formula is as follows:

[0053] Where D is the spectral dataset, x i represents the spectral data of the i-th sample, y i represents its category label, and n is the total number of samples.

[0054] Step 2: Preprocess the characteristic spectrum data, including normalization of the working spectrum, smoothing of the spectrum data, first-order derivative of the characteristic spectrum, etc.

[0055] The data preprocessing scheme includes but is not limited to the above operation process, and preprocessing adjustment can be made according to the spectral preprocessing method for different disease detection.

[0056] Further, the spectrum normalization calculation formula is:

[0057]

[0058] Further, the spectrum normalization calculation formula is:

[0059]

[0060] Further, the first-order derivative calculation formula is:

[0061]

[0062] Among them, S norm is the preprocessing result of spectral normalization, S is the number of input pest characteristic spectra, S min With S max are the minimum and maximum values ​​of the input characteristic spectrum, S smooth is the smoothed data point, S i+j S is the spectral data point around the smoothing point, and r is the smoothing window radius. i+1 and S i are the spectral data values ​​of two adjacent wavelength points, and Δλ is the wavelength difference between the two adjacent wavelength points.

[0063] A crop pest and disease detection model is constructed. The crop pest and disease detection model is based on a random forest machine learning method, including randomly extracting subsamples from a training sample set and constructing a disease anomaly decision tree. Furthermore, the decision tree is calculated as follows:

[0064]

[0065] Among them, h(x) is the prediction result of disease characteristics, N is the number of decision trees, T i is the prediction result of the i-th decision tree.

[0066] The data verification of the disease abnormal point is specifically as follows:

[0067] According to the position information of the abnormal disease points in the target crop detected by the crop pest and disease point target detection module, the characteristic spectrum data of the abnormal disease points are obtained; the obtained characteristic spectrum data of the abnormal disease points are subjected to the spectrum preprocessing in step 2; the preprocessed characteristic spectrum data of the abnormal disease points are input into the crop pest and disease detection model to perform abnormal disease point feature verification. The abnormal disease point feature verification calculation method is specifically as follows:

[0068] Z=model(X processed ), where model represents the detection model trained in step 3, and X processed It is the characteristic spectrum data of disease abnormal points after preprocessing.

Claims

1. A point target-based pest detection system, characterized in that: The system includes a spectral image data acquisition module, a spectral image data preprocessing module, a crop disease and insect pest point target detection module, and a crop disease and insect pest point target verification module; The spectral image data acquisition module is used to obtain crop spectral image data and transmit the crop spectral image data to the spectral image preprocessing module in real time; The spectral image data preprocessing module is used to preprocess the input crop spectral image data and extract a spectral band image with a specific wavelength as the central wavelength; The crop pest and disease point target detection module is used to receive the spectral band image data transmitted by the spectral image data preprocessing module and detect the location information of abnormal disease points in the target crop; The crop pest and disease point target verification module is used to construct a crop pest and disease detection model based on the spectral data of healthy crops and diseased crops, perform feature detection on the disease abnormal points output by the crop pest and disease point target detection module, and complete data verification of the disease abnormal points.

2. The point target-based pest detection system according to claim 1, characterized in that: The spectral image acquisition device in the spectral image data acquisition module is a snapshot imaging spectral camera, and the spectral band range covers the visible light to near infrared light band.

3. The point target-based pest detection system according to claim 2, characterized in that: The spectral image data preprocessing module includes a spectral feature preprocessing module and a spectral band image output module in sequence. The spectral feature preprocessing module performs background subtraction and mean fusion processing on the crop spectral image data, and the spectral band image output module extracts a spectral band image with a specific wavelength as the center wavelength according to the characteristics of the pests and diseases to be detected.

4. The point target-based pest detection system according to claim 3, characterized in that: The following operations are performed in the crop pest and disease point target detection module: segmenting the disease feature image; detecting disease abnormal points; fusing the disease feature image segmentation result with the disease abnormal point detection result to obtain the position coordinates of the disease abnormal point after fusion processing.

5. The point target-based pest detection system according to claim 4, characterized in that: The disease feature image segmentation is specifically as follows: The deep residual neural network is used to train the spectral band image data, and the gradient descent strategy is used to update the weight parameters and feature gradients. The gradient of the disease abnormal point is calculated as follows: Among them, W new represents the updated weight, W old represents the weight before update, η represents the learning rate, which is used to control the step size of weight update, Represents the gradient of the loss function L with respect to the current feature weight w; the feature gradient is calculated as: Among them, T represents the gradient of the feature weight of the disease abnormal point, represents the gradient of the abnormal disease point bias, Represents the weight gradient of the disease anomaly points extracted by the neuron; constructs the Mask of the crop contour and stores the position coordinates of the crop contour.

6. The point target-based pest detection system according to claim 5, characterized in that: The specific detection disease abnormal points are: Based on the disease feature image segmentation prediction results, the formula Calculate the gradient amplitude between the target crop and the disease abnormal point, where G x is the horizontal gradient, G y is the vertical gradient; the abnormal disease points are screened by setting the gradient detection threshold. The calculation method of the gradient threshold is: T = G mean +α·(G max -G min )·β, where G mean is the mean gradient, G max and G min They represent the maximum and minimum values ​​of the gradient respectively, α is the adaptive factor, and β is the proportional factor.

7. The point target-based pest detection system according to claim 6, characterized in that: When the disease feature image segmentation results are fused with the disease anomaly point detection results, the crop contour area is deducted from the disease anomaly point detection results according to the crop contour Mask, the disease anomaly points after the crop contour deduction are constructed as a Mask matrix, and the position coordinates of the disease anomaly points after feature fusion processing are saved.

8. The point target-based pest detection system according to claim 7, characterized in that: The crop pest and disease detection model is constructed based on the spectral data of healthy crops and diseased crops as follows: Step 1: Build a database of crop pest and disease detection models based on the characteristic spectral data of crop health and disease areas, and calibrate the characteristic spectral data to the category to which it belongs; Step 2: Preprocess the characteristic spectrum data, including normalization of the working spectrum and smoothing of the spectrum data; Step 3: Perform spectral data modeling based on the characteristic spectral data preprocessed in step 2 to construct a crop disease and pest detection model, wherein the crop disease and pest detection model is based on a random forest machine learning method, including randomly extracting subsamples from a training sample set and constructing a disease anomaly decision tree.

9. The point target-based pest detection system according to claim 8, characterized in that: The data verification of the disease abnormal point is specifically as follows: According to the position information of the abnormal disease points in the target crop detected by the crop pest and disease point target detection module, the characteristic spectrum data of the abnormal disease points are obtained; the obtained characteristic spectrum data of the abnormal disease points are subjected to the spectrum preprocessing in step 2; the preprocessed characteristic spectrum data of the abnormal disease points are input into the crop pest and disease detection model to perform abnormal disease point feature verification. The abnormal disease point feature verification calculation method is specifically as follows: Z=model(X processed ), where model represents the detection model trained in step 3, and X processed It is the characteristic spectrum data of disease abnormal points after preprocessing.

Images