Landscape plant precision maintenance control method based on intelligent perception and growth fitting

Abstract

The application discloses a landscape plant precision maintenance control method based on intelligent sensing and growth fitting, and particularly relates to the field of self-adaptive control and image processing, and is used for solving the control strategy degradation problem caused by the plant growth state being a hidden variable; the method is realized by the following steps: acquiring environment and image data of landscape plants and processing the data into a feature vector, obtaining a growth hidden state through state classification, calculating the distance between the feature vector and preset prototype points of each category in a feature space, determining that the growth hidden state is an ambiguous state if the feature vector is adjacent to at least two different category prototype points at the same time, subsequently extracting historical observation data before and after the execution of the last maintenance action, analyzing the time sequence change mode and calculating the deviation degree feature between the actual change track and the health expected response track, finally correcting the growth hidden state in combination with the time sequence change mode and the deviation degree feature, and generating a precise maintenance equipment control instruction according to the correction result.

Description

Technical Field

[0001] This invention relates to the fields of adaptive control and image processing technology, and more specifically, to a method for precise maintenance and control of landscape plants based on intelligent perception and growth fitting. Background Technology

[0002] In the refined maintenance and management of landscape plants, to improve the automation and precision of maintenance operations, existing technologies generally employ monitoring methods based on environmental sensing and image acquisition, combined with data-driven models to assist decision-making. Specifically, by deploying sensor networks within the maintenance area to acquire environmental parameters such as soil and weather, and using image acquisition devices to periodically capture images of the plant's appearance, this multi-source data is input into a pre-established growth analysis model or empirical rule base for processing. This process infers the plant's growth status and generates corresponding control commands for maintenance equipment such as irrigation and fertilization. The aim is to replace the traditional maintenance model that relies on manual experience with an automated system, thereby optimizing resource allocation and maintaining the plant's health.

[0003] However, in the existing technical solutions based on the fitting of sensor data and growth models, the core growth states of plants, such as physiological vitality, stress type and degree, are internal latent variables that cannot be directly measured by sensors. The system can only rely on observable external indirect data, such as image visual features and environmental sensor readings, to estimate them. This fundamental limitation caused by the characteristics of latent variables leads to the technical problem of control strategy degradation when the system derives control strategies based on observation data. Many different actual internal growth states may correspond to similar or identical observation data appearances, but the maintenance control commands generated by existing methods based on such observation data may tend to be the same, making it impossible to make effective distinctions and accurate responses. This causes the system's control actions to deviate from the plant's actual needs in key scenarios, thus restricting further improvement in maintenance accuracy. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for precise maintenance and control of landscape plants based on intelligent perception and growth fitting to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Precision maintenance and control methods for landscape plants based on intelligent sensing and growth fitting include: S1. Obtain observation data of landscape plants, including environmental data and appearance image data; S2. Process the appearance image data to extract visual feature parameters, combine them with environmental data to form an observation data sequence, and convert the observation data sequence into a corresponding feature vector; S3. Classify the feature vectors to obtain the current latent growth state of the landscape plants; S4. Calculate the distance between the feature vector and the preset prototype point of each growth latent state category in the feature space. If the distance between the feature vector and the prototype point of at least two different categories simultaneously satisfies the preset proximity condition, then the growth latent state is determined to be in an ambiguous state. S5. When in an ambiguous state, extract historical observation data within the time period before and after the last preset maintenance control action, analyze the temporal change pattern of the historical observation data, and calculate the deviation characteristics between the actual change trajectory of the historical observation data and the expected response trajectory based on the normal growth latent state. S6. Correct the latent growth state by combining the temporal change pattern and deviation characteristics, and generate maintenance equipment control instructions based on the corrected latent growth state.

[0006] Furthermore, observational data on landscape plants are acquired, including environmental data and visual image data, including: Obtain environmental data of the environment in which the landscape plants are located, including soil moisture data, ambient temperature data, and light intensity data; The appearance image data of the landscape plants is acquired by an image acquisition device. Environmental data and visual imagery data together constitute the observation data.

[0007] Furthermore, the image acquisition device acquires appearance image data according to a preset cycle, wherein the preset cycle is dynamically adjusted according to the growth stage of the landscape plants. A first acquisition frequency is used during the vigorous growth stage, and a second acquisition frequency is used during the slow growth stage. The first acquisition frequency is higher than the second acquisition frequency.

[0008] Furthermore, the appearance image data is processed to extract visual feature parameters, which are then combined with environmental data to form an observation data sequence. This observation data sequence is then converted into a corresponding feature vector, including: Color space transformation and region segmentation were performed on the appearance image data to extract color histogram features representing plant canopy color and texture features representing leaf structure. The color histogram features and texture features were used as visual feature parameters. The visual feature parameters, along with soil moisture data, ambient temperature data, and light intensity data from the environmental data, are arranged in chronological order of collection time to form an observation data sequence. The visual feature parameters and environmental data in the observation data sequence are normalized, and the normalized data are spliced ​​together in a fixed order to form a feature vector.

[0009] Furthermore, the feature vectors are classified into states to obtain the current latent growth states of the landscape plants, including: Input the feature vectors into a pre-trained classification model; The classification model outputs a feature vector that belongs to multiple preset latent growth state categories; The category of the latent growth state with the highest membership degree is determined as the current latent growth state of the landscape plant.

[0010] Further, the distance between the feature vector and the preset prototype point of each growth latent state category in the feature space is calculated. If the distance between the feature vector and the prototype point of at least two different categories simultaneously satisfies the preset proximity condition, the growth latent state is determined to be in an ambiguous state, including: Obtain the pre-defined prototype points of each latent growth state category in the feature space; Calculate the Euclidean distance between the feature vector and each prototype point; If the Euclidean distance between the feature vector and at least two prototype points of different classes is less than a preset distance threshold, then the growth hidden state is determined to be in an ambiguous state.

[0011] Furthermore, the prototype points are determined as follows: historical observation data are collected and their corresponding growth latent state categories are labeled; cluster analysis is performed on the feature vectors corresponding to the historical observation data belonging to the same category; and the cluster center points obtained from the cluster analysis are used as the prototype points of the corresponding growth latent state categories.

[0012] Furthermore, when in an ambiguous state, historical observation data from the period before and after the last preset maintenance control action is extracted, the temporal change pattern of the historical observation data is analyzed, and the deviation characteristics between the actual change trajectory of the historical observation data and the expected response trajectory based on the normal growth latent state are calculated, including: Extract a sequence of historical observation data from the historical records, within a continuous period before and after the last irrigation or fertilization action; Time series analysis was performed on soil moisture data and visual characteristic parameters in historical observation data series to obtain the trends of soil moisture change and canopy color change. Based on the plant growth model corresponding to the normal growth latent state, the expected response trajectories of soil moisture and canopy color are generated within the same historical period. The first deviation between the soil moisture change trend and the corresponding expected response trajectory, and the second deviation between the canopy color change trend and the corresponding expected response trajectory are calculated separately. The first deviation and the second deviation are used together as the deviation feature.

