An artificial intelligence-based coastal residential area plant health state prediction method

Abstract

The present application relates to the technical field of wisdom agriculture, in particular to a plant health state prediction method for coastal residential areas based on artificial intelligence, which specifically comprises the following steps: deploying sensors to collect real-time related parameters of plants, and dividing the preprocessed parameters into a training set and a test set; defining each plant in the training set as a graph node and calculating a weighted dynamic adjacency matrix; constructing a deep-coupled space-time dependence model, inputting the feature sequence of each graph node and the weighted dynamic adjacency matrix sequence of each time step into the model, and obtaining a plant health state prediction result; training and optimizing the deep-coupled space-time dependence model using the data in the training data set, and generating an optimized model; and testing the data in the test set to obtain a final plant health state prediction result. The present application can improve the accuracy of the plant health state prediction result for coastal residential areas by constructing a model that reflects the deep coupling of dynamic correlation and space-time information of real physical processes.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to an artificial intelligence-based method for predicting the health status of plants in coastal residential areas. Background Technology

[0002] As the core space for population concentration in coastal cities, coastal residential areas rely heavily on their plant landscapes. These landscapes not only enhance the quality of the living environment and regulate the regional microclimate (such as cooling, humidifying, and purifying the air), but also serve as crucial barriers against the unique coastal environment (such as salt spray and storm surges) and for maintaining ecosystem stability. Studying the health status of plants in coastal residential areas essentially aims to achieve synergistic development between society and ecology by ensuring the health of these "green carriers," thereby improving the living quality of coastal residents and realizing the sustainable governance of coastal cities.

[0003] Existing plant health prediction methods are mostly designed based on the growth environment of common terrestrial plants, and are not fully adapted to the complex and special scenarios of coastal residential areas, such as high salt spray, strong sea winds, and soil salinization. Therefore, how to effectively utilize the complex spatiotemporal dependent data inherent in coastal residential areas is the key to adapting plant health prediction methods to the specific conditions of coastal residential areas. Shifting from "post-event response" to "pre-event prediction" in plant health prediction is a research hotspot in this field. Researchers typically use spatiotemporal graph neural networks as the mainstream technical solution. These models usually abstract each plant or monitoring unit as a node in a graph, using graph neural networks to capture the spatial interactions between plants, and combining this with recurrent neural networks for modeling. The "GNN+RNN" architecture, which leverages the inherent temporal evolution of each node, has achieved significant success in areas such as traffic flow prediction and weather forecasting, and has been initially applied to plant health prediction. However, directly applying existing spatiotemporal graph network models to scenarios with specific environmental stresses, such as coastal residential areas, reveals fundamental limitations in spatial dimension, model architecture, and the fundamental paradigm of model-driven approaches. These limitations lead to deviations in the model's simulation of real risk propagation paths, an inability to perceive the historical dependencies of nodes, and an over-reliance on large-scale, high-quality labeled datasets. Consequently, this results in insufficient prediction accuracy, high deployment costs, and limitations on the model's depth of characterization of complex spatiotemporal interactions. In summary, existing technologies cannot spatially reflect environment-driven dynamic anisotropic relationships and neglect the influence of historical states on current spatial interactions in terms of spatiotemporal coupling mechanisms.

[0004] Therefore, this invention proposes an artificial intelligence-based method for predicting the health status of plants in coastal residential areas to solve the above problems. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by developing an artificial intelligence-based method for predicting the health status of plants in coastal residential areas. By performing dynamic spatial correlation, historical perception spatiotemporal coupling, and calculating the salt spray-soil salinity accumulation loss term, it can solve the problems of existing methods being unsuitable for coastal scenarios, having prediction lag, and being dependent on data, thereby improving the prediction of the health status of plants in coastal residential areas.

[0006] The technical solution of this invention to solve the technical problem is a method for predicting the health status of plants in coastal residential areas based on artificial intelligence, comprising the following steps:

[0007] S1. Deploy various types of sensors to collect physiological and biochemical parameters and environmental physical parameters of plants in coastal residential areas and perform standardized processing to label the true health status of each plant, and then divide the data into training and test sets.

[0008] S2. Define each plant as a graph node, and calculate the weighted dynamic adjacency matrix based on the Gaussian kernel weights of the static distance between different graph nodes and the learnable multi-path dynamic modulation module.

[0009] S3. Construct a deeply coupled spatiotemporal dependency model, which includes an input layer, a spatial attention aggregation module for multi-head historical state perception, a stacked gating temporal evolution module, and a prediction output layer. Input the temporal feature sequence and weighted dynamic adjacency matrix sequence of each graph node into the model to obtain the plant health status prediction result.

[0010] S4. Calculate the composite loss function based on the loss-driven loss term and the salt accumulation physical constraint term. Use the training set to train and optimize the deeply coupled spatiotemporal dependency model to generate the optimized model.

[0011] S5. Input the data from the test set into the optimized, deeply coupled spatiotemporal dependency model to obtain the final prediction results of plant health status.

[0012] S1 is as follows:

[0013] Physiological and biochemical parameters include plant growth and metabolism indicators, normalized difference vegetation index, photochemical vegetation index, canopy temperature, leaf area index, and salt stress index.

[0014] Environmental physical parameters include wind speed vector, soil electrical conductivity, soil moisture, plant geographical location, and distance of plants from the coastline;

[0015] The sensors include remote sensing sensors, near-ground sensors, anemometers, soil conductivity and soil moisture sensors, plant physiological sensors, biochemical index detection sensors, salt spray concentration sensors, and tidal water level sensors.

