Urban greening system based on big data and optimization method thereof

Abstract

The invention discloses an urban greening system based on big data and an optimization method thereof, relates to the technical field of artificial intelligence, and solves the problems of difficulty in heterogeneous data integration, insufficient real-time analysis capability of dynamic data and low management efficiency in an existing urban greening system. Multi-source heterogeneous data are collected through an API interface and Internet of Things communication, and are cleaned, aligned, converted and fused through a distributed real-time data stream processing platform; image recognition and a machine learning algorithm are used for monitoring diseases and insect pests in real time, and a growth prediction model is trained in combination with an LSTM network, so that the quick response capability of the system to environmental changes is enhanced; according to the method, target weights are dynamically adjusted and optimized through the deep Q network, a multi-target optimization model is constructed in combination with mixed integer linear programming, context awareness and a genetic algorithm, intelligent optimization of irrigation, fertilization, pest control and greening layout is achieved, and the management efficiency and the resource utilization rate of an urban greening system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically to an urban greening system based on big data and an optimization method thereof. Background Art

[0002] With the acceleration of urbanization, urban greening has become an important means to improve urban livability and sustainable development. Modern urban greening not only involves plant planting and maintenance, but also includes ecological environmental protection, water resources management, public participation and other aspects. The development of big data technology has brought new opportunities for urban greening management. By collecting and analyzing various types of data, such as meteorological data, soil moisture data, plant growth data, remote sensing images, etc., refined management can be achieved, resource utilization efficiency can be improved, and greening effects can be optimized. In addition, the application of Internet of Things technology makes data collection more convenient and real-time, providing strong technical support for urban greening management. However, although big data technology has shown great potential in urban greening management, it still faces many challenges in practical applications and urgently needs further research and improvement.

[0003] Existing urban greening management systems based on big data collect a large amount of data by integrating multiple data sources, such as weather stations, soil moisture sensors, remote sensing satellites, and drones, to support refined management and decision-making. These systems usually use data fusion technology to integrate data from different sources for comprehensive analysis. Some systems can more accurately predict plant growth conditions and irrigation needs by integrating meteorological data and soil moisture data. In addition, existing systems also use machine learning and artificial intelligence technologies for data mining and predictive analysis.

[0004] However, the existing urban greening management system still has some technical problems in practical application. First, heterogeneous data integration is difficult. The urban greening management system involves a wide range of data types, including but not limited to environmental sensor data (such as temperature and humidity, light intensity), geographic information data (such as soil type and distribution), plant growth cycle data, and human activity data (such as human density and activity area). These data come from different collection sources, with different formats, units, sampling frequencies, etc., which makes the cleaning, alignment, and conversion steps in the data fusion process complex and time-consuming; second, the real-time analysis capability of dynamic data is insufficient. The urban greening management system needs to monitor and respond to environmental changes in real time, but the existing systems are often based on static models and fixed rules and cannot effectively respond to dynamic changes; for example, the system may be based on a model established with historical data and cannot adjust the irrigation plan in time to cope with extreme weather such as drought or heavy rain. Finally, the management efficiency is low. The existing system has deficiencies in multi-objective optimization and it is difficult to optimize multiple objectives at the same time, such as plant growth, water resource utilization, environmental protection, etc., resulting in resource conflicts and low management efficiency.

[0005] Therefore, the present invention discloses an urban greening system based on big data and an optimization method thereof. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention discloses an urban greening system based on big data and an optimization method thereof, aiming to solve the problems raised in the above background technology.

[0007] The present invention adopts the following technical solutions:

[0008] A method for optimizing urban greening based on big data, comprising the following steps:

[0009] S1. Obtain temperature and humidity, light intensity, soil moisture and air quality data from environmental sensors through API interfaces and IoT wireless communication methods, obtain plant species and growth cycle data from plant growth cycle databases, obtain crowd density and activity area data from public data sources, and regularly obtain image data of green areas from drone inspection systems, and store them in temporary buffers;

[0010] S2. Input the data into a distributed real-time data stream processing platform to clean, align, convert and fuse the collected data; the distributed real-time data stream processing platform integrates a unified format data set through a deep heterogeneous fusion algorithm, including unified format environmental parameters, vegetation growth status, and human activity patterns;

[0011] S3, based on the image data of green areas of the drone inspection system, identifies the symptoms of pests and diseases through machine learning algorithms based on image recognition, and automatically triggers corresponding emergency response measures through event-driven mechanisms when abnormal environmental data is detected, including increasing the monitoring frequency and adjusting the irrigation strategy;

