Machine vision-based intelligent management system, equipment, and media for the health of garden plants.

Abstract

This invention discloses a machine vision-based intelligent management system, device, and medium for the health of garden plants, relating to the fields of machine vision and image processing technology. The system includes: a machine vision module for acquiring a set of garden plant images containing visible light, infrared, and spectral images; a plant risk first, second, and third map acquisition module for tracing growth accidents, heat stress, and nutritional status risks based on visible light, infrared, and spectral images, respectively, and obtaining corresponding risk maps; and a garden plant health management execution module for performing garden plant health management based on these three maps. This invention solves the technical problem that traditional methods struggle to comprehensively and accurately capture the multi-dimensional health status of garden plants, thus hindering precise management. It achieves the technical effect of comprehensive and accurate capture and intelligent management of multi-dimensional health risks related to growth accidents, heat stress, and nutritional status in garden plants.

Description

Technical Field

[0001] This invention relates to the field of machine vision and image processing technology, specifically to a machine vision-based intelligent management system, equipment, and medium for the health of garden plants. Background Technology

[0002] In reality, the health management of garden plants is of great significance to ecological maintenance and landscape effects, and accurately grasping the health status of plants is crucial. Current technologies largely rely on manual inspections and single sensors to monitor plant health. While these methods have proven effective in enclosed or small-scale garden settings, their limitations become apparent in large and complex garden applications as garden management requirements increase. Traditional methods struggle to comprehensively capture the multi-dimensional states of plant growth, heat stress, and nutrition, and are constrained by human experience and single data points, leading to inaccurate plant health assessments and failing to meet the needs of intelligent and refined management of garden plants. Machine vision technology offers a potential solution to this problem. Summary of the Invention

[0003] This application provides a machine vision-based intelligent management system, equipment, and media for the health of garden plants, which addresses the technical problem of the difficulty in comprehensively and accurately capturing the multi-dimensional health status of garden plants using traditional methods, thereby achieving precise management.

[0004] In view of the above problems, this application provides a machine vision-based intelligent management system, equipment and media for the health of garden plants.

[0005] In a first aspect, this application provides a machine vision-based intelligent management system for the health of garden plants, the system comprising:

[0006] A machine vision module is used to acquire a set of garden plant images, including a set of visible light images, a set of infrared images, and a set of spectral images. A first plant risk map acquisition module is used to trace growth accident risk characteristics based on the visible light image set to obtain a first plant risk map. A second plant risk map acquisition module is used to trace heat stress risk characteristics based on the infrared image set to obtain a second plant risk map. A third plant risk map acquisition module is used to trace nutrient status risk characteristics based on the spectral image set to obtain a third plant risk map. A garden plant health management execution module is used to perform garden plant health management based on the first, second, and third plant risk maps.

[0007] Secondly, this application provides an electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is used to execute the machine vision-based intelligent management system for garden plant health provided in this application.

[0008] Thirdly, this application provides a computer-readable storage medium storing a computer program for executing the machine vision-based intelligent management system for garden plant health provided in this application.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application acquires visible light, infrared, and spectral images of garden plants through a multimodal visual terminal. After processing such as growth status feature recognition, pixel-level temperature distribution extraction, and nutrient feature interpretation, it traces the risk characteristics of growth accidents, heat stress, and nutrient status and generates corresponding risk maps. Combining these three types of maps for health management, it comprehensively grasps the multi-dimensional health risks of garden plants, making garden plant health management more intelligent and precise. This achieves the technical effect of comprehensive and accurate capture and intelligent management of multi-dimensional health risks of garden plants, including growth accidents, heat stress, and nutrient status. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the structure of a machine vision-based intelligent management system for the health of garden plants provided in an embodiment of this application.

[0013] Figure 2 A schematic diagram of the structure of the electronic device provided in this application.

[0014] Explanation of reference numerals in the attached figures: Machine vision module 10, Plant risk first map acquisition module 20, Plant risk second map acquisition module 30, Plant risk third map acquisition module 40, Garden plant health management execution module 50, Bus 300, Receiver 301, Processor 302, Transmitter 303, Memory 304, Bus interface 305. Detailed Implementation

[0015] This application provides a machine vision-based intelligent management system, equipment, and media for the health of garden plants, which addresses the technical problem of traditional methods failing to comprehensively and accurately capture the multi-dimensional health status of garden plants and thus achieve precise management.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] Example 1, as Figure 1 As shown, this application provides a machine vision-based intelligent management system for the health of garden plants, the system comprising:

[0018] The machine vision module 10 is used to obtain a set of garden plant images, which includes a set of visible light images of plants, a set of infrared images of plants, and a set of spectral images of plants.

[0019] Specifically, the machine vision module 10 obtains a monitoring set of garden plants through a multimodal vision terminal that includes a visible light imaging group, an infrared thermal imaging group, and a multispectral imaging group. Then, it performs multi-level preprocessing on the monitoring set, including image enhancement, plant target segmentation, and perceptual modality classification, to obtain a garden plant image set. The specific steps are described in detail in the garden plant monitoring set acquisition unit and the garden plant image set acquisition unit.

[0020] The first plant risk map acquisition module 20 is used to trace the risk characteristics of growth accidents based on the plant visible light image set and obtain the first plant risk map.

[0021] Specifically, based on a set of visible light images of plants, the first risk map of plants is generated by identifying the growth state vectors of each plant, constructing multiple plant feature labels, building a plant growth accident tree model, inferring growth accident paths and tracing risk characteristics. The specific steps are explained in detail in the plant growth state vector acquisition unit to the plant risk map acquisition unit.

[0022] The second plant risk map acquisition module 30 is used to trace the characteristics of heat stress risk based on the plant infrared image set and obtain the second plant risk map.

[0023] Specifically, based on the plant infrared image set, a second plant risk map is generated by extracting temperature distribution maps of each plant, mining the normal temperature range of each plant, detecting the temperature deviation matrix of each plant, and tracing the source of heat stress risk. The specific steps are explained in detail in the plant temperature distribution map acquisition unit to the plant risk second map acquisition unit.

[0024] The third plant risk map acquisition module 40 is used to trace the nutritional status risk characteristics based on the plant spectral image set and obtain the third plant risk map.

[0025] Specifically, based on the plant spectral image set, the risk of nutrient deficiency, excess, and imbalance is evaluated by interpreting the nutrient feature vectors of each plant, the inducing factors are traced, and a third plant risk map is generated. The specific steps are explained in detail in the plant nutrient feature vector acquisition unit to the plant risk third map acquisition unit.

[0026] The garden plant health management execution module 50 is used to perform garden plant health management based on the first plant risk map, the second plant risk map, and the third plant risk map.