[0016] S2 is as follows:

[0017] S2.1, Define graph nodes:

[0018] Define each plant as a graph node, and set the total number of graph nodes to [value]. , Indicates the first Each graph node The feature sequence of each graph node at each time step is represented as follows: , , Indicates the total monitoring period. Indicates time step The feature sequence of each graph node Indicates time step First Features of each graph node ;

[0019] S2.2 Calculate the dynamic adjacency matrix:

[0020] 1) Calculate the Gaussian kernel weights for the static distance between any two graph nodes, and calculate the Euclidean distance between any pair of nodes based on the geographical location of each graph node; then, using the Euclidean distance as the independent variable and the preset scale hyperparameter as the bandwidth, map the distance to Gaussian kernel weights between 0 and 1 using the Gaussian kernel function.

[0021] 2) At each time step, the anisotropic effects brought about by different physical media are calculated through the learnable multi-path dynamic modulation module to obtain the multi-path dynamic modulation value of the node pair formed by any two graph nodes at each time step. The physical media are specifically the acquired environmental physical parameters.

[0022] The multi-path dynamic modulation value is specifically calculated by determining the degree of consistency between the connection direction of any node pair and the dominant direction of each environmental physical parameter, and then processed by a linear rectification function. Calculate the strength of the directional influence of each physical medium on any node pair; simultaneously input the environmental physical parameter vector to the linear mapping of the learnable parameters to obtain the modulation coefficient between 0 and 1; then scale the strength of the directional influence using the modulation coefficient to obtain the modulation value of any physical medium on any node pair.

[0023] 3) At each time step, the Gaussian kernel weight of the static distance is weighted and superimposed with the multi-path dynamic modulation values ​​obtained from each physical medium to obtain the time-varying correlation strength between nodes, that is, the elements of the dynamic adjacency matrix generated between any two graph nodes at each time step. ,when Or when the static distance between node pairs is greater than the set maximum influence distance, , and Both represent the indices of graph nodes. Specifically, this is expressed in time steps. No. The graph node and the first Elements of the dynamic adjacency matrix generated between nodes in the graph;

[0024] The dynamic adjacency matrix sequence within a monitoring time period is represented as follows: , No. Dynamic adjacency matrix at each time step It is expressed as follows:

[0025] ;

[0026] Then process the generated dynamic adjacency matrix To perform symmetric normalization, first calculate the degree matrix of the nodes in both the row and column directions, and then use the square root of the degree matrix as the normalization factor for the dynamic adjacency matrix. Weighting is performed on both the left and right sides to generate a normalized dynamic adjacency matrix. , The weighted adjacency matrix sequence within each monitoring time period is represented as follows: , Indicates the normalized i-th A dynamic adjacency matrix for each time step.

[0027] S3 is as follows:

[0028] The input layer receives the feature sequence of each graph node at each time step. and the normalized weighted dynamic adjacency matrix sequence ;in, , The feature dimension is represented by the number of physiological and biochemical parameters included; the weighted dynamic adjacency matrix sequence is specifically introduced as a relational attribute in the spatial attention aggregation module of multi-head historical state perception.

[0029] Then, a spatial attention aggregation module with multi-head historical state awareness is used to extract spatial interaction features between graph nodes, and a stacked gating temporal evolution module is used to capture temporal evolution relationships. Finally, the prediction output layer evaluates the model at the last monitoring time step. Decode the hidden state and output the future. The predicted feature values ​​of each node within a time step are used to determine the plant health status based on the predicted feature values ​​and a preset threshold, thus obtaining the plant health status prediction result.

[0030] In the prediction output layer, the model will monitor the last time step at each graph node. The final hidden state of the output The input is decoded in a multilayer perceptron, which consists of several fully connected layers, including an input layer and multiple processing layers. The activation function has a hidden layer and an output layer, the output layer using... Activation function The future is directly output through a multi-layer sensor. The predicted feature values ​​for each time step are then used to determine the plant health status based on these predicted feature values ​​and a preset threshold, thus obtaining the plant health status prediction result, as shown below:

[0031] ,

[0032] in, This represents the operation of a multilayer perceptron. The dimension of the output sequence is , Indicates the first Each graph node in the future Predicted feature values ​​for each time step , This represents the output of the last layer of the stacked gated timing evolution module. Each graph node at the last monitoring time step The hidden state of the output. This indicates the total number of layers in the stacked gated timing evolution module.

[0033] The operation of the spatial attention aggregation module for multi-head historical state perception is as follows:

[0034] The spatial attention aggregation module with multi-head historical state perception calculates the correlation between the target node and its neighboring nodes in terms of feature representation and historical hidden state with respect to each physical medium at each time step to obtain attention weights. It then uses multiple attention heads to weight and aggregate the features of neighboring nodes to obtain spatial interaction features that fuse historical state information. The calculation formula is as follows:

[0035] ,

[0036] ,

[0037] ,

[0038] in, Indicates the first A physical medium in Time steps on the first The graph node and the first Attention weights between the nodes of the graph, the th The first graph node is used as the target node, and the second... Each graph node serves as a neighbor node, and the total number of physical media is [number]. ; This represents the operation of the linear rectification activation function. This represents the operation of the linear exponential unit activation function; Represented as the first A scoring vector calibrated by a physical medium; and They represent the first A linear transformation of physical medium characteristics and historical hidden states; Indicates the first The weighting coefficient of each physical medium; Indicates the first The historical hidden state of each graph node Indicates the first The historical hidden state of each graph node; Indicates the first The graph node and the first Each graph node at time step The weighted dynamic adjacency matrix; express Time step The set of neighboring nodes of a graph node; This represents a vector concatenation operation; Represents the target node Regarding the first A physical medium in time step Spatial interaction characteristics, Represents the target node At time step The spatial interaction characteristics are summarized, which are the spatial interaction characteristics of various physical media.

[0039] The specific operations in the stacked gating timing evolution module are as follows:

[0040] Stacked gated timing evolution module consists of One gated loop unit Composed of stacked layers;

[0041] First layer The input is the output of the spatial attention module. The historical hidden state of the previous time step ;

[0042] In the stacked gating timing evolution module, the first layer The state update process receives the first The output of the layer As input, and update its own hidden state to ,in, ;

[0043] The last layer in the stacked gating timing evolution module Hidden state of output As this time step The final spatiotemporal feature representation of the node.

