Method and system for predicting diseases and insect pests of saline-alkaline tolerant rice based on artificial intelligence

Abstract

The invention discloses a method and a system for predicting diseases and insect pests of saline-alkaline tolerant rice based on artificial intelligence. The method comprises the steps of collecting rice data, optimizing the rice data, establishing a prediction model for the diseases and insect pests of the saline-alkaline tolerant rice and predicting the diseases and insect pests of the saline-alkaline tolerant rice. The invention belongs to the technical field of agricultural information, and particularly relates to a method and a system for predicting diseases and insect pests of saline-alkali tolerant rice based on artificial intelligence. According to the scheme, rhombic plot segmentation is dynamically adjusted according to the irregular terrain and the saline-alkali accumulation area boundary of the saline-alkali tolerant rice field, and the collaborative weight is calculated based on the saline-alkali stress degree and the disease and insect pest risk level; key region learning is enhanced in combination with a plot total score and a selection probability formula; a plot score, a neighborhood pest and disease density mean value and a neighborhood saline-alkaline gradient are introduced into a spatial branch, so that the capturing capability of spatial correlation is improved; on the basis of the dynamic adjustment tolerance threshold, the influence of the extreme value is limited through the loss upper limit coefficient, a pest and disease prediction error regulation and control function is constructed, and the pest and disease prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of agricultural information technology, in particular to a salt-tolerant paddy disease and pest prediction method and system based on artificial intelligence. BACKGROUND

[0002] The salt-tolerant paddy disease and pest prediction method refers to a technical means for predicting the risk of future disease and pest occurrence by combining the soil salt and alkali characteristics of salt-tolerant paddy fields, the crop growth state and environmental factors, data collection and analysis and model construction; the core is to associate salt and alkali stress, field environment and disease and pest occurrence rules to provide a basis for prevention and control decisions. However, the general salt-tolerant paddy disease and pest prediction method has the problems of being unable to accurately capture the spatial correlation of diseases and pests, insufficient samples in high-occurrence areas, lack of intermediate transition state samples, insufficient capture of dynamic changes from light to heavy, and thus large prediction error of critical state, and finally low reliability of disease and pest prediction. The general salt-tolerant paddy disease and pest prediction method has the problems of poor capture ability of the nonlinear relationship among salt and alkali stress, growth state and disease and pest, poor data error tolerance in high-salt and alkali areas, prediction results being biased by extreme values, and thus poor disease and pest prediction accuracy. SUMMARY

[0003] In view of the above, in order to overcome the defects of the prior art, the application provides a salt-tolerant rice disease and pest prediction method and system based on artificial intelligence, which aims at the problems that the general salt-tolerant rice disease and pest prediction method cannot accurately capture the spatial correlation of diseases and pests, the sample of high-incidence area is insufficient, the intermediate transition state sample is lacking, the dynamic change from light to heavy is not captured, and the prediction error of critical state is large, finally the reliability of disease and pest prediction is low. According to the irregular terrain and salt and alkali accumulation area boundary of salt-tolerant rice field, the rhombus block segmentation is dynamically adjusted, the disease and pest area, the normal area, the disputed area and the empty area are distinguished, the empty area with all neighboring areas being disease and pest areas is merged, the disease and pest sample of the disputed area with all neighboring areas being normal areas is removed, the field noise filtering is realized, and the invalid disease and pest data interference is reduced. The salt and alkali stress degree and the disease and pest risk grade are calculated to obtain the collaborative weight, the total score of the block and the selection probability formula are combined to preferentially select the disease and pest high-incidence area and the high-salt block sample, and the key area learning is strengthened. Through gradient interpolation, the intermediate state sample is generated based on the low, medium and high risk samples. The reliability of the final disease and pest prediction is improved. In view of the problems that the general salt-tolerant rice disease and pest prediction method has poor capture ability for the nonlinear relationship among salt and alkali stress, growth state and disease and pest, poor error tolerance for high-salt area data, and prediction results are biased by extreme values, which leads to poor disease and pest prediction accuracy, the application designs four feature branches, introduces block scoring, neighborhood disease and pest density average and neighborhood salt and alkali gradient in the spatial branch to improve the capture ability of spatial correlation. Based on the dynamically adjusted tolerance threshold, the influence of extreme values is limited by the loss upper limit coefficient, and then the disease and pest prediction error regulation function is constructed to solve the problem of poor error tolerance for high-salt area data, and prevent abnormal values caused by extreme weather from dominating the prediction. The disease and pest prediction accuracy is improved.

[0004] The technical scheme adopted by the application is as follows: the application provides a salt-tolerant rice disease and pest prediction method based on artificial intelligence, which comprises the following steps:

[0005] Step S1: rice data collection;

[0006] Step S2: rice data optimization;

[0007] Step S3: establishing a salt-tolerant rice disease and pest prediction model;

[0008] Step S4: salt-tolerant rice disease and pest prediction.