[0027] Specifically, the garden plant health management execution module 50 first integrates the data from the first, second, and third plant risk maps, extracting key risk information from each map. Growth accident risks (such as pests and diseases, branch dieback) in the first map, heat stress risks (such as localized high-temperature areas) in the second map, and nutrient risks (such as nitrogen deficiency, phosphorus excess) in the third map are integrated into a unified plant spatial coordinate system. Coordinate matching is used to map the association between different risks of the same plant. For example, the eastern branch of a cherry tree is marked as having a high risk of aphid infestation in the first map, a medium risk of heat stress in the second map, and a low nitrogen content risk in the third map.

[0028] Based on the fused risk data, the module quantifies and prioritizes various risks. The analytic hierarchy process (AHP) is used to assign weights to different risk types: growth accident risk directly affects plant survival, with a weight of 0.4; heat stress risk affects physiological functions, with a weight of 0.3; and nutritional risk affects growth status, with a weight of 0.3. For each risk region, the individual risk values ​​(range 0-1) are weighted to calculate the comprehensive risk value. For example, the growth accident risk value of the cherry blossom's eastern branch is 0.8, the heat stress risk value is 0.6, and the nutritional risk value is 0.5, resulting in a comprehensive risk value of 0.8×0.4+0.6×0.3+0.5×0.3=0.65, classifying it as medium-high risk and prioritizing its inclusion in the intervention plan. The risk region ranges defined based on the comprehensive risk value are shown in Table 1.

[0029] Table 1: Comprehensive Risk Value Classification Table

[0030]

[0031] Based on the risk type and level, the module matches corresponding health management measures. For growth accident risks, such as aphid infestation, the pest and disease database recommends imidacloprid spray at a concentration of 1000 times dilution, applied every 7 days. For heat stress risks, such as localized high temperatures, it is recommended to provide intermittent shading between 12-2 PM, combined with foliar spraying, twice daily, 500ml each time. To address nutritional risks, such as nitrogen deficiency, a topdressing plan should be developed, with urea applied at a rate of 15g / g. Combine this with watering to dissolve the pesticides. At the same time, measures should consider synergistic effects; for example, avoid using heat-sensitive pesticides when controlling aphids to prevent exacerbating the effects of heat stress.

[0032] The module also generates an execution schedule and monitoring plan, clearly defining the implementation order, frequency, and responsible personnel for each measure. For example, measures in medium- and high-risk areas are initiated within 3 days, while those in low-risk areas are arranged within 7 days. After implementation, images are collected regularly every 3 days using the machine vision module 10 to regenerate a risk map and assess the effectiveness. If the overall risk value drops below 0.3, the measures are adjusted to preventative maintenance, such as monthly foliar fertilizer application. If the risk is not mitigated, the plan is optimized, such as increasing the pesticide concentration or changing the pesticide variety.

[0033] By integrating multi-dimensional risk information, quantitative assessment, and precise policy implementation, the scientific and dynamic management of garden plant health has been achieved, effectively improving the efficiency and pertinence of risk intervention.

[0034] In one possible implementation, the plant risk first map acquisition module 20 further includes:

[0035] The system includes: a plant growth state vector acquisition unit, which identifies growth state features based on the plant visible light image set to obtain growth state vectors for each plant; a plant feature label construction unit, which identifies species and stage features based on the plant visible light image set to construct multiple plant feature labels; a plant growth accident tree model construction unit, which mines and learns growth accident events based on the multiple plant feature labels to construct multiple plant growth accident tree models; a plant growth accident path acquisition unit, which guides the multiple plant growth accident tree models based on the multiple plant feature labels to perform growth accident deduction on each plant growth state vector to obtain multiple plant accident paths; and a plant risk first atlas acquisition unit, which traces risk features based on the multiple plant accident paths to generate the plant risk first atlas.

[0036] Specifically, clear images of individual plants are selected from a set of visible light plant images, and growth status features are first identified. The images are processed by grayscale conversion and edge detection to extract leaf contour information. Features such as single leaf area and number of leaves are calculated, with the calculation error for single leaf area controlled within ±2%. Using reference objects of known size in the image, such as standard-length markers, proportional conversion is performed to obtain spatial features such as plant height and crown width, with the plant height measurement deviation for each plant not exceeding 5cm. Simultaneously, texture analysis algorithms are used to extract detailed features such as leaf texture density and branching angle. Finally, a multi-dimensional growth status vector is generated for each plant, covering key information such as leaf morphology, plant size, and structural distribution. For example, the growth status vector of a certain rose plant might include specific values ​​such as: single leaf area 8.5cm², plant height 60cm, and number of branches 5.

[0037] Then, a ResNet-50 plant classification model was constructed based on a deep residual network architecture. This model includes residual blocks composed of convolutional layers, batch normalization layers, and ReLU activation functions. Skip connections alleviate the gradient vanishing problem during deep network training. The model consists of 50 trainable layers. Finally, global average pooling layers and fully connected layers output the classification results. The number of neurons in the fully connected layers matches the number of garden plant species to be identified; for example, 80 neurons are used for 80 species. During training, a set of visible light images labeled with plant species was used as training data. After data augmentation processing such as random rotation, scaling, and brightness adjustment, the images were divided into training and validation sets in an 8:2 ratio. The Adam optimizer was used, with an initial learning rate of 0.001, decaying by 10% every 10 epochs. The difference between the predicted and labeled categories was calculated using the cross-entropy loss function. The training was iterated for 100 epochs. After each epoch, the accuracy was evaluated using the validation set. Training stopped when the validation set accuracy stabilized above 92%. The model input is a preprocessed plant visible light image, such as one that has been resized to 224×224 pixels and normalized to the range of [0,1]. The output is the probability distribution of the plant species corresponding to the image. The category with the highest probability is taken as the final recognition result, such as outputting specific plant species labels like apricot and crape myrtle.

[0038] After identifying growth status features, species stage features were identified based on the same batch of visible light images. A pre-trained ResNet-50 plant classification model was used to identify plant species in the images. By extracting deep features from the images and comparing them with sample features in the plant species database, the species identification accuracy reached over 92%, distinguishing more than 80 common garden plants such as ginkgo, crape myrtle, and liriope. Simultaneously, plant growth characteristics were analyzed, combining leaf color (e.g., light green corresponds to the seedling stage, dark green to the mature stage) and flowering status (e.g., the presence of flower buds or flowers corresponds to the flowering period) to determine the growth stage, dividing it into seedling stage, vegetative growth stage, reproductive growth stage, bud stage, flowering stage, and full bloom stage. Finally, feature labels containing species and growth stage were constructed for each plant, such as ginkgo - vegetative growth stage, crape myrtle - flowering stage, and rose - bud stage, forming a plant feature label set.