[0044] S4 is as follows:

[0045] During the training and optimization process of the model, the Adam optimizer is used. The composite loss function is minimized by using the mini-batch gradient method combined with the backpropagation algorithm along time. Training termination conditions are set until the conditions are met to terminate the training of the model and generate the optimized model.

[0046] The calculation process of the composite loss function is as follows:

[0047] 1) Loss-driven loss term:

[0048] The mean squared error of all nodes at all prediction time steps is used as the loss function. To penalize features across different dimensions on an equal scale, the loss function is also averaged across the feature dimensions, calculated as follows:

[0049] ,

[0050] in, The model represents the first The graph node in the future Feature prediction values ​​at time point This represents the corresponding actual observed value. Represents the square of the L2 norm;

[0051] 2) Physical constraint term for salt accumulation:

[0052] The salt accumulation physical constraint term is calculated based on the squared difference between the partial derivative of the salt stress index contained in the feature prediction value output by the model and the salt input flux. The salt input flux includes air salt spray input flux and soil salt input flux.

[0053] The air salt spray input flux is calculated from the wind speed vector and the unit vector along the coastline at each graph node, while also considering the attenuation of the distance from each graph node to the coastline; the soil salinity input flux is calculated from the soil conductivity and soil moisture; and weights are assigned to the salinity input flux, including the air salt spray input flux and the soil salinity input flux.

[0054] Among them, the salt stress index is a physiological and biochemical parameter, while the wind speed vector, soil electrical conductivity, and soil moisture are all environmental physical parameters.

[0055] The formula for calculating the air salt spray input flux is as follows:

[0056] ,

[0057] The formula for calculating soil salinity input flux is as follows:

[0058] ,

[0059] in, express Time step Air salt spray input flux at each graph node. Learnable weighting coefficients representing the input flux of air salt spray. express Wind speed vector at time step, Indicates the first The unit vector in the direction of the coastline at each graph node. Indicates the first The distance from each graph node to the coastline. Represents the distance attenuation constant; express Time step Soil salinity input flux for each graph node Learnable weighting coefficients representing soil salinity input flux. express Time step Soil conductivity collected at each graph node express Time step Soil moisture collected at each map node;

[0060] The formula for calculating the composite loss function is as follows:

[0061] ,

[0062] in, Represents the composite loss function. This represents the mean squared error loss based on the loss-driven loss term. This represents the loss based on the physical constraint term of salt accumulation. This represents the set of all learnable parameters of the model. This represents an adjustable hyperparameter used to balance the importance of the two losses.

[0063] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0064] (1) Multi-source heterogeneous information fusion to improve prediction accuracy: This invention integrates the physiological and biochemical parameters of plants in coastal residential areas with environmental physical parameters through graph node features, comprehensively considers static spatial distance, dynamic environmental factors and the directional influence of physical media, and constructs a weighted dynamic adjacency matrix sequence, which can more realistically depict the spatiotemporal interaction relationship between plants and significantly improve the accuracy and stability of plant health status prediction.

[0065] (2) Dynamic spatiotemporal dependency modeling to capture complex evolutionary patterns: This invention adopts a deeply coupled spatiotemporal dependency model, which includes a spatial attention aggregation module for multi-head historical state perception and a stacked gating temporal evolution module. It can fully model the interaction of neighboring nodes in the spatial dimension and capture the dynamic evolutionary patterns of plant health status in the temporal dimension, making the model more robust to non-stationary environments.

[0066] (3) Introducing physical constraints to ensure the rationality of prediction results: This invention introduces a salt accumulation physical constraint term in the composite loss function, explicitly models the air salt spray input flux and soil salt input flux and assigns learnable weights, so that the model is more suitable for the prediction of plant health status in coastal residential areas, realizes the physical consistency constraint of plant health prediction, effectively avoids the physically unreasonable results that may occur in the pure data-driven model, and improves the interpretability and reliability of the prediction.

[0067] In summary, this invention constructs time-varying spatial adjacency relationships and designs a deeply coupled graph cyclic network model to calculate the salt spray-soil salinity accumulation loss term. This enables accurate modeling of complex and dynamic spatiotemporal dependencies, adapts to the prediction of plant health status in coastal scenarios, solves the problems of prediction lag and data dependence, and thus improves the accuracy of plant health status prediction results in coastal residential areas. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0069] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0070] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0071] Example 1

[0072] like Figure 1 As shown, an artificial intelligence-based method for predicting the health status of plants in coastal residential areas includes the following steps:

[0073] S1. Deploy various types of sensors within the monitoring area to collect physiological and biochemical parameters and environmental physical parameters of each plant in the coastal residential area in real time. Standardize the collected multidimensional raw data and label the true health status of each plant. Then divide the processed multidimensional raw data into training set and test set.

[0074] S2. Define each plant in the training set as a graph node. Calculate the weighted dynamic adjacency matrix based on the Gaussian kernel weights of the static distances between different graph nodes and the learnable multi-path dynamic modulation module, thereby generating... The weighted dynamic adjacency matrix sequence within a monitoring time period;

[0075] S3. Construct a deeply coupled spatiotemporal dependency model, which includes an input layer, a spatial attention aggregation module for multi-head historical state perception, a stacked gated temporal evolution module, and a prediction output layer. Input the feature sequence of each graph node at each time step and the weighted dynamic adjacency matrix sequence into the model to obtain the plant health status prediction result.

[0076] S4. Use the concentrated data in the training dataset to train and optimize the deeply coupled spatiotemporal dependent model. Calculate the composite loss function based on the combination of loss-driven loss term and salt accumulation physical constraint term. Use the Adam optimizer, mini-batch gradient method, and backpropagation algorithm along time to minimize the composite loss function. Set training termination conditions until the conditions are met to terminate the training of the model and generate the optimized model.

