An orchard disease identification and water, fertilizer and pesticide intelligent decision-making method based on 5G Internet of Things

Abstract

The application discloses a kind of orchard disease identification and water fertilizer medicine intelligent decision-making method based on 5G Internet of Things, belong to orchard wisdom planting technical field.This method includes: collecting fruit tree canopy visible light image, soil parameter and meteorological environmental data, and adopting superpixel pre-segmentation is combined with shape constraint extraction branch leaf fruit organ area, combined with fruit tree growth period health feature library and mahalanobis distance screening suspected abnormal area;Rely on D-S evidence theory fusion multi-feature determination suspected abnormal area's disease and pest attribution type, and determine the disease and pest species of suspected abnormal area;Combined with single organ damage degree and whole plant damage organ proportion determined disease and pest severity grade, based on expert rule generation water fertilizer medicine basic strategy, and introduce machine learning correction model to water fertilizer medicine basic strategy is self-adapting correction, output water fertilizer medicine intelligent decision-making and issue execution.The application disease and pest identification precision is high, can realize orchard disease and pest precision identification and water fertilizer medicine variable intelligent control.

Description

Technical Field

[0001] This invention relates to the field of smart orchard planting technology, specifically to a method for orchard disease identification and intelligent decision-making regarding water, fertilizer, and pesticide application based on 5G Internet of Things. Background Technology

[0002] With the rapid development of smart agriculture technology, IoT-based refined management of orchards has become an important direction for improving fruit yield and quality and reducing production inputs. In orchard production management, accurate identification of pests and diseases and coordinated decision-making regarding water, fertilizer, and pesticide application are core aspects. Traditional methods relying on manual inspections and experience-based application of pesticides and fertilizers suffer from low efficiency, insufficient control precision, and the risk of excessive application of pesticides and fertilizers, making them unsuitable for the management needs of large-scale orchards. Existing technologies have seen the emergence of various orchard management solutions combining sensing devices and intelligent control. For example, Chinese invention patent application CN120712977A discloses an intelligent orchard control irrigation, fertilization, and spraying method and system. This system generates irrigation, fertilization, and spraying plans by collecting data on fruit tree pests, plants, and soil. It also determines the compatibility of mixing fertilizers and pesticides based on their chemical properties and physical states, enabling targeted fertilization and spraying, thus improving the automation level of orchard management to a certain extent.

[0003] However, the aforementioned existing technologies still have many shortcomings: In terms of disease and pest identification, they do not perform precise segmentation of fruit tree organs and matching of health benchmarks according to growth stages, but directly identify symptoms based on the entire image. This is easily affected by complex field backgrounds, changes in light, and differences in fruit tree growth stages, resulting in insufficient accuracy and stability. Furthermore, they lack a hierarchical identification logic for diseases and pests, and lack an effective secondary judgment mechanism for atypical abnormal areas, easily leading to confusion and misjudgment of diseases and pests. At the level of water, fertilizer, and pesticide decision-making, existing solutions mostly generate control strategies based on preset rules and fixed models, failing to adaptively optimize and correct adjustable parameters of the strategies by incorporating real-time soil and meteorological environmental parameters. This limits the adaptability of the control solutions to the field environment, making it difficult to achieve truly precise variable control. Summary of the Invention

[0004] The purpose of this invention is to provide a method for orchard disease identification and intelligent decision-making on water, fertilizer and pesticides based on 5G Internet of Things, so as to solve the technical defects of existing orchard disease and pest identification, such as low accuracy, easy confusion of diseases and pests, and inability of water, fertilizer and pesticide decisions to adapt to field environment as mentioned in the background.

