Early identification and early warning method and system for tobacco diseases

Abstract

The invention belongs to the technical field of early-stage tobacco disease early warning, and particularly relates to an early-stage identification and early-warning method and system for tobacco diseases. The method comprises the following steps: acquiring a data set T1 of hyperspectral data in each stage of the whole growth period of tobacco, screening diseased tobacco plants with obvious disease characteristics in each stage from the data set T1, acquiring corresponding hyperspectral data of the diseased tobacco plants in each stage in the previous stage, marking the hyperspectral data with the obvious disease characteristics in the previous stage as diseased, and marking the hyperspectral data with the obvious disease characteristics in the next stage as diseased tobacco plants; marking the hyperspectral data without obvious disease characteristics in the previous stage as pre-disease, and training the machine learning model by using a data set comprising the hyperspectral data of the diseased tobacco plants and the pre-diseased tobacco plants to obtain a tobacco disease early recognition and early warning model, so as to recognize the tobacco plants without obvious disease characteristics in the pre-disease period. Precise prevention and control are carried out on field tobacco plants which are difficult to distinguish disease characteristics by naked eyes before diseases occur, and timely early warning can be carried out before the disease characteristics of the tobacco plants occur.

Description

Technical Field

[0001] The present invention belongs to the technical field of early warning of tobacco leaf diseases, and particularly relates to a method and a system for early identification and warning of tobacco diseases. Background Art

[0002] Tobacco is an important cash crop in China and plays an important role in the national economy. However, due to various factors such as climate and environment, various diseases frequently occur in field tobacco, and the disease occurrence environment is complex and the disease onset speed is fast, which seriously affects the stability of the yield and quality of tobacco leaf raw materials and the income of tobacco farmers. For this reason, the Chinese patent application with publication number CN116297311A discloses a method for identifying early tobacco leaf diseases. First, tobacco experts identify tobacco leaves of different disease types, and then use a handheld near-infrared spectrometer to obtain the near-infrared spectral data of healthy and diseased tobacco leaves with different infection degrees. Based on the near-infrared spectral data of healthy and diseased tobacco leaves with different infection degrees, a three-layer convolutional structure is trained to obtain a tobacco leaf disease spectral prediction model. Finally, the real-time near-infrared spectral data of tobacco leaves is input into the tobacco leaf disease spectral prediction model for prediction, and the early tobacco leaf disease prediction result is output. The prediction result is whether the tobacco leaf is healthy or the type of early tobacco leaf disease, and the disease types include powdery mildew, black blast, mosaic disease, and brown spot. This method uses diseased tobacco leaves with visually distinguishable different infection degrees to establish a prediction model, which can identify in time at the early stage when the tobacco leaves show diseases, so as to treat the tobacco leaves when the disease degree is relatively low, realizing the early prevention and control of tobacco leaf diseases. However, this method uses near-infrared light data for disease identification, and can only identify tobacco leaves with obvious disease characteristics in a local area, and cannot realize large-area disease monitoring. The disease prevention and control only takes a point as a surface; and it can only identify and warn after the tobacco leaves show obvious disease characteristics that can be visually identified. At this time, the tobacco leaves have been inoculated with pathogens and suffered from diseases, and a large amount of drugs for disease prevention and control need to be invested for treatment. The prevention and control of tobacco field diseases is still not timely, and the prevention and control effect is relatively low. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and a system for early identification and warning of tobacco diseases, so as to solve the problem that the existing methods for early identification of tobacco diseases result in untimely prevention and control of tobacco field diseases and relatively low prevention and control effects.

[0004] The present invention provides a method for early identification and warning of tobacco diseases to solve the above technical problems, including: Obtaining hyperspectral data of the tobacco plants to be measured; Input the hyperspectral data of the tobacco plants to be tested into the early identification and warning model for tobacco diseases, and output the disease identification results; the disease identification results include pre-disease or existing disease; the early identification and warning model for tobacco diseases is obtained by training a machine learning model using a dataset including the hyperspectral data of diseased and pre-diseased tobacco plants; the construction process of the dataset is as follows: obtain the dataset T1 of the hyperspectral data at each stage of the whole growth period of tobacco, screen out the diseased tobacco plants with obvious disease characteristics at each stage from the dataset T1, obtain the hyperspectral data corresponding to the diseased tobacco plants at the previous stage at each stage, label the hyperspectral data with obvious disease characteristics at the previous stage as existing disease, and label the hyperspectral data without obvious disease characteristics at the previous stage as pre-disease.