[0039] Next, the first plant feature label is extracted based on multiple plant feature labels, and the first growth accident event set is retrieved. Through top-level features, intermediate factors, and basic factors, the causal relationship path of the accident is constructed and the causal co-occurrence confidence is strengthened to generate the first plant growth accident tree model. The specific steps are detailed in the process of obtaining sub-units from the first growth accident event set and constructing sub-units in the first plant growth accident tree model.

[0040] Then, tags matching the current plant are selected from multiple plant feature labels. For example, if a plant is labeled as a rose in the bud stage, a pre-built rose-bud stage plant growth fault tree model is invoked. These feature labels act as retrieval keys, accurately locating the model corresponding to the plant species and growth stage, ensuring the relevance of the inference. The plant's growth state vector is input into the model, containing multi-dimensional growth features such as leaf color, bud fullness, and branch growth behavior. The model then performs analysis based on these features.

[0041] The model internally stores various causal pathways related to the rose's bud stage, such as fungal infection → calyx rot → bud drop, nutrient imbalance → poor growth point development → bud deformity, etc. During the simulation, the model compares the features in the growth state vector with the factors in the pathway. For example, if the vector shows the feature of brown spots at the base of the bud, it will trigger the activation of the fungal infection-related pathway; if the bud diameter is detected to be smaller than the average for the same period, it will trigger the activation of the nutrient imbalance pathway. By matching the features in the vector with the intermediate and basic factors in the pathway layer by layer, the model will select pathways that match the current plant growth state, forming multiple plant growth accident pathways. Each pathway clearly presents the complete chain from the initial cause to the final accident manifestation.

[0042] After identifying multiple growth accident pathways, risk characteristic tracing began. First, key elements were extracted from each pathway, including basic factors triggering the accident, such as fungal infection and nutrient imbalance; intermediate factors driving the accident, such as calyx rot and poor growth point development; and the final accident manifestations, such as bud drop and bud deformity. Next, the strength of the correlation between these factors was analyzed. For example, the close association between fungal infection and calyx rot in a particular pathway, and the frequency of the same factor appearing in different pathways, were analyzed to determine which factors are the main sources of risk.

[0043] Subsequently, these risk factors were categorized by their scope and severity, and their spatial distribution characteristics on the plants were analyzed. For example, spots caused by fungal infections were concentrated on the eastern buds, and nutrient imbalances affected overall bud development. This information was then integrated into a visual map. The map marks the location, correlation, and potential accident types of each risk factor, forming a primary risk map for the plant, which visually presents the growth accident risk characteristics faced by the rose during the bud stage.

[0044] By using plant feature tags to guide the fault tree model to perform targeted deduction of growth state vectors, and then sorting and integrating risk factors, a first map that clearly reflects the source and manifestation of plant growth accident risks is generated, providing clear guidance for accurate identification and intervention of growth risks.

[0045] In one possible implementation, the plant growth accident tree model building unit further includes:

[0046] The first growth accident event set acquisition subunit extracts a first plant feature label based on the multiple plant feature labels and performs growth accident event retrieval based on the first plant feature label to obtain a first growth accident event set; the accident top-level feature acquisition subunit performs top-level feature tracing based on the first growth accident event set to obtain the top-level features of each accident; the accident intermediate factor acquisition subunit performs intermediate triggering factor tracing based on the top-level features of each accident based on the first growth accident event set to obtain the intermediate factors of each accident; the accident basic factor acquisition subunit performs basic triggering factor tracing based on the intermediate factors of each accident based on the first growth accident event set to obtain the basic factors of each accident; the accident causal association path construction subunit constructs multiple accident causal association paths based on the top-level features of each accident, the intermediate factors of each accident, and the basic factors of each accident; and the causal co-occurrence confidence of the causal association paths is strengthened to generate a first plant growth accident tree model.

[0047] Specifically, firstly, a primary plant feature label is extracted from multiple plant feature tags, for example, cherry blossom-peak bloom period is selected as the target label. Based on this label, a search is performed in a plant growth accident event database. The database covers records of diseases, pests, wilting, and abnormal growth of different plant species and growth stages over the past 10 years. A label matching algorithm is used to filter out growth accident events related to cherry blossom-peak bloom period, forming a primary growth accident event set. These events include specific cases such as premature petal drop, branch wilting, and flower deformity.

[0048] Top-level feature tracing was conducted on the first set of growth accident events. Each event in the set was analyzed one by one, and the abnormal plant state ultimately presented by each event was extracted as the top-level feature. For example, from events related to premature petal drop, top-level features such as large-scale petal drop and scorched petal edges were summarized; from events related to branch withering, top-level features such as withering of branch tips and wrinkling of the bark were extracted, thereby clarifying the final manifestation of various growth accidents.

[0049] For each top-level feature of an incident, we return to the first set of growth incident events and trace the intermediate triggering factors that led to the appearance of the top-level feature by reviewing the development process in the event records. Taking the large-scale petal shedding as an example, we analyze the physiological changes before the petals fall off in the incident and identify intermediate factors such as petal cell dehydration and pedicel abscission. These factors are the key links connecting the basic cause and the top-level feature.

[0050] Based on the first set of growth accident events, the basic triggering factors of each identified intermediate factor are further traced. For example, for the intermediate factor of petal cell dehydration, by analyzing information such as environmental records and maintenance operations in the event, the root causes of this factor are identified, such as continuous high temperature and drought, and decreased root water absorption capacity. These basic factors are the initial triggers that lead to the start of the entire accident chain.

[0051] Next, based on the identified top-level characteristics, intermediate factors, and basic factors of each accident, causal relationship paths are constructed. By drawing a logical relationship diagram, it is clarified how each basic factor triggers the corresponding intermediate factor, and then how the intermediate factor leads to the appearance of the top-level characteristic, forming a complete chain of basic factor → intermediate factor → top-level characteristic. For example, continuous high temperature and drought → petal cell dehydration → large-scale petal drop is a typical causal relationship path of an accident, and thus multiple causal relationship paths of accidents are constructed.

[0052] Finally, the confidence of causal co-occurrence was strengthened for the constructed causal relationship paths of multiple accidents. This was achieved by statistically analyzing the frequency of co-occurrence of basic factors, intermediate factors, and top-level features in the event set for each path, combined with supplementary analysis of events involving similar plants at similar growth stages, to adjust the confidence of the paths. Paths with high co-occurrence frequency and strong logical connections had their confidence weight increased; paths that were sporadic or logically weak were corrected, ultimately generating the first plant growth accident tree model. This model clearly presents the logical relationship between the causes, development process, and final manifestation of cherry blossom-peak bloom growth accidents. Similarly, the above steps were performed on the remaining plant feature labels to construct corresponding plant growth accident tree models.

