Artificial Intelligence-Based Monitoring and Early Warning Methods and Systems for Landscape Greening Maintenance

Abstract

This invention discloses an artificial intelligence-based method and system for monitoring and early warning of garden greening maintenance, involving the technical field of smart gardening and artificial intelligence intersection, including: a data acquisition module, an intelligent analysis module, and a decision-making and early warning module. It collects multi-source monitoring data and performs feature extraction and fusion processing to construct a garden health causal graph. A graph neural network is used for state analysis to generate maintenance instructions. This invention effectively solves the technical challenge of effectively fusing above-ground images and underground sensor data at spatiotemporal scales by combining spatiotemporal alignment and causal analysis, utilizing an attention-based fusion network. It introduces a causal discovery algorithm to construct a garden health causal graph, achieving a leap from shallow data correlation to deep internal mechanism mining, enabling accurate identification of abnormal plant states and interpretable diagnosis of root causes.

Description

Technical Field

[0001] This invention relates to the technical field of the intersection of smart landscaping and artificial intelligence, and in particular to a method and system for monitoring and early warning of landscaping maintenance based on artificial intelligence. Background Technology

[0002] With the acceleration of urbanization, the level of precision and intelligence in the maintenance and management of landscaping, as an important component of the urban ecosystem, needs to be improved. Traditional maintenance methods mainly rely on manual inspections and experience-based judgment, which suffers from low efficiency, strong subjectivity, and difficulty in early warning and diagnosis of root causes. Currently, although some studies have attempted to introduce technologies such as remote sensing and the Internet of Things, such as using drones to collect vegetation images or deploying soil sensors for preliminary monitoring, existing technologies still have significant limitations: it is difficult to effectively align and fuse above-ground images and underground sensor data at spatiotemporal scales, resulting in insufficient information utilization; most analysis methods are based solely on data correlation, lacking causal relationship mining, and unable to distinguish the root causes of abnormal plant conditions, such as water shortage, nutrient deficiency, or disease; at the same time, existing systems mostly remain at the level of data display and simple threshold alarms, failing to combine plant health mechanisms for state analysis and intelligent maintenance decision support.

[0003] Currently, Chinese invention patent CN107248127A discloses a landscape greening monitoring system. This system includes interconnected data acquisition and conversion modules. The data acquisition module further includes an information acquisition, publishing, and management module, and is connected to a monitoring platform and database. This invention aims to achieve the digitalization, networking, and automation of landscape greening planning, design, construction, management, and maintenance through real-time monitoring and historical data analysis. However, the system architecture in related technologies focuses heavily on data acquisition, transmission, and presentation, lacking the ability to deeply integrate and spatiotemporally align multi-source heterogeneous data (such as above-ground images and underground sensor data). Furthermore, it fails to introduce a causal analysis mechanism at the analysis level, making it unable to determine the root causes and provide interpretable diagnoses of plant health conditions. In addition, the system's decision support function is relatively weak, relying heavily on preset rules or historical comparisons, and has not yet achieved accurate early warning and adaptive maintenance strategy generation based on dynamic inference of health status. Therefore, its intelligence level and decision-making effectiveness still need improvement. Summary of the Invention

[0004] The technical problem solved by this invention is that in existing garden greening maintenance and monitoring technologies, it is difficult to effectively align and integrate above-ground image data and underground sensor data in terms of spatiotemporal scale, forming "data silos." This leads to a one-sided perception of plant health status. Existing analysis methods mostly rely on statistical correlations on the surface of data, lacking the exploration of the intrinsic causal mechanisms between plant physiology and soil environment, making it impossible to provide an explainable root cause diagnosis for abnormal states such as "yellowing leaves." The system's early warning and decision-making capabilities have a low level of intelligence, mostly remaining at the level of simple threshold alarms and data display, failing to generate accurate and differentiated maintenance strategies based on dynamic inference of health status.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: First aspect, an artificial intelligence-based garden greening maintenance monitoring and early warning system, comprising: a data acquisition module, an intelligent analysis module, and a decision-making and early warning module;

[0006] The data acquisition module is used to collect multi-source monitoring data of the garden and green area, including above-ground image data and underground sensor data;

[0007] The intelligent analysis module is used to extract and fuse features from multi-source monitoring data to obtain a spatiotemporal grid representation. It constructs a garden health causal graph using a PC causal discovery algorithm and performs state analysis on the spatiotemporal grid representation based on the garden health causal graph to obtain the probability value of abnormal root causes.

[0008] The decision-making and early warning module is used to analyze the probability values ​​of abnormal root causes and generate maintenance instructions.

[0009] As a preferred embodiment of the artificial intelligence-based garden greening maintenance monitoring and early warning system of the present invention, the data acquisition module includes an above-ground data acquisition unit and an underground data acquisition unit.

[0010] The ground data acquisition unit is used to acquire ground image data using image acquisition equipment mounted on a drone or fixed tower;

[0011] The image acquisition equipment includes a multispectral camera, a hyperspectral camera, a visible light camera, and an infrared thermal imager;

[0012] The above-ground image data includes multispectral image data, visible light image data, and infrared thermal image data;

[0013] The underground data acquisition unit is used to acquire underground sensor data through a sensor network deployed in the soil;

[0014] The sensor network includes soil moisture sensors, soil temperature sensors, and soil nutrient sensors;

[0015] The underground sensor data includes soil moisture, soil temperature, and soil nutrient content.

[0016] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system of the present invention, the intelligent analysis module includes a graph construction unit, a feature extraction unit, a spatiotemporal alignment unit, and a causal analysis unit.

[0017] The graph construction unit is used to construct a causal graph of garden health using a PC causal discovery algorithm;

[0018] The feature extraction unit is used to extract features from the above-ground image data and the underground sensor data to obtain image features and sensor features.

[0019] The spatiotemporal alignment unit is used to map the image features and sensor features to a spatiotemporal grid representation through an attention-based fusion network;

[0020] The causal analysis unit is used to perform state analysis on the garden health causal map using a graph neural network, taking a spatiotemporal grid representation as input, and outputting the probability value of abnormal root causes.

[0021] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system of the present invention, the feature extraction of above-ground image data and underground sensor data includes the following processing logic:

[0022] The above-ground image data is preprocessed, including image denoising, distortion correction, and illumination equalization.

[0023] A convolutional neural network is used to extract aboveground image features from the preprocessed aboveground image data to obtain aboveground image feature vectors;

[0024] The aboveground image features include those characterizing the morphological structure and physiological state of the plant canopy;

[0025] Time series smoothing and anomaly removal were performed on the underground sensor data, and the time change rates of soil moisture, soil temperature and soil nutrient content were calculated.

[0026] A recurrent neural network is used to extract underground sensor features from the processed underground sensor data to obtain underground sensor feature vectors;

[0027] The features of the underground sensors include the time-varying rate of soil moisture, the time-varying rate of soil temperature, and the time-varying rate of soil nutrient content.