[0013] Furthermore, the latent growth state is corrected by combining the temporal variation pattern and deviation characteristics, and control instructions for maintenance equipment are generated based on the corrected latent growth state, including: A comprehensive feature vector is constructed based on the soil moisture change trend, the canopy color change trend, the first deviation, and the second deviation. The comprehensive feature vector is input into a preset decision rule, which defines the mapping relationship between different feature combinations and each latent state category. Based on the output of the decision rule, the current growth hidden state is updated to the growth hidden state category pointed to by the mapping relationship, and the corrected growth hidden state is obtained. Based on the corrected latent growth state, the system queries the preset control strategy table to obtain the corresponding irrigation or fertilization parameters and generates control commands.

[0014] Furthermore, the pre-defined decision rules are established as follows: historical data samples containing soil moisture change trends, canopy color change trends, first deviation, second deviation, and finally confirmed latent growth state categories are collected; based on the historical data samples, a classifier is trained to learn the mapping relationship from the comprehensive feature vector to the latent growth state categories; and the discrimination logic contained in the trained classifier is solidified into decision rules.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing a dynamic determination mechanism for ambiguous states based on feature space and combining it with the fusion correction of multi-source temporal evidence, the reliability of latent state estimation for landscape plant growth is significantly improved. After preliminary state classification, the method actively assesses the position of the estimated state in feature space and its proximity to prototype points of each category, thereby identifying ambiguous states caused by the similarity of observation data. When an ambiguous state is determined, the method automatically triggers a deep diagnostic process based on historical control action response analysis, extracts the actual response pattern of the plant to the previous maintenance intervention, and compares it with the expected theoretical trajectory under healthy conditions to obtain quantified deviation features. This effectively distinguishes the intrinsic growth states that are similar in static observation but have different dynamic responses, avoiding the problem of control strategy degradation caused by state misjudgment. It can generate differentiated control instructions for different real physiological conditions, enhancing the robustness and accuracy of the state identification process.

[0016] 2. Capable of autonomously switching operating modes based on the confidence level of state estimation: When the estimation results are clear, an efficient path is adopted; when the estimation results are ambiguous, a more complex and multi-source verification mechanism is automatically invoked for state correction. This dynamic decision-making logic based on confidence assessment demonstrates excellent adaptability and stability when facing complex and ever-changing natural environments and individual plant differences. By using image visual features as one of the key dimensions of state perception and performing spatiotemporal correlation and fusion analysis with soil environmental sensor data, a comprehensive understanding of plant phenotypes and growth microenvironment is obtained, supporting more accurate model fitting and prediction. This not only improves the accuracy of single control actions but also optimizes the adaptive adjustment capability of long-term maintenance strategies through continuous state correction, achieving a dynamic optimal match between maintenance resource allocation and the actual needs of plants. Attached Figure Description

[0017] Figure 1 This is a flowchart of the landscape plant precision maintenance and control method based on intelligent sensing and growth fitting according to the present invention. Figure 2 This is a flowchart for determining whether a growth latent state is in an ambiguous state according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Figure 1 This invention presents a precise maintenance and control method for landscape plants based on intelligent sensing and growth fitting, comprising: S1. Obtain observation data of landscape plants, including environmental data and appearance image data; S2. Process the appearance image data to extract visual feature parameters, combine them with environmental data to form an observation data sequence, and convert the observation data sequence into a corresponding feature vector; S3. Classify the feature vectors to obtain the current latent growth state of the landscape plants; S4. Calculate the distance between the feature vector and the preset prototype point of each growth latent state category in the feature space. If the distance between the feature vector and the prototype point of at least two different categories simultaneously satisfies the preset proximity condition, then the growth latent state is determined to be in an ambiguous state. S5. When in an ambiguous state, extract historical observation data within the time period before and after the last preset maintenance control action, analyze the temporal change pattern of the historical observation data, and calculate the deviation characteristics between the actual change trajectory of the historical observation data and the expected response trajectory based on the normal growth latent state. S6. Correct the latent growth state by combining the temporal change pattern and deviation characteristics, and generate maintenance equipment control instructions based on the corrected latent growth state.

[0020] S1. Obtain observation data of landscape plants, including environmental data and appearance image data. The specific implementation is as follows: To acquire observation data of landscape plants, soil moisture sensors, ambient temperature sensors, and light intensity sensors are installed in the area where the landscape plants are located to form an environmental data acquisition network. The soil moisture sensors are either capacitive or frequency domain reflectometers. The probes of the soil moisture sensors are vertically inserted into the soil layer where the main root system of the plants is distributed, at a depth of, for example, 15 to 20 centimeters, to sense the volumetric water content of the soil in the root activity area. The soil moisture sensors measure at fixed time intervals, for example, 10 minutes. The output signal of the soil moisture sensors is converted from analog to digital to obtain soil moisture data in percentage form. The ambient temperature sensor is a digital temperature sensor. The sensors include an ambient temperature sensor installed near the plant canopy in a well-ventilated and shaded area, with a measurement interval set to 10 minutes, and an output temperature data unit of degrees Celsius. A light intensity sensor, using a photoresistor or photodiode sensor, is installed above the plant canopy in an unobstructed location, with a measurement interval set to 10 minutes, and an output light intensity data unit of lux. The soil moisture sensor, ambient temperature sensor, and light intensity sensor transmit the collected soil moisture data, ambient temperature data, and light intensity data to the central processing unit or data storage unit in real time via wired or wireless communication.

[0021] The acquisition of landscape plant appearance images is achieved through an image acquisition device. This device utilizes a webcam or industrial camera with autofocus, and is fixedly mounted on a support capable of clearly capturing the entire canopy of the target plant. The acquisition process is controlled by a preset periodic trigger signal. This preset period is dynamically adjusted according to the growth stage of the landscape plant. The growth stage is determined based on the phenological calendar corresponding to the plant species, or by analyzing the rate of change in plant height and canopy size in historical image sequences. For example, for garden shrubs, the rapid shoot growth period in spring and the vigorous growth period in summer are classified as the vigorous growth stage. During the vigorous growth stage, the image acquisition device uses a higher initial acquisition frequency. The image acquisition device operates at a higher frequency. For example, the first acquisition frequency is set to collect data twice a day, and the trigger time for the first acquisition frequency is set at 9:00 AM and 3:00 PM respectively. The autumn growth slowdown period and the winter dormancy period are divided into slow growth stages. During the slow growth stage, the image acquisition device operates at a lower second acquisition frequency, for example, set to collect data once a day, and the trigger time for the second acquisition frequency is set around noon. The specific value of the first acquisition frequency is higher than the specific value of the second acquisition frequency. After receiving the trigger signal, the image acquisition device automatically captures color digital images of the landscape plants. The image acquisition device stores the captured color digital images as appearance image data along with timestamp information.