[0009] Further, in step S1, the rice data collection is to collect salt-tolerant rice disease and pest data; the salt-tolerant rice disease and pest data includes soil physical and chemical data, field environment data, crop growth data and disease and pest data; the original data is converted into standard normal distribution based on Z-score standardization; and the label is marked as disease and pest risk; and the initial rice disease and pest data set is obtained.

[0010] Furthermore, in step S2, the rice data optimization is to optimize the initial rice pest and disease dataset to obtain a final rice pest and disease dataset; specifically, the following steps are included:

[0011] Step S21: Diamond plot segmentation; the salt-alkali tolerant rice field is regarded as a continuous space to generate a diamond grid: a starting point is selected within the segmentation range as a vertex of the first diamond; the starting point is used as a reference, and the coordinates of the other three vertices of the diamond are determined according to the initial plot size and the angle of the diamond; the adjacent vertex of the first diamond is used as a new starting point, and the above steps are repeated to generate adjacent diamonds in sequence until the entire segmentation range is covered; during the generation process, if the rice field boundary is encountered, the shape and size of the diamond are adjusted. The specific operation is: the diamond plot size is adaptively determined according to the spatial distance characteristics of the pest and disease samples, and the diamond plot size is Expressed as: ;in, is the average minimum spatial distance of all samples; a is the adaptive factor; ; is the average distance between samples of pest and disease occurrence; is the average distance of all samples;

[0012] Step S22: Field noise filtering; Classification of the field into pest and disease areas G according to the sample type in the field pest , normal area G normal , controversial areas with both pests and diseases and normal samples G dis and the empty region G without samples zero If all the neighbors of the empty area are pest-infested areas, the empty area will be merged into the pest-infested area; if all the neighbors of the disputed area are normal areas, the pest-infested samples in the disputed area will be regarded as noise;

[0013] Step S23: Pest and disease sample selection; introduction of salinity-pest and disease synergy weight , expressed as: ;in, is the i-th plot degree of salinity and alkali stress; is the pest and disease risk level of the i-th plot; for the target plot and the four neighboring plots, calculate the total score of the diamond plot , expressed as: ; ; Calculate selection probability , expressed as: ; B is the parameter in the selection probability, ; is the average salinity value of the plot; is the maximum salinity value;

[0014] Step S24: Sample optimization: Randomly select pest and disease samples in the selected plot and its neighborhood, and let Indicates low-risk diseased plants, Indicates medium-risk disease strains, Represents high-risk diseased plants, and generates intermediate state samples through gradient interpolation and , expressed as: ; ;in, and All obey random numbers, which are independent of each other.

[0015] Furthermore, in step S3, the establishment of the salt-alkali tolerant rice pest and disease prediction model is based on the rice pest and disease dataset, extracting features respectively through a multi-branch architecture, and constructing a pest and disease prediction error control function, thereby realizing the establishment of the salt-alkali tolerant rice pest and disease prediction model; specifically, the following contents are included:

[0016] Step S31: Model architecture design; feature branches include soil-environment branch, crop growth branch, time series branch, and space branch; the soil-environment branch processes soil salinity, pH value, humidity, and temperature features through a fully connected network, strengthens the salt × humidity interaction term through the BatchNorm layer, and outputs the environmental feature vector The crop growth branch processes plant height, tiller number, and leaf moisture content characteristics through a fully connected network and outputs a growth feature vector The time series branch processes the past disease index and pest density of historical pest data through the LSTM network and outputs a time series feature vector The spatial branch introduces the spatial characteristics of plot scores, mean density of neighborhood pests and diseases, and neighborhood salinity gradient , is the average salinity of the western neighborhood, is the average salinity value of the eastern neighborhood. The spatial association is extracted through the fully connected layer and the spatial feature vector is output. , expressed as: ; is a fully connected layer; is the mean density of pests and diseases in the neighborhood; the feature vectors of the four branches are weighted and fused, and the weights are adaptively adjusted through training; the prediction head uses a fully connected layer to map the fused features into the future disease index prediction value;

[0017] Step S32: loss function design; construct pest and disease prediction error control function, through The fault tolerance interval tolerates small errors and is calculated by the loss upper limit coefficient Limiting the upper limit of loss; loss function Expressed as: ;

[0018] wherein y is the true label; is the predicted label; is the tolerance threshold, which varies with the salinity stress level , and is expressed as: ; is the adaptive punishment strength; ; is the average value of the true label in the training sample; is the average value of the predicted label in the training sample;

[0019] Step S33: model training; the rice pest and disease data set is divided into a training set, a validation set and a test set; the training set is used for model parameter learning, the validation set is used for monitoring overfitting in the training process, and the test set is used for evaluating the final performance of the model; RAdam is adopted; the disease index prediction MSE of the validation set is used as an index, and after each round of training in the training process, the validation set MSE is calculated, if the index does not decrease for 5 consecutive rounds, the training is stopped; a test threshold is set, if the prediction accuracy of the test set at this time is higher than the test threshold, the salt-tolerant rice pest and disease prediction model is established, otherwise the parameters are adjusted and retrained.