[0053] By performing hierarchical tracing, causal analysis, and confidence optimization of growth accident events associated with plant feature tags, a growth accident tree model that accurately reflects the accident occurrence mechanism at a specific plant growth stage was generated, providing a structured analytical framework for subsequent growth accident deduction and risk assessment.

[0054] In one possible implementation, the plant risk second map acquisition module 30 further includes:

[0055] The plant temperature distribution map acquisition unit extracts pixel-level temperature distribution from the plant infrared image set to obtain temperature distribution maps for each plant; the plant normal temperature range acquisition unit mines normal temperature ranges based on multiple plant feature tags to obtain normal temperature ranges for each plant; the plant temperature deviation matrix acquisition unit detects deviations in the temperature distribution maps based on the normal temperature ranges for each plant to obtain temperature deviation matrices for each plant; and the plant risk second atlas acquisition unit traces the source of heat stress risk based on the temperature deviation matrices for each plant to generate the plant risk second atlas.

[0056] Specifically, infrared images of individual plants are selected from a set of plant infrared images, and pixel-level temperature distribution is extracted. First, the infrared images are preprocessed by using Gaussian filtering to remove noise interference and ensure the stability of the temperature data. Then, based on the principles of infrared thermal imaging, the grayscale value of each pixel in the image is converted into a corresponding temperature value, where the pixel coordinates correspond one-to-one with the spatial location on the plant surface. In this way, the entire infrared image is converted into a matrix containing temperature information. Each element in the matrix represents the real-time temperature at a corresponding location on the plant surface, ultimately generating a temperature distribution map that visually reflects the temperature differences in different parts of the plant. The temperatures of different regions in the map are distinguished by color gradients, facilitating the observation of temperature distribution characteristics.

[0057] Next, based on the constructed plant feature tags, normal temperature ranges are mined to obtain the normal temperature intervals for each plant. For each plant feature tag, which includes information on plant species and growth stage, temperature records of the corresponding plant under normal growth conditions during historical periods are retrieved. By statistically analyzing these historical temperature data, the mean temperature and its fluctuation range are calculated. Combined with research findings on suitable temperatures for the corresponding growth stage in plant physiology, a reasonable normal temperature range is determined. For example, for the tag "camphor tree - mature stage," by analyzing temperature data from the same period over the past three years without heat stress, and combining this with the physiological characteristics of mature camphor trees, its normal temperature range is determined. This range reflects the temperature fluctuation range of the plant under healthy conditions.

[0058] Then, using the obtained normal temperature ranges for each plant, deviation detection is performed on the corresponding plant temperature distribution maps to obtain temperature deviation matrices for each plant. The temperature value of each pixel in the temperature distribution map is compared with the normal temperature range for that plant. If the temperature value is within the normal range, the deviation value is recorded as 0; if the temperature value is higher than the upper limit of the range, the deviation value is the difference between the actual temperature and the upper limit; if the temperature value is lower than the lower limit of the range, the deviation value is the difference between the lower limit and the actual temperature, which is recorded as a negative value. Through this pixel-by-pixel deviation calculation, the temperature distribution map is converted into a matrix containing deviation values. Each element in the matrix corresponds to the degree of temperature deviation at a certain location on the plant surface, thereby quantifying the temperature anomalies in different parts of the plant.

[0059] Finally, multiple plant experts conducted heat stress risk assessments on the historical temperature deviation set of plants to obtain multiple evaluation sets. Through ensemble value calculation, a reliable heat stress risk set was obtained. Based on this, an evaluation coordinate system was constructed with the historical deviation set. Temperature deviation information was input to obtain the evaluation space. Then, the second plant risk map was generated by tracing the initiating factors. The specific steps are explained in detail in the section from obtaining sub-units from the heat stress risk assessment set to obtaining sub-units from the second plant risk map.

[0060] By extracting temperature from infrared images, determining normal ranges based on feature labels, and calculating temperature deviations, a deviation matrix that accurately reflects the degree of temperature anomalies in different parts of the plant was obtained, providing a quantitative basis for tracing the source of heat stress risks and generating a second risk map of plants.

[0061] In one possible implementation, the plant risk second map acquisition unit further includes:

[0062] The heat stress risk assessment set acquisition subunit involves multiple plant experts assessing the heat stress risk of plant temperature history deviation sets to obtain multiple heat stress risk assessment sets. The reliable heat stress risk set acquisition subunit calculates the heat stress risk concentration value based on each heat stress risk assessment set to obtain a reliable heat stress risk set. The heat stress risk assessment coordinate system construction subunit constructs a heat stress risk assessment coordinate system based on the plant temperature history deviation sets and the reliable heat stress risk sets. The heat stress risk assessment space acquisition subunit inputs each temperature deviation information within the plant temperature deviation matrix into the heat stress risk assessment coordinate system to obtain a heat stress risk assessment space. The plant risk second atlas acquisition subunit traces the initiating factors based on the heat stress risk assessment space to obtain the plant risk second atlas.

[0063] Specifically, firstly, an evaluation team composed of multiple plant experts was organized to conduct a heat stress risk assessment on a set of historical temperature deviations in plants. This set of historical deviations includes records of different plants deviating from the normal temperature range at different growth stages, covering various types of deviations such as excessively high or low temperatures and fluctuations. Based on principles of plant physiology, the correlation between heat stress manifestations and temperature deviations, and their own experience in diagnosing plant heat damage, the plant experts independently conducted a risk assessment for each historical deviation record, including determining the risk level (e.g., mild, moderate, severe), the possible scope of impact, and the potential type of damage. This resulted in multiple heat stress risk assessment sets, with each historical temperature deviation record corresponding to a set of evaluation results from different experts; that is, each historical temperature deviation in plants corresponds to a heat stress risk assessment set.

[0064] Then, a central value is calculated for each heat stress risk assessment set to eliminate individual evaluation differences and obtain a reliable heat stress risk set. For multiple expert evaluation results corresponding to the same historical deviation record, a weighted average method is used to integrate the risk level values. For example, mild, moderate, and severe are quantified as 1, 2, and 3 respectively. At the same time, the weights are adjusted with reference to the degree of consistency of evaluations. If the expert evaluation disagreement is small, the arithmetic mean is taken; if the disagreement is large, the expert seniority weight is introduced, and the evaluation of senior experts is given higher weight. Finally, a single reliable risk value corresponding to each historical deviation record is obtained, which is then aggregated to form a reliable heat stress risk set.