[0022] The acquired soil moisture data, ambient temperature data, light intensity data, and appearance image data are aligned and correlated using their respective timestamp information. The central processing unit or data recording unit packages the soil moisture data, ambient temperature data, light intensity data, and appearance image data collected at the same time point or within the same time period into a set of time-stamped raw data packets. These raw data packets together constitute the raw observation data set for subsequent processing. Soil moisture data, ambient temperature data, and light intensity data serve as quantitative indicators directly characterizing the microenvironmental conditions for plant growth, while appearance image data serves as a visual information source recording the plant's morphology. Together, these data provide a multi-dimensional data foundation for a comprehensive assessment of plant growth status. All observation data are stored in non-volatile memory, forming a time-indexed historical database. Soil moisture data... The specific values ​​are based on the actual measurements of the soil moisture sensor, the environmental temperature data, and the light intensity data. The specific content of the appearance image data is based on the pixel matrix actually captured by the image acquisition device. The composition of the observation data is the result of logically associating and storing the soil moisture data, environmental temperature data, light intensity data, and appearance image data sets that have a time synchronization relationship. The process of the image acquisition device acquiring appearance image data according to a preset cycle is based on the dynamic adjustment logic of the preset cycle. That is, the first acquisition frequency and the second acquisition frequency are switched according to an external growth stage determination signal. The growth stage determination signal comes from a preset calendar time or from a stage marker generated after analyzing historical growth data, thereby realizing the adaptive matching between the data acquisition frequency and the intensity of the plant's own physiological activities.

[0023] S2. Process the appearance image data to extract visual feature parameters, combine it with environmental data to form an observation data sequence, and convert the observation data sequence into a corresponding feature vector. The specific implementation is as follows: The process of processing landscape plant appearance image data to extract visual feature parameters begins with digital image processing of the acquired landscape plant appearance image data. Landscape plant appearance image data is a color digital image matrix stored in the RGB color space. Color space conversion transforms the landscape plant appearance image data from the RGB color space to the HSV or Lab color space. The purpose of color space conversion is to better distinguish the green vegetation of the plant canopy from the soil background or other non-plant areas. In the converted HSV color space, the value range of the hue component H is used to distinguish green vegetation pixels. The identification condition for green vegetation pixels is, for example, set to a hue component H value between 60 and 180. Based on the vegetation pixel mask obtained from color segmentation, the original landscape plant appearance image data is used for region segmentation. Region segmentation extracts the region of interest image that mainly contains the plant canopy. Region segmentation avoids interference from soil and background in subsequent feature extraction.

[0024] Color histogram features are extracted from the segmented plant canopy region image. The calculation of color histogram features targets the hue component H channel in the HSV color space. The calculation process divides the hue component H value range of 0 to 360 degrees into several equally spaced intervals, for example, 36 intervals, with each interval being 10 degrees wide. The calculation process counts the number of pixels within the plant canopy region whose hue component H value falls within each interval. The count is then divided by the total number of pixels in the plant canopy region to obtain a normalized hue distribution histogram. This normalized hue distribution histogram is the color histogram feature. The color histogram feature is a multi-dimensional vector; its dimension is equal to the number of intervals, for example, 36. The color histogram feature characterizes the composition and distribution of colors in the plant canopy.

[0025] Texture features are extracted from the segmented plant canopy region image. The gray-level co-occurrence matrix (GLCM) method is used to calculate the texture features. First, the plant canopy region image is converted to a grayscale image. A GLCM is then calculated on the grayscale image, describing the joint probability of a pair of gray levels occurring in the image under a specific spatial relationship. The spatial relationship is defined as the distance of one pixel in the 0-degree direction. Gray levels are quantized into multiple levels, for example, 16. Based on the calculated GLCM, a set of statistics is calculated as texture features. These statistics include contrast, correlation, energy, and homogeneity. Contrast reflects the sharpness of the image and the depth of texture grooves; correlation measures the consistency of the image texture; energy reflects the uniformity of the image texture; and homogeneity describes the local uniformity of the image texture. The values ​​of these four statistics—contrast, correlation, energy, and homogeneity—are combined into a four-dimensional vector, which serves as the texture feature. The texture features characterize the structural roughness and pattern regularity of the plant leaf surface.

[0026] The extracted color histogram features and texture features are used as visual feature parameters. A visual feature parameter is a composite feature vector that combines color and texture information. For example, a visual feature parameter can be constructed by concatenating a 36-dimensional vector of color histogram features with a 4-dimensional vector of texture features to form a 40-dimensional visual feature parameter vector.

[0027] Visual feature parameters, along with soil moisture, ambient temperature, and light intensity data from the environmental data, are arranged chronologically to form an observation data sequence. The specific process involves retrieving the corresponding soil moisture, ambient temperature, light intensity, and visual feature parameters from the historical database for each specific data collection point. Soil moisture is represented as a percentage, ambient temperature as a degree Celsius, light intensity as a lux value, and the visual feature parameter as a multidimensional vector. Data from the same time point are grouped into a single data unit. Multiple consecutive time point data units are then arranged chronologically to form the observation data sequence. This observation data sequence is a multivariate time series data set that varies over time.

[0028] Normalization was performed on visual feature parameters and environmental data in the observation data sequence. The purpose of normalization is to eliminate differences in units and numerical ranges among different feature data. Normalization was performed separately for soil moisture data, ambient temperature data, light intensity data, and visual feature parameters. For soil moisture data, a minimum-maximum normalization method was used. Soil moisture data normalization requires pre-calculating the minimum and maximum values ​​of soil moisture data from historical data. For example, the minimum soil moisture data might be 10%, and the maximum might be 60%. The normalization operation involves subtracting the minimum value of 10% from the current soil moisture data and then dividing by the difference between the maximum value (60%) and the minimum value (10%) (50%). This normalization operation ensures that the normalized soil moisture data falls within the range of 0 to 1.

[0029] For ambient temperature data, a minimum-maximum normalization method is used. Normalization requires pre-calculating the minimum and maximum values ​​of the ambient temperature data from historical data. For example, if the minimum ambient temperature is 5 degrees Celsius and the maximum is 40 degrees Celsius, the normalization operation involves subtracting the minimum value of 5 degrees Celsius from the current ambient temperature data and then dividing by the difference between the maximum value of 40 degrees Celsius and the minimum value of 5 degrees Celsius (35 degrees Celsius). This normalization operation ensures that the normalized ambient temperature data falls within the range of 0 to 1.

[0030] For light intensity data, a minimum-maximum normalization method is used. Normalization requires pre-calculating the minimum and maximum values ​​of light intensity data from historical data. For example, the minimum value of light intensity data might be 0 lux, and the maximum value might be 100,000 lux. The normalization operation involves subtracting the minimum value of 0 lux from the current light intensity data and then dividing by the difference between the maximum value of 100,000 lux and the minimum value of 0 lux, which is 100,000 lux. This normalization operation ensures that the normalized light intensity data falls within the range of 0 to 1.