[0020] Further, in step S4, the salt-tolerant rice pest and disease prediction is real-time collection of salt-tolerant rice pest and disease data, which is input into the salt-tolerant rice pest and disease prediction model after standardization processing, and the disease index prediction value corresponding to the pest and disease risk is taken as the pest and disease prediction result.

[0021] The salt-tolerant rice pest and disease prediction system based on artificial intelligence provided by the application comprises a rice data acquisition module, a rice data optimization module, a salt-tolerant rice pest and disease prediction model establishment module and a salt-tolerant rice pest and disease prediction module.

[0022] The rice data acquisition module collects salt-tolerant rice pest and disease data, which is labeled with pest and disease risk labels after Z-score standardization processing, to form an initial rice pest and disease data set.

[0023] The rice data optimization module performs rhombic plot segmentation, field noise filtering, pest and disease sample selection based on rating rules and gradient type deduction sample optimization on the initial rice pest and disease data set to obtain a final rice pest and disease data set.

[0024] The salt-tolerant rice pest and disease prediction model establishment module extracts features and fuses them based on the rice pest and disease data set through a multi-branch architecture, constructs a pest and disease prediction error regulation function, and establishes a salt-tolerant rice pest and disease prediction model through training.

[0025] The salt-tolerant paddy disease and pest prediction module predicts the disease and pest based on the salt-tolerant paddy disease and pest prediction model and the real-time collected salt-tolerant paddy disease and pest data.

[0026] The beneficial effects achieved by the above-mentioned scheme are as follows:

[0027] (1) For the problem that the general salt-tolerant paddy disease and pest prediction method cannot accurately capture the spatial correlation of the disease and pest, the high-incidence area sample is insufficient, the intermediate transition state sample is lacking, the dynamic change from light to heavy is not captured, and the prediction error of the critical state is large, finally the reliability of the disease and pest prediction is low, the scheme dynamically adjusts the rhombus block segmentation according to the irregular terrain and salt and alkali aggregation area boundary of the salt-tolerant paddy field, distinguishes the disease and pest area, the normal area, the disputed area and the empty area, merges the empty area whose neighborhood is all disease and pest areas, removes the disease and pest sample of the disputed area whose neighborhood is all normal areas, realizes the noise filtering in the field, and reduces the invalid disease and pest data interference; the salt and alkali stress degree and the disease and pest risk grade are calculated to obtain the cooperative weight, the total score of the block and the selection probability formula are combined to preferentially select the disease and pest high-incidence area and the high-salt block sample, and the key area learning is strengthened; and through gradient interpolation, the intermediate state sample is generated based on the low, medium and high risk samples; and the reliability of the final disease and pest prediction is improved.

[0028] (2) For the problem that the general salt-tolerant paddy disease and pest prediction method has poor capture ability for the nonlinear relationship among salt and alkali stress, growth state and disease and pest, poor error tolerance for high-salt area data, and prediction results are biased by extreme values, and the disease and pest prediction accuracy is poor, the scheme improves the capture ability of spatial correlation by designing four feature branches and introducing block scoring, neighborhood disease and pest density mean and neighborhood salt and alkali gradient in the spatial branch; based on the dynamic adjustment of the tolerance threshold, the influence of the extreme value is limited by the loss upper limit coefficient, and then the disease and pest prediction error regulation function is constructed, the problem of poor error tolerance for high-salt area data is solved, and the abnormal value caused by extreme weather is prevented from dominating the prediction; and the disease and pest prediction accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 A flowchart of the salt-tolerant paddy disease and pest prediction method based on artificial intelligence provided by the present application is provided.

[0030] Figure 2 A schematic diagram of the salt-tolerant paddy disease and pest prediction system based on artificial intelligence provided by the present application is provided.

[0031] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. DETAILED DESCRIPTION

[0032] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely in the description of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application; based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0033] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the systems or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0034] Embodiment one, refer to Figure 1 The present application provides a salt-tolerant rice pest and disease prediction method based on artificial intelligence, which comprises the following steps:

[0035] Step S1: collecting rice data; collecting salt-tolerant rice pest and disease data, labeling pest and disease risk labels after Z-score standardization processing, and forming an initial rice pest and disease data set;

[0036] Step S2: rice data optimization; performing rhombic plot segmentation, field noise filtering, pest and disease sample selection based on rating rules, and gradient type sample optimization on the initial rice pest and disease data set to obtain a final rice pest and disease data set;

[0037] Step S3: establishing a salt-tolerant rice pest and disease prediction model; based on the rice pest and disease data set, extracting features and fusing through a multi-branch architecture, constructing a pest and disease prediction error control function, and training to establish a salt-tolerant rice pest and disease prediction model;

[0038] Step S4: salt-tolerant rice pest and disease prediction; based on the salt-tolerant rice pest and disease prediction model, real-time collected salt-tolerant rice pest and disease data are predicted for pest and disease.