[0077] S5. Using the same calculation method on the data in the test set, the feature sequence and weighted dynamic adjacency matrix sequence of each graph node at each time step are obtained. These are then input into the optimized deep-coupled spatiotemporal dependency model to predict the plant health status of the coastal residential area and obtain the final plant health status prediction result.

[0078] In a specific implementation, S1 is as follows:

[0079] Physiological and biochemical parameters include plant growth and metabolism indicators, normalized difference vegetation index, photochemical vegetation index, canopy temperature, leaf area index, salt stress index, etc.

[0080] Environmental physical parameters include wind speed vector, soil electrical conductivity, soil moisture, plant geographical location, and distance of plants from the coastline.

[0081] Sensors include remote sensing sensors, near-ground sensors, anemometers, soil conductivity and soil moisture sensors, plant physiological sensors, biochemical index detection sensors, salt spray concentration sensors, and tidal water level sensors.

[0082] In a specific implementation, S2 is as follows:

[0083] S2.1, Define graph nodes:

[0084] Define each plant as a graph node, and set the total number of graph nodes to [value]. , Indicates the first Each graph node The feature sequence of each graph node at each time step is represented as follows: , , Indicates the total monitoring period. Indicates time step The feature sequence of each graph node Indicates time step First Features of each graph node ;

[0085] S2.2 Calculate the dynamic adjacency matrix:

[0086] 1) Calculate the Gaussian kernel weights for the static distance between any two graph nodes, and calculate the Euclidean distance between any pair of nodes based on the geographical location of each graph node; then, using the Euclidean distance as the independent variable and the preset scale hyperparameter as the bandwidth, map the distance to Gaussian kernel weights between 0 and 1 using the Gaussian kernel function.

[0087] The calculation formula is as follows:

[0088] ,

[0089] in, Indicates the first Each graph node and the Each graph node Gaussian kernel weights for static distances between them; Indicates the first Each graph node and the Each graph node The Euclidean distance between them, i.e., the static distance. ; This represents the preset scale hyperparameter, which sets a physically intuitive benchmark for the strength of the association between graph nodes. The closer the nodes are, the stronger the potential mutual influence, and it can provide a stable prior knowledge for the model.

[0090] 2) At each time step, the anisotropic effects brought about by different physical media are calculated through the learnable multi-path dynamic modulation module to obtain the multi-path dynamic modulation value of the node pair formed by any two graph nodes at each time step. The physical media are specifically the acquired environmental physical parameters.

[0091] The multi-path dynamic modulation value is specifically calculated by determining the degree of consistency between the connection direction of any node pair and the dominant direction of each environmental physical parameter, and then processed by a linear rectification function. Calculate the strength of the directional influence of each physical medium on any node pair; simultaneously input the environmental physical parameter vector to the linear mapping of the learnable parameters to obtain the modulation coefficient between 0 and 1; then scale the strength of the directional influence using the modulation coefficient to obtain the modulation value of any physical medium on any node pair.

[0092] The calculation formula is as follows:

[0093] ,

[0094] in, Indicates the first Each graph node and the Each graph node In the In the time step about the first The modulation value of a physical medium, , , This indicates the total number of physical media types, which include wind speed, wind direction, irrigation system status, and surface runoff direction, etc. Indicates the first Each time step contains A vector of environmental factors for a physical medium , Indicates the first In the first time step A vector of environmental factors for a physical medium; Indicates from the first Each graph node Pointing to the Each graph node The position vector; Indicates the first The weight vector of each physical medium Indicates the first Learnable bias of a physical medium; The Euclidean norm of a vector; This represents the Sigmoid activation function; Represents the linear rectified function;

[0095] 3) At each time step, the Gaussian kernel weight of the static distance is weighted and superimposed with the multi-path dynamic modulation values ​​obtained from each physical medium to obtain the time-varying correlation strength between nodes, that is, the elements of the dynamic adjacency matrix generated between any two graph nodes at each time step. ,when Or when the static distance between node pairs is greater than the set maximum influence distance, , and Both represent the indices of graph nodes. Specifically, this is expressed in time steps. No. The graph node and the first Elements of the dynamic adjacency matrix generated between nodes in the graph;

[0096] The formula for calculating the elements of the dynamic adjacency matrix is ​​as follows:

[0097] ,

[0098] in, Indicates the first In the first time step Each graph node and the Each graph node Elements of the generated dynamic adjacency matrix;

[0099] The dynamic adjacency matrix sequence within a monitoring time period is represented as follows: , No. Dynamic adjacency matrix at each time step It is expressed as follows:

[0100] ;

[0101] Then process the generated dynamic adjacency matrix To perform symmetric normalization, first calculate the degree matrix of the nodes in both the row and column directions, and then use the square root of the degree matrix as the normalization factor for the dynamic adjacency matrix. Weighting is performed on both the left and right sides to generate a normalized dynamic adjacency matrix. , The weighted adjacency matrix sequence within each monitoring time period is represented as follows: , Indicates the normalized i-th A dynamic adjacency matrix for each time step;

[0102] The calculation process for dynamic adjacency matrix normalization is as follows:

[0103] ,

[0104] in, express a diagonal matrix; Represents the identity matrix.

[0105] It should be noted that the weighted adjacency matrix sequence is not only used to characterize the spatiotemporal topological relationship between graph nodes, but also serves as a relational attribute input in the subsequent spatial attention modeling step in plant health status prediction. This allows for the explicit introduction of environment-driven dynamic adjacency information in spatiotemporal dependency modeling. In the attention weight calculation, the weighted adjacency matrix sequence explicitly influences the correlation score between nodes through the relation modulation factor, enabling the spatial aggregation result to simultaneously reflect the time-varying correlation driven by the environment and the historical evolutionary characteristics of the plant itself.