[0031] For the color histogram feature component in the visual feature parameters, the color histogram feature component is already a normalized pixel probability distribution. The value of each dimension of the color histogram feature component ranges from 0 to 1 and the sum is 1. Therefore, the color histogram feature component does not undergo additional normalization. For the texture feature component in the visual feature parameters, each statistic of the texture feature component is normalized by minimum and maximum values. The normalization of each texture statistic needs to be based on the minimum and maximum values ​​of the historical data for that statistic.

[0032] Normalized soil moisture data, ambient temperature data, light intensity data, color histogram feature components, and texture feature components are concatenated in a fixed order to form a feature vector. This fixed order is defined as follows: first, normalized soil moisture data; second, normalized ambient temperature data; third, normalized light intensity data; fourth, all dimensions of the color histogram feature components; and finally, all dimensions of the texture feature components. For example, if the color histogram feature is 36-dimensional and the texture feature is 4-dimensional, the total dimension of the concatenated feature vector is 1+1+1+36+4=43. This feature vector serves as a structured numerical representation of the overall state of landscape plants at a specific time point and is used for subsequent state classification processing.

[0033] S3. Classify the feature vectors to obtain the current latent growth state of the landscape plants. The specific implementation is as follows: The process of classifying feature vectors to obtain the current latent growth state of landscape plants relies on a pre-trained classification model. A pre-trained classification model is a machine learning model that has already learned parameters using historical data and can classify newly input feature vectors. Obtaining a pre-trained classification model requires a model training phase. This phase begins with collecting historical observation data and labeling the corresponding latent growth state categories. Historical observation data comes from stored data accumulated during the long-term operation of the landscape plant maintenance system. This data includes soil moisture, ambient temperature, light intensity, and appearance image data collected at multiple past time points. The appearance image data undergoes the same processing procedure as in step S2 to extract visual feature parameters. These parameters, along with the environmental data, constitute a historical observation data sequence. This sequence then undergoes the same transformation process as in step S2 to obtain historical feature vectors. Labeling the latent growth state categories requires the intervention of horticultural experts. These experts determine a true latent growth state category for each historical feature vector based on historical records of actual plant growth, pest and disease occurrences, and maintenance logs. The preset latent growth state categories are a set of mutually exclusive state labels predefined based on common landscape plant maintenance problems. The preset latent growth state categories include, for example, healthy state, mild water shortage state, severe water shortage state, fertilizer deficiency state, and pest and disease infection state.

[0034] The model training phase requires preparing a training dataset, which consists of a large number of labeled historical feature vectors and their corresponding latent state class labels. The training dataset is divided into a training subset and a validation subset. The training subset is used to directly adjust the internal parameters of the classification model. The validation subset is used to evaluate the generalization performance of the classification model and prevent overfitting. The classification model can be implemented using various types of machine learning algorithms, such as Support Vector Machines, Random Forests, or Feedforward Neural Networks. Taking the Random Forest algorithm as an example, it completes the classification task by constructing multiple decision trees and performing ensemble voting. During training, the Random Forest algorithm requires configuring hyperparameters. Hyperparameters include, for example, the number of decision trees in the forest, the maximum number of features considered when splitting nodes in each decision tree, and the maximum depth of the trees. The specific values ​​of the hyperparameters are determined using a grid search on the validation subset using cross-validation. The grid search attempts different combinations of hyperparameters and evaluates the classification accuracy of each combination on the validation subset. The grid search selects the hyperparameter combination with the highest classification accuracy on the validation subset as the final hyperparameter setting.

[0035] The random forest model is trained using a training subset and defined hyperparameters. During training, each decision tree independently samples with replacement from the training subset to form a bootstrap sample set. Each decision tree grows using the bootstrap sample set. At each node of the decision tree, a subset of features is randomly selected from all features. The optimal split point is chosen from the randomly selected feature subset to generate the branch of the decision tree. After training, a random forest model containing multiple decision trees is obtained. This trained random forest model is the pre-trained classification model. The pre-trained classification model is persistently saved as a data file for loading and use when the landscape plant precision maintenance control method is implemented.

[0036] When a feature vector is input into a pre-trained classification model, the feature vector is the 43-dimensional numerical vector output from step S2. The classification model receives the feature vector as input. The internal computation process of the classification model varies depending on the model type. For a trained random forest model, the input feature vector is fed into each decision tree in the forest. Each decision tree makes judgments layer by layer from the root node according to its internal splitting rules. The decision tree eventually assigns the feature vector to a leaf node. This leaf node corresponds to a vote for a growing hidden state class. The classification voting results of all decision trees in the random forest for the feature vector are collected. The number of votes received by each growing hidden state class is counted.

[0037] The classification model outputs the membership degree of a feature vector belonging to multiple pre-defined latent growth state categories. Membership degree is a probability value or score. It represents the likelihood that a feature vector belongs to a specific latent growth state category. For the random forest model, membership degree can be calculated by dividing the number of votes received for each latent growth state category by the total number of decision trees. For example, if a random forest contains 100 decision trees and receives 80 votes for the healthy state category, then the membership degree of the feature vector belonging to the healthy state category is 80 / 100 = 0.8. The classification model ultimately outputs a membership degree vector. The length of the membership degree vector is equal to the number of pre-defined latent growth state categories. Each element of the membership degree vector corresponds to the membership degree value of a category. The sum of the membership degree values ​​of all categories is 1.

[0038] The latent growth state category with the highest membership degree is determined as the current latent growth state of the landscape plant. This determination process involves comparing all membership degree values ​​in the membership degree vector. The element with the largest membership degree value is found. The latent growth state category corresponding to this element is the category with the highest membership degree. For example, the preset latent growth state categories include five types: healthy, slightly water-deficient, severely water-deficient, nutrient-deficient, and pest / disease-infected. The classification model outputs a membership degree vector of 0.7 for healthy, 0.1 for slightly water-deficient, 0.1 for severely water-deficient, 0.05 for nutrient-deficient, and 0.05 for pest / disease-infected. The healthy state has the highest membership degree of 0.7. Therefore, the healthy state is determined as the current latent growth state of the landscape plant. In rare cases, two or more categories may have the same highest membership degree value. In such cases, a predefined rule can be used for selection. This predefined rule could be, for example, selecting the category ranked higher in the preset category list, or marking this situation as a tie requiring special handling. The final determined latent growth state is output as a discrete state label. This latent growth state represents the classification model's optimal assessment of the landscape plant's health status based on the current feature vector. This latent growth state will serve as the basis for subsequent steps involving ambiguity resolution and potential correction.