[0039] Embodiment two, refer to Figure 1The embodiment is based on the above embodiment, in step S1, the rice data collection is collecting salt-tolerant rice pest data; the salt-tolerant rice pest data includes soil physical and chemical data, field environment data, crop growth data and pest data; the soil physical and chemical data includes soil salinity, pH value and quantitative saline-alkali stress degree; the field environment data includes field humidity, daily average temperature and sunshine duration; the crop growth data includes plant height, tiller number and leaf water content; the pest data includes leaf lesion number, pest density and disease index; based on Z-score standardization, the original data is converted to standard normal distribution, and the scale influence is eliminated; and the label is marked as pest risk, including no risk, low risk, medium risk and high risk; an initial rice pest data set is obtained;

[0040] Data acquisition:

[0041] Soil salinity: a portable soil salinity tester is used to collect soil samples at a 5m x 5m grid in a rice field, 0-20cm plough layer soil samples are taken at each point, impurities are removed, distilled water is added, the soil-water ratio is 1:5, stirring is placed for 30 minutes, the electrode is inserted into the supernatant to directly read the salinity value, 30-50 samples are collected from each field to take the average value;

[0042] pH value: a precision pH meter is used to measure the soil extract, each sample is measured repeatedly for 3 times, and the arithmetic mean value is taken as the pH value of the point;

[0043] Quantitative saline-alkali stress degree: the comprehensive stress index is calculated by the formula:

[0044] Field humidity and daily average temperature: based on the temperature and humidity sensor, the air relative humidity is recorded every hour, the average value of 8:00-18:00 of each day is taken as the field humidity of the day; the temperature data of each day is recorded for 24 hours, and the arithmetic mean value is taken as the daily average temperature;

[0045] Sunshine duration: through the sunshine duration sensor, the cumulative duration of solar radiation ≥120W / m² per day is directly recorded;

[0046] Plant height: 10 rice plants are randomly selected from each sampling plot, the vertical distance from the stem base to the highest leaf tip is measured with a tape measure, and the average value is taken;

[0047] Tiller number: the above 10 rice plants are counted one by one, the tiller with stem diameter ≥0.5cm and complete leaf is effective tiller, and the average tiller number per plant is taken;

[0048] Leaf water content: 3 healthy functional leaves are picked, the fresh weight Wf is weighed with an electronic balance, after blanching at 105℃ for 30 minutes, the dry weight Wd is weighed after drying at 80℃ to constant weight, and the leaf water content is calculated ​​

[0049] Leaf lesion number: Unmanned aerial vehicle inspection was used, carrying a 50 million pixel camera to shoot field images from 10:00 to 14:00 on sunny days. The image processing software OpenCV was used to automatically identify and count the lesions. At the same time, 20 rice plants were randomly selected, and 3 functional leaves of each plant were investigated. The number of lesions was counted manually for calibration, and the deviation between the two was controlled within 10%;

[0050] Pest density: Five-point sampling method was used, and one insect trap was placed at each point. The number of pests was counted every morning, and the pest density per unit area was calculated.

[0051] Disease index: According to the proportion of lesion coverage area to total leaf area, the disease index was classified as follows: no risk corresponds to no lesion, low risk corresponds to 0-20%, medium risk corresponds to 20-50%, and high risk corresponds to more than 50%.

[0052] Example three, see Figure 1 This example is based on the above example, in step S2, the rice data optimization is performed on the initial rice pest data set to obtain the final rice pest data set; specific contents include the following:

[0053] Step S21: diamond plot segmentation; the salt-tolerant rice field is regarded as a continuous space, and the occurrence of pests and diseases has spatial aggregation. The humidity is high in low-lying areas, and pests and diseases are prone to outbreak. The salt-tolerant rice field often forms irregular plots due to irrigation, topography and other factors. The flexibility of the diamond is higher, and the size and angle of the plot can be adjusted to better fit the complex topography of the low-lying and high-slope fields. Especially for irregular salt and alkali distribution areas formed by salt and alkali aggregation, diamond segmentation can more accurately delineate plots with similar salt and alkali stress levels. Generate a diamond grid: select a starting point as one vertex of the first diamond in the segmentation range; based on the starting point, determine the coordinates of the other three vertices of the diamond according to the initial plot size and the angle of the diamond; take the adjacent vertex of the first diamond as a new starting point, repeat the above steps, and generate adjacent diamonds in turn until the entire segmentation range is covered. During the generation process, if the boundary of the rice field is encountered, the shape and size of the diamond are adjusted to ensure that the diamond grid fits the plot boundary; the specific operation is as follows: the size of the diamond plot is determined adaptively according to the spatial distance characteristics of the disease and pest samples to ensure that each diamond plot contains samples with similar disease and pest risks. The size of the diamond plot is represented as: ; wherein, is the average minimum spatial distance of all samples; a is the adaptive factor; ; is the average distance between disease and pest occurrence samples; is the average distance of all samples, ensuring that the plot size matches the aggregation degree of the disease and pests;