[0028] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system of the present invention, the processing logic of mapping image features and sensor features to a spatiotemporal grid representation based on an attention mechanism fusion network includes:

[0029] Obtain the acquisition timestamps corresponding to the feature vectors of the above-ground images and the feature vectors of the underground sensors, and establish a time index sequence;

[0030] Based on the time index sequence, time interpolation or sliding window matching is used to pair the time-correspondence features of the above-ground image and the time-series features of the underground sensor to obtain a time synchronization feature set.

[0031] The processing logic for time-correspondence pairing of above-ground image features and underground sensor features includes:

[0032] When the sampling time interval of the surface image features is greater than the sampling time interval of the underground sensor, the interpolated features corresponding to the acquisition time of the surface image features are obtained by linear interpolation based on the underground sensor features at adjacent time points.

[0033] When the sampling interval of multi-source monitoring data is uneven or there is missing data, multiple data points in the corresponding time period are extracted within a preset time range based on a sliding time window, and matching features are obtained through aggregation calculation.

[0034] The processing logic for the aggregation calculation includes:

[0035] Calculate the acquisition timestamp t for the feature vector of each underground sensor. s Timestamp of image data acquisition above ground t image The absolute time difference Δt between them is calculated using the following formula:

[0036] ;

[0037] Where Δt represents the absolute time difference, t s Indicates the collection timestamp, t image Indicates the timestamp of the above-ground image data acquisition;

[0038] The weight of each underground sensor feature vector is calculated using a Gaussian function, and the formula is as follows:

[0039] ;

[0040] Where w represents the weight, and σ represents the scale parameter that controls the decay rate;

[0041] The weighted average of the feature vectors of the underground sensors is calculated using the following formula:

[0042] ;

[0043] in, Let v represent the fused feature value of the j-th dimension, n represent the dimension of the feature vector, and v i,j w represents the j-th eigenvalue of the i-th eigenvector. i This represents the decay weight corresponding to the i-th eigenvector;

[0044] Extract the spatial coordinates of the above-ground image acquisition points and the spatial coordinates of the underground sensor deployment points;

[0045] A spatial weight matrix is ​​used to interpolate and weight the features of the above-ground image and the features of the underground sensor, and the features of the above-ground image and the features of the underground sensor are mapped to a spatial grid coordinate system to generate a preliminary spatiotemporal matching feature matrix.

[0046] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system described in this invention, the specific logic for interpolating and weighting the above-ground image features and underground sensor features includes:

[0047] The garden green area is divided into a spatial grid composed of three-dimensional voxels according to a preset resolution using a regular grid division algorithm.

[0048] For each spatial grid point, obtain the corresponding above-ground image feature vector and underground sensor feature vector;

[0049] Preliminary feature values ​​for each spatial grid point are calculated by using an inverse distance weighting method based on the spatial distance between the feature vectors of the above-ground image and the feature vectors of the underground sensors.

[0050] The fused feature value for each spatial grid point is obtained by weighted averaging the feature vectors of the above-ground image and the underground sensor. The calculation formula is as follows:

[0051] ;

[0052] Where F represents the fused feature value of each spatial grid point, and V a V represents the feature value of the aboveground part of the image. b Indicates the characteristic values ​​of underground sensors. V represents a The first preset weighting coefficient, V represents b The second preset weighting coefficient;

[0053] The initial spatiotemporal matching feature matrix is ​​constructed using the fused feature values ​​of all grid points as matrix element values, and the initial spatiotemporal matching feature matrix is ​​input into the fusion network of the attention mechanism.

[0054] The fusion network of the attention mechanism includes a multi-head spatiotemporal attention layer, a feature fusion coding layer, and an output normalization layer;

[0055] The multi-head spatiotemporal attention layer is used to model the spatiotemporal dependency relationship of the preliminary spatiotemporal matching feature matrix. By calculating the attention weights between features of different grid points, it outputs an enhanced spatiotemporal feature representation.

[0056] The feature fusion coding layer is used to perform nonlinear transformation and dimensionality compression on the enhanced spatiotemporal feature representation, and output the compressed spatiotemporal feature representation;

[0057] The output normalization layer is used to perform layer normalization on the compressed spatiotemporal feature representation to obtain the processed spatiotemporal feature representation, and then reorganizes the processed spatiotemporal feature representation in the spatiotemporal dimension to generate a spatiotemporal grid representation.

[0058] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system described in this invention, the processing logic for constructing a landscape health causal graph includes:

[0059] Candidate causal characteristic variables were extracted from multi-source monitoring data, including plant physiological state variables and soil environmental variables.

[0060] The soil environmental variables include soil volumetric water content, soil temperature, and soil nutrient content change rate, while the plant physiological state variables include chlorophyll content index, canopy temperature, and vegetation cover.

[0061] The PC causal discovery algorithm is used to learn the causal structure of candidate causal feature variables and construct an initial directed acyclic graph, which consists of nodes representing variables and edges representing causal relationships.

[0062] The nodes of the initial directed acyclic graph correspond to candidate causal feature variables extracted from multi-source monitoring data, and the edges of the initial directed acyclic graph represent the causal relationships of the candidate causal feature variables and their weights are determined based on the RESIT algorithm.

[0063] By setting a causal effect threshold for the edges connecting each node in the initial directed acyclic graph, and removing edges whose effect intensity is lower than the causal effect threshold, a causal graph of garden health is obtained.

[0064] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system described in this invention, the system uses a spatiotemporal grid representation as input and performs state analysis on a landscape health causal graph using a graph neural network to obtain the probability value of abnormal root causes. The processing logic includes:

[0065] The spatiotemporal grid representation is loaded into the garden health causal graph as the initial feature vector for each node in the graph;

[0066] The node features are iteratively updated by stacking multiple graph convolutional layers. In each layer, each node aggregates its own features and the features of neighboring nodes in the garden health causal graph through graph convolution operations to obtain the updated node state representation.

[0067] Through multi-layer propagation, node embedding vectors are learned and output, while the state estimate of each node is encoded as the probability of a potential root cause.

[0068] The node embedding vector is input into a specific output layer to perform regression calculations and output the probability values ​​of the abnormal root causes of the six nodes.

[0069] The six nodes include soil volumetric water content node, soil temperature node, soil nutrient content change rate node, chlorophyll content index node, canopy temperature node, and vegetation coverage node.

[0070] As a preferred embodiment of the artificial intelligence-based landscape greening maintenance monitoring and early warning system of the present invention, the decision-making and early warning module includes an anomaly detection unit and a maintenance decision unit.

[0071] The anomaly detection unit is used to determine the anomaly type based on the anomaly root cause probability values ​​of each of the six nodes. The anomaly types include water shortage anomalies, pest and disease anomalies, nutrient imbalance anomalies, and environmental stress anomalies.