[0065] Subsequently, a coordinate system for assessing heat stress risk was constructed based on the historical temperature deviation set and the reliable heat stress risk set. The horizontal axis represents the degree to which the temperature deviates from the normal range, while the vertical axis represents the risk value corresponding to that deviation, indicating the potential level of heat stress risk. By mapping the two sets of data to coordinate points, distribution points were formed in the coordinate system. Then, a risk trend line was generated through curve fitting, enabling the coordinate system to intuitively reflect the correspondence between temperature deviation and heat stress risk.

[0066] Subsequently, each temperature deviation information within the temperature deviation matrix of each plant is input into the heat stress risk assessment coordinate system. For the temperature deviation value corresponding to each pixel in the matrix, the corresponding position is found on the horizontal axis of the coordinate system, and mapped along the vertical axis to the risk trend line to obtain the heat stress risk coefficient corresponding to that deviation value. The higher the coefficient, the more severe the risk. After summing the risk coefficients corresponding to all deviation information, a heat stress risk assessment space is formed. This space is presented in matrix form, corresponding one-to-one with the spatial distribution of the plant temperature deviation matrix, clearly reflecting the degree of heat stress risk of each part of the plant.

[0067] Finally, based on the heat stress risk assessment space, the initiating factors were traced to generate a second plant risk map. The distribution areas of high-risk coefficients in the assessment space were analyzed, and combined with environmental monitoring data (such as air temperature, light intensity, and humidity) and plant physiological information (such as leaf water content and stomatal conductance) from the same period, the causes of abnormal temperatures in these areas were traced, such as a sudden increase in local temperature due to strong direct sunlight, or transpiration imbalance caused by insufficient root water absorption.

[0068] These triggering factors are associated with and labeled with risk areas in the map. High-risk areas are marked with dark colors, medium-risk areas with medium-dark colors, and low-risk areas with light colors, visually representing the spatial distribution of different risk levels through color gradients. For triggering factors, a combination of text labels and arrows is used: the corresponding triggering factor, such as direct sunlight, poor ventilation, or soil drought, is labeled near each risk area, and an arrow connects it to the specific area affected by that factor. For example, if a high-risk area at the leaf tip is caused by direct sunlight, "direct sunlight" is labeled next to that area with an arrow pointing to it; if a medium-risk area at the base of the plant is caused by insufficient ventilation, "poor ventilation" is labeled and connected to the corresponding area. Simultaneously, a color legend and factor description table are attached to the right side of the map, clearly indicating the risk level corresponding to different colors and the meaning of each triggering factor. This ensures that the second plant risk map clearly displays the spatial distribution and severity of risks, and visually links the causes of each risk area, forming a complete visualization of heat stress risk.

[0069] Through expert evaluation and integration, coordinate system construction, risk mapping, and factor tracing, a second atlas was generated that can accurately present the distribution of plant heat stress risks and their causes, providing a clear basis for targeted mitigation of heat stress.

[0070] In one possible implementation, the plant risk third map acquisition module 40 further includes:

[0071] The system comprises the following components: a plant nutrient feature vector acquisition unit, which interprets nutrient features based on the plant spectral image set to obtain various plant nutrient feature vectors; a plant nutrient deficiency risk set acquisition unit, which assesses the risk of nutrient deficiency based on the plant nutrient feature vectors to obtain a plant nutrient deficiency risk set; a plant nutrient excess risk set acquisition unit, which assesses the risk of nutrient excess based on the plant nutrient feature vectors to obtain a plant nutrient excess risk set; a plant nutrient balance risk set acquisition unit, which assesses the risk of nutrient balance based on the plant nutrient feature vectors to obtain a plant nutrient balance risk set; and a plant risk third map acquisition unit, which traces inducing factors based on the plant nutrient deficiency risk set, the plant nutrient excess risk set, and the plant nutrient balance risk set to generate the plant risk third map.

[0072] Specifically, spectral images of individual plants are selected from a collection of plant spectral images and preprocessed to eliminate interference. Savitzky-Golay filtering is used to smooth the spectral curves, reducing the impact of high-frequency noise and improving the signal-to-noise ratio of the spectral data. Standard normal transformation is employed to correct baseline drift caused by differences in plant surface scattering, ensuring consistency across different samples. The preprocessed spectral images cover the 400-1000 nm band, containing reflectance and absorption information of plant leaves in the visible to near-infrared region.

[0073] Next, based on the principles of plant nutrition, characteristic wavelength bands related to nutrient elements were extracted. For example, nitrogen has characteristic absorption in the 700-900 nm wavelength band, phosphorus has an absorption peak in the 550-600 nm wavelength band, and chlorophyll (related to nitrogen content) has strong absorption near 680 nm and a high reflectance peak near 750 nm. For each plant sample, 300 spectral reflectance data points were extracted within these characteristic wavelength bands to form the original spectral feature matrix.

[0074] Then, a partial least squares regression algorithm was used to construct a correlation model between spectral features and nutrient content. Spectral data of plant samples with known nutrient contents, such as nitrogen, phosphorus, and potassium contents obtained through laboratory chemical analysis, were used as the training set and input into the model for parameter optimization. This enabled the model to invert nutrient element content through spectral reflectance, achieving a coefficient of determination (R²) of 0.88 and a prediction error controlled within 5%. The preprocessed plant spectral feature matrix was then input into the model to obtain the content values ​​of 12 major nutrients, including nitrogen, phosphorus, potassium, calcium, and magnesium, for the corresponding plants.

[0075] These nutrient content values ​​are then arranged in a preset order to form a vector containing 12 dimensions, namely the nutrient characteristic vector of each plant. For example, the nutrient characteristic vector of a certain shrub may include specific values ​​such as nitrogen: 3.2%, potassium: 2.8%, and phosphorus: 0.5%, comprehensively reflecting the nutritional status of the plant.

[0076] Next, the first plant nutrient feature vector is extracted based on each plant nutrient feature vector. Combined with its corresponding plant feature label, the historical record set of nutrient deficiency risk assessment is retrieved. Multiple assessment models are trained and fused to obtain the first nutrient deficiency risk assessment channel. The first plant nutrient deficiency risk coefficient is obtained by inputting this vector and added to the plant nutrient deficiency risk set. The specific steps are explained in detail from obtaining the sub-unit from the first plant nutrient feature vector to obtaining the sub-unit of the first plant nutrient deficiency risk coefficient.

[0077] Next, the first plant nutrient feature vector is extracted from each plant nutrient feature vector. The nutrient data of the individual plants that need to be analyzed are selected. This vector contains the content and ratio information of 12 nutrients such as nitrogen, phosphorus, and potassium. For example, in the vector of a certain red-leaf plum, the phosphorus content is 1.2% and the potassium content is 3.5%, clearly showing the current nutrient composition.