[0005] To achieve the above objectives, this invention provides a method for orchard disease identification and intelligent decision-making regarding water, fertilizer, and pesticide application based on 5G Internet of Things, comprising the following steps: The field sensing nodes deployed by 5G IoT collect visible light images of the fruit tree canopy, and simultaneously collect soil parameters and meteorological environmental data. The visible light image of the fruit tree canopy is pre-segmented by superpixel and classified by multi-color space features. Combined with the morphological constraints of the fruit tree organs, organ instance identification is completed, and image blocks of leaves, fruits and branches are extracted to form an organ region set. Determine the current fruit tree's growth stage, match it with the standard feature library of healthy organs corresponding to the fruit tree's growth stage, and use Mahalanobis distance to measure the feature deviation of each organ region to screen and obtain a feature set of suspected abnormal regions. Based on a pre-constructed knowledge base of disease and pest symptom mechanisms, the DS evidence theory is used to perform multi-feature evidence fusion on each suspected abnormal area to determine the disease confidence and pest confidence of each suspected abnormal area, and to determine the disease and pest classification of each suspected abnormal area accordingly. The disease and pest classification includes disease type, pest type and uncertain type. For suspected abnormal areas of diseases and pests, weighted feature similarity matching is performed with the corresponding sub-databases in the disease and pest symptom mechanism knowledge base to obtain the types of diseases and pests in the suspected abnormal areas; For suspected anomalous regions of uncertain class, frequency domain feature analysis is used to complete a secondary determination to obtain the types of pests and diseases in the suspected anomalous regions; Based on the types of pests and diseases in each suspected abnormal area, and combined with the severity level of pests and diseases determined by the degree of damage to a single organ and the proportion of damaged organs in the whole plant, a pest and disease identification result is constructed. By combining soil parameters and meteorological environmental data, the basic water, fertilizer and pesticide management strategy determined based on the results of pest and disease identification and the fruit tree growth period is adaptively modified to obtain intelligent water, fertilizer and pesticide decisions and issue them for execution.

[0006] Preferably, the method for superpixel pre-segmentation and multi-color space feature classification is as follows: The SLIC algorithm was used to divide the visible light image of the fruit tree canopy into superpixels, resulting in several superpixel blocks. Each superpixel block was then converted to the Lab and YCbCr color spaces, and the multi-channel color features of each superpixel block were extracted. A classification criterion is constructed based on the prior color of fruit tree organs. A Gaussian mixture model is then used to cluster and classify the color features of all superpixel blocks to determine the region type of each superpixel block. The region types include leaf region, fruit region, branch region, and background region.

[0007] Preferably, the method for identifying organ instances is as follows: By introducing morphological prior constraints corresponding to leaves, fruits and branches, and performing region growth and merging on superpixel blocks of the same type, preliminary organ connectivity regions are obtained. For preliminary organ connectivity regions without overlap or adhesion, the organ regions are directly retained. For preliminary organ connectivity regions with overlap or adhesion, the distance transformation combined with the marker watershed algorithm is used to split the regions into instances and separate them into organ regions. Based on all organ regions, corresponding organ mask images are generated. The visible light images of the fruit tree canopy are cropped according to each organ mask image, and image blocks of leaves, fruits and branches are extracted respectively. All organ image blocks are integrated to form an organ region set.

[0008] Preferably, the method for measuring the feature deviation of each organ region using Mahalanobis distance is as follows: A sliding window is used to traverse the organ regions, and the color and texture features of each window are extracted to construct a multi-dimensional window feature vector. Calculate the Mahalanobis distance between the multidimensional window feature vector and the baseline feature corresponding to the organ category to which the organ region belongs in the standard feature library of healthy organs; Windows whose Mahalanobis distance exceeds a preset deviation threshold are marked as abnormal windows. Connected abnormal windows are merged to obtain suspected abnormal sub-regions. Multidimensional features of each suspected abnormal sub-region are extracted to form a suspected abnormal region feature set. The multidimensional features include color features, texture features, edge features, and internal uniformity features.

[0009] Preferably, the method for determining the confidence levels for diseases and pests is as follows: The multidimensional features of each suspected abnormal region are used as multidimensional identification features; Each identification feature is used as an independent piece of evidence. Typical features in the disease and pest symptom mechanism knowledge base are matched to calculate the basic probability allocation of diseases and pests corresponding to each independent piece of evidence. The basic probability assignments of all independent evidence were fused using Dempster's combination rule to obtain the disease and pest confidence scores for each suspected anomalous area.

[0010] Preferably, the rule for determining the type of pest or disease is as follows: If the confidence level of pests is not less than the preset high confidence threshold and the difference between the confidence level of pests and the confidence level of diseases is not less than the preset confidence difference threshold, then the suspected abnormal area is determined to be a pest-related area. If the confidence level of the disease is not less than the preset high confidence threshold and the difference between the confidence level of the disease and the confidence level of the pest is not less than the preset confidence difference threshold, then the suspected abnormal area is determined to be a disease. Otherwise, the suspected abnormal area is classified as uncertain.