[0005] Further, the dataset used for training the machine learning model also includes the hyperspectral data of healthy tobacco plants, and healthy tobacco plants refer to those that do not show obvious disease characteristics during each stage of the whole growth period of tobacco.

[0006] Further, the hyperspectral data input into the early identification and warning model for tobacco diseases and used for training the early identification and warning model for tobacco diseases is the extracted sensitive band data.

[0007] Further, the sensitive bands include the sensitive bands of pre-disease, existing disease, and health, and the sensitive bands of pre-disease, existing disease, and health are determined by analyzing the contribution rates of the reflectance of each band of the hyperspectral data of tobacco plants to the discrimination of pre-diseased tobacco plants, existing diseased tobacco plants, and healthy tobacco plants using machine learning algorithms.

[0008] Further, the hyperspectral data is pre-processed data, and the pre-processing includes lens calibration, reflectance calibration, and atmospheric correction.

[0009] Further, the machine learning model is a Stacking ensemble learning model including N base learners and one meta-learner, where N≥2.

[0010] Further, N is 4, and the corresponding base learners respectively adopt the random forest model, gradient boosting decision tree model, extreme gradient boosting tree model, and lightweight gradient boosting tree model, and the meta-learner adopts the multiple linear regression model.

[0011] Further, the hyperspectral data of the tobacco plants to be tested input into the early identification and warning model for tobacco diseases is the hyperspectral data of the complete field obtained by splicing the hyperspectral data of each tobacco plant to be tested.

[0012] The beneficial effects of the above technical solution are as follows: The present invention is an exploratory invention. A dataset T1 containing hyperspectral data of each stage of the entire tobacco growth period is established. After determining the diseased tobacco plants with obvious disease characteristics at each stage from T1, the diseased tobacco plants are traced back. The hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage are obtained from the dataset T1, and the hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage are labeled. If the diseased tobacco plants do not have obvious disease characteristics in the previous stage, the hyperspectral data corresponding to the previous stage is marked as pre-diseased, indicating that the tobacco plants are already in the pre-diseased state and will immediately show obvious disease characteristics and enter the diseased state in the next stage; if the diseased tobacco plants have obvious disease characteristics in the previous stage, the hyperspectral data corresponding to the previous stage is marked as diseased, indicating that the tobacco plants are already in the diseased state. According to the above judgment criteria, the hyperspectral data of the diseased tobacco plants in the previous stage are labeled to form a dataset including the hyperspectral data of the diseased and pre-diseased tobacco plants. The labeled dataset is used to train a machine learning model to obtain an early identification and warning model for tobacco diseases. The hyperspectral data of the tobacco plants in the field are input into the early identification and warning model for tobacco diseases, and the tobacco plants without obvious disease characteristics in the pre-diseased period can be identified, realizing the precise prevention and control of the tobacco plants in the field before the occurrence of diseases with disease characteristics that are difficult to distinguish by the naked eye. It can provide timely warnings before the tobacco plants show disease characteristics, guide tobacco farmers and scientific researchers to spray pesticides reasonably on the pre-diseased tobacco plants in advance, greatly reducing the amount of drugs used for disease prevention and control, and providing a strong reference for preventing the occurrence of diseases.

[0013] To solve the above technical problems, the present invention also provides an early identification and warning system for tobacco diseases, including a processor, and the processor is used to execute computer program instructions to implement.

[0014] Furthermore, it further includes a hyperspectral data acquisition unit, and the hyperspectral data acquisition unit is a hyperspectral camera mounted on an unmanned aerial vehicle remote sensing platform.