[0078] Based on the plant feature tags corresponding to the vector, such as red-leaf plum - growth period, the database of nutrient excess risk assessment is searched to select nutrient excess cases and assessment results of the same type of plant and the same stage, forming the first nutrient excess risk assessment record set. Each record has a nutrient feature vector and the corresponding excess risk level, such as mild, moderate and severe.

[0079] Next, the dataset was used to supervise the training of three models: Random Forest, Support Vector Machine, and Backpropagation Neural Network. 70% of the data served as the training set, and 30% as the validation set. Nutrient feature vectors were used as inputs, and excess risk levels were used as outputs. Training was stopped when the Random Forest achieved 90% accuracy, and the Support Vector Machine and Backpropagation Neural Network achieved 85% accuracy, resulting in three nutrient excess risk assessment models.

[0080] The model was trained by output fusion, with weights assigned according to accuracy: Random Forest 0.4, Support Vector Machine 0.3, and Backpropagation Neural Network 0.3. After cross-validation adjustments, the fusion evaluation accuracy reached 90%, forming the first nutrient excess risk assessment channel. The first plant nutrient feature vector was input into this channel to calculate the first plant nutrient excess risk coefficient, which was then added to the plant nutrient excess risk set.

[0081] The process for assessing plant nutrient balance risk is similar: After extracting the first plant nutrient feature vector, the historical data set of nutrient balance risk assessment is retrieved based on the corresponding tags. The three models are trained until the accuracy on the validation set reaches 85%-90%, and then fused to form the first nutrient balance risk assessment channel. After inputting the vector, the first plant nutrient balance risk coefficient is calculated based on the proportions between elements and added to the plant nutrient balance risk set.

[0082] Finally, by integrating three risk sets—nutrient deficiency, excess, and balance—the inducing factors were traced. The nutritional characteristics corresponding to each risk coefficient were analyzed, and combined with soil testing data and fertilization records, the causes were traced: deficiency risk may be caused by insufficient soil fertility and decreased root absorption function; excess risk is mostly related to excessive fertilization and excessive soil nutrient retention; balance risk is often caused by antagonistic effects between elements (such as excessive phosphorus inhibiting zinc absorption). These inducing factors and risk areas were associated and marked on the map, with different colors used to distinguish risk areas: deficiency was indicated in yellow, excess in red, and balance in blue. Inducing factor labels (such as nitrogen deficiency in soil, excessive phosphorus application, etc.) and corresponding risk coefficients were attached, generating a third plant risk map.

[0083] By employing a methodology consistent with nutrient deficiency risk assessment, we completed the risk assessment of nutrient excess and imbalance. Combined with the tracing of the causes of multi-dimensional risks, we generated a third atlas that comprehensively reflects the risks and underlying causes of plant nutrient status, providing a scientific basis for the precise regulation of plant nutrition.

[0084] In one possible implementation, the plant nutrient deficiency risk set acquisition unit further includes:

[0085] The first plant nutrient feature vector acquisition subunit extracts a first plant nutrient feature vector based on the plant nutrient feature vectors. The first nutrient deficiency risk assessment record set acquisition subunit performs a historical nutrient deficiency risk assessment retrieval based on the plant feature tags corresponding to the first plant nutrient feature vectors to obtain a first nutrient deficiency risk assessment record set. The nutrient deficiency risk assessment model acquisition subunit performs supervised training on multiple learning models based on the first nutrient deficiency risk assessment record set to obtain multiple nutrient deficiency risk assessment models. The first nutrient deficiency risk assessment channel acquisition subunit performs output fusion training based on the multiple nutrient deficiency risk assessment models to obtain a first nutrient deficiency risk assessment channel. The first plant nutrient deficiency risk coefficient acquisition subunit inputs the first plant nutrient feature vector into the first nutrient deficiency risk assessment channel to obtain a first plant nutrient deficiency risk coefficient, and adds the first plant nutrient deficiency risk coefficient to the plant nutrient deficiency risk set.

[0086] Specifically, firstly, the first plant nutrient feature vector is extracted from each plant nutrient feature vector, and representative single plant nutrient data is selected as the analysis object. This vector contains the content information of 12 nutrient elements such as nitrogen, phosphorus, potassium, and calcium. For example, in the nutrient feature vector of a certain osmanthus plant, the nitrogen content is 2.8% and the phosphorus content is 0.6%, reflecting the current nutrient composition of the plant.

[0087] Next, based on the plant feature tags corresponding to the first plant nutrient feature vector, such as Osmanthus - peak blooming period, a search is performed in the historical database of nutrient deficiency risk assessment. This database stores nutrient deficiency cases and corresponding assessment results for different plant species and growth stages. By matching tags, historical records related to Osmanthus - peak blooming period are filtered out to form the first nutrient deficiency risk assessment record set. This set contains nutrient feature vectors of multiple similar cases and corresponding deficiency risk levels, such as mild, moderate, and severe.

[0088] Then, supervised training of multiple learning models was conducted using the first nutrient deficiency risk assessment record set. Three models were selected: random forest, support vector machine, and backpropagation neural network. 70% of the data in the record set was used as the training set, and 30% as the validation set. The models used nutrient feature vectors as input and deficiency risk levels as output, and the model parameters were iteratively optimized. During training, the evaluation accuracy of the validation set was monitored in real time. Training was stopped when the accuracy of the random forest model reached 90%, and the accuracy of the support vector machine and backpropagation neural network reached over 85%, resulting in three nutrient deficiency risk assessment models.

[0089] Subsequently, multiple nutritional deficiency risk assessment models were trained by output fusion, and the evaluation results of each model were integrated using a weighted voting method. Weights were assigned based on the models' accuracy on the validation set: 0.4 for Random Forest, 0.3 for Support Vector Machine, and 0.3 for Backpropagation Neural Network. Through multiple rounds of cross-validation, the weight ratios were adjusted to improve the accuracy of the fused assessment to over 90%, forming the first nutritional deficiency risk assessment channel. This channel integrates the advantages of each model and reduces the evaluation bias of a single model.

[0090] Finally, the first plant nutrient feature vector is input into the first nutrient deficiency risk assessment channel. This channel calculates the deviation of each nutrient element content from the historical deficiency threshold and, combined with the evaluation logic of the three fusion models, outputs the first plant nutrient deficiency risk coefficient. For example, if the nitrogen content in an osmanthus sample is lower than the normal threshold of osmanthus during its full bloom period by 15%, the risk coefficient is calculated to be 0.72, ranging from 0 to 1, with higher values ​​indicating higher risk. This coefficient is then added to the plant nutrient deficiency risk set to complete the nutrient deficiency risk assessment for that plant.