[0011] Preferably, the suspected abnormal regions of uncertain class are further determined through frequency domain feature analysis, specifically as follows: The suspected abnormal regions of the uncertain class are converted into grayscale images. A two-dimensional discrete Fourier transform is performed on the grayscale images and the spectrum is centered to obtain the amplitude spectrum. The amplitude spectrum is divided into annular partitions, and the energy proportions of low, medium and high frequency regions are calculated. At the same time, the radial distribution entropy and angular distribution entropy of the spectrum are extracted, and frequency domain features are constructed accordingly. The cosine similarity between the frequency domain features and the pre-constructed standard spectrum templates for diseases and pests is calculated. The confidence scores for diseases and pests are then updated. Based on the updated confidence scores for diseases and pests, the types of diseases and pests corresponding to the suspected abnormal area are determined.

[0012] Preferably, the method for determining the severity level of the pests and diseases is as follows: Based on the identified pests and diseases, calculate the proportion of the total area of ​​the suspected abnormal area corresponding to the pest or disease to the total area of ​​the affected organ, and divide the single organ damage level according to the preset proportion range. The proportion of affected organ areas within the statistical monitoring range to the total number of organ areas is used to calculate the overall severity index of the plant, which is then combined with the severity level of individual organ damage. Based on this overall severity index, the severity level of the pest or disease is determined.

[0013] Preferably, the method for obtaining the intelligent water, fertilizer, and pesticide decision-making is as follows: based on the pest and disease identification results and the fruit tree growth stage, a basic water, fertilizer, and pesticide management strategy is generated by matching a preset expert rule base. Each adjustable parameter in the basic water, fertilizer, and pesticide management strategy is used as a variable to be optimized. Simultaneously, real-time collected soil parameters and meteorological environmental parameters are used as environmental feature vectors, which are input into a pre-trained machine learning correction model for adaptive correction. The resulting intelligent water, fertilizer, and pesticide decision-making is adapted to the current fruit tree growth status, pest and disease situation, and field environment. Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses SLIC superpixel segmentation combined with morphological constraints to separate fruit tree organ instances. It also uses a health feature database based on growth stage and Mahalanobis distance to screen abnormal areas, thus solving the problem of traditional whole-image recognition being easily affected by background interference, and making the location of pest and disease areas more accurate.

[0014] 2. Based on the DS evidence theory, multi-feature fusion is used to distinguish between diseases, pests, and uncertain areas. For fuzzy samples, frequency domain features are added for secondary discrimination, which greatly reduces the probability of confusion and misjudgment of diseases and pests, and makes the identification more robust.

[0015] 3. By adopting a machine learning correction model that integrates soil and meteorological data, the adjustable parameters of water, fertilizer and pesticide strategies are adaptively corrected. Compared with fixed rule schemes, the management scheme is more in line with the real-time field environment, realizing precise intelligent decision-making on water, fertilizer and pesticides. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0017] The following is combined Figure 1 The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] like Figure 1 As shown in the figure, this embodiment provides a method for orchard disease identification and intelligent decision-making on water, fertilizer, and pesticides based on 5G Internet of Things. The method is executed sequentially according to the following steps: data collection, organ region extraction, suspected abnormal area screening, disease and pest attribution determination, disease and pest type identification, severity level assessment, and intelligent decision generation on water, fertilizer, and pesticides. The specific implementation steps are as follows: S1. Field sensing nodes deployed via 5G IoT are used to collect visible light images of the fruit tree canopy, simultaneously collecting soil parameters and meteorological environmental data. Specifically, field sensing nodes are deployed at preset intervals within the target orchard, each equipped with a 5G communication module and connected to the orchard's 5G IoT system. Each field sensing node includes a canopy visual acquisition unit, a soil multi-parameter sensing unit, and a meteorological environmental monitoring unit. The canopy visual acquisition unit collects visible light images of the fruit tree canopy at a set resolution, with the acquisition angle and frequency preset according to the canopy morphology. The soil multi-parameter sensing unit collects soil parameters from the root zone of the fruit trees, including at least soil moisture content, effective nitrogen, phosphorus, and potassium content, and soil pH. The meteorological environmental monitoring unit collects meteorological environmental data from the surrounding area of ​​the canopy, including air temperature and humidity, light intensity, and wind speed. All data collected through field sensing nodes is synchronously uploaded to the cloud processing platform via 5G communication modules. The cloud processing platform is responsible for processing subsequent image analysis, pest and disease identification, and water, fertilizer and pesticide decision-making calculations for S2 to S7. During the synchronization process, a unified timestamp is added to each data through a 5G edge gateway to achieve time alignment of fruit tree canopy visible light images, soil parameters and meteorological environmental data.