[0106] In a specific implementation, S3 is as follows:

[0107] The input layer receives the feature sequence of each graph node at each time step. and the normalized weighted dynamic adjacency matrix sequence ;in, , The feature dimension is represented by the number of physiological and biochemical parameters included; the weighted dynamic adjacency matrix sequence is specifically introduced as a relational attribute in the spatial attention aggregation module of multi-head historical state perception.

[0108] Then, a spatial attention aggregation module with multi-head historical state awareness is used to extract spatial interaction features between graph nodes, and a stacked gating temporal evolution module is used to capture temporal evolution relationships. Finally, the prediction output layer evaluates the model at the last monitoring time step. Decode the hidden state and output the future. The predicted feature values ​​of each node within a time step are used to determine the plant health status based on the predicted feature values ​​and a preset threshold, thus obtaining the plant health status prediction result.

[0109] In the prediction output layer, the model will monitor the last time step at each graph node. The final hidden state of the output The input is decoded in a multilayer perceptron, which consists of several fully connected layers, including an input layer and multiple processing layers. The activation function has a hidden layer and an output layer, the output layer using... Activation function The future is directly output through a multi-layer sensor. The predicted feature values ​​for each time step are then used to determine the plant health status based on these predicted feature values ​​and a preset threshold, thus obtaining the plant health status prediction result, as shown below:

[0110] ,

[0111] in, This represents the operation of a multilayer perceptron. The dimension of the output sequence is , Indicates the first Each graph node in the future Predicted feature values ​​for each time step , This represents the output of the last layer of the stacked gated timing evolution module. Each graph node at the last monitoring time step The hidden state of the output. This indicates the total number of layers in the stacked gated timing evolution module.

[0112] In a specific implementation, the spatial attention aggregation module for multi-head historical state perception operates as follows:

[0113] The spatial attention aggregation module with multi-head historical state perception calculates the correlation between the target node and its neighboring nodes in terms of feature representation and historical hidden state with respect to each physical medium at each time step to obtain attention weights. It then uses multiple attention heads to weight and aggregate the features of neighboring nodes to obtain spatial interaction features that fuse historical state information. The calculation formula is as follows:

[0114] ,

[0115] ,

[0116] ,

[0117] in, Indicates the first A physical medium in Time steps on the first The graph node and the first Attention weights between the nodes of the graph, the th The first graph node is used as the target node, and the second... Each graph node serves as a neighbor node, and the total number of physical media is [number]. ; This represents the operation of the linear rectification activation function. This represents the operation of the linear exponential unit activation function; Represented as the first A scoring vector calibrated by a physical medium; and They represent the first A linear transformation of physical medium characteristics and historical hidden states; Indicates the first The weighting coefficient of each physical medium; Indicates the first The historical hidden state of each graph node Indicates the first The historical hidden state of each graph node; Indicates the first The graph node and the first Each graph node at time step The weighted dynamic adjacency matrix; express Time step The set of neighboring nodes of a graph node; This represents a vector concatenation operation; Represents the target node Regarding the first A physical medium in time step Spatial interaction characteristics, Represents the target node At time step The spatial interaction characteristics are summarized, which are the spatial interaction characteristics of various physical media.

[0118] In a specific implementation, the operations in the stacked gating timing evolution module are as follows:

[0119] Stacked gated timing evolution module consists of One gated loop unit Composed of stacked layers;

[0120] First layer The input is the output of the spatial attention module. The historical hidden state of the previous time step ;

[0121] In the stacked gating timing evolution module, the first layer The state update process receives the first The output of the layer As input, and update its own hidden state to ,in, ;

[0122] The last layer in the stacked gating timing evolution module Hidden state of output As this time step The final spatiotemporal feature representation of the node.

[0123] In a specific implementation, S4 is as follows:

[0124] 1) Loss-driven loss term:

[0125] The mean squared error of all nodes at all prediction time steps is used as the loss function. To penalize features across different dimensions on an equal scale, the loss function is also averaged across the feature dimensions, calculated as follows:

[0126] ,

[0127] in, The model represents the first The graph node in the future Feature prediction values ​​at time point This represents the corresponding actual observed value. Represents the square of the L2 norm;

[0128] 2) Physical constraint term for salt accumulation:

[0129] The cumulative physical constraint term for salt is calculated based on the squared difference between the partial derivative of the salt stress index included in the model's output feature predictions and the salt input flux. The salt input flux includes air salt spray input flux and soil salinity input flux, and the calculation formula is as follows:

[0130] ,

[0131] in, express Time step Salt stress index at each node in the graph Represents the L2 norm;

[0132] The air salt spray input flux is calculated from the wind speed vector and the unit vector along the coastline at each graph node, while also considering the attenuation of the distance from each graph node to the coastline; the soil salinity input flux is calculated from the soil conductivity and soil moisture; and weights are assigned to the salinity input flux, including the air salt spray input flux and the soil salinity input flux.

[0133] Among them, the salt stress index is a physiological and biochemical parameter, while the wind speed vector, soil electrical conductivity, and soil moisture are all environmental physical parameters.

[0134] The formula for calculating the air salt spray input flux is as follows:

[0135] ,

[0136] The formula for calculating soil salinity input flux is as follows:

[0137] ,

[0138] in, express Time step Air salt spray input flux at each graph node. Learnable weighting coefficients representing the input flux of air salt spray. express Wind speed vector at time step, Indicates the first The unit vector in the direction of the coastline at each graph node. Indicates the first The distance from each graph node to the coastline. Represents the distance attenuation constant; express Time step Soil salinity input flux for each graph node Learnable weighting coefficients representing soil salinity input flux. express Time step Soil conductivity collected at each graph node express Time step Soil moisture collected at each map node;

[0139] The formula for calculating the composite loss function is as follows:

[0140] ,

[0141] in, Represents the composite loss function. This represents the mean squared error loss based on the loss-driven loss term. This represents the loss based on the physical constraint term of salt accumulation. This represents the set of all learnable parameters of the model. This represents an adjustable hyperparameter used to balance the importance of the two losses.