[0039] Figure 2 The flowchart for determining whether a growth latent state is in an ambiguous state is given in this invention. S4: Calculate the distance between the feature vector and the preset prototype point of each growth latent state category in the feature space. If the distance between the feature vector and the prototype point of at least two different categories simultaneously satisfies the preset proximity condition, then the growth latent state is determined to be in an ambiguous state. The specific implementation is as follows: The process of calculating the distance between the feature vector and the pre-defined prototype points of each latent growth state category in the feature space to determine whether a latent growth state is in an ambiguous state first requires obtaining the pre-defined prototype points of each latent growth state category in the feature space. The pre-defined prototype points of each latent growth state category in the feature space refer to a representative point for each latent growth state category in the feature space. This representative point summarizes the core position of typical samples belonging to that category in the feature space. Obtaining the pre-defined prototype points of each latent growth state category in the feature space requires pre-determined and stored prototype point coordinate data. The prototype point coordinate data is generated through an offline prototype point determination process and stored in the system for later retrieval. The prototype point determination process begins by collecting historical observation data and labeling its corresponding latent growth state category. The source of the historical observation data is the same as the historical data source used in the model training phase of step S3. The historical observation data is a long-term accumulated observation data record with annotations from horticultural experts by the landscape plant maintenance system. For each set of historical observation data, it is converted into the corresponding historical feature vector by performing the same data acquisition and processing flow as steps S1 and S2. Each historical feature vector is associated with a latent growth status category label confirmed by a horticultural expert based on the original records. This latent growth status category label is one of several pre-defined mutually exclusive categories, such as a healthy status label, a slightly water-deficient status label, a severely water-deficient status label, a nutrient-deficient status label, or a pest or disease infection status label.

[0040] Cluster analysis is performed on feature vectors corresponding to historical observation data belonging to the same category. Cluster analysis is an unsupervised machine learning technique. The purpose of cluster analysis is to divide a set of feature vectors into several internally similar subgroups based on their similarity to each other. Cluster analysis first requires choosing a specific clustering algorithm, such as K-means clustering. For each latent growth state category, the expected number of clusters K for cluster analysis in that category needs to be set. The expected number of clusters K can be set empirically or determined based on historical data using heuristic methods such as the elbow rule. For example, for the healthy state category, the expected number of clusters K can be set to 1. For the nutrient deficiency category, the expected number of clusters K can be set to 2. The K-means clustering algorithm is used to analyze all historical feature vectors belonging to the same category. The K-means clustering algorithm randomly initializes K cluster centroids. The K-means clustering algorithm then iteratively executes two steps until the cluster centroids no longer change significantly. The first step is to assign each historical feature vector to the cluster represented by its nearest cluster centroid. The second step is to recalculate the average of all historical feature vectors in each cluster. The average value is used as the new cluster centroids. After iteration, the K-means clustering algorithm outputs K final cluster centroids and the cluster label to which each historical feature vector belongs. These cluster centroids represent different data sub-patterns under this latent state category.

[0041] Cluster centroids obtained from cluster analysis are used as prototype points for the corresponding latent growth states. This means that one latent growth state can correspond to one or more prototype points. For example, the healthy state might generate only one prototype point after cluster analysis, while the nutrient deficiency state might generate two. Each prototype point is a numerical vector with the same dimension as the feature vector. The coordinates of the prototype point are the coordinates of its corresponding cluster centroid. The coordinates of all prototype points for all preset latent growth states are organized and stored in a prototype point lookup table. The prototype point lookup table records a unique identifier for each prototype point, its corresponding latent growth state, and its specific coordinates in the feature space. Obtaining the preset prototype points for each latent growth state in the feature space involves retrieving the coordinate information of all prototype points from this prototype point lookup table when needed.

[0042] Calculate the Euclidean distance between the feature vector and each prototype point. The feature vector is a multidimensional numerical vector output from step S2, such as a 43-dimensional vector. Each prototype point is also a multidimensional numerical vector with the same dimension. The Euclidean distance is the straight-line distance between two points in multidimensional space. Calculating the Euclidean distance between the feature vector and a prototype point requires performing a series of mathematical operations. First, calculate the difference between the value of each dimension of the feature vector and the corresponding value of the prototype point. Then, square the difference of each dimension. Next, sum the squared differences of all dimensions to obtain a total. Finally, take the square root of this sum. The result of the square root operation is the Euclidean distance between the feature vector and the prototype point. The Euclidean distance is a non-negative scalar value. The smaller the Euclidean distance value, the closer the feature vector is to the prototype point in multidimensional space. The larger the Euclidean distance value, the farther the feature vector is from the prototype point. For each prototype point in the prototype point lookup table, the above calculation process needs to be repeated to obtain a set of distance values. Each distance value in this set corresponds to the Euclidean distance between the feature vector and a specific prototype point.

[0043] If the Euclidean distance between a feature vector and at least two prototype points of different classes is less than a preset distance threshold, the latent state is considered ambiguous. The preset distance threshold is a critical value used to determine proximity. Setting the preset distance threshold requires consideration of the scale of the feature space and a business understanding of the concept of proximity. The preset distance threshold can be determined by analyzing the distance distribution from feature vectors to their correct class prototype points in historical data. For example, a large number of historically correctly classified samples can be collected. The minimum Euclidean distance from the feature vector of each sample to all prototype points of its class can be calculated. The statistical distribution of these minimum Euclidean distances can be analyzed, such as calculating their mean and standard deviation. The preset distance threshold can be set to the mean plus a certain multiple of the standard deviation. For example, the preset distance threshold can be set to the mean plus one standard deviation. Alternatively, the preset distance threshold can be set to the distance value corresponding to a certain percentile, such as the distance value corresponding to the 75th percentile. The preset distance threshold is a parameter that needs to be preset and may be adjusted empirically; for example, it might be set to 0.5.

[0044] The specific execution of the judgment logic is as follows: Iterate through all calculated Euclidean distances. Identify distance items whose Euclidean distance values ​​are less than a preset distance threshold. Check the prototype points corresponding to these distance items that meet the condition of being less than the preset distance threshold. Analyze the growth latent state category to which these prototype points belong. If these prototype points belong to two or more different growth latent state categories, the judgment condition is met. For example, the Euclidean distance between the feature vector and a prototype point belonging to the healthy state category is 0.4. The Euclidean distance between the feature vector and a prototype point belonging to the mildly dehydrated state category is 0.45. The preset distance threshold is 0.5. Then 0.4 and 0.45 are both less than 0.5. The prototype points corresponding to 0.4 and 0.45 belong to the two different categories of healthy state and mildly dehydrated state. Therefore, the growth latent state is judged to be in an ambiguous state. This means that the current feature vector is close to the typical regions of two different categories in the feature space at the same time, making it difficult for the classification model in step S3 to clearly distinguish them. Its output growth latent state estimation results have low reliability and are prone to confusion. The result of determining that the latent state of growth is ambiguous is output as a logical flag. This logical flag will trigger a more in-depth analysis process in subsequent step S5. If the feature vector is only close to multiple prototype points of the same class, or only close to prototype points of one class, it is not determined to be an ambiguous state.