[0019] S2. Perform superpixel pre-segmentation and multi-color space feature classification on the visible light image of the fruit tree canopy, and complete organ instance labeling by combining fruit tree organ morphology constraints, extracting image patches of leaves, fruits, and branches to form organ region sets. This step can be specifically divided into two stages: superpixel pre-segmentation and classification, and organ instance labeling.

[0020] The first stage is superpixel pre-segmentation and classification, namely superpixel pre-segmentation and multi-color space feature classification. Specifically, the SLIC (Simple Linear Iterative Clustering) algorithm is used to divide the visible light image of the fruit tree canopy into superpixels. The initial size and compactness coefficient of the superpixels are set to obtain several superpixel blocks with boundaries that fit the image texture. The image region corresponding to each superpixel block is converted from the RGB color space to the Lab color space and the YCbCr color space, respectively. The pixel mean and variance of the L, a, b channels and Y, Cb, Cr channels of each superpixel block are extracted as multi-channel color features. A classification benchmark is constructed based on the pre-statistical prior color of fruit tree organs. The classification benchmark is used as the initial cluster center of the Gaussian mixture model. The multi-channel color features of all superpixel blocks are clustered and classified in combination with the Gaussian mixture model. The category label of each superpixel block is output to obtain the region type of each superpixel block. The region type includes leaf region, fruit region, branch region and background region. Superpixel blocks labeled as background regions are removed. The prior color of the fruit tree organs is based on the distribution range of pixel values ​​of leaves, fruits, and branches in the Lab and YCbCr color spaces, obtained from a large number of field samples.

[0021] The second stage combines fruit tree organ morphology constraints to complete organ instance labeling. Specifically, it involves pre-setting morphological prior constraints for leaves, fruits, and branches. Leaf morphological prior constraints include aspect ratio ranges and edge smoothness thresholds; fruit morphological prior constraints include roundness thresholds and area ranges; and branch morphological prior constraints include slenderness ratio thresholds and width ranges. Using these morphological prior constraints as termination criteria for region growth, superpixel blocks of the same region type are merged according to spatial adjacency to obtain preliminary organ connectivity regions. Adhesion is then determined for these preliminary organ connectivity regions. If the preliminary organ connectivity is... If the number of local maxima of the distance transform within a region is greater than 1, then the preliminary organ connectivity region is determined to have overlap and adhesion. Preliminary organ connectivity regions without overlap and adhesion are directly retained as organ regions. For preliminary organ connectivity regions with overlap and adhesion, the distance transform combined with the label watershed algorithm is used to split the instances and separate them into independent individual organ regions. Based on all organ regions, corresponding binary organ mask images are generated, and the original visible light image of the fruit tree canopy is cropped according to each organ mask image to extract image blocks of leaves, fruits and branches respectively. All image blocks with organ category labels are integrated to form an organ region set.

[0022] S3. Determine the current fruit tree's growth stage, match it with the standard feature library of healthy organs corresponding to the fruit tree's growth stage, and use Mahalanobis distance to measure the feature deviation of each organ region to filter out a set of suspected abnormal regions. Specifically, first, determine the current fruit tree's growth stage by combining the fruit tree variety's phenological calendar with the current date, or by identifying the organ morphology and color features of the concentrated organ regions, to determine the current fruit tree growth stage, such as the budding stage, new shoot growth stage, flowering and fruit setting stage, etc. In addition, a standard feature library of healthy organs for different fruit tree growth stages is pre-constructed. The standard feature library of healthy organs stores the vector form of the baseline features of healthy leaves, healthy fruits, and healthy branches under different fruit tree growth stages and the covariance matrix of the corresponding feature dimensions. The baseline features include color features and texture features.