[0072] The probability values ​​of the root causes of the anomalies of the six nodes are compared with the probability thresholds to obtain the corresponding maintenance instructions. The probability thresholds include the first probability threshold and the second probability threshold.

[0073] Wherein, the second probability threshold is less than the first threshold probability;

[0074] When P water When θ1 > θ1, a water shortage anomaly is triggered;

[0075] When P temp >θ1 or P canopy When the value is greater than θ1, an environmental stress exception is triggered.

[0076] When P nutrient When the value is greater than θ1, nutritional imbalance-related abnormalities are triggered.

[0077] When P chloro >θ1 and P coverage >θ1 or P chloro >θ1 and P water When <θ2, pest and disease-related anomalies are triggered;

[0078] Among them, Pwater P represents the probability value of anomalies at the soil volumetric water content node. temp P represents the probability value of anomalous root causes at soil temperature nodes. nutrient P represents the probability value of anomaly root causes at the nodes representing the rate of change in soil nutrient content. chloro P represents the probability value of abnormal root causes for nodes in the chlorophyll content index. canopy P represents the probability value of the anomalous root cause of the canopy temperature node. coverage θ1 represents the probability value of the abnormal root cause of the vegetation cover node, and θ2 represents the first probability threshold and the second probability threshold.

[0079] The maintenance decision unit is used to generate corresponding maintenance instructions for abnormal types. The maintenance instructions include irrigation control instructions, fertilization control instructions, pest and disease control instructions, and environmental regulation instructions.

[0080] When an abnormality is identified as water shortage, an irrigation control command is triggered to control the irrigation system to perform quantitative water replenishment.

[0081] When an abnormality is identified as a pest or disease, a pest or disease control command is triggered.

[0082] When an abnormality is identified as a nutrient imbalance, a fertilizer regulation instruction is triggered.

[0083] When an anomaly is identified as an environmental stressor, an environmental adjustment command is triggered.

[0084] Secondly, the artificial intelligence-based monitoring and early warning method for landscaping maintenance includes the following steps:

[0085] Step S1: Collect multi-source monitoring data of the garden and green area;

[0086] Step S2: Perform feature extraction and fusion processing on the multi-source monitoring data to obtain a spatiotemporal grid representation;

[0087] Step S3: Construct a cause-and-effect graph of garden health, and perform state analysis on the spatiotemporal grid representation based on the cause-and-effect graph of garden health to obtain the probability value of abnormal root causes;

[0088] Step S4: Analyze the probability values ​​of abnormal root causes and generate maintenance instructions.

[0089] The beneficial effects of this invention are as follows: By combining spatiotemporal alignment and causal analysis, and utilizing an attention-based fusion network, the technical challenge of effectively fusing above-ground images and underground sensor data at spatiotemporal scales is effectively solved. This achieves efficient fusion and unified representation of multi-source heterogeneous data. By introducing a causal discovery algorithm to construct a causal graph of garden health, a leap from shallow data correlation to deep internal mechanism mining is realized, enabling accurate identification of abnormal plant states and interpretable diagnosis of root causes. Based on this, the system automatically generates and executes differentiated maintenance strategies according to the above-mentioned deep analysis results, thereby transforming the traditional passive processing mode that relies on human experience into proactive and precise regulation based on causal analysis, significantly improving the intelligence level, accuracy, and overall efficiency of garden maintenance management. Attached Figure Description

[0090] Figure 1 This is a schematic diagram of the basic process of an artificial intelligence-based landscape greening maintenance monitoring and early warning system provided in one embodiment of the present invention. Detailed Implementation

[0091] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0092] Example 1, referring to Figure 1 As an embodiment of the present invention, an artificial intelligence-based landscape greening maintenance monitoring and early warning system is provided, including: a data acquisition module, an intelligent analysis module, and a decision-making and early warning module;

[0093] The data acquisition module is used to collect multi-source monitoring data of the garden and green area. The multi-source monitoring data includes above-ground image data and underground sensor data.

[0094] The intelligent analysis module is used to extract and fuse features from multi-source monitoring data to obtain a spatiotemporal grid representation. A garden health causal graph is constructed using the PC causal discovery algorithm. Based on the garden health causal graph, the spatiotemporal grid representation is analyzed to obtain the probability value of abnormal root causes.

[0095] The decision-making and early warning module is used to analyze the probability values ​​of abnormal root causes and generate maintenance instructions.

[0096] This invention combines spatiotemporal alignment with causal analysis, effectively addressing the technical challenge of seamlessly integrating above-ground images and underground sensor data across spatiotemporal scales using an attention-based fusion network. It achieves efficient fusion and unified representation of multi-source heterogeneous data, introduces a causal discovery algorithm to construct a causal graph of garden health, and leaps from shallow data correlation to deep-seated mechanism mining. This enables precise identification of abnormal plant conditions and interpretable diagnosis of root causes. Based on this deep analysis, the system automatically generates and executes differentiated maintenance strategies, transforming the traditional passive processing model reliant on human experience into proactive and precise regulation based on causal analysis. This significantly improves the intelligence, accuracy, and overall efficiency of garden maintenance management.

[0097] The data acquisition module includes above-ground data acquisition units and underground data acquisition units;

[0098] The ground data acquisition unit is used to acquire ground-based image data using image acquisition equipment mounted on drones or fixed towers;

[0099] Image acquisition equipment includes multispectral cameras, hyperspectral cameras, visible light cameras, and infrared thermal imagers;

[0100] The above-ground image data includes multispectral image data, visible light image data, and infrared thermal image data;

[0101] The underground data acquisition unit is used to acquire underground sensor data through a sensor network deployed in the soil;

[0102] The sensor network includes soil moisture sensors, soil temperature sensors, and soil nutrient sensors;

[0103] The data from the underground sensors include soil moisture, soil temperature, and soil nutrient content.

[0104] In practice, the data acquisition module achieves comprehensive data collection of garden vegetation by constructing an integrated air-ground three-dimensional monitoring network. The ground-based data acquisition unit utilizes a DJI M300RTK drone equipped with a Sentera 6X multispectral camera with a resolution of 1280×960. This camera covers five bands: blue, green, red, red-edge, and near-infrared. It is also equipped with a FLIR T1050sc infrared thermal imager with a thermal sensitivity better than 20mk, covering a temperature range of -40℃ to 2000℃. Hikvision DS-2TD6237 series fixed thermal imaging gimbals are deployed in key ground areas. This device has a detection range of up to 150 meters and a horizontal rotation range of 360 degrees. Through a combination of preset flight path navigation and fixed-point monitoring, multispectral images of the vegetation are collected twice daily, acquiring image data with a spatial resolution of 5 cm across five bands. Simultaneously, visible light images with a resolution of 20 megapixels and infrared thermal imaging data with a resolution of 1024×768 and a temperature measurement accuracy of ±2℃ are also acquired.