[0015] The beneficial effects of the above technical solution are as follows: The present invention is an exploratory invention. A dataset T1 containing hyperspectral data of each stage of the entire tobacco growth period is established. After determining the diseased tobacco plants with obvious disease characteristics at each stage from T1, the diseased tobacco plants are traced back. The hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage is obtained from the dataset T1, and the hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage is labeled. If the diseased tobacco plants do not have obvious disease characteristics in the previous stage, the hyperspectral data corresponding to the previous stage is marked as pre-diseased, indicating that the tobacco plants are already in the pre-diseased state at this time and will immediately show obvious disease characteristics and enter the diseased state in the next stage; if the diseased tobacco plants have obvious disease characteristics in the previous stage, the hyperspectral data corresponding to the previous stage is marked as diseased, indicating that the tobacco plants are already in the diseased state at this time. According to the above judgment criteria, the hyperspectral data of the diseased tobacco plants in the previous stage is labeled to form a dataset including the hyperspectral data of the diseased and pre-diseased tobacco plants. The labeled dataset is used to train a machine learning model to obtain an early identification and warning model for tobacco diseases. By inputting the hyperspectral data of the tobacco plants in the field into the early identification and warning model for tobacco diseases, the tobacco plants without obvious disease characteristics in the pre-diseased period can be identified, realizing precise prevention and control of the tobacco plants in the field before the occurrence of diseases with disease characteristics that are difficult to distinguish by the naked eye. It can give early warnings before the tobacco plants show disease characteristics, guide tobacco farmers and scientific research personnel to spray pesticides reasonably on the pre-diseased tobacco plants in advance, greatly reducing the amount of drugs used for disease prevention and control, and providing a strong reference for preventing the occurrence of diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of early identification and warning of tobacco diseases in an embodiment of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] In order to make the purpose, technical solution and advantages of the present invention clearer, the specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0018] The present invention is an exploratory invention. A dataset T1 containing hyperspectral data of each stage of the entire tobacco growth period is established. After determining the diseased tobacco plants with obvious disease characteristics at each stage from T1, trace back the diseased tobacco plants, obtain the hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage from the dataset T1, and label the hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage. If the diseased tobacco plants do not have obvious disease characteristics in the previous stage, label the hyperspectral data corresponding to the previous stage as pre-diseased, indicating that the tobacco plants are already in the pre-diseased state at this time and will immediately show obvious disease characteristics and enter the diseased state in the next stage; if the diseased tobacco plants have obvious disease characteristics in the previous stage, label the hyperspectral data corresponding to the previous stage as diseased, indicating that the tobacco plants are already in the diseased state at this time. Label the hyperspectral data of the diseased tobacco plants in the previous stage according to the above judgment criteria to form a dataset including the hyperspectral data of diseased and pre-diseased tobacco plants. Use the labeled dataset to train a machine learning model to obtain an early identification and warning model for tobacco diseases. Input the hyperspectral data of the tobacco plants in the field into the early identification and warning model for tobacco diseases, and the tobacco plants without obvious disease characteristics in the pre-diseased period can be identified, realizing the precise prevention and control of the tobacco plants in the field before the occurrence of diseases with disease characteristics that are difficult to distinguish by the naked eye. It can give early warnings before the tobacco plants show disease characteristics, guide tobacco farmers and researchers to spray pesticides reasonably on the pre-diseased tobacco plants in advance, greatly reduce the dosage of drugs for disease prevention and control, provide a strong reference for preventing the occurrence of diseases, and realize the efficient, non-destructive, fast and real-time detection of the pre-diseased conditions of the tobacco leaves in the field, providing a strong reference for preventing the occurrence of diseases.

[0019] Method Embodiment An early identification and warning method for tobacco diseases according to the present invention, as Figure 1 shown, includes the following processes: 1. Construct a dataset for training an early identification and warning model for tobacco diseases.