[0045] When step S4 determines that the growth latent state is not ambiguous, it indicates that the current feature vector clearly belongs to a typical region of a single growth latent state category in the feature space, and the preliminary state classification result of step S3 has high confidence. In this case, the system will not initiate the deep analysis process of step S5, but will directly adopt the growth latent state output by step S3 as the final valid state. Subsequently, the system directly queries the preset control strategy table based on the growth latent state, obtains the corresponding control parameters, and generates maintenance equipment control commands, thereby completing an efficient decision-making path.

[0046] S5. When in an ambiguous state, extract historical observation data from the time periods before and after the last preset maintenance control action, analyze the temporal change pattern of the historical observation data, and calculate the deviation characteristics between the actual change trajectory of the historical observation data and the expected response trajectory based on the normal growth latent state. Specifically, this is implemented as follows: When step S4 determines that the latent growth state is ambiguous, the system initiates in-depth analysis of historical observation data to obtain additional discriminative basis. The system needs to extract a sequence of historical observation data from the historical records, representing a continuous period before and after the execution time of the last preset maintenance control action. Preset maintenance control actions refer to standardized maintenance operations automatically executed by the system or triggered by operators through the system. Types of preset maintenance control actions include irrigation and fertilization. Historical records are a collection of observation data indexed by time, stored in the system's non-volatile memory. The extraction process requires determining two key time points. The first key time point is the precise moment of the last preset maintenance control action execution. This precise moment is obtained from the system's control command execution log. The second key time point is the start and end point of an analysis time window defined around this execution moment. The span of the analysis time window needs to be set based on the typical response time of plants to maintenance actions. For example, for irrigation actions, the analysis time window can be set to 2 hours before the action execution to 6 hours after. For fertilization actions, the analysis time window can be set to 1 day before the action execution to 3 days after. The system retrieves and extracts all observation data entries whose timestamps fall within the analysis time window from historical records, based on the start and end times of that window. These observation data entries are arranged in chronological order by timestamp, forming a continuous historical observation data sequence. Each data entry in the historical observation data sequence contains soil moisture data, ambient temperature data, light intensity data, and visual characteristic parameters for that moment.

[0047] Time series analysis was performed on soil moisture data and visual characteristic parameters from historical observation data sequences. The purpose of time series analysis is to extract meaningful trends or patterns from observations that change over time. For soil moisture data, time series analysis first arranges the soil moisture data in chronological order to form a soil moisture time series. The soil moisture time series may contain fluctuations due to sensor noise or transient interference. To clearly reveal the overall direction of soil moisture change within the analysis time window, the soil moisture time series needs to be smoothed. Smoothing can be done using the moving average method. The window size of the moving average is set, for example, to three consecutive data points. The moving average calculates the smoothed value for each time point, which is the average of the soil moisture data for itself and its immediate and next-to-last data points. Trend fitting is then performed on the smoothed soil moisture time series. Trend fitting can be done using the linear regression method. The linear regression method uses time as the independent variable and soil moisture as the dependent variable to fit a straight line that best represents the distribution trend of the data points. The slope of this fitted line represents the trend of soil moisture change. A positive slope indicates an increasing trend in soil moisture, while a negative slope indicates a decreasing trend. The absolute value of the slope reflects the rate of change. The resulting trend in soil moisture is a quantified trend value.

[0048] The goal of this study is to perform time-series analysis on visual feature parameters from historical observation data to determine the trend of canopy color change. Visual feature parameters are multi-dimensional vectors. First, it's necessary to extract dimensions representing canopy color or calculate a comprehensive color index from these parameters. For example, the sum of probability values ​​corresponding to hue intervals in healthy green areas can be extracted from the color histogram feature components of the visual feature parameters to create a canopy green index. The canopy green index is calculated for each time point and arranged chronologically to form a canopy green index time series. The canopy green index time series undergoes the same smoothing and trend fitting operations as the soil moisture time series. Smoothing uses a moving average method, with a window size set, for example, for three consecutive data points. Trend fitting uses linear regression to obtain the slope of the fitted line. This slope represents the trend of canopy color change. A positive slope indicates an upward trend in the canopy green index, meaning the color is transitioning to a healthier green. A negative slope indicates a downward trend in the canopy green index, meaning the color is yellowing or fading. The absolute value of the slope reflects the rate of color change. The final obtained trend of canopy color change is a quantified trend value.

[0049] Based on a plant growth model corresponding to the latent state of normal growth, the expected response trajectories of soil moisture and canopy color are generated within the same historical time period. The plant growth model corresponding to the latent state of normal growth is a mathematical or empirical model describing how key state parameters of a healthy plant should change over time after being subjected to maintenance control actions. The construction of the plant growth model can be based on principles of plant physiology. These principles include, for example, the transport patterns of water in the soil, plant, and atmospheric continuum, and the relationship between chlorophyll synthesis and nutrient supply. Alternatively, the plant growth model can be an input-output relationship model trained using a large amount of historical health state observation data. The inputs to the plant growth model include the type of maintenance control action, the execution parameters of the action, and the initial environmental conditions at the time of action execution. For irrigation actions, execution parameters include, for example, the amount of irrigation water. For fertilization actions, execution parameters include, for example, the type and amount of fertilizer applied. Initial environmental conditions include soil moisture data, ambient temperature data, and light intensity data at the time of action execution. The outputs of the plant growth model are the expected values ​​of soil moisture and canopy green index at each time point within the analysis time window.

[0050] The specific process for generating the expected response trajectory is as follows: First, the type and execution parameters of the last preset maintenance control action are retrieved from historical records. Then, the initial soil moisture, initial ambient temperature, and initial light intensity data at the time of the action's execution are obtained. These input parameters are substituted into the plant growth model. Based on its inherent calculation rules, the plant growth model simulates the curves showing the expected changes in soil moisture and canopy green index over time, from the moment the action was executed until the end of the analysis time window. These two curves are collectively referred to as the expected response trajectory. The expected response trajectory represents the theoretically ideal state change process that should occur after this maintenance action, assuming the plant is in perfect health.

[0051] The first deviation between the soil moisture change trend and the corresponding expected response trajectory, and the second deviation between the canopy color change trend and the corresponding expected response trajectory, are calculated separately. Calculating the first deviation requires comparing the actual soil moisture change trend obtained from the analysis with the expected soil moisture change trend derived from the expected response trajectory. The expected soil moisture value curve in the expected response trajectory can be derived using the same method as the actual data analysis. That is, a linear trend fitting of the expected soil moisture value over time is performed to obtain an expected soil moisture change trend value. The first deviation is defined as the absolute value of the difference between the actual soil moisture change trend value and the expected soil moisture change trend value. The absolute value of the difference equals the actual soil moisture change trend value minus the expected soil moisture change trend value, and then the absolute value of the difference is taken. The first deviation is a non-negative value. The larger the first deviation value, the greater the deviation between the actual soil moisture change direction or rate and the expected value of the healthy model.