[0054] Step S22: Field noise filtering: The field data of saline-alkali land has spatial heterogeneity caused by the accumulation of salt and alkali in low-lying areas and error noise caused by salt corrosion of sensors. It is necessary to improve the data quality through plot segmentation and sample screening. The specific operation is: divide the samples into pest and disease area G according to the type of samples in the plot pest , normal area G normal , controversial areas with both pests and diseases and normal samples G dis and the empty region G without samples zero If all the neighbors of the empty area are pest and disease areas, there is a risk of missed detection, and the empty area is merged into the pest and disease area; if the neighbors of the disputed area are all normal areas, there is a risk of misjudgment, and the pest and disease samples in the disputed area are regarded as noise;

[0055] Step S23: Pest and disease sample selection: In pest and disease prediction, high-value samples from clearly defined high-incidence areas should be prioritized to reduce the impact of low-value samples from disputed areas with ambiguous edges. The value of plots should be quantified through a scoring mechanism to provide a probabilistic basis for sample selection.

[0056] The specific operation is: define the scoring rules: introduce the salt-alkali-pest and disease-injury synergy weight , expressed as: ;in, is the i-th plot degree of salinity and alkali stress; is the pest and disease risk level of the i-th plot; for the target plot and the four neighboring plots, calculate the total score of the diamond plot , expressed as: ; ; Calculate selection probability , expressed as: ; The higher the score, the higher the probability of the diamond plot being selected, ensuring that the synthetic samples are concentrated in the areas with high incidence of pests and diseases; B is the parameter in the selection probability, ; is the average salinity value of the plot; is the maximum salinity value; the higher the salinity, the larger B is, and the higher the probability that the high-salinity plot sample will be selected;

[0057] Step S24: Sample optimization: Randomly select pest and disease samples in the selected plot and its neighborhood, and let Indicates low-risk diseased plants, Indicates medium-risk disease strains, Represents high-risk diseased plants, and uses gradient interpolation to simulate the gradual change of pests and diseases from mild to severe, generating intermediate state samples and , expressed as: ; ;in, and are subject to random numbers, independent of each other; the labels of the intermediate state samples are based on the proportion of the average lesion coverage area of the parent to the total area of the leaf for grading annotation.

[0058] By performing the above operations, the general salt-tolerant rice disease and pest prediction method cannot accurately capture the spatial correlation of diseases and pests, the sample in the high-occurrence area is insufficient, there is a lack of intermediate transition state samples, the dynamic changes from light to heavy are not captured, which leads to large prediction error of the critical state, and finally the reliability of the disease and pest prediction is low. According to the irregular terrain and salt and alkali aggregation area boundary of the salt-tolerant rice field, the rhombus block segmentation is dynamically adjusted, the disease and pest area, the normal area, the disputed area, and the empty area are distinguished, the empty area with all neighboring areas being disease and pest areas is merged, the disease and pest samples in the disputed area with all neighboring areas being normal areas are removed, the field noise filtering is realized, and the invalid disease and pest data interference is reduced. Based on the salt and alkali stress degree and the disease and pest risk level calculation cooperative weight, the total score and the selection probability formula are combined, the disease and pest high-occurrence area and the high-salt block sample are preferentially selected, and the key area learning is strengthened. And through gradient interpolation, the intermediate state samples are generated based on low, medium and high risk samples. Thus, the reliability of the final disease and pest prediction is improved.

[0059] Embodiment four, refer to Figure 1 This embodiment is based on the above-mentioned embodiment. In step S3, the salt-tolerant rice disease and pest prediction model is established based on the rice disease and pest data set, the features are extracted through the multi-branch architecture, the nonlinear relationship of salt and alkali stress-growth state-disease and pest is learned and fused, and a disease and pest prediction error regulation function is constructed, thereby realizing the establishment of the salt-tolerant rice disease and pest prediction model. Specifically, the following contents are included:

[0060] Step S31: model architecture design; the feature branches include soil-environment branch, crop growth branch, time sequence branch and space branch; the soil-environment branch processes the soil salt, pH value, humidity and temperature features through the fully connected network, and strengthens the salt x humidity interaction item through the BatchNorm layer. High salt + high humidity will accelerate the spread of diseases. The output environment feature vector is represented as: ; is a Sigmoid activation function; W and b are the weight matrix and bias term of the fully connected network; salt and hum represent the features of soil salt and humidity; the crop growth branch processes the plant height, tiller number and leaf water content features through the fully connected network, and focuses on capturing the growth index-resistance correlation. Plants with fewer tillers are more susceptible to pests. The output growth feature vector ; the time sequence branch processes the past disease index and pest density of the historical disease and pest data through the LSTM network, captures the dynamic trend, and the pest density increases exponentially after rain. The output time sequence feature vector ; the spatial branch introduces the plot score, the average spatial feature of the neighborhood pest and disease density, and the neighborhood salinity gradient , improves the spatial correlation capture ability, is the average salinity value of the west neighborhood, is the average salinity value of the east neighborhood, the spatial correlation is extracted through the fully connected layer, the plot pest and disease is significantly affected by the west neighborhood, and the spatial feature vector is output, which is expressed as: ; is the fully connected layer; is the average neighborhood pest and disease density; the feature vectors of the four branches are fused by weighting, the weight is adaptively adjusted through training, and the gradient correlation of soil salinity rising → leaf water content decreasing → disease index rising is emphasized; the prediction head adopts a fully connected layer to map the fused features to the future disease index prediction value; which is expressed as: ; is the disease index prediction value in the next 30 days; and are the weights and biases of the prediction head; the pest and disease risk is obtained according to the prediction value; the prediction value equal to 0 corresponds to no risk; the prediction value belongs to 0-20 corresponds to low risk; the prediction value belongs to 20-50 corresponds to medium risk; the prediction value greater than 50 corresponds to high risk;

[0061] Step S32: loss function design; there are two typical problems in salt-tolerant rice pest and disease data: one is the small error caused by manual counting of disease spots, and the other is the abnormal value caused by sudden insect pests caused by extreme weather; the traditional loss function example least square loss is sensitive to small error or cannot suppress abnormal value influence, such as SVR's ε-insensitive loss; therefore, a pest and disease prediction error control function is constructed, which tolerates small error through the fault tolerance interval, limits the loss upper limit through the loss upper limit coefficient , and suppresses the abnormal value disturbance caused by extreme weather;

[0062] The loss function is expressed as: ;

[0063] Wherein, y is the true label; is the predicted label; is the tolerance threshold, which changes with the salt stress level , which is expressed as: , the high salinity area increases to 5, which tolerates slightly larger measurement errors caused by salt interference; is the adaptive punishment strength; ; is the average value of the true label in the training sample; is the average value of the predicted label in the training sample;

[0064] Step S33: model training; the rice pest and disease data set is divided into a training set, a validation set and a test set; the training set is used for model parameter learning, the validation set is used for monitoring overfitting in the training process, and the test set is used for evaluating the final performance of the model; RAdam is adopted, the initial learning rate is 0.001; the disease index prediction MSE of the validation set is taken as an index, and the training is stopped after 5 rounds of continuous decline to prevent overfitting; and the model is compressed through parameter pruning; the prediction MSE of the validation set is taken as a core monitoring index, and the validation set MSE is calculated after each round of training in the training process; if the index does not decrease for 5 rounds in a row, the training is stopped to prevent the model from overfitting the training data; a test threshold is set, if the prediction accuracy of the test set at this time is higher than the test threshold, the salt-tolerant rice pest and disease prediction model is established, otherwise the parameters are adjusted and retrained.

[0065] By performing the above operation, the general salt-tolerant rice pest and disease prediction method has poor ability to capture the nonlinear relationship between salt stress, growth state and pest and disease, poor error tolerance for high-salt area data, and prediction results are biased by extreme values, which leads to poor pest and disease prediction accuracy. The present scheme improves the capture ability of spatial correlation by designing 4 feature branches and introducing plot scores, neighborhood pest and disease density averages and neighborhood salt gradient in the spatial branch; based on the dynamic adjustment of the tolerance threshold, the influence of extreme values is limited by the loss upper limit coefficient, and then the pest and disease prediction error control function is constructed, which solves the problem of poor error tolerance for high-salt area data, prevents abnormal values caused by extreme weather from dominating the prediction, and improves the pest and disease prediction accuracy.

[0066] Embodiment five, refer to Figure 1 This embodiment is based on the above-mentioned embodiment. In step S4, the salt-tolerant rice pest and disease prediction is real-time collection of salt-tolerant rice pest and disease data, which is input into the salt-tolerant rice pest and disease prediction model after standardization processing, and the obtained future 30-day corresponding pest and disease risk is taken as the pest and disease prediction result; if there is a medium risk and a high risk, the management personnel are given early warning treatment.