[0105] The underground data acquisition unit employs a Decagon 5TE three-parameter sensor network, covering volumetric water content (0–100%), temperature (-40–60℃), and electrical conductivity (0–23 dS / m). Sensors are deployed in a 20m × 20m grid at a soil depth of 15–25 cm, transmitting data via LoRa wireless protocol with a transmission distance of at least 2 km and power consumption maintained below 0.1 watts. The system collects soil volumetric water content data every 30 minutes with an accuracy of ±3%, and soil temperature with an accuracy of ±1℃. Simultaneously, nitrogen, phosphorus, and potassium content data are calculated based on electrical conductivity, with calibration errors controlled within 5%. This achieves multi-dimensional, high-precision data acquisition of the growth status of garden vegetation, significantly improving the comprehensiveness and timeliness of green space maintenance monitoring, providing a reliable data foundation for precise maintenance decisions, and effectively enhancing the scientific level of garden maintenance management.

[0106] The intelligent analysis module includes a graph construction unit, a feature extraction unit, a spatiotemporal alignment unit, and a causal analysis unit;

[0107] The graph construction unit is used to construct a causal graph of garden health using a PC-based causal discovery algorithm.

[0108] The feature extraction unit is used to extract features from above-ground image data and underground sensor data to obtain image features and sensor features;

[0109] The spatiotemporal alignment unit is used to map image features and sensor features to a spatiotemporal grid representation through an attention-based fusion network;

[0110] The causal analysis unit is used to perform state analysis on the garden health causal graph using a graph neural network, taking a spatiotemporal grid representation as input, and outputting the probability value of abnormal root causes.

[0111] In practical implementation, the system achieves a composite modeling approach combining knowledge-driven and data-verified methods through the aforementioned garden health causal graph construction method. The feature extraction unit employs deep learning to mine features from multi-source monitoring data. For aboveground image data, the preprocessing stage first uses a Gaussian filtering algorithm (kernel size 5×5) for image denoising, then performs distortion correction using camera calibration parameters, and finally uses histogram equalization to achieve light equalization. The preprocessed image is input into a convolutional neural network based on the ResNet-50 architecture. Five convolutional modules extract multi-level features, ultimately outputting a 1024-dimensional aboveground image feature vector. This feature vector fully characterizes the morphological and structural features of the plant canopy (including leaf area index and canopy coverage) and physiological state indicators (including chlorophyll content and water stress index).

[0112] In the causal analysis phase, the system demonstrated exceptional interpretability in its diagnostic capabilities. Taking the symptom of "yellowing leaves" as an example, the system can accurately attribute the cause based on the constructed causal graph: when the causal path shows a strong causal relationship between soil moisture and leaf yellowing, while the correlation with nutrient indicators is weak, the system identifies it as water stress; when nitrogen content shows a significant causal relationship with yellowing, and soil moisture is within the normal range, the system diagnoses it as nutrient deficiency; when pest and disease monitoring indicators form a clear causal link with yellowing, it identifies it as disease invasion. This causal-based diagnostic approach allows the system not only to identify abnormal states but also to deeply analyze their formation mechanisms, providing a reliable basis for precise intervention. This causal analysis scheme, by establishing an interpretable causal network, achieves accurate diagnosis and attribution analysis of abnormal states in gardens, significantly improving the scientific rigor and relevance of maintenance decisions and providing strong technical support for the refined management of garden greening.

[0113] Feature extraction is performed on above-ground image data and underground sensor data. The processing logic includes:

[0114] The above-ground image data is preprocessed, including image denoising, distortion correction, and illumination equalization.

[0115] A convolutional neural network is used to extract aboveground image features from the preprocessed aboveground image data to obtain aboveground image feature vectors;

[0116] Aboveground image features include those characterizing the morphological structure and physiological state of the plant canopy;

[0117] Time series smoothing and anomaly removal were performed on the underground sensor data, and the time change rates of soil moisture, soil temperature and soil nutrient content were calculated.

[0118] A recurrent neural network is used to extract underground sensor features from the processed underground sensor data to obtain underground sensor feature vectors;

[0119] The characteristics of the underground sensors include the time-varying rate of soil moisture, the time-varying rate of soil temperature, and the time-varying rate of soil nutrient content.

[0120] In practice, for the underground sensor data, the preprocessing stage employs a moving average filter (window width of 7 sampling points) for time series smoothing, and outlier data points are removed based on the 3σ criterion. Subsequently, the time-varying rates of soil moisture, soil temperature, and soil nutrient content (nitrogen, phosphorus, and potassium) are calculated to form a standardized sensor sequence. This sequence is input into a bidirectional LSTM recurrent neural network, which contains two hidden layers (128 neurons per layer). Through temporal dependency mining, the final output is an underground sensor feature vector containing the time-varying rates of soil moisture, soil temperature, and soil nutrient content, with a feature dimension of 256. The entire feature extraction process is completed on a computing platform equipped with an NVIDIA Tesla V100, with a single processing time of no more than 2 seconds, effectively improving feature extraction accuracy. Intelligent feature mining of multi-source monitoring data is achieved through convolutional neural networks, significantly improving the accuracy and comprehensiveness of plant state characterization, providing high-quality feature input for subsequent garden health status assessment, and effectively enhancing the monitoring accuracy and reliability of the system.

[0121] The attention-based fusion network's processing logic for mapping image features and sensor features to a spatiotemporal grid representation includes:

[0122] Obtain the acquisition timestamps corresponding to the feature vectors of the above-ground images and the feature vectors of the underground sensors, and establish a time index sequence;

[0123] Based on the time index sequence, time interpolation or sliding window matching is used to pair the time-correspondence features of the above-ground image and the time-series features of the underground sensor to obtain a time synchronization feature set.

[0124] The processing logic for time-correspondence pairing of above-ground image features and underground sensor features includes:

[0125] When the sampling time interval of the surface image features is greater than the sampling time interval of the underground sensor, the interpolated features corresponding to the acquisition time of the surface image features are obtained by linear interpolation based on the underground sensor features at adjacent time points.

[0126] When the sampling interval of multi-source monitoring data is uneven or there is missing data, multiple data points in the corresponding time period are extracted within a preset time range based on a sliding time window, and matching features are obtained through aggregation calculation.