[0052] Calculating the second deviation requires comparing the actual canopy color change trend obtained from the analysis with the expected canopy color change trend derived from the expected response trajectory. The expected canopy color change trend value is obtained by performing a linear trend fitting from the expected canopy green index curve in the expected response trajectory. The second deviation is defined as the absolute value of the difference between the actual and expected canopy color change trend values. The absolute value of the difference equals the actual canopy color change trend value minus the expected canopy color change trend value, and then taking the absolute value of the difference. The second deviation is also a non-negative value. A larger second deviation value indicates a greater deviation between the actual canopy color change direction or rate and the expected value of the healthy model.

[0053] The first and second deviations are combined as a deviation feature. A deviation feature is a feature pair containing two scalar values. The deviation feature comprehensively reflects the degree of difference between the actual dynamic response and the expected health response of landscape plants to soil moisture status and canopy visual characteristics after the most recent maintenance intervention. When latent growth states are classified as ambiguous, the deviation feature provides crucial quantitative evidence from a dynamic response perspective to distinguish different potential latent states. For example, for plants with impaired water absorption due to root diseases, the actual soil moisture change trend may show a slow recovery after irrigation, resulting in a large first deviation, while the canopy color may remain somewhat green, resulting in a small second deviation. The combined pattern of deviation features helps distinguish between water shortage states and disease states.

[0054] S6. Correct the latent growth state by combining the temporal change pattern and deviation characteristics, and generate maintenance equipment control commands based on the corrected latent growth state. The specific implementation is as follows: The process of correcting the latent state of growth by combining temporal variation patterns and deviation characteristics begins with the construction of a comprehensive feature vector. The construction of this comprehensive feature vector is based on four input quantities: soil moisture change trend, canopy color change trend, first deviation, and second deviation. The soil moisture change trend is a quantified trend value obtained in step S5 through time-series analysis of soil moisture data from historical observation data sequences. This value is a signed real number; for example, 0.05 indicates a 0.05% increase in soil moisture per hour. The canopy color change trend is also a quantified trend value obtained in step S5 through time-series analysis of visual feature parameters from historical observation data sequences. This value is also a signed real number; for example, -0.02 indicates a 0.02 decrease in the canopy green index per hour. The first deviation is the deviation between the soil moisture change trend calculated in step S5 and the expected response trajectory; this value is a non-negative real number, for example, 0.03. The second deviation is also a non-negative real number, for example, 0.01. The specific operation of constructing the comprehensive feature vector involves concatenating four values—soil moisture change trend, canopy color change trend, first deviation, and second deviation—in a fixed order into a 4-dimensional vector. For example, the fixed order could be set as follows: the first dimension is the soil moisture change trend, the second is the canopy color change trend, the third is the first deviation, and the fourth is the second deviation. The comprehensive feature vector serves as a compact numerical representation of the difference between the plant's dynamic response and the expected outcome under the current ambiguous state, and it is used as input for subsequent decision rules.

[0055] The comprehensive feature vector is input into a predefined decision rule. The predefined decision rule is a set of rules that defines the mapping relationship between different feature combinations and various latent growth state categories. The establishment of the predefined decision rule requires an offline rule learning process. The establishment of the predefined decision rule begins with the collection of historical data samples. The historical data samples are derived from numerous past cases of ambiguity in latent growth states. For each historical ambiguity case, it is necessary to collect its corresponding soil moisture change trend, canopy color change trend, first deviation, second deviation, and the finally confirmed latent growth state category. The finally confirmed latent growth state category refers to the true plant state category determined by horticultural experts through detailed post-hoc examination or longer-term observation in that historical case, such as a final confirmation of severe water shortage or root disease. These historical data samples constitute a training dataset, in which each sample contains four feature values ​​and one category label.

[0056] A classifier is trained based on historical data samples to learn the mapping relationship from the comprehensive feature vector to the latent growth state category. The classifier can be a decision tree algorithm, a support vector machine algorithm, or a Naive Bayes algorithm, etc. Taking the decision tree algorithm as an example, the training process aims to construct a tree structure that can correctly classify samples based on four feature values. Starting from the root node, the decision tree algorithm recursively selects the best feature and splitting threshold at each node to divide the samples into different child nodes until a stopping condition is met. The stopping condition may be, for example, that the number of samples in a node is less than a preset minimum number of samples, or that all samples in a node belong to the same category. After training, the decision tree classifier contains the discrimination logic from the comprehensive feature vector to the latent growth state category. This discrimination logic is represented by paths from the root node to the leaf nodes, each path corresponding to a combination of feature judgment conditions, ultimately pointing to a latent growth state category.

[0057] The optimal feature is the feature that best distinguishes different latent growth states at a given node. The selection process evaluates each dimension of the comprehensive feature vector sequentially, calculating the purity improvement of the child node set after splitting by that feature. The quantification of purity improvement typically uses information gain or Gini impurity reduction, specifically calculated based on the change in the distribution of training sample classes under that feature. The feature that optimizes the metric is selected as the splitting feature for the current node. The splitting threshold is a numerical boundary point determined for the selected optimal feature. The setting process searches across all possible values ​​of the feature, traversing candidate values ​​or using a binary search method to find the specific value that maximizes the purity improvement metric. This threshold divides the samples in the node into two subsets; for example, samples with soil moisture variation trends less than or equal to 0.01 are assigned to the left child node, and samples greater than 0.01 are assigned to the right child node.

[0058] The discriminative logic inherent in the trained classifier is solidified into decision rules. This solidification process transforms the decision tree structure into a series of explicit conditional judgment rules. For example, a rule can be extracted from a path in the decision tree: if the soil moisture change trend is less than 0.01 and the first deviation is greater than 0.05, then the latent growth state category is root disease. Converting all paths in the decision tree into such rules forms a rule base. This rule base is the preset decision rule. Preset decision rules can also be stored in the system in the form of lookup tables or logical expressions. Inputting the comprehensive feature vector into the preset decision rules involves applying these rules one by one to the current comprehensive feature vector, checking which rule's condition the comprehensive feature vector satisfies, thereby determining the latent growth state category pointed to by the mapping relationship.

[0059] Based on the output of the decision rule, the current latent growth state is updated to the latent growth state category pointed to by the mapping relationship, resulting in the corrected latent growth state. The current latent growth state is the result initially estimated in step S3 but determined to be ambiguous in step S4. The output of the decision rule specifies a new latent growth state category. For example, if the current latent growth state is mild water shortage, but the decision rule determines based on the comprehensive feature vector that the mapping relationship points to root disease, then the current latent growth state is updated to root disease. This updated state is the corrected latent growth state. If the output of the decision rule is consistent with the current latent growth state, the corrected latent growth state remains unchanged. The corrected latent growth state represents the system's final judgment on the plant state after integrating additional evidence such as temporal change patterns and deviation characteristics.