[0020] The dataset includes a dataset of hyperspectral data of diseased and pre-diseased tobacco plants. The construction process is as follows: Obtain a dataset T1 of hyperspectral data of each stage of the entire tobacco growth period, screen out the diseased tobacco plants with obvious disease characteristics at each stage from the dataset T1, obtain the hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage, label the hyperspectral data with obvious disease characteristics in the previous stage as diseased, and label the hyperspectral data without obvious disease characteristics in the previous stage as pre-diseased. Further, the dataset used for training the machine learning model during training also includes the hyperspectral data of healthy tobacco plants. Healthy tobacco plants refer to those that do not show obvious disease characteristics during each stage of the entire tobacco growth period. Specifically, it includes the following steps: 1) Obtain the hyperspectral data of each stage of the entire tobacco growth period, and establish a dataset T1 based on the hyperspectral data of each stage of the entire growth period.

[0021] The remote sensing time-series data (i.e., hyperspectral data) of each stage of the whole growth period of tobacco are collected by using a hyperspectral camera mounted on an unmanned aerial vehicle (UAV) remote sensing platform. Preferably, in order to increase the sample data, the remote sensing time-series data of the whole growth period of tobacco in two fields are collected respectively, and the tobacco plants in the two fields include the diseased stage, the pre-diseased stage and healthy tobacco plants. For example, a DJI M350 equipped with a Gaiasky-mini3-VN hyperspectral camera is used to obtain the hyperspectral images of tobacco plants in two fields respectively. The hyperspectral wavelength range is 400 - 1000 nm, the flight altitude is 40 m, the waypoint overlap and the flight line overlap are both 60%, and the resolution is 5.5 nm.

[0022] To improve the prediction accuracy, the hyperspectral data are preprocessed, and the preprocessing includes lens calibration, reflectance calibration and atmospheric correction. The SpecView software can be used to perform lens calibration, reflectance calibration and atmospheric correction on the hyperspectral images. Among them, in the SpecView software, the camera lens is selected to be aligned with the gray board for automatic exposure to perform reflectance calibration, and the matching gray cloth curve is imported in the SpecView software for atmospheric correction of the hyperspectral image data.

[0023] Furthermore, in order to enable large-area disease monitoring, the hyperspectral data are also mosaicked to obtain the hyperspectral image of the complete field after mosaicking. In a preferred embodiment, the preprocessed hyperspectral data are mosaicked, and the tobacco whole-growth-period tobacco leaf growth dataset T1 is established by using the hyperspectral image of the complete field obtained after mosaicking. Specifically, the HiRegistrator software is used to perform geometric precise correction on the preprocessed hyperspectral data, perform rough registration and full-band registration on the data, and then the Photoscan software is used for image mosaicking. The mosaicking process includes: importing the POS information of each preprocessed hyperspectral data, performing aerial triangulation, calculating the dense point cloud, generating the surface model, generating the DEM elevation information, importing the mask image, and generating the orthoimage. Finally, the hyperspectral images of each complete field after mosaicking are obtained, and the dataset T1 is established according to the hyperspectral data of the complete field obtained after mosaicking.

[0024] 2) Select the diseased tobacco plants with obvious disease characteristics at each stage from the dataset T1, and obtain the hyperspectral data corresponding to the diseased tobacco plants at each stage in the previous stage from the dataset T1 to establish a reverse tobacco leaf disease occurrence process dataset T2. Having obvious disease characteristics means that the disease can be discriminated by the human eye.

[0025] 3) According to the principle that the hyperspectral data with obvious disease characteristics in the previous stage is defined as diseased, and the hyperspectral data without obvious disease characteristics in the previous stage is defined as pre-diseased, the hyperspectral data in dataset T2 is image-labeled. The hyperspectral data of healthy tobacco plants that did not show obvious disease characteristics during the whole growth period of tobacco is screened out from dataset T1 and labeled as healthy, forming a dataset including hyperspectral data of three types of tobacco plants: diseased, pre-diseased, and healthy.

[0026] 2. Use the dataset including hyperspectral data of three types of tobacco plants: diseased, pre-diseased, and healthy to train the machine learning model, and obtain an early identification and warning model for tobacco diseases. The training process includes the following steps: 1) Divide the dataset including hyperspectral data of three types of tobacco leaves: diseased, pre-diseased, and healthy into a training set, a test set, and a validation set.