[0023] Subsequently, feature deviation is measured and suspected abnormal regions are screened. A size-adaptive sliding window is used to traverse each organ region in the organ region set, extracting color and texture features within each window to construct a multidimensional window feature vector. The Mahalanobis distance between this multidimensional window feature vector and the baseline features of the corresponding organ category and current reproductive stage in the healthy organ standard feature library is calculated. The Mahalanobis distance calculation process incorporates the covariance matrix of the corresponding baseline features to eliminate correlation interference between feature dimensions in the multidimensional window feature vector. A preset deviation threshold is set, and windows whose Mahalanobis distance exceeds the preset threshold are marked as abnormal windows. Spatially connected abnormal windows are merged to obtain suspected abnormalities. Sub-regions are defined, and spatially disconnected abnormal windows are merged into the same suspected abnormal sub-region if they belong to the same organ region and their feature similarity is higher than a preset similarity threshold; otherwise, they are treated as independent suspected abnormal sub-regions. Multidimensional features of each suspected abnormal sub-region are extracted to form a suspected abnormal region feature set. The multidimensional features include color features, texture features, edge features, and internal uniformity features. Among them, color features include the mean, variance, and skewness of each color channel; texture features include the energy, contrast, and entropy of the gray-level co-occurrence matrix; edge features include gradient magnitude, edge density, and contour curvature; and internal uniformity features include regional gray-level variance, gray-level range, and pixel gray-level dispersion.

[0024] S4. Based on a pre-constructed knowledge base of disease and pest symptom mechanisms, the DS evidence theory is used to perform multi-feature evidence fusion on each suspected abnormal area to determine the disease confidence and pest confidence of each suspected abnormal area, and to determine the disease and pest classification of each suspected abnormal area accordingly. Specifically, a knowledge base of disease and pest symptom mechanisms is pre-constructed, which includes a disease symptom sub-base and a pest symptom sub-base, respectively storing typical features of the symptom mechanisms of various fruit tree diseases and fruit tree pests, including lesion color distribution, lesion texture regularity, and insect hole morphology characteristics, which can form a one-to-one matching dimension with the multi-dimensional features of the above-mentioned suspected abnormal sub-regions.

[0025] Subsequently, confidence level fusion calculation and pest / disease attribution type determination are performed. The confidence level fusion calculation specifically involves: using the multidimensional features of each suspected anomalous region as multidimensional identification features; matching each identification feature with typical features in the pest / disease symptom mechanism knowledge base to obtain the feature matching degree of the corresponding disease and pest; normalizing the feature matching degree to obtain the basic probability allocation of each independent piece of evidence corresponding to the disease proposition, pest proposition, and uncertain proposition; and using Dempster's combination rule to perform orthogonal and fusion operations on the basic probability allocation of all independent evidence to obtain the final disease confidence and pest confidence for each suspected anomalous region. The specific determination of the pest / disease classification is as follows: a preset high confidence threshold and a confidence difference threshold are set; if the pest confidence level is not less than the preset high confidence threshold, and the difference between the pest confidence level and the disease confidence level is not less than the preset confidence difference threshold, then the suspected abnormal area is classified as a pest; if the disease confidence level is not less than the preset high confidence threshold, and the difference between the disease confidence level and the pest confidence level is not less than the preset confidence difference threshold, then the suspected abnormal area is classified as a disease; if neither of the above two determination conditions is met, then the suspected abnormal area is classified as an uncertain area.

[0026] S51. For suspected abnormal regions classified as diseases and pests, weighted feature similarity matching is performed with the corresponding sub-databases in the disease and pest symptom mechanism knowledge base to obtain the disease and pest types in the suspected abnormal regions. Specifically: For suspected abnormal regions classified as diseases, their multidimensional features are compared with the typical features of various fruit tree diseases in the disease symptom sub-database of the disease and pest symptom mechanism knowledge base using weighted cosine similarity calculation. Color features, texture features, edge features, and internal uniformity features are each assigned a pre-defined weight. The fruit tree disease type with the highest cosine similarity exceeding the pre-defined similarity threshold is selected as the disease and pest type for the suspected abnormal region. For suspected abnormal regions classified as pests, the same weighted feature similarity matching method as for diseases is used to match them with the pest symptom sub-database in the disease and pest symptom mechanism knowledge base to obtain the corresponding pest types, which are then used as the disease and pest types for the suspected abnormal regions.