[0127] The processing logic for aggregation calculations includes:

[0128] Calculate the acquisition timestamp t for the feature vector of each underground sensor. s Timestamp of image data acquisition above ground t image The absolute time difference Δt between them is calculated using the following formula:

[0129] ;

[0130] Where Δt represents the absolute time difference, t s Indicates the collection timestamp, t image Indicates the timestamp of the above-ground image data acquisition;

[0131] The weight of each underground sensor feature vector is calculated using a Gaussian function, and the formula is as follows:

[0132] ;

[0133] Where w represents the weight, and σ represents the scale parameter that controls the decay rate;

[0134] The weighted average of the feature vectors of the underground sensors is calculated using the following formula:

[0135] ;

[0136] in, Let v represent the fused feature value of the j-th dimension, n represent the dimension of the feature vector, and v i,j w represents the j-th eigenvalue of the i-th eigenvector. i This represents the decay weight corresponding to the i-th eigenvector;

[0137] Extract the spatial coordinates of the above-ground image acquisition points and the spatial coordinates of the underground sensor deployment points;

[0138] A spatial weight matrix is ​​used to interpolate and weight the features of the above-ground image and the features of the underground sensor, and the features of the above-ground image and the features of the underground sensor are mapped to a spatial grid coordinate system to generate a preliminary spatiotemporal matching feature matrix.

[0139] In practical implementation, the attention-based fusion network achieves effective fusion of multi-source features through refined spatiotemporal alignment processing. The system first obtains the acquisition timestamps corresponding to the feature vectors of the above-ground images and the feature vectors of the underground sensors, establishing a complete time index sequence. To address the time synchronization issue, an adaptive time pairing strategy is adopted: when the sampling time interval of the above-ground image features (typically 2 hours) is greater than the sampling time interval of the underground sensors (30 minutes), linear interpolation is used based on the underground sensor features at adjacent time points to calculate interpolated features that precisely correspond to the acquisition time of the above-ground image features; when the sampling interval of multi-source monitoring data is uneven or there are missing data points, a 30-minute sliding time window is set, and multiple data points are extracted within the window for aggregation calculation.

[0140] During the aggregation calculation process, the system first calculates the acquisition timestamp t of the feature vector of each underground sensor. s Timestamp of image data acquisition above ground t image The absolute time difference between them is Δt = |t s -t image The timestamps use the UNIX timestamp format, accurate to the second. A Gaussian function is used to calculate the weight of each feature, with the scaling parameter σ set to 3600 seconds (1 hour) to ensure that time differences decay smoothly within a reasonable range. Finally, a weighted average is used to calculate the fused feature values, where the feature vector dimension n is set to 6 dimensions based on the sensor type, including humidity, temperature, and key nutrient indicators.

[0141] In terms of spatial alignment, the system extracts the GPS coordinates (accuracy ±0.1 meters) of the above-ground image acquisition points and the precise spatial coordinates of the underground sensor deployment points. An inverse distance-weighted spatial interpolation algorithm is used to construct a 50m × 50m spatial grid coordinate system. Features from different sources are mapped to a unified grid using a spatial weight matrix, generating a preliminary spatiotemporal matching feature matrix containing 256 feature dimensions. This matrix serves as the input to the subsequent attention mechanism network, effectively improving the spatiotemporal alignment accuracy of multi-source data. Through refined temporal pairing and spatial interpolation algorithms, the system effectively solves the inconsistency problem of multi-source monitoring data at spatiotemporal scales, significantly improving the accuracy and reliability of feature fusion, and laying a solid data foundation for the accurate assessment of the health status of gardens.

[0142] The specific logic for interpolating and weighting the features of the above-ground image and the features of the underground sensor includes:

[0143] The garden green area is divided into a spatial grid composed of three-dimensional voxels according to a preset resolution using a regular grid division algorithm.

[0144] For each spatial grid point, obtain the corresponding above-ground image feature vector and underground sensor feature vector;

[0145] Preliminary feature values ​​for each spatial grid point are calculated by using an inverse distance weighting method based on the spatial distance between the feature vectors of the above-ground image and the feature vectors of the underground sensors.

[0146] The fused feature value for each spatial grid point is obtained by weighted averaging the feature vectors of the above-ground image and the underground sensor. The calculation formula is as follows:

[0147] ;

[0148] Where F represents the fused feature value of each spatial grid point, and V a V represents the feature value of the aboveground part of the image. b Indicates the characteristic values ​​of underground sensors. V represents a The first preset weighting coefficient, V represents b The second preset weighting coefficient;

[0149] The initial spatiotemporal matching feature matrix is ​​constructed using the fused feature values ​​of all grid points as matrix element values, and the initial spatiotemporal matching feature matrix is ​​input into the fusion network of the attention mechanism.

[0150] The fused feature values ​​of all grid points in the spatial grid are arranged according to the index order of their spatial coordinates in three-dimensional space, and organized into a four-dimensional tensor T, whose dimensions are represented as (X, Y, Z, F).

[0151] Where X represents the resolution of the spatial grid on the X-axis, Y represents the resolution of the spatial grid on the Y-axis, Z represents the resolution of the spatial grid on the Z-axis, and F represents the dimension of the fused feature value for each grid point.

[0152] The four-dimensional tensors T generated in different time slices of the same geographical region are stacked in time series order to add a time dimension, forming a five-dimensional preliminary spatiotemporal matching feature matrix M, whose dimensions are represented as (Time, X, Y, Z, F).

[0153] The attention mechanism fusion network includes a multi-head spatiotemporal attention layer, a feature fusion encoding layer, and an output normalization layer;

[0154] The multi-head spatiotemporal attention layer is used to model the spatiotemporal dependency relationship of the initial spatiotemporal matching feature matrix. By calculating the attention weights between features of different grid points, it outputs an enhanced spatiotemporal feature representation.

[0155] The feature fusion coding layer is used to perform nonlinear transformation and dimensionality compression on the enhanced spatiotemporal feature representation, and outputs the compressed spatiotemporal feature representation;

[0156] The output normalization layer is used to perform layer normalization on the compressed spatiotemporal feature representation to obtain the processed spatiotemporal feature representation, and then reorganizes the processed spatiotemporal feature representation in the spatiotemporal dimension to generate a spatiotemporal grid representation.

[0157] In its implementation, this system employs a layered fusion architecture to achieve spatial integration and spatiotemporal modeling of multi-source features. First, based on a regular grid partitioning algorithm, the garden area is divided into a 1m × 1m × 0.5m three-dimensional voxel spatial grid. Each grid point is fused with surrounding monitoring data using an inverse distance weighting algorithm: with the target grid point as the center, a 10-meter search radius is set, and spatial weights are calculated for above-ground image features and underground sensor features within the radius, with the weight value inversely proportional to the square of the distance. In the feature fusion stage, the weight coefficient for above-ground image features is set to 0.6, and the weight coefficient for underground sensor features is set to 0.4. The fused feature value for each grid point is calculated using a weighting formula. After completing the spatial grid feature calculation, the system organizes the fused feature values ​​into a four-dimensional tensor based on spatial coordinate indices. The X and Y axes have resolutions representing the number of grids in the region's length and width directions, respectively. The Z axis has three layers representing features at different heights, and the feature dimension F is fixed at 64 dimensions. In the time-series processing stage, a five-dimensional spatiotemporal matrix of 30 consecutive time slices (with a 2-hour time interval) is input into a multi-head spatiotemporal attention network. This network contains eight attention heads, each calculating self-attention weights in the spatiotemporal dimension, and establishing long-range dependencies through a query-key-value mechanism. The feature fusion encoding layer employs a three-layer fully connected network (with 256, 128, and 64 neurons respectively) for nonlinear transformation and dimensionality compression, and finally outputs a 32-dimensional spatiotemporal grid representation through layer normalization. Through spatial grid partitioning and the multi-head attention mechanism, deep spatiotemporal fusion of multi-source monitoring data is achieved, significantly improving the accuracy and completeness of feature representation, and providing reliable technical support for the accurate monitoring and early warning of garden health status.