[0027] Specifically, use ENVI software to establish a region of interest ROI, extract the spectral reflectance of 224 bands in the hyperspectral data of three types of tobacco plants: diseased, pre-diseased, and healthy, and form a spectral information database of three types of tobacco plants: diseased, pre-diseased, and healthy. Divide the spectral information database of three types of tobacco plants: diseased, pre-diseased, and healthy into a training set, a test set, and a validation set. In this embodiment, use python software to divide the spectral information database of three types of tobacco plants: diseased, pre-diseased, and healthy into a training set, a test set, and a validation set according to a certain ratio.

[0028] 2) Use a variety of machine learning models to construct an early identification and warning model for tobacco diseases.

[0029] Train a machine learning model using the partitioned dataset to obtain an early identification and warning model for tobacco diseases. The machine learning model can use models such as Random Forest, Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGB), or LightGBM. As a preferred implementation, the machine learning model is a Stacking ensemble learning model (Stack-model) that includes N base learners and one meta-learner, where N≥2. As a preferred implementation, N is 4, and the corresponding base learners respectively use the Random Forest model, the Gradient Boosting Decision Tree model, the Extreme Gradient Boosting model, and the LightGBM model, and the meta-learner uses a multiple linear regression model. In this embodiment, the smote data augmentation algorithm in the python software is used to make the sample size of the training set reach 630, the sample size of the test set reach 270, with a total of 900 samples. Use the Random Forest model, the Gradient Boosting Decision Tree model, the Extreme Gradient Boosting model, the LightGBM model, and the Stacking ensemble learning model to respectively construct an early identification and warning model for tobacco diseases based on unmanned aerial vehicle hyperspectral data. The results are shown in Table 1. The accuracy of the Random Forest model is 83%, the accuracy of the Gradient Boosting Decision Tree model is 81%, the accuracy of the Extreme Gradient Boosting model is 86%, the accuracy of the LightGBM model is 86%, and the accuracy of the Stacking ensemble learning model is 88%. The Stacking ensemble learning model can accurately identify 3 categories of tobacco leaves and can also accurately determine and identify pre-diseased tobacco plants.

[0030] Table 1 Model Accuracy Kappa Recall F1 Random Forest 0.83 0.74 0.83 0.83 GBDT 0.81 0.72 0.81 0.81 XGB 0.86 0.79 0.86 0.86 LGBM 0.86 0.79 0.86 0.86 Stack-model 0.88 0.82 0.88 0.88 Optionally, to improve the recognition accuracy and efficiency, the hyperspectral data used for training the early identification and warning model for tobacco diseases is the extracted sensitive band data. The sensitive bands include the sensitive bands of pre-diseased, diseased, and healthy plants. Use machine learning algorithms to analyze the contribution rate of the reflectance of each band of the tobacco plant hyperspectral data to the discrimination of pre-diseased, diseased, and healthy tobacco plants. Specifically, according to the obtained spectral information database of 3 types of tobacco plants, use the Random Forest model and the Decision Tree model to analyze the contribution rate of 224 bands to the discrimination of 3 types of tobacco plants, rank the spectral reflectance of 224 bands according to the contribution rate, and use the bands with higher contribution rates to the discrimination of each type of tobacco plant as the corresponding sensitive bands. Since the wavelength of 695.65nm has the highest contribution rate to pre-diseased tobacco plants, the wavelength of 695.65nm can be used as the sensitive band of pre-diseased tobacco plants.

[0031] 3. Input the hyperspectral data of the tobacco plants to be measured into the early identification and warning model for tobacco diseases, and the disease identification results can be automatically output. When the early identification and warning model for tobacco diseases is trained using a dataset containing pre-diseased and diseased plants, the output disease identification results are pre-diseased or diseased. When the early identification and warning model for tobacco diseases is trained using a dataset containing healthy, pre-diseased, and diseased plants, the output disease identification results include healthy, pre-diseased, or diseased.