[0027] S52. For suspected anomalous regions of uncertain class, a secondary determination is made through frequency domain feature analysis to obtain the pest and disease types in the suspected anomalous region. Specifically, the suspected anomalous region of uncertain class is converted into a grayscale image, a two-dimensional discrete Fourier transform is performed on the grayscale image, and the spectrum obtained by the transform is centered to obtain an amplitude spectrum map. With the spectrum center as the origin, the amplitude spectrum map is divided into low-frequency, mid-frequency, and high-frequency regions in increasing radius. The proportion of energy in each region to the total spectrum energy is calculated to obtain the energy proportion of each frequency band. At the same time, the radial distribution entropy and angular distribution entropy of the amplitude spectrum map are extracted, and the frequency domain is constructed by combining the energy proportion of each frequency band. Feature vectors; calculate cosine similarity between the frequency domain feature vectors and the pre-constructed standard spectrum templates for diseases and pests, respectively. Update the disease and pest confidence scores for the suspected abnormal region based on the cosine similarity results. Specifically, use the similarity obtained from frequency domain feature matching as a correction coefficient to weight and update the original disease and pest confidence scores, and then re-normalize them. Based on the updated disease and pest confidence scores, repeat the disease and pest classification rules in S4 to determine the disease and pest types in the suspected abnormal region. If it is still determined to be an uncertain category, mark the suspected region as a sample to be reviewed and push it to human experts for review and judgment.

[0028] S6. Based on the types of pests and diseases in each suspected abnormal area, and combined with the severity level of pests and diseases determined by the degree of damage to a single organ and the proportion of damaged organs on the whole tree, a pest and disease identification result is constructed. This step is specifically as follows: First, determine the degree of damage to a single organ. Based on the types of pests and diseases on the fruit tree determined in S51 and S52, for each type of pest and disease, calculate the proportion of the total area of ​​the suspected abnormal area corresponding to that type of pest and disease to the total area of ​​the affected organ area. The total area of ​​the suspected abnormal area is the sum of the total areas of all suspected abnormal areas identified as belonging to that type of pest and disease. The total area of ​​the affected organ area is the sum of the organ area areas of the affected organs parasitized by that pest and disease. The suspected abnormal area is extracted from the organ area. Multiple sets of proportion intervals are preset, and each set of proportion intervals corresponds to a single organ damage level. Based on the proportion of the total area of ​​the suspected abnormal area corresponding to that type of pest and disease to the total area of ​​the affected organ area, the single organ damage level of that type of pest and disease relative to the fruit tree is determined. Subsequently, the proportion of affected organ areas within the monitoring range to the total number of organ areas is statistically analyzed. This represents the percentage of affected organs on the entire fruit tree relative to the pest or disease. The number of affected organ areas is the total number of organ areas exhibiting this type of pest or disease, while the total number of organ areas is the total number of all organ areas within the monitoring range of a single fruit tree. Based on this, a weighted index of the overall tree severity is calculated using the individual organ damage severity level. Finally, a severity index grading threshold is preset. Based on the threshold range of the overall tree severity index, the severity level of the pest or disease is determined. The severity level is divided into three levels: mild, moderate, and severe. Finally, the pest or disease types and corresponding severity levels of all suspected abnormal areas on a single fruit tree are integrated to construct a complete pest or disease identification result.