[0012] S4. Train historical environmental data, plant growth data, and pest and disease symptoms through long short-term memory networks to build a growth prediction model;

[0013] S5, inputting the real-time collected environmental data into the growth prediction model to obtain the growth prediction results of future environmental changes, and generating optimization strategies including irrigation plans, fertilization strategies, pest and disease control measures and greening area layout based on the prediction results;

[0014] S6. Define optimization objectives and initial weight values, including plant growth conditions, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects, and dynamically adjust the weight value of each objective according to environmental changes through the deep Q network algorithm;

[0015] S7, constructing a situation-based multi-objective optimization model through mixed integer linear programming, situational awareness and genetic algorithm, inputting real-time environmental data, pest and disease data, growth prediction data and optimization strategy into the situation-based multi-objective optimization model, obtaining the optimal comprehensive management strategy, including irrigation duration, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type, and storing the output results in a database;

[0016] S8. Send the output results of the scenario-based multi-objective optimization model to the control system, and the control system dynamically adjusts the operating parameters of the irrigation system and fertilization equipment and facilities through an adaptive control algorithm according to the optimization strategy.

[0017] As a further technical solution of the present invention, the working method of the deep heterogeneous fusion algorithm is: filling missing values ​​through mean interpolation, median interpolation and K nearest neighbor interpolation, and identifying and processing outliers through the Z-Score method; timestamp alignment through linear interpolation and spline interpolation; spatial coordinate alignment through geocoding and coordinate transformation; based on the aligned data, the deep heterogeneous fusion algorithm fuses the features of data from different sources through multimodal data; the multimodal data fusion extracts the features of image and time series data through convolutional neural networks and recurrent neural networks, and adopts mutual information method, principal component analysis and feature engineering, and generates a feature vector F through a dynamic weight adjustment mechanism. The fused feature vector F is used as the input of the deep neural network, and the nonlinear mapping relationship between input data and output data is learned through a multilayer perceptron and an autoencoder to generate a data set in a unified format; the multilayer perceptron learns the nonlinear mapping relationship between input data and output data through forward propagation and backpropagation algorithms, introduces nonlinear characteristics through the activation function ReLU, and improves the expression ability of the model; the autoencoder learns the low-dimensional representation of data through an encoder and a decoder, and optimizes the model parameters through the mean square error MSE.

[0018] As a further technical solution of the present invention, the working principle of the machine learning algorithm based on image recognition is:

[0019] Step 301: pre-processing the image data collected by the drone through data cleaning and image enhancement, wherein the data cleaning removes noise in the image through Gaussian filtering, and the image enhancement enhances the details of the image through histogram equalization and contrast adjustment;

[0020] Step 302: extract features from the image through a convolutional neural network; the convolutional neural network performs a convolution operation on the image through a convolution layer to extract local features, performs a maximum pooling operation on the feature map through a pooling layer to reduce computational complexity; and performs a nonlinear transformation on the output of the convolutional layer through an activation function;

[0021] Step 303: locate and identify the symptoms of pests and diseases in the extracted image features through a target detection algorithm; the working method of the target detection algorithm is:

[0022] r1. Generate candidate regions of pest and disease symptoms through sliding windows and region proposal networks.

[0023] r2. Extract features of the candidate area through a convolutional neural network and generate a feature vector;

[0024] r3. Classify the feature vector and perform bounding box regression through the fully connected layer and cross entropy loss function to determine the location and category of the pest and disease symptoms;

[0025] r4. Ensure the accuracy and reliability of the machine learning algorithm through data labeling and model training; the data labeling is to label the pest and disease symptoms in the image through manual labeling or semi-supervised learning to generate a training data set; the model training is to train the model through the back propagation algorithm and the optimization algorithm SGD;

[0026] r5. The identified pest and disease symptoms are transmitted to the central server in real time through an event-driven mechanism.

[0027] As a further technical solution of the present invention, the working method of the event-driven mechanism is: reading pest and disease symptom data from the message queue through an event processor, generating response measure events, and pushing the response measure events to the message queue, triggering emergency response measures through a rule engine and a decision tree algorithm, including increasing the monitoring frequency, adjusting the irrigation strategy, and initiating pest and disease control measures.