[0158] The processing logic for constructing a cause-and-effect diagram of garden health includes:

[0159] Candidate causal characteristic variables were extracted from multi-source monitoring data, including plant physiological state variables and soil environmental variables.

[0160] Soil environmental variables include soil volumetric water content, soil temperature, and the rate of change of soil nutrient content; plant physiological state variables include chlorophyll content index, canopy temperature, and vegetation cover.

[0161] The PC causal discovery algorithm is used to learn the causal structure of candidate causal feature variables and construct an initial directed acyclic graph, which consists of nodes representing variables and edges representing causal relationships.

[0162] The nodes of the initial directed acyclic graph correspond to candidate causal feature variables extracted from multi-source monitoring data, and the edges of the initial directed acyclic graph represent the causal relationships of the candidate causal feature variables and their weights are determined based on the RESIT algorithm.

[0163] By setting a causal effect threshold for the edges connecting each node in the initial directed acyclic graph, and removing edges whose effect intensity is lower than the causal effect threshold, a causal graph of garden health is obtained.

[0164] In practice, candidate causal feature variables were extracted from multi-source monitoring data. Soil environmental variables included soil volumetric water content (measurement range 0-100%), soil temperature (measurement range -40-60℃), and soil nutrient content change rate (based on conductivity conversion). Plant physiological state variables included chlorophyll content index (calculated based on NDVI), canopy temperature (infrared thermometry), and vegetation cover (based on image segmentation). A PC causal discovery algorithm was used for causal structure learning. An initial directed acyclic graph was constructed using conditional independence tests with a significance level of α=0.05 and orientation rules. Nodes in the graph corresponded to 12 feature variables, and edges represented causal relationships. The RESIT algorithm was used to calculate causal weights, and a causal effect threshold of 0.15 was set. Edges with effect strengths below this threshold were removed, resulting in a garden health causal graph containing 28 valid causal relationships. Through systematic causal discovery and effect quantification, a visual representation of the intrinsic causal relationships within the garden ecosystem was achieved, significantly improving the interpretability and accuracy of state diagnosis and providing a scientific basis for precise maintenance decisions.

[0165] Using a spatiotemporal grid representation as input, a graph neural network is used to perform state analysis on the garden health causal graph to obtain the probability values ​​of abnormal root causes. The processing logic includes:

[0166] The spatiotemporal grid representation is loaded into the causal graph of garden health as the initial feature vector of each node in the graph;

[0167] The node features are iteratively updated by stacking multiple graph convolutional layers. In each layer, each node aggregates its own features and the features of neighboring nodes in the garden health causal graph through graph convolution operations to obtain the updated node state representation.

[0168] Through multi-layer propagation, node embedding vectors are learned and output, while the state estimate of each node is encoded as the probability of a potential root cause.

[0169] The node embedding vector is input into a specific output layer to perform regression calculations and output the probability values ​​of the abnormal root causes of the six nodes.

[0170] The six nodes include soil volumetric water content, soil temperature, soil nutrient content change rate, chlorophyll content index, canopy temperature, and vegetation cover.

[0171] In practical implementation, the state analysis module uses a graph neural network to achieve accurate root cause analysis of anomalies on the garden health causal graph. The system loads a 256-dimensional spatiotemporal grid representation as initial feature vectors into six key nodes in the garden health causal graph: soil volumetric water content node, soil temperature node, soil nutrient content change rate node, chlorophyll content index node, canopy temperature node, and vegetation cover node. A three-layer graph convolutional network is used for feature propagation, with each layer containing 128 hidden units and using the GELU activation function. During graph convolution, each node aggregates the features of its neighboring nodes through weighted aggregation. The weights are determined by pre-calculated causal strength; for example, the soil volumetric water content node aggregates the features of the chlorophyll content index node with a weight of 0.35, and the soil temperature node aggregates the features of the canopy temperature node with a weight of 0.41.

[0172] After three layers of iterative propagation, the system learns node embedding vectors of dimension 64. These vectors simultaneously encode the state estimate of each node and its probability of being a potential root cause. The final output layer uses the sigmoid activation function for regression calculation, outputting the probability values ​​of anomalous root causes for each of the six nodes: soil volumetric water content, soil temperature, soil nutrient content change rate, chlorophyll content index, canopy temperature, and vegetation cover. The output layer outputs the probability values ​​of each node as an anomalous root cause, ranging from 0 to 1. By fusing graph neural networks with a garden health causal graph, accurate assessment and root cause analysis of garden health status are achieved, significantly improving the scientific rigor and targeted nature of maintenance decisions.

[0173] The decision-making and early warning module includes an anomaly detection unit and a maintenance decision-making unit;

[0174] The anomaly detection unit is used to determine the anomaly type based on the anomaly root cause probability values ​​of each of the six nodes. The anomaly types include water shortage anomalies, pest and disease anomalies, nutrient imbalance anomalies, and environmental stress anomalies.

[0175] The probability values ​​of the root causes of the anomalies of the six nodes are compared with the probability thresholds to obtain the corresponding maintenance instructions. The probability thresholds include the first probability threshold and the second probability threshold.

[0176] Wherein, the second probability threshold is less than the first threshold probability;

[0177] When P water When θ1 > θ1, a water shortage anomaly is triggered;

[0178] When P temp >θ1 or P canopy When the value is greater than θ1, an environmental stress exception is triggered.

[0179] When P nutrientWhen the value is greater than θ1, nutritional imbalance-related abnormalities are triggered.

[0180] When P chloro >θ1 and P coverage >θ1 or P chloro >θ1 and P water When <θ2, pest and disease-related anomalies are triggered;

[0181] Among them, P water P represents the probability value of anomalies at the soil volumetric water content node. temp P represents the probability value of anomalous root causes at soil temperature nodes. nutrient P represents the probability value of anomaly root causes at the nodes representing the rate of change in soil nutrient content. chloro P represents the probability value of abnormal root causes for nodes in the chlorophyll content index. canopy P represents the probability value of the anomalous root cause of the canopy temperature node. coverage θ1 represents the probability value of the abnormal root cause of the vegetation cover node, and θ2 represents the first probability threshold and the second probability threshold.