[0142] Example 2

[0143] This embodiment uses simulation experiments to construct the dataset and scenario. Let the monitoring area be... A two-dimensional continuous space, with uniform or clustered sampling within the region. Plant location Each position corresponds to a graph node. The time axis is discretized into Each monitoring time step At each time step, a feature vector containing physiological, biochemical, and environmental physical quantities is generated for each node. And generate accurate health status labels. Node characteristics include physiological and biochemical quantities such as relative chlorophyll content, transpiration rate, and photosynthetic rate, as well as environmental physical quantities such as wind speed vector, soil electrical conductivity, soil moisture, distance from the node to the coastline, and unit vector of the coastline normal. Health labels are determined by potential "salt stress intensity" and threshold rules.

[0144] To verify the effectiveness of the method, our method was compared with several other methods, namely GRU-Only (which does not explicitly construct a graph, but only uses node-independent time series GRU regression and judgment) and STGCN (which uses a static Gaussian kernel graph). Spatiotemporal graph convolutional networks (methods without dynamic modulation and physical constraints), GAT-Static (graph attention networks based on static graphs, where attention does not introduce historical hidden states), and Ours w / o Phys (representing the method after removing physical constraint terms in this invention).

[0145] The evaluation indicators include both regression and classification indicators. The regression side uses MAE and RMSE to measure future performance. The numerical error of the key channel is analyzed; the health status is assessed by precision, recall, and F1 score at the classification end, and the mean and standard deviation obtained based on the test set are given.

[0146] During the experiment, the following settings were made: , , , Attention count , number of floors Hidden Dimensions , The experimental results are shown in Table 1.

[0147] Table 1: Main Results of Simulation Experiment

[0148]

[0149] As shown in Table 1, the multi-head historical state perception attention introduced by the dynamic graph can significantly reduce regression error. The present invention (Ours) with the addition of physical constraint terms can further improve accuracy and improve the F1 score of health determination.

[0150] To further understand the contributions of each component, ablation experiments were also conducted. When wind direction consistency modulation was removed, Precision and F1 decreased by approximately 1.6% and 1.4%, respectively; when soil diffusion modulation was removed, MAE increased by approximately 5.1%. When the value is set from 0.2 to 0, the RMSE increases by about 6.5%, which shows that physical consistency regularization can suppress unreasonable oscillations and improve generalization.

[0151] Example 3

[0152] To demonstrate that the method of this invention can improve the quality of life and the efficiency of plant health management in coastal residential areas, a "garden community" in a coastal area of ​​Qingdao is used as an example to illustrate the complete implementation process from sensor deployment and data collection to model prediction and intervention implementation.

[0153] Located in the northern temperate monsoon climate zone, the "garden community" is significantly influenced by the maritime climate. It is about 150-300 meters from the coastline, with an average annual salt spray deposition of about 380 mg / m². During the typhoon season in summer, the sea breeze speed can reach 12-16 m / s. The soil is mainly coastal tidal soil, with an average surface soil electrical conductivity of about 3.8 dS / m. The "garden community" is mainly planted with trees and shrubs, with 120 trees and 180 shrubs.

[0154] The method of this invention is applied to the typhoon season of the "garden community" to reduce the mortality rate and maintenance costs of plants during the typhoon season by providing early warning of the health status of trees and shrubs 7 days in advance, thereby improving the greening and aesthetics of the "garden community" and protecting the living environment of residents.

[0155] The area was divided into three zones based on distance from the coastline: near-coastal zone A (within 150 meters from the coastline), intermediate zone B (150-250 meters from the coastline), and inner zone C (250-300 meters from the coastline). In each zone, 20 trees and 15 shrubs were randomly selected as monitoring targets, for a total of 105 plants, or 105 map nodes.

[0156] Near-ground spectral sensors were deployed in the middle of the canopy of each plant to collect normalized difference vegetation index (NDE), photochemical vegetation index (PDV), and canopy temperature. Soil sensors were deployed in the root zone of each plant to collect soil conductivity and soil moisture. A miniature anemometer was deployed in each of zones A, B, and C to measure wind speed vector and wind direction. Two salt spray concentration sensors were deployed in zone A and one in zone B to measure airborne salt spray concentration. Three healthy leaves from each plant were selected and marked, and plant physiological sensors were deployed to collect salt stress index (SSI) and leaf area index (NDVI).

[0157] The collected data were normalized, and the feature dimension F=8 for each plant was determined based on the number of collected parameters. Health status was labeled, with horticultural experts assessing plant health weekly and setting labeling standards based on physiological parameters. The standards for a healthy state were: SSI≤0.3, NDVI≥0.6, and no yellow spots on leaves; a sub-healthy state was defined as: 0.3<SSI≤0.6, 0.4≤NDVI<0.6, and a small number of yellow spots on leaves (≤5%); and a stressed state was defined as: SSI>0.6, NDVI<0.4, and yellow spots accounting for >5% of leaves.

[0158] Data was collected from July to September. The collected data was divided into training and test sets in a 7:3 ratio based on time series data, and used to predict future trends. SSI and NDVI values ​​for the day;

[0159] After defining the graph nodes, calculate the dynamic adjacency matrix. First, calculate the static distance Gaussian kernel weight between any two nodes. Taking node i=1 in region A and node j=20 in region B as an example, the coordinates of node i=1 are N36°04', E120°22', and the coordinates of node j=20 are N36°04', E120°23'. The Euclidean distance between the two nodes is 92 meters. The scale hyperparameter is known. =45 meters, the weight is calculated to be 0.15 according to the Gaussian kernel function; considering the interaction range of plant roots, if the node spacing is greater than 90 meters, the weight is 0; then, wind speed vector and soil conductivity are selected as the core physical medium, and wind speed modulation value and soil conductivity modulation value are calculated; finally, the dynamic adjacency matrix is ​​calculated based on the weight and modulation value and normalized.