[0091] By extracting feature vectors, retrieving historical data, and training and fusing multiple models, a precise risk assessment mechanism for nutrient deficiency was constructed, providing a reliable method for the quantitative assessment of plant nutrient deficiency risk and effectively supporting the generation of the third risk map of plants.

[0092] In one possible implementation, the machine vision module 10 further includes:

[0093] The garden plant monitoring set acquisition unit obtains a garden plant monitoring set based on a multimodal visual terminal, wherein the multimodal visual terminal includes a visible light imaging group, an infrared thermal imaging group, and a multispectral imaging group; the garden plant image set acquisition unit performs multi-level preprocessing based on the garden plant monitoring set to obtain the garden plant image set, wherein the multi-level preprocessing includes image enhancement, plant target segmentation, and perceptual modality classification.

[0094] Specifically, firstly, the multimodal vision terminal begins operation. The visible light imaging group collects data on the appearance characteristics of garden plants, such as leaf morphology and plant height, forming initial visible light monitoring data. Simultaneously, the infrared thermal imaging group captures the temperature distribution on the plant surface, generating a 1024×768 pixel temperature matrix, updated every 30 minutes to record the plant's real-time temperature status. The multispectral imaging group focuses on spectral information in the 400-1000nm wavelength band, collecting 300 spectral feature points for each plant to form spectral monitoring data. This real-time data from the three imaging groups is aggregated to form the garden plant monitoring set, comprehensively covering the plant's appearance, temperature, and spectral characteristics.

[0095] After obtaining the monitoring set of garden plants, image enhancement processing was first performed. For visible light images, a histogram equalization algorithm was used to improve the uneven illumination problem and enhance the clarity of leaf texture; for infrared images, Gaussian filtering was used to reduce noise and control the error of temperature data within ±0.5℃; for multispectral images, band calibration was performed to correct spectral drift caused by equipment deviation and ensure the consistency of spectral data.

[0096] After image enhancement, a plant target segmentation process is initiated for images in the garden plant monitoring set containing background elements such as soil, buildings, and rocks. First, the preprocessed images are input into the trained U-Net semantic segmentation model. This model performs multi-scale downsampling on the images through the encoder, gradually extracting key features such as plant edge contours, leaf textures, and branch morphology, while retaining spatial location information at different levels. Subsequently, the decoder upsamples based on the feature map output by the encoder, while fusing high-resolution features at the corresponding level to accurately restore the detailed information of the plant region. The model classifies each pixel in the image as either a plant or background through pixel-level classification, generating a binary segmentation mask, with plant region pixels marked as target value 1 and background pixels marked as background value 0. Then, morphological erosion and dilation processing removes small noise points and smooths edges, ultimately completely separating the target plant regions such as tree branches and leaves, and dense branches and leaves of shrubs from the complex background, providing clean image data without redundant interference for subsequent feature recognition and analysis of plants.

[0097] The U-Net semantic segmentation model is built on an encoder-decoder architecture: the encoder consists of multiple convolutional and pooling layers, which progressively downsamples and extracts multi-scale features from garden plant images, including detailed features such as leaf edges and branch textures, as well as high-level features such as the overall outline of the plant, while retaining feature maps at each level for skip connections; the decoder progressively upsamples through deconvolutional layers to restore low-resolution feature maps to the input image size, and fuses high-resolution features from the corresponding encoder levels through skip connections to enhance the accuracy of segmentation boundaries, finally outputting a segmentation mask with the same size as the input image through a 1x1 convolutional layer. During training, a set of visible light images of garden plants labeled with plants (including trees and shrubs) and background (including soil and buildings) is used as training data. After data augmentation such as rotation, scaling, and brightness adjustment, the images are input into the model. The difference between the predicted mask and the labeled mask is calculated using the cross-entropy loss function, and the model parameters are iteratively updated using the Adam optimizer. The segmentation accuracy is monitored in real time using a validation set, and training stops when the accuracy stabilizes above 95%. The model takes an enhanced visible light image of garden plants as input and outputs a binary segmentation mask, where the pixel value of the plant area is marked as 1 and the pixel value of the background area is marked as 0, thus achieving accurate separation of plants from the background.

[0098] Finally, perceptual modality classification is performed. Based on imaging principles and data characteristics, the processed images are classified as follows: appearance feature images are classified into the visible light image set of plants, temperature matrix images are classified into the infrared image set of plants, and spectral feature data are classified into the spectral image set of plants, ultimately forming a garden plant image set containing visible light images, infrared images, and spectral data.

[0099] Through the collaborative operation of real-time acquisition and multi-level preprocessing of multimodal vision terminals, a high-quality image set covering multiple dimensions of plant appearance, temperature, and spectrum was obtained, providing reliable data support for the system to subsequently trace risk characteristics such as growth accidents, heat stress, and nutrient status.

[0100] Example 2: Based on the inventive concept of the machine vision-based intelligent management system for garden plant health in the foregoing embodiments, this application also provides an electronic device, including: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the machine vision-based intelligent management system for garden plant health as described in any one of Examples 1 above.

[0101] Appendix Figure 2 This is a schematic diagram of the structure of an exemplary electronic device of this application. Figure 2In this document, the bus architecture is represented by bus 300. Bus 300 may include any number of interconnected buses and bridges, and bus 300 connects various circuits including one or more processors represented by processor 302 and memory represented by memory 304. Bus 300 may also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 305 provides an interface between bus 300 and receiver 301 and transmitter 303. Receiver 301 and transmitter 303 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 302 is responsible for managing bus 300 and general processing, while memory 304 can be used to store data used by processor 302 during operation.

[0102] In Embodiment 3, based on the same inventive concept as the machine vision-based intelligent management system for garden plant health in the foregoing embodiments, this application also provides a computer-readable storage medium storing a computer program, which, when executed, implements the machine vision-based intelligent management system for garden plant health described in any one of Embodiment 1 above.