[0182] The maintenance decision unit is used to generate corresponding maintenance instructions for abnormal types. The maintenance instructions include irrigation control instructions, fertilization control instructions, pest and disease control instructions, and environmental regulation instructions.

[0183] When an abnormality is identified as water shortage, an irrigation control command is triggered to control the irrigation system to perform quantitative water replenishment.

[0184] When an abnormality is identified as a pest or disease, a pest or disease control command is triggered.

[0185] When an abnormality is identified as a nutrient imbalance, a fertilizer regulation instruction is triggered.

[0186] When an anomaly is identified as an environmental stressor, an environmental adjustment command is triggered.

[0187] In practical implementation, the decision-making and early warning module achieves accurate anomaly identification and maintenance decisions through a multi-level probability threshold judgment mechanism. The system sets the first probability threshold θ1=0.7 and the second probability threshold θ2=0.2. The anomaly detection unit monitors the anomaly root cause probability values ​​of six nodes in real time: when the probability value P of the soil volumetric water content node... water When the soil temperature node probability value P > 0.7, a water shortage anomaly is triggered; temp >0.7 or the probability value P of the canopy temperature node canopy When the value is greater than 0.7, an environmental stress anomaly is triggered; when the probability value of the soil nutrient content change rate node P... nutrient When the chlorophyll content index value is greater than 0.7, it triggers nutrient imbalance-related abnormalities; when the probability value of the chlorophyll content index node P is greater than 0.7, it triggers nutrient imbalance-related abnormalities. chloro >0.7 and the probability value P of vegetation coverage nodes coverage>0.7, or the probability value of the chlorophyll content index node P chloro The probability value P of soil volumetric water content nodes is greater than 0.7. water When the value is less than 0.2, pest and disease-related anomalies are triggered.

[0188] The maintenance decision-making unit generates targeted instructions based on the type of anomaly: water shortage anomalies trigger irrigation control instructions, controlling the drip irrigation system to replenish water at a rate of 1.5–2 liters per square meter; pest and disease anomalies trigger control instructions, activating the targeted pesticide application system and pushing early warning information to the management personnel's terminal; nutrient imbalance anomalies trigger fertilization control instructions, precisely proportioning nitrogen, phosphorus, and potassium fertilizers according to the type of element deficiency; environmental stress anomalies trigger environmental adjustment instructions, automatically activating shade nets, misting cooling, or ventilation equipment. Through multi-level threshold judgment and precise instruction generation, intelligent assessment of the garden's health status and automated execution of maintenance decisions are achieved, significantly improving the accuracy and timeliness of green space maintenance and providing a complete technical solution for refined management of garden greening.

[0189] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0190] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. An artificial intelligence-based garden maintenance monitoring and early warning system, characterized in that, The application relates to a greenbelt health monitoring system. The data acquisition module is used for acquiring multi-source monitoring data of a greenbelt area, and the multi-source monitoring data comprises aboveground image data and underground sensor data. The intelligent analysis module is used for feature extraction and fusion processing of the multi-source monitoring data, obtaining space-time grid representation, constructing a greenbelt health causal diagram through a PC causal discovery algorithm, and performing state analysis on the space-time grid representation based on the greenbelt health causal diagram to obtain an abnormal root cause probability value. The decision-making and early warning module is used for analyzing the abnormal root cause probability value and generating maintenance instructions. The intelligent analysis module comprises a space-time alignment unit and a causal analysis unit. The space-time alignment unit is used for mapping image features and sensor features to space-time grid representation through a fusion network based on an attention mechanism, and the processing logic comprises the following steps: obtaining time stamps corresponding to aboveground image feature vectors and underground sensor feature vectors, and establishing a time index sequence; based on the time index sequence, time corresponding pairing of the aboveground image features and the underground sensor time sequence features is performed in a time interpolation or sliding window matching mode to obtain a time synchronization feature set, and the processing logic comprises the following steps: when the sampling time interval of the aboveground image features is greater than that of the underground sensors, interpolation features corresponding to the aboveground image feature acquisition time are calculated based on the underground sensor features of adjacent time points by linear interpolation; when the sampling intervals of the multi-source monitoring data are uneven or there is data loss, a plurality of data points in a corresponding time period are extracted in a preset time range based on a sliding time window, and matching features are obtained through aggregation calculation, and the processing logic comprises the following steps: a Gaussian function is used to calculate the weight of each underground sensor feature vector, and the calculation formula is as follows: Calculate the acquisition timestamp t for the feature vector of each underground sensor. s Timestamp of image data acquisition above ground t image The absolute time difference Δt between them is calculated using the following formula: ； wherein Δt represents an absolute time difference, t s represents a collection time stamp, t image represents an above-ground part image data collection time stamp; wherein w represents the weight, and sigma represents a scale parameter for controlling the decay rate; ； weighted average calculation is performed on the underground sensor feature vectors, and the calculation formula is as follows: spatial coordinates of aboveground image acquisition points and spatial coordinates of underground sensor layout points are extracted; ； wherein, denotes the fusion feature value of the jth dimension, n denotes the dimension of the feature vector, v i,j denotes the jth feature value of the ith feature vector, w i denotes the attenuation weight corresponding to the ith feature vector; a spatial weight matrix is used for interpolation and weighted average calculation of the aboveground image features and the underground sensor features, the aboveground image features and the underground sensor features are mapped into a spatial grid coordinate system to generate a preliminary space-time matching feature matrix; the preliminary space-time matching feature matrix is input into a fusion network of an attention mechanism to obtain space-time grid representation; the causal analysis unit is used for inputting the space-time grid representation, performing state analysis on the greenbelt health causal diagram through a graph neural network, and outputting an abnormal root cause probability value. The data acquisition module comprises an aboveground data acquisition unit and an underground data acquisition unit.

2. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 1, characterized in that: The aboveground data acquisition unit is used for acquiring aboveground image data through an image acquisition device carried by a unmanned aerial vehicle or a fixed tower; the image acquisition device comprises a multispectral camera, a hyperspectral camera, a visible light camera and an infrared thermal imager; the aboveground image data comprises multispectral image data, visible light image data and infrared thermal image data; the underground data acquisition unit is used for acquiring underground sensor data through a sensor network arranged in soil. ​ The sensor network comprises a soil humidity sensor, a soil temperature sensor, and a soil nutrient sensor. The underground sensor data comprises soil humidity, soil temperature, and soil nutrient content.

3. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 1, characterized in that: The intelligent analysis module further comprises a graph construction unit and a feature extraction unit. The graph construction unit is configured to construct a garden health causal graph by using a PC causal discovery algorithm. The feature extraction unit is configured to perform feature extraction on the aboveground image data and the underground sensor data to obtain image features and sensor features.

4. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 3, characterized in that: The processing logic for feature extraction on the aboveground image data and the underground sensor data comprises: The aboveground image data is preprocessed, and the preprocessing comprises image denoising, distortion correction, and illumination equalization. A convolutional neural network is used to extract aboveground image features from the preprocessed aboveground image data to obtain an aboveground image feature vector. The aboveground image features comprise features representing plant canopy morphological structure and physiological state. The underground sensor data is subjected to time series smoothing and abnormal value elimination processing, and the time variation rates of soil humidity, soil temperature, and soil nutrient content are calculated. A recurrent neural network is used to extract underground sensor features from the processed underground sensor data to obtain an underground sensor feature vector. The underground sensor features comprise soil humidity time variation rate, soil temperature time variation rate, and soil nutrient content time variation rate.

5. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 4, characterized in that: The specific logic for interpolation and weighted average of the aboveground image features and the underground sensor features comprises: The garden greening area is divided into a spatial grid composed of three-dimensional voxels by using a regular grid division algorithm at a preset resolution; For each spatial grid point, the corresponding aboveground image feature vector and underground sensor feature vector are obtained; The inverse distance weighting method is used to calculate the preliminary feature value of each spatial grid point according to the spatial distance of the aboveground image feature vector and the underground sensor feature vector; The weighted average calculation is performed based on the aboveground image feature vector and the underground sensor feature vector to obtain the fusion feature value of each spatial grid point, and the calculation formula is: ； wherein F represents a fusion feature value of each spatial grid point, V a represents an above-ground part image feature value, V b represents an underground part sensor feature value, represents a first preset weight coefficient of V a represents a second preset weight coefficient of V b .​ The fusion feature values of all grid points are used as matrix element values to form a preliminary spatio-temporal matching feature matrix, and the preliminary spatio-temporal matching feature matrix is input into the fusion network of the attention mechanism; The fusion network of the attention mechanism comprises a multi-head spatio-temporal attention layer, a feature fusion encoding layer, and an output normalization layer. The multi-head spatio-temporal attention layer is configured to model the spatio-temporal dependency relationship of the preliminary spatio-temporal matching feature matrix, calculate the attention weights between different grid point features, and output an enhanced spatio-temporal feature representation. The feature fusion encoding layer is configured to perform nonlinear transformation and dimension compression on the enhanced spatio-temporal feature representation to output a compressed spatio-temporal feature representation. The output normalization layer is configured to perform layer normalization on the compressed spatio-temporal feature representation to obtain a spatio-temporal grid representation.

6. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 3, characterized in that: The processing logic for constructing the garden health causal graph comprises: candidate causal feature variables are extracted from the multi-source monitoring data, and the candidate causal feature variables comprise plant physiological state variables and soil environmental variables; The soil environment variables include soil volumetric water content, soil temperature and soil nutrient content change rate, and the plant physiological state variables include chlorophyll content index, canopy temperature and vegetation coverage; The PC causal discovery algorithm is used for causal structure learning of the candidate causal feature variables to construct an initial directed acyclic graph, and the initial directed acyclic graph is composed of nodes representing variables and edges representing causal relationships; The nodes of the initial directed acyclic graph correspond to the candidate causal feature variables extracted from the multi-source monitoring data, and the edges of the initial directed acyclic graph represent the causal relationships of the candidate causal feature variables and determine the weights based on the RESIT algorithm; A causal effect threshold is set for the edges connecting the nodes in the initial directed acyclic graph, and edges with an effect intensity lower than the causal effect threshold are removed to obtain the garden health causal graph.

7. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 1, characterized in that: The spatiotemporal grid representation is loaded into the garden health causal graph as the initial feature vector of each node in the graph. The node features are iteratively updated through stacked multi-layer graph convolution layers, and in each layer, each node aggregates its own features and neighbor node features in the garden health causal graph through graph convolution operations to obtain updated node state representations. Through multi-layer propagation, node embedding vectors are learned and output, while encoding the state estimates of each node as the probability of a latent root cause. The node embedding vectors are input into a specific output layer for regression calculation to output six abnormal root cause probability values for the nodes respectively. The six nodes include soil volumetric water content node, soil temperature node, soil nutrient content change rate node, chlorophyll content index node, canopy temperature node and vegetation coverage node. The decision warning module includes an anomaly detection unit and a maintenance decision unit.

8. The artificial intelligence-based garden maintenance monitoring and early warning system according to claim 1, characterized in that: The anomaly detection unit is used to determine the anomaly type according to the abnormal root cause probability values of the six nodes, and the anomaly type includes water shortage class anomaly, disease and pest class anomaly, nutrient imbalance class anomaly and environmental stress class anomaly. The abnormal root cause probability values of the six nodes are compared with the probability threshold to obtain the corresponding maintenance instructions, and the probability threshold includes a first probability threshold and a second probability threshold. The second probability threshold is less than the first threshold probability. The maintenance decision unit is used to generate corresponding maintenance instructions for the anomaly type, and the maintenance instructions include irrigation control instructions, fertilization control instructions, disease and pest control instructions and environmental regulation instructions. When P water > θ1, trigger water shortage anomaly; When P temp > θ1 or P canopy > θ1, trigger environmental stress-like anomalies; When P nutrient > θ1, trigger the nutritional imbalance class exception; When P chloro > θ1 and P coverage > θ1 or P chloro > θ1 and P water < θ2, trigger the pest abnormality; wherein P water represents the abnormal root cause probability value of the soil volume water content node, P temp represents the abnormal root cause probability value of the soil temperature node, P nutrient represents the abnormal root cause probability value of the soil nutrient content change rate node, P chloro represents the abnormal root cause probability value of the chlorophyll content index node, P canopy represents the abnormal root cause probability value of the canopy temperature node, P coverage represents the abnormal root cause probability value of the vegetation coverage node, θ1represents a first probability threshold value, and θ2represents a second probability threshold value. When it is determined as a water shortage class anomaly, the irrigation control instruction is triggered to control the irrigation system to perform quantitative water replenishment. When it is determined as a disease and pest class anomaly, the disease and pest control instruction is triggered. When it is determined as a nutrient imbalance class anomaly, the fertilization control instruction is triggered. When it is determined as an environmental stress class anomaly, the environmental regulation instruction is triggered. The method comprises the following steps:

9. A method for monitoring and early warning of landscaping maintenance based on artificial intelligence, comprising the artificial intelligence-based landscaping maintenance monitoring and early warning system as described in any one of claims 1-8, characterized in that, Step S1: collecting multi-source monitoring data of the landscaping area; Step S2: performing feature extraction and fusion processing on the multi-source monitoring data to obtain a spatiotemporal grid representation; Step S3: constructing a garden health causal graph, and performing state analysis on the spatiotemporal grid representation based on the garden health causal graph to obtain abnormal root cause probability values; ​ Step S4: analyzing the abnormal root cause probability value and generating a maintenance instruction.

Images