[0160] The time-series features and dynamic adjacency matrix of each graph node are input into a deeply coupled spatiotemporal dependency model to obtain the predicted health status. The model is trained and optimized based on the data in the training set, and then the data in the test set is input into the optimized model to obtain the final detection results. During the testing process, an intervention effect comparison experiment was also conducted. One plant from each of regions A, B, and C was selected as the intervention group, and one plant from each of regions A, B, and C was selected as the non-intervention group. After the experiment, the average SSI of the intervention group was 0.52 (sub-healthy), the average NDVI was 0.51 (borderline healthy), the average proportion of yellow spots on leaves was 3%, and all three plants survived. The average SSI of the non-intervention group was 0.78 (stressed state), the average NDVI was 0.35 (stressed state), the average proportion of yellow spots on leaves was 12%, two plants survived, and one plant withered.

[0161] The intervention measures for the intervention group were to irrigate the plants under stress with fresh water to reduce soil salinity, and to spray the leaves with 2% propylene glycol to enhance the plants’ salt tolerance.

[0162] In summary, by using a dynamic adjacency matrix and a deep spatiotemporal coupling model to predict the health status of plants in coastal residential areas, and by optimizing the model with a salt spray-soil salinity accumulation loss term, the health status of plants can be accurately predicted. Based on the prediction results, interventions can be implemented to reduce plant mortality, decrease maintenance costs, and improve the ecological maintenance efficiency of coastal residential areas.

[0163] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for predicting the health status of plants in coastal residential areas based on artificial intelligence, characterized in that, Includes the following steps: S1. Deploy various types of sensors to collect physiological and biochemical parameters and environmental physical parameters of plants in coastal residential areas and perform standardized processing to label the true health status of each plant, and then divide the data into training and test sets. S2. Define each plant as a graph node, and calculate the weighted dynamic adjacency matrix based on the Gaussian kernel weights of the static distance between different graph nodes and the learnable multi-path dynamic modulation module. S2 is as follows: S2.1, Define graph nodes: Define each plant as a graph node, and set the total number of graph nodes to [value]. , Indicates the first Each graph node The feature sequence of each graph node at each time step is represented as follows: , , Indicates the total monitoring period. Indicates time step The feature sequence of each graph node Indicates time step First Features of each graph node ; S2.2 Calculate the dynamic adjacency matrix: 1) Calculate the Gaussian kernel weight for the static distance between any two graph nodes, and calculate the Euclidean distance between any pair of nodes based on the geographical location of each graph node; then, using the Euclidean distance as the independent variable and the preset scale hyperparameter as the bandwidth, map the distance to a Gaussian kernel weight between 0 and 1 using the Gaussian kernel function. 2) At each time step, the anisotropic effects brought about by different physical media are calculated through the learnable multi-path dynamic modulation module to obtain the multi-path dynamic modulation value of the node pair formed by any two graph nodes at each time step. The physical media are specifically the acquired environmental physical parameters. The multi-path dynamic modulation value is specifically calculated by determining the degree of consistency between the connection direction of any node pair and the dominant direction of each environmental physical parameter, and then processed by a linear rectification function. Calculate the strength of the directional influence of each physical medium on any node pair; simultaneously input the environmental physical parameter vector to the linear mapping of the learnable parameters to obtain the modulation coefficient between 0 and 1; then scale the strength of the directional influence using the modulation coefficient to obtain the modulation value of any physical medium on any node pair. 3) At each time step, the Gaussian kernel weight of the static distance is weighted and superimposed with the multi-path dynamic modulation values ​​obtained from each physical medium to obtain the time-varying correlation strength between nodes, that is, the elements of the dynamic adjacency matrix generated between any two graph nodes at each time step. ,when Or when the static distance between node pairs is greater than the set maximum influence distance, , and Both represent the indices of graph nodes. Specifically, this is expressed in time steps. No. The graph node and the first Elements of the dynamic adjacency matrix generated between nodes in the graph; The dynamic adjacency matrix sequence within a monitoring time period is represented as follows: , No. Dynamic adjacency matrix at each time step It is expressed as follows: ； Then process the generated dynamic adjacency matrix To perform symmetric normalization, first calculate the degree matrix of the nodes in both the row and column directions, and then use the square root of the degree matrix as the normalization factor for the dynamic adjacency matrix. Weighting is performed on both the left and right sides to generate a normalized dynamic adjacency matrix. , The weighted adjacency matrix sequence within each monitoring time period is represented as follows: , Indicates the normalized i-th A dynamic adjacency matrix for each time step; S3. Construct a deeply coupled spatiotemporal dependency model, which includes an input layer, a spatial attention aggregation module for multi-head historical state perception, a stacked gating temporal evolution module, and a prediction output layer. Input the temporal feature sequence and weighted dynamic adjacency matrix sequence of each graph node into the model to obtain the plant health status prediction result. The operation of the spatial attention aggregation module for multi-head historical state perception is as follows: The spatial attention aggregation module with multi-head historical state perception calculates the correlation between the target node and its neighboring nodes in terms of feature representation and historical hidden state with respect to each physical medium at each time step to obtain attention weights. It then uses multiple attention heads to weight and aggregate the features of neighboring nodes to obtain spatial interaction features that fuse historical state information. The calculation formula is as follows: ， ， ， in, Indicates the first A physical medium in Time steps on the first The graph node and the first Attention weights between the nodes of the graph, the th The first graph node is used as the target node, and the second... Each graph node serves as a neighbor node, and the total number of physical media is [number]. ; This represents the operation of the linear rectification activation function. This represents the operation of the linear exponential unit activation function; Represented as the first A scoring vector calibrated by a physical medium; and They represent the first A linear transformation of physical medium characteristics and historical hidden states; Indicates the first The weighting coefficient of each physical medium; Indicates the first The historical hidden state of each graph node Indicates the first The historical hidden state of each graph node; Indicates the first The graph node and the first Each graph node at time step The weighted dynamic adjacency matrix; express Time step The set of neighboring nodes of a graph node; This represents a vector concatenation operation; Represents the target node Regarding the first A physical medium in time step Spatial interaction characteristics, Represents the target node At time step The spatial interaction characteristics are summarized, which are the spatial interaction characteristics of various physical media. S4. Calculate the composite loss function based on the loss-driven loss term and the salt accumulation physical constraint term. Use the training set to train and optimize the deeply coupled spatiotemporal dependency model to generate the optimized model. During the training and optimization process of the model, the Adam optimizer is used. The composite loss function is minimized by using the mini-batch gradient method combined with the backpropagation algorithm along time. Training termination conditions are set until the conditions are met to terminate the training of the model and generate the optimized model. The calculation process of the composite loss function is as follows: 1) Loss-driven loss term: The mean squared error of all nodes at all prediction time steps is used as the loss function. To penalize features across different dimensions on an equal scale, the loss function is also averaged across the feature dimensions, calculated as follows: ， in, The model represents the first The graph node in the future Feature prediction values ​​at time point This represents the corresponding actual observed value. Representing feature dimension, , Represents the square of the L2 norm; 2) Physical constraint term for salt accumulation: The salt accumulation physical constraint term is calculated based on the squared difference between the partial derivative of the salt stress index contained in the feature prediction value output by the model and the salt input flux. The salt input flux includes air salt spray input flux and soil salt input flux. The air salt spray input flux is calculated from the wind speed vector and the unit vector along the coastline at each graph node, while also considering the attenuation of the distance from each graph node to the coastline; the soil salinity input flux is calculated from the soil conductivity and soil moisture; and weights are assigned to the salinity input flux, including the air salt spray input flux and the soil salinity input flux. Among them, the salt stress index is a physiological and biochemical parameter, while the wind speed vector, soil electrical conductivity, and soil moisture are all environmental physical parameters. The formula for calculating the air salt spray input flux is as follows: ， The formula for calculating soil salinity input flux is as follows: ， in, express Time step Air salt spray input flux at each graph node. Learnable weighting coefficients representing the input flux of air salt spray. express Wind speed vector at time step, Indicates the first The unit vector in the direction of the coastline at each graph node. Indicates the first The distance from each graph node to the coastline. Represents the distance attenuation constant; express Time step Soil salinity input flux for each graph node Learnable weighting coefficients representing soil salinity input flux. express Time step Soil conductivity collected at each graph node express Time step Soil moisture collected at each map node; The formula for calculating the composite loss function is as follows: ， in, Represents the composite loss function. This represents the mean squared error loss based on the loss-driven loss term. This represents the loss based on the physical constraint term of salt accumulation. This represents the set of all learnable parameters of the model. This represents an adjustable hyperparameter used to balance the importance of the two losses; S5. Input the data from the test set into the optimized, deeply coupled spatiotemporal dependency model to obtain the final prediction results of plant health status.