[0103] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0104] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A machine vision-based intelligent management system for garden plant health, characterized in that, The system includes: The machine vision module is used to obtain a set of garden plant images, which includes a set of visible light images of plants, a set of infrared images of plants, and a set of spectral images of plants. The first plant risk map acquisition module is used to trace the risk characteristics of growth accidents based on the plant visible light image set and obtain the first plant risk map. The second plant risk map acquisition module is used to trace the heat stress risk characteristics based on the plant infrared image set and obtain the second plant risk map. The third plant risk map acquisition module is used to trace the nutritional status risk characteristics based on the plant spectral image set and obtain the third plant risk map. The garden plant health management execution module is used to perform garden plant health management based on the first plant risk map, the second plant risk map, and the third plant risk map. The step of tracing growth accident risk characteristics based on the plant visible light image set to obtain a first plant risk map includes: The plant growth state vector acquisition unit performs growth state feature recognition based on the plant visible light image set to obtain the growth state vector of each plant. The plant feature label construction unit performs species and stage feature recognition based on the plant visible light image set and constructs multiple plant feature labels. The plant growth accident tree model construction unit performs growth accident event mining and learning based on the multiple plant feature labels to construct multiple plant growth accident tree models. The plant growth accident path acquisition unit guides the multiple plant growth accident tree models to perform growth accident deduction on each plant growth state vector based on the multiple plant feature labels, and obtains multiple plant growth accident paths. The plant risk first map acquisition unit traces risk characteristics based on the multiple plant growth accident paths and generates the plant risk first map. 2.The machine vision-based intelligent garden plant health management system of claim 1, wherein, Based on the aforementioned multiple plant feature labels, growth accident event mining and learning are performed to construct multiple plant growth accident tree models, including: The first growth accident event set acquisition subunit extracts the first plant feature tag based on the multiple plant feature tags, and performs growth accident event retrieval based on the first plant feature tag to obtain the first growth accident event set. The accident top-level feature acquisition subunit performs top-level feature tracing based on the first grown accident event set to obtain the top-level features of each accident; The accident intermediate factor acquisition subunit traces the intermediate triggering factors of each accident based on the first growth accident event set to obtain the intermediate factors of each accident. The accident basic factor acquisition subunit traces the basic triggering factors of each accident intermediate factor based on the first growth accident event set to obtain each accident basic factor. The accident causal relationship path construction subunit constructs multiple accident causal relationship paths based on the top-level features of each accident, the intermediate factors of each accident, and the basic factors of each accident. The first plant growth accident tree model construction subunit strengthens the causal co-occurrence confidence based on the multiple accident causal relationship paths to generate the first plant growth accident tree model. 3.The machine vision-based intelligent garden plant health management system of claim 1, wherein, Based on the aforementioned plant infrared image set, heat stress risk characteristics are traced to obtain a second plant risk atlas, including: The plant temperature distribution map acquisition unit extracts pixel-level temperature distribution based on the plant infrared image set to obtain temperature distribution maps for each plant. The plant normal temperature range acquisition unit mines the normal temperature range based on multiple plant feature tags to obtain the normal temperature range for each plant. The plant temperature deviation matrix acquisition unit performs deviation detection on the temperature distribution map of each plant according to the normal temperature range of each plant, and obtains the temperature deviation matrix of each plant. The plant risk second map acquisition unit performs heat stress risk tracing based on the temperature deviation matrix of each plant and generates the plant risk second map.

4. The intelligent management system for garden plant health based on machine vision as described in claim 3, characterized in that, Based on the temperature deviation matrices of each plant, the source of heat stress risk is traced, and a second risk map of the plants is generated, including: The heat stress risk assessment set acquisition subunit was conducted by multiple plant experts to assess the heat stress risk of plant temperature history deviation sets, resulting in multiple heat stress risk assessment sets. The reliable thermal stress risk set acquisition sub-unit calculates the thermal stress risk concentration value based on each thermal stress risk assessment set to obtain the reliable thermal stress risk set; The heat stress risk assessment coordinate system construction sub-unit constructs a heat stress risk assessment coordinate system based on the plant temperature history deviation set and the reliable heat stress risk set. The heat stress risk assessment space acquisition subunit inputs each temperature deviation information in the temperature deviation matrix of each plant into the heat stress risk assessment coordinate system to obtain the heat stress risk assessment space; The second plant risk map acquisition subunit traces the initiating factors based on the heat stress risk assessment space to obtain the second plant risk map.

5. The intelligent management system for garden plant health based on machine vision as described in claim 1, characterized in that, Based on the plant spectral image set, nutrient status risk characteristics are traced to obtain a third plant risk map, including: The plant nutrient feature vector acquisition unit interprets nutrient features based on the plant spectral image set to obtain nutrient feature vectors for each plant. The plant nutrient deficiency risk set acquisition unit performs a nutrient deficiency risk assessment based on the plant nutrient feature vectors to obtain a plant nutrient deficiency risk set. The plant nutrient excess risk set acquisition unit performs a nutrient excess risk assessment based on each plant nutrient feature vector to obtain a plant nutrient excess risk set. The plant nutrient balance risk set acquisition unit performs a nutrient balance risk assessment based on the plant nutrient feature vectors to obtain the plant nutrient balance risk set. The plant risk third map acquisition unit traces the inducing factors based on the plant nutrient deficiency risk set, the plant nutrient excess risk set, and the plant nutrient balance risk set to generate the plant risk third map.

6. The intelligent management system for garden plant health based on machine vision as described in claim 5, characterized in that, Nutrient deficiency risk assessment is performed based on the aforementioned plant nutrient characteristic vectors to obtain a plant nutrient deficiency risk set, including: The first plant nutrient feature vector acquisition subunit extracts the first plant nutrient feature vector based on the plant nutrient feature vectors. The first nutrient deficiency risk assessment record set acquisition sub-unit performs a historical nutrient deficiency risk assessment retrieval based on the plant feature tags corresponding to the first plant nutrient feature vector to obtain the first nutrient deficiency risk assessment record set. The nutrient deficiency risk assessment model acquisition sub-unit performs supervised training on multiple learning models based on the first nutrient deficiency risk assessment record set to obtain multiple nutrient deficiency risk assessment models. The first nutritional deficiency risk assessment channel acquisition subunit performs output fusion training based on the multiple nutritional deficiency risk assessment models to obtain the first nutritional deficiency risk assessment channel. The first plant nutrient deficiency risk coefficient acquisition subunit inputs the first plant nutrient feature vector into the first nutrient deficiency risk assessment channel to obtain the first plant nutrient deficiency risk coefficient, and adds the first plant nutrient deficiency risk coefficient to the plant nutrient deficiency risk set.

7. The intelligent management system for garden plant health based on machine vision as described in claim 1, characterized in that, Obtain a collection of images of garden plants, including: The garden plant monitoring set acquisition unit obtains a garden plant monitoring set based on a multimodal visual terminal, wherein the multimodal visual terminal includes a visible light imaging group, an infrared thermal imaging group, and a multispectral imaging group. The garden plant image set acquisition unit performs multi-level preprocessing based on the garden plant monitoring set to obtain the garden plant image set. The multi-level preprocessing includes image enhancement, plant target segmentation, and perceptual modality classification.

8. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store the processor's executable instructions; The processor is used to execute the machine vision-based intelligent management system for the health of garden plants as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is used to execute the machine vision-based intelligent management system for the health of garden plants as described in any one of claims 1 to 7.

Images