[0067] Embodiment six, refer to Figure 2 This embodiment is based on the above-mentioned embodiment. The present application provides a salt-tolerant rice pest and disease prediction system based on artificial intelligence, which comprises a rice data acquisition module, a rice data optimization module, a salt-tolerant rice pest and disease prediction model establishment module and a salt-tolerant rice pest and disease prediction module.

[0068] The rice data acquisition module acquires salt-tolerant rice pest and disease data, which is labeled with pest and disease risk tags after Z-score standardization processing to form an initial rice pest and disease data set.

[0069] The rice data optimization module performs rhombus plot segmentation, field noise filtering, pest and disease sample selection based on rating rules, and gradient deduction sample optimization on the initial rice pest and disease data set to obtain a final rice pest and disease data set.

[0070] The salt-tolerant rice pest and disease prediction model establishment module extracts features and fuses them based on the rice pest and disease data set through a multi-branch architecture, constructs a pest and disease prediction error control function, and establishes a salt-tolerant rice pest and disease prediction model through training.

[0071] The salt-tolerant rice pest and disease prediction module predicts pests and diseases based on the salt-tolerant rice pest and disease prediction model and real-time collected salt-tolerant rice pest and disease data.

[0072] It should be noted that, in this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0073] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application.

[0074] The above describes the present application and its embodiments, which are not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed and belong to the protection scope of the present application.

Claims

1. A salt-tolerant water rice disease and pest prediction method based on artificial intelligence, characterized in that: The method comprises the following steps: Step S1: rice data collection; collecting salt-tolerant rice pest and disease data, marking the pest and disease risk label after Z-score standardization processing, and forming an initial rice pest and disease data set; Step S2: rice data optimization; the initial rice pest and disease data set is subjected to rhombus plot segmentation, field noise filtering, pest and disease sample selection based on rating rules, and gradient type sample optimization, to obtain a final rice pest and disease data set; Step S3: establishing a salt-tolerant rice pest and disease prediction model; based on the rice pest and disease data set, features are extracted and fused through a multi-branch architecture, a pest and disease prediction error control function is constructed, and a salt-tolerant rice pest and disease prediction model is established after training; Step S4: salt-tolerant rice pest and disease prediction; based on the salt-tolerant rice pest and disease prediction model, real-time collected salt-tolerant rice pest and disease data are subjected to pest and disease prediction. 2.The salt-tolerant and alkali-tolerant rice disease and pest prediction method based on artificial intelligence according to claim 1, characterized in that: In step S2, the rice data optimization is rice data optimization on the initial rice pest and disease data set to obtain a final rice pest and disease data set; specifically including the following contents: Step S21: rhombus plot segmentation; taking the saline-alkaline rice field as a continuous space, a rhombus grid is generated: selecting a starting point as a vertex of the first rhombus in the segmentation range; taking the starting point as a reference, determining the coordinates of the other three vertices of the rhombus according to the initial plot size and the angle of the rhombus; taking the adjacent vertex of the first rhombus as a new starting point, repeating the above steps to generate adjacent rhombuses in turn until the entire segmentation range is covered; in the generation process, if the rice field boundary is encountered, the shape and size of the rhombus are adjusted, and the specific operation is: the rhombus plot size is adaptively determined according to the spatial distance characteristics of the pest sample, and the rhombus plot size is represented as: ; wherein, is the average minimum spatial distance of all samples; a is an adaptive factor; ; is the average distance between pest occurrence samples; is the average distance of all samples; Step S22: field noise filtering; Step S23: pest and disease sample selection; Step S24: sample optimization. 3.The salt-tolerant and alkali-tolerant rice disease and pest prediction method based on artificial intelligence according to claim 2, characterized in that: In step S2, the field noise filtering is according to the sample type in the plot into the pest and disease area G pest , normal area G normal , controversial area G dis with both pest and disease samples and normal samples, and empty area G zero without samples; if all the neighborhoods of the empty area are the pest and disease area, the empty area is merged into the pest and disease area; if the neighborhoods of the controversial area are all normal areas, the pest and disease samples in the controversial area are regarded as noise.

4. The salt-tolerant and alkali-tolerant rice disease and pest prediction method based on artificial intelligence according to claim 3, characterized in that: In step S2, the disease and pest sample selection is introduced with a salt-alkali-disease and pest coordination weight , is expressed as: ; wherein, is the salt-alkali stress degree of the ith plot; is the disease and pest risk level of the ith plot; for the target plot and the four neighbors, the total score of the diamond plot is calculated , is expressed as: ; ; the selection probability is calculated , is expressed as: ; B is a parameter in the selection probability, ; is the average salt value of the plot; is the maximum salt value.​ 5.The salt-tolerant and alkali-tolerant rice disease and pest prediction method based on artificial intelligence according to claim 4, characterized in that: In step S2, the sample optimization is to randomly select disease and pest samples in the selected plot and neighborhood, and let represent low-risk disease strains, represent medium-risk disease strains, represent high-risk disease strains, and intermediate state samples are generated by gradient interpolation and , represented as: ; ; wherein, and are random numbers subject to and are independent of each other.