2. The method for predicting the health status of plants in coastal residential areas based on artificial intelligence according to claim 1, characterized in that, S3 is as follows: The input layer receives the feature sequence of each graph node at each time step. and the normalized weighted dynamic adjacency matrix sequence ;in, , The feature dimension is represented by the number of physiological and biochemical parameters included; the weighted dynamic adjacency matrix sequence is specifically introduced as a relational attribute in the spatial attention aggregation module of multi-head historical state perception. Then, a spatial attention aggregation module with multi-head historical state awareness is used to extract spatial interaction features between graph nodes, and a stacked gating temporal evolution module is used to capture temporal evolution relationships. Finally, the prediction output layer evaluates the model at the last monitoring time step. Decode the hidden state and output the future. The predicted feature values ​​of each node within a time step are used to determine the plant health status based on the predicted feature values ​​and a preset threshold, thus obtaining the plant health status prediction result. In the prediction output layer, the model will monitor the last time step at each graph node. The final hidden state of the output The input is decoded in a multilayer perceptron, which consists of several fully connected layers, including an input layer and multiple processing layers. The activation function has a hidden layer and an output layer, the output layer using... Activation function The future is directly output through a multi-layer sensor. The predicted feature values ​​for each time step are then used to determine the plant health status based on these predicted feature values ​​and a preset threshold, thus obtaining the plant health status prediction result, as shown below: ， in, This represents the operation of a multilayer perceptron. The dimension of the output sequence is , Indicates the first Each graph node in the future Predicted feature values ​​for each time step , This represents the output of the last layer of the stacked gated timing evolution module. Each graph node at the last monitoring time step The hidden state of the output. This indicates the total number of layers in the network within the stacked gated timing evolution module.

3. The method for predicting the health status of plants in coastal residential areas based on artificial intelligence according to claim 2, characterized in that, The specific operations in the stacked gating timing evolution module are as follows: Stacked gated timing evolution module consists of One gated loop unit Composed of stacked layers; First layer The input is the output of the spatial attention module. The historical hidden state of the previous time step ; In the stacked gating timing evolution module, the first layer The state update process receives the first The output of the layer As input, and update its own hidden state to ,in, ; The last layer in the stacked gating timing evolution module Hidden state of output As this time step The final spatiotemporal feature representation of the node.

4. The method for predicting the health status of plants in coastal residential areas based on artificial intelligence according to claim 1, characterized in that: Physiological and biochemical parameters include plant growth and metabolism indicators, normalized difference vegetation index, photochemical vegetation index, canopy temperature, leaf area index, and salt stress index. Environmental physical parameters include wind speed vector, soil electrical conductivity, soil moisture, plant geographical location, and distance of plants from the coastline; The sensors include remote sensing sensors, near-ground sensors, anemometers, soil conductivity and soil moisture sensors, plant physiological sensors, biochemical index detection sensors, salt spray concentration sensors, and tidal water level sensors.

Images