[0029] S7. Combining soil parameters and meteorological environmental data, the basic water, fertilizer, and pesticide management strategy determined based on pest and disease identification results and fruit tree growth stage is adaptively corrected to obtain intelligent water, fertilizer, and pesticide decisions and execute them. Specifically, this step involves: First, based on the pest and disease types and severity levels in the pest and disease identification results, combined with the current fruit tree growth stage, a preset expert rule base is matched to generate a basic water, fertilizer, and pesticide management strategy including pesticide application parameters, fertilization parameters, and irrigation parameters. Then, each adjustable parameter in the basic water, fertilizer, and pesticide management strategy is used as a variable to be optimized. Simultaneously, real-time collected soil parameters and meteorological environmental parameters are concatenated into an environmental feature vector, which is input into a pre-trained machine learning correction model. The machine learning correction model uses a gradient boosting tree model, trained based on historical orchard environmental data, pest and disease occurrence data, and water, fertilizer, and pesticide operation effect data, to fit the nonlinear mapping relationship between field environmental parameters and the optimal parameters for water, fertilizer, and pesticide operations. The machine learning correction model adaptively optimizes and corrects adjustable parameters in the basic water, fertilizer, and pesticide management strategy, such as application rate, application frequency, nitrogen, phosphorus, and potassium fertilizer ratio, fertilizer application rate, irrigation volume, and irrigation cycle. It eliminates strategy parameters that are not suitable for the current field temperature, humidity, soil moisture, and soil fertility, and outputs intelligent water, fertilizer, and pesticide decisions that are suitable for the current fruit tree growth status, pest and disease situation, and field environment. Finally, the intelligent water, fertilizer, and pesticide decisions are distributed to intelligent pesticide application equipment, intelligent fertilizer application equipment, and intelligent irrigation equipment in the field for execution via 5G Internet of Things.

[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for orchard disease identification and intelligent decision-making regarding water, fertilizer, and pesticide application based on 5G Internet of Things, characterized in that, Includes the following steps: The field sensing nodes deployed by 5G IoT collect visible light images of the fruit tree canopy, and simultaneously collect soil parameters and meteorological environmental data. The visible light image of the fruit tree canopy is pre-segmented by superpixel and classified by multi-color space features. Combined with the morphological constraints of the fruit tree organs, organ instance identification is completed, and image blocks of leaves, fruits and branches are extracted to form an organ region set. Determine the current fruit tree's growth stage, match it with the standard feature library of healthy organs corresponding to the fruit tree's growth stage, and use Mahalanobis distance to measure the feature deviation of each organ region to screen and obtain a feature set of suspected abnormal regions. Based on a pre-constructed knowledge base of disease and pest symptom mechanisms, the DS evidence theory is used to perform multi-feature evidence fusion on each suspected abnormal area to determine the disease confidence and pest confidence of each suspected abnormal area, and to determine the disease and pest classification of each suspected abnormal area accordingly. The disease and pest classification includes disease type, pest type and uncertain type. For suspected abnormal areas of diseases and pests, weighted feature similarity matching is performed with the corresponding sub-databases in the disease and pest symptom mechanism knowledge base to obtain the types of diseases and pests in the suspected abnormal areas; For suspected anomalous regions of uncertain class, frequency domain feature analysis is used to complete a secondary determination to obtain the types of pests and diseases in the suspected anomalous regions; Based on the types of pests and diseases in each suspected abnormal area, and combined with the severity level of pests and diseases determined by the degree of damage to a single organ and the proportion of damaged organs in the whole plant, a pest and disease identification result is constructed. By combining soil parameters and meteorological environmental data, the basic water, fertilizer and pesticide management strategy determined based on the results of pest and disease identification and the fruit tree growth period is adaptively modified to obtain intelligent water, fertilizer and pesticide decisions and issue them for execution.

2. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 1, characterized in that, The method for superpixel pre-segmentation and multi-color space feature classification is as follows: The SLIC algorithm was used to divide the visible light image of the fruit tree canopy into superpixels, resulting in several superpixel blocks. Each superpixel block was then converted to the Lab and YCbCr color spaces, and the multi-channel color features of each superpixel block were extracted. A classification criterion is constructed based on the prior color of fruit tree organs. A Gaussian mixture model is then used to cluster and classify the color features of all superpixel blocks to determine the region type of each superpixel block. The region types include leaf region, fruit region, branch region, and background region.

3. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 2, characterized in that, The method for identifying organ instances is as follows: By introducing morphological prior constraints corresponding to leaves, fruits and branches, and performing region growth and merging on superpixel blocks of the same type, preliminary organ connectivity regions are obtained. For preliminary organ connectivity regions without overlap or adhesion, the organ regions are directly retained. For preliminary organ connectivity regions with overlap or adhesion, the distance transformation combined with the marker watershed algorithm is used to split the regions into instances and separate them into organ regions. Based on all organ regions, corresponding organ mask images are generated. The visible light images of the fruit tree canopy are cropped according to each organ mask image, and image blocks of leaves, fruits and branches are extracted respectively. All organ image blocks are integrated to form an organ region set.

4. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 3, characterized in that, The method for measuring the characteristic deviation of each organ region using Mahalanobis distance is as follows: A sliding window is used to traverse the organ regions, and the color and texture features of each window are extracted to construct a multi-dimensional window feature vector. Calculate the Mahalanobis distance between the multidimensional window feature vector and the baseline feature corresponding to the organ category to which the organ region belongs in the standard feature library of healthy organs; Windows whose Mahalanobis distance exceeds a preset deviation threshold are marked as abnormal windows. Connected abnormal windows are merged to obtain suspected abnormal sub-regions. Multidimensional features of each suspected abnormal sub-region are extracted to form a suspected abnormal region feature set. The multidimensional features include color features, texture features, edge features, and internal uniformity features.

5. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 4, characterized in that, The methods for determining the confidence levels for diseases and pests are as follows: The multidimensional features of each suspected abnormal region are used as multidimensional identification features; Each identification feature is used as an independent piece of evidence. Typical features in the disease and pest symptom mechanism knowledge base are matched to calculate the basic probability allocation of diseases and pests corresponding to each independent piece of evidence. The basic probability assignments of all independent evidence were fused using Dempster's combination rule to obtain the disease and pest confidence scores for each suspected anomalous area.

6. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 5, characterized in that, The rules for determining the type of pest or disease are as follows: If the confidence level of pests is not less than the preset high confidence threshold and the difference between the confidence level of pests and the confidence level of diseases is not less than the preset confidence difference threshold, then the suspected abnormal area is determined to be a pest-related area. If the confidence level of the disease is not less than the preset high confidence threshold and the difference between the confidence level of the disease and the confidence level of the pest is not less than the preset confidence difference threshold, then the suspected abnormal area is determined to be a disease. Otherwise, the suspected abnormal area is classified as uncertain.

7. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 6, characterized in that, The suspected abnormal regions of uncertain categories are further determined through frequency domain feature analysis, specifically as follows: The suspected abnormal regions of the uncertain class are converted into grayscale images. A two-dimensional discrete Fourier transform is performed on the grayscale images and the spectrum is centered to obtain the amplitude spectrum. The amplitude spectrum is divided into annular partitions, and the energy proportions of low, medium and high frequency regions are calculated. At the same time, the radial distribution entropy and angular distribution entropy of the spectrum are extracted, and frequency domain features are constructed accordingly. The cosine similarity between the frequency domain features and the pre-constructed standard spectrum templates for diseases and pests is calculated. The confidence scores for diseases and pests are then updated. Based on the updated confidence scores for diseases and pests, the types of diseases and pests corresponding to the suspected abnormal area are determined.

8. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 1, characterized in that, The method for determining the severity level of the pests and diseases is as follows: Based on the identified pests and diseases, calculate the proportion of the total area of ​​the suspected abnormal area corresponding to the pest or disease to the total area of ​​the affected organ, and divide the single organ damage level according to the preset proportion range. The proportion of affected organ areas within the statistical monitoring range to the total number of organ areas is used to calculate the overall severity index of the plant, which is then combined with the severity level of individual organ damage. Based on this overall severity index, the severity level of the pest or disease is determined.

9. The orchard disease identification and intelligent decision-making method for water, fertilizer, and pesticide application according to claim 1, characterized in that, The method for obtaining the intelligent water, fertilizer, and pesticide decision-making is as follows: based on the results of pest and disease identification and the fruit tree growth period, a basic water, fertilizer, and pesticide management strategy is generated by matching a preset expert rule base. Each adjustable parameter in the basic water, fertilizer, and pesticide management strategy is used as a variable to be optimized. At the same time, the real-time collected soil parameters and meteorological environmental parameters are used as environmental feature vectors and input into a pre-trained machine learning correction model for adaptive correction. The intelligent water, fertilizer, and pesticide decision-making is then output to adapt to the current fruit tree growth status, pest and disease situation, and field environment.

Images