[0060] Based on the corrected latent growth states, the system queries a pre-defined control strategy table to obtain the corresponding irrigation or fertilization parameters and generates control commands. The pre-defined control strategy table is a predefined mapping table that associates each possible latent growth state category with a set of specific maintenance equipment control parameters. The formulation of the pre-defined control strategy table requires horticultural knowledge and maintenance experience. For example, for a healthy state, the control strategy might be to perform no maintenance actions. For a severely water-deficient state, the control strategy might be to perform a large irrigation, for example, setting the irrigation volume to 10 liters per square meter. For a nutrient-deficient state, the control strategy might be to perform a fertilization, for example, setting the fertilization volume to 50 grams of nitrogen fertilizer per square meter. For a pest or disease infestation state, the control strategy might be to trigger an alarm and recommend the use of a specific pesticide. The pre-defined control strategy table is stored in the system as a data table, with each row recording a latent growth state category and its corresponding action type and parameters.

[0061] When querying the preset control strategy table, the corrected latent growth state is used as the index key to retrieve matching records, thereby obtaining the corresponding irrigation or fertilization parameters. For example, if the corrected latent growth state is severe water shortage, the query result will show an action type of irrigation and a parameter of 10 liters per square meter. Generating control commands involves encapsulating the acquired action type and parameters into a device-recognizable command format. This command format may include, for example, a target device identifier, an action type code, and an action quantity value. The generated command is sent to the corresponding irrigation valve controller or fertilization pump controller via a communication interface, thereby driving the maintenance equipment to perform precise maintenance operations. Through this process, in-depth analysis and correction are completed when there is ambiguity in the state estimation, and targeted control commands are generated based on the correction results.

[0062] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0063] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0064] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. Computer-readable storage media can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0065] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0066] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0067] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0068] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0069] If a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0071] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting, characterized in that, include: S1. Obtain observation data of landscape plants, including environmental data and appearance image data; S2. Process the appearance image data to extract visual feature parameters, combine them with environmental data to form an observation data sequence, and convert the observation data sequence into a corresponding feature vector; S3. Classify the feature vectors to obtain the current latent growth state of the landscape plants; S4. Calculate the distance between the feature vector and the preset prototype point of each growth latent state category in the feature space. If the distance between the feature vector and the prototype point of at least two different categories simultaneously satisfies the preset proximity condition, then the growth latent state is determined to be in an ambiguous state. S5. When in an ambiguous state, extract historical observation data from the period before and after the last preset maintenance control action, analyze the temporal change pattern of the historical observation data, and calculate the deviation characteristics between the actual change trajectory of the historical observation data and the expected response trajectory based on the normal growth latent state. S6. Correct the latent growth state by combining the temporal change pattern and deviation characteristics, and generate maintenance equipment control instructions based on the corrected latent growth state.

2. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 1, characterized in that, Acquire observational data on landscape plants, including environmental data and visual image data, including: Obtain environmental data of the environment in which the landscape plants are located, including soil moisture data, ambient temperature data, and light intensity data; The appearance image data of the landscape plants is acquired by an image acquisition device. Environmental data and visual imagery data together constitute the observation data.

3. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 2, characterized in that, The image acquisition device acquires appearance image data according to a preset cycle. The preset cycle is dynamically adjusted according to the growth stage of the landscape plants. A first acquisition frequency is used during the vigorous growth stage, and a second acquisition frequency is used during the slow growth stage. The first acquisition frequency is higher than the second acquisition frequency.

4. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 1, characterized in that, The process involves processing appearance image data to extract visual feature parameters, combining this data with environmental data to construct an observation data sequence, and then converting the observation data sequence into a corresponding feature vector, including: Color space transformation and region segmentation were performed on the appearance image data to extract color histogram features representing plant canopy color and texture features representing leaf structure. The color histogram features and texture features were used as visual feature parameters. The visual feature parameters, along with soil moisture data, ambient temperature data, and light intensity data from the environmental data, are arranged in chronological order of collection time to form an observation data sequence. The visual feature parameters and environmental data in the observation data sequence are normalized, and the normalized data are spliced ​​together in a fixed order to form a feature vector.

5. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 1, characterized in that, The feature vectors are classified into states to obtain the current latent growth states of the landscape plants, including: Input the feature vectors into a pre-trained classification model; The classification model outputs a feature vector that belongs to multiple preset latent growth state categories; The category of the latent growth state with the highest membership degree is determined as the current latent growth state of the landscape plant.

6. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 1, characterized in that, Calculate the distance between the feature vector and the preset prototype points of each latent state category in the feature space. If the distance between the feature vector and the prototype points of at least two different categories simultaneously satisfies a preset proximity condition, then the latent state is determined to be in an ambiguous state, including: Obtain the pre-defined prototype points of each latent growth state category in the feature space; Calculate the Euclidean distance between the feature vector and each prototype point; If the Euclidean distance between the feature vector and at least two prototype points of different classes is less than a preset distance threshold, then the growth hidden state is determined to be in an ambiguous state.

7. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 6, characterized in that, The prototype points are determined as follows: historical observation data are collected and their corresponding growth latent state categories are labeled; cluster analysis is performed on the feature vectors corresponding to the historical observation data belonging to the same category; and the cluster center points obtained from the cluster analysis are used as the prototype points of the corresponding growth latent state categories.

8. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 1, characterized in that, When in an ambiguous state, historical observation data from the period before and after the last preset maintenance control action is extracted, the temporal change pattern of the historical observation data is analyzed, and the deviation characteristics between the actual change trajectory of the historical observation data and the expected response trajectory based on the normal growth latent state are calculated, including: Extract a sequence of historical observation data from the historical records, within a continuous period before and after the last irrigation or fertilization action; Time series analysis was performed on soil moisture data and visual characteristic parameters in historical observation data series to obtain the trends of soil moisture change and canopy color change. Based on the plant growth model corresponding to the normal growth latent state, the expected response trajectories of soil moisture and canopy color are generated within the same historical period. The first deviation between the soil moisture change trend and the corresponding expected response trajectory, and the second deviation between the canopy color change trend and the corresponding expected response trajectory are calculated separately. The first deviation and the second deviation are used together as the deviation feature.

9. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 1, characterized in that, The latent growth state is corrected by combining temporal variation patterns and deviation characteristics, and maintenance equipment control commands are generated based on the corrected latent growth state, including: A comprehensive feature vector is constructed based on the soil moisture change trend, the canopy color change trend, the first deviation, and the second deviation. The comprehensive feature vector is input into a preset decision rule, which defines the mapping relationship between different feature combinations and each latent state category. Based on the output of the decision rule, the current growth hidden state is updated to the growth hidden state category pointed to by the mapping relationship, and the corrected growth hidden state is obtained. Based on the corrected latent growth state, the system queries the preset control strategy table to obtain the corresponding irrigation or fertilization parameters and generates control commands.

10. The method for precise maintenance and control of landscape plants based on intelligent sensing and growth fitting according to claim 9, characterized in that, The pre-defined decision rules are established as follows: collect historical data samples including soil moisture change trends, canopy color change trends, first deviation, second deviation, and finally confirmed latent growth state categories; based on the historical data samples, train a classifier to learn the mapping relationship from the comprehensive feature vector to the latent growth state categories; and solidify the discrimination logic contained in the trained classifier into decision rules.

Images