[0032] If the hyperspectral data used for training the early identification and warning model for tobacco diseases is the hyperspectral data of the complete field after splicing, then the hyperspectral data of the tobacco plants to be measured input into the early identification and warning model for tobacco diseases is the hyperspectral data of the complete field obtained by splicing the hyperspectral data of each tobacco plant to be measured. If the hyperspectral data used for training the early identification and warning model for tobacco diseases is the extracted sensitive band data, then the hyperspectral data of the tobacco plants to be measured input into the early identification and warning model for tobacco diseases also corresponds to the extracted sensitive band data. If the hyperspectral data used for training the early identification and warning model for tobacco diseases is the preprocessed data, then the hyperspectral data input into the early identification and warning model for tobacco diseases is also the preprocessed data.

[0033] System embodiment An early identification and warning system for tobacco diseases according to the present invention includes a processor, which is used to execute computer program instructions to implement an early identification and warning method for tobacco diseases. This method is the early identification and warning method described in the above method embodiment and will not be elaborated here.

[0034] The system further includes a hyperspectral data acquisition unit, which is a hyperspectral camera mounted on an unmanned aerial vehicle remote sensing platform and is used to acquire the hyperspectral data of the tobacco plants to be measured.

Claims

1. A method for early identification and warning of tobacco diseases, characterized in that: include: Obtaining hyperspectral data of the tobacco strain to be tested; Input the hyperspectral data of the tobacco plants to be tested into the tobacco disease early identification and warning model, and output the disease identification results; The disease identification results include pre-disease or diseased disease; the early identification and warning model for tobacco diseases is obtained by training the machine learning model using a data set including hyperspectral data of diseased and pre-disease tobacco plants; The process of constructing the dataset is as follows: obtain the dataset T1 of hyperspectral data of each stage of the tobacco growth period, screen out the diseased tobacco plants with obvious disease characteristics in each stage from the dataset T1, obtain the hyperspectral data corresponding to the diseased tobacco plants in each stage in the previous stage, mark the hyperspectral data with obvious disease characteristics in the previous stage as diseased, and mark the hyperspectral data without obvious disease characteristics in the previous stage as pre-disease.

2. The method for early identification and warning of tobacco diseases according to claim 1, characterized in that: The data set used to train the machine learning model also includes hyperspectral data of healthy tobacco plants, which means that tobacco plants do not show obvious signs of disease at any stage of their entire growth period.

3. The method for early identification and warning of tobacco diseases according to claim 2, characterized in that: The hyperspectral data input into the tobacco disease early identification and warning model and used in training the tobacco disease early identification and warning model are the extracted sensitive band data.

4. The method for early identification and warning of tobacco diseases according to claim 3, characterized in that: The sensitive bands include pre-disease, diseased and healthy sensitive bands, which are determined by analyzing the contribution rates of the reflectance of each band of tobacco plant hyperspectral data to the discrimination of pre-disease tobacco plants, diseased tobacco plants and healthy tobacco plants by using a machine learning algorithm.

5. The method for early identification and warning of tobacco diseases according to claim 1, characterized in that: The hyperspectral data is pre-processed data, and the pre-processing includes lens calibration, reflectivity calibration and atmosphere correction.

6. The method for early identification and warning of tobacco diseases according to claim 1, characterized in that: The machine learning model is a Stacking ensemble learning model including N base learners and one meta learner, where N≥2.

7. The method for early identification and warning of tobacco diseases according to claim 6, characterized in that: The N is 4, and the corresponding base learners respectively use the random forest model, the gradient boosting decision tree model, the extreme gradient boosting tree model and the lightweight gradient boosting tree model, and the meta learner uses the multivariate linear regression model.

8. The method for early identification and warning of tobacco diseases according to claim 1, characterized in that: The hyperspectral data of the tobacco plants to be tested input into the tobacco disease early recognition and warning model is the hyperspectral data of the complete field obtained by splicing the collected hyperspectral data of each tobacco plant to be tested.

9. A tobacco disease early identification and warning system, comprising a processor, characterized in that: The processor is used to execute computer program instructions to implement the steps in the tobacco disease early identification and warning method according to any one of claims 1-8.

10. The tobacco disease early identification and warning system according to claim 9, characterized in that: It also includes a hyperspectral data acquisition unit, which is a hyperspectral camera carried on the UAV remote sensing platform.

Images