6. The artificial intelligence-based salt-alkali-tolerant rice pest prediction method according to claim 5, characterized in that: In step S3, the establishment of the salt-tolerant rice pest and disease prediction model is based on the rice pest and disease data set, features are extracted through a multi-branch architecture respectively, and a pest and disease prediction error control function is constructed, thereby realizing the establishment of the salt-tolerant rice pest and disease prediction model; specifically including the following contents: Step S31: model architecture design; feature branches include soil-environment branch, crop growth branch, time sequence branch and space branch; the soil-environment branch processes soil salinity, pH value, humidity and temperature features through a fully connected network, and strengthens the salt x humidity interaction item through a BatchNorm layer, and outputs an environment feature vector ; the crop growth branch processes plant height, tillering number and leaf water content features through a fully connected network, and outputs a growth feature vector ; the time sequence branch processes past disease index and pest density of historical pest data through an LSTM network, and outputs a time sequence feature vector ; the space branch introduces a plot score, a neighbor pest density mean space feature, and a neighbor salinity gradient , , which is an average salinity value of the west neighbor, , which is an average salinity value of the east neighbor, extracts spatial correlation through a fully connected layer, and outputs a spatial feature vector , which is expressed as: ; , which is a fully connected layer; , which is a neighbor pest density mean; the feature vectors of the four branches are fused with weights, and the weights are adaptively adjusted through training; a prediction head adopts a fully connected layer to map the fused features to a future disease index prediction value; Step S32: loss function design; Step S33: model training; the rice pest and disease data set is divided into a training set, a validation set and a test set; the training set is used for model parameter learning, the validation set is used for monitoring overfitting in the training process, and the test set is used for evaluating the final performance of the model; RAdam is adopted; the disease index prediction MSE of the validation set is taken as an index, after each round of training in the training process, the validation set MSE is calculated, if the index does not decrease for 5 consecutive rounds, the training is stopped; a test threshold is set, if the prediction accuracy on the test set at this time is higher than the test threshold, the salt-tolerant rice pest and disease prediction model is established, otherwise the parameters are adjusted and retrained. 7.The salt-tolerant and alkali-tolerant rice disease and pest prediction method based on artificial intelligence according to claim 6, characterized in that: In step S3, the loss function design is to construct a pest prediction error regulation function, through The fault tolerance interval tolerates small errors, and the loss upper limit coefficient limits the loss upper limit; Loss function is represented as: ; where y is the true label; is the predicted label; is the tolerance threshold, which varies with the salinity stress level is expressed as: ; is the adaptive punishment strength; ; is the average value of the true label in the training sample; is the average value of the predicted label in the training sample. 8.The salt-tolerant and alkali-tolerant rice disease and pest prediction method based on artificial intelligence according to claim 7, characterized in that: In step S1, the rice data collection is the collection of salt-tolerant rice pest and disease data; the salt-tolerant rice pest and disease data includes soil physicochemical data, field environment data, crop growth data and pest and disease data; the original data is converted to a standard normal distribution based on Z-score standardization; and the label is marked as pest and disease risk; an initial rice pest and disease data set is obtained.

9. The artificial intelligence-based salt-alkali-tolerant rice pest and disease prediction method according to claim 8, characterized in that: In step S4, the salt-tolerant rice pest and disease prediction is the real-time collection of salt-tolerant rice pest and disease data, which is input into the salt-tolerant rice pest and disease prediction model after standardization processing, and the pest and disease risk corresponding to the disease index prediction value obtained is taken as the pest and disease prediction result.

10. The AI-based salt-alkali tolerant paddy disease and pest prediction system for implementing the AI-based salt-alkali tolerant paddy disease and pest prediction method according to any one of claims 1-9, characterized in that: It comprises a rice data collection module, a rice data optimization module, a salt-tolerant rice pest and disease prediction model establishment module and a salt-tolerant rice pest and disease prediction module; The rice data collection module collects salt-tolerant rice pest data, and labels the pest risk label after Z-score standardization processing to form an initial rice pest data set; The rice data optimization module performs rhombus plot segmentation, field noise filtering, pest sample selection based on rating rules, and gradient-based sample optimization on the initial rice pest data set to obtain a final rice pest data set; The salt-tolerant rice pest prediction model establishment module establishes a salt-tolerant rice pest prediction model based on the rice pest data set, extracts features through a multi-branch architecture, fuses them, constructs a pest prediction error control function, and trains the model. The salt-tolerant rice pest prediction module predicts the pests based on the salt-tolerant rice pest prediction model and real-time collected salt-tolerant rice pest data.

Images