[0028] As a further technical solution of the present invention, the growth prediction model includes an input layer, an embedding layer, a long short-term memory layer, an attention mechanism layer, a fully connected layer and an output layer; the operation process steps of the growth prediction model are:

[0029] Step 501: receiving and processing input data through the input layer, including historical environmental data, plant growth data, and pest and disease symptoms;

[0030] Step 502: The input data is converted into a high-dimensional vector representation by the embedding layer to capture data features; the embedding layer converts plant species and pest types into continuous vector representations by a word embedding method;

[0031] Step 503: receiving the high-dimensional vector output by the embedding layer through the long short-term memory layer, capturing the long-term dependency in the time series data, and extracting the time series features; the long short-term memory layer dynamically adjusts the retention and discarding of information through a gating mechanism;

[0032] Step 504: The attention mechanism layer is used to enhance the model's attention to important time steps; the attention mechanism uses the attention mechanism to calculate the weight of each time step, and the weighted sum is used as the supplementary information output by the long short-term memory layer to enhance the model's sensitivity to key time steps;

[0033] Step 505: The fully connected layer performs a nonlinear transformation on the extracted features to generate a prediction result; the fully connected layer performs a nonlinear transformation on the outputs of the long short-term memory layer and the attention mechanism layer through a ReLU activation function to map the features extracted by the long short-term memory layer and the attention mechanism layer to the output space;

[0034] Step 506: Output the final growth prediction result through the output layer, including the plant growth status and pest and disease prediction under future environmental changes; the output layer generates a binary classification prediction result through the Sigmoid activation function and generates a multi-classification prediction result through the Softmax activation function.

[0035] As a further technical solution of the present invention, the method for generating an irrigation plan, a fertilization strategy, pest and disease control measures and an optimization strategy for the layout of green areas based on the prediction results in S5 is as follows: based on the vegetation growth requirements and soil moisture data, an irrigation plan is generated through an expert system, including the time, frequency and amount of irrigation; based on the vegetation growth status and soil nutrient data, a fertilization plan is generated through a nutritional model, including the type, amount and frequency of fertilizer; based on the pest and disease warning information and the vegetation growth status, control measures are formulated through an expert system, including spraying pesticides and pruning diseased branches; based on the human activity pattern and the vegetation growth status, the layout of the green area and the type of vegetation are generated through a spatial planning method.

[0036] As a further technical solution of the present invention, the dynamic adjustment method of the deep Q network algorithm is:

[0037] Step 701, defining optimization objectives, including plant growth conditions, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects, and setting initial weight values ​​for each optimization objective;

[0038] Step 702: construct a deep neural network through a convolutional neural network and a fully connected layer; including an online network and a target network, wherein the online network is used to update and calculate the Q value of the current strategy in real time; the weight of the target network is regularly copied from the online network to calculate the target Q value to stabilize the training process; the convolutional neural network extracts the features of the input data through the convolutional layer and the pooling layer to generate a feature vector; the fully connected layer maps the feature vector to the Q value, i.e., the expected reward value of each action, through a multi-layer perceptron;

[0039] Step 703: collect environmental data in real time through the sensor network and the drone inspection system, including temperature, humidity, light intensity, and soil moisture, and pre-process the collected environmental data through normalization and standardization processing methods to construct an initialized environmental state space;

[0040] Step 704: Combine the collected environmental data and the state information of the optimization target into a state vector as the input of the deep Q network; the state vector includes the current environmental data and the state information of the optimization target;

[0041] Step 705: Select an action through the adaptive decay mechanism of the greedy strategy, wherein the greedy strategy randomly selects an action or selects the action with the highest Q value in the current state according to the exploration rate; the exploration rate is gradually reduced during the training process to balance exploration and utilization;

[0042] Step 706: Execute the selected action, and collect environmental data and status information of the optimization target after the action is executed through the sensor network and the drone inspection system. If the plant growth condition improves, a positive reward is given, and if the pest and disease symptoms worsen, a negative reward is given;

[0043] Step 707: Update the Q value through the Bellman equation, wherein the Bellman equation calculates the new Q value through the discount factor γ and the immediate reward r, calculates the difference between the current Q value and the target Q value through the mean square error loss function MSE, and updates the network parameters through the back propagation mechanism;

[0044] Step 708: Periodically copy the weights of the online network to the target network through weight repetition to maintain the stability of the target value and reduce fluctuations during the training process;

[0045] Step 709, continuously execute steps 704 to 708, interact with the environment, collect data, update network weights, and stop when the predetermined number of training rounds or convergence condition is met.

[0046] As a further technical solution of the present invention, the context-based multi-objective optimization model includes a data reading layer, a feature extraction layer, a context perception layer, a goal setting layer, an optimization modeling layer and an algorithm solving layer; the working method based on context perception and multi-objective optimization model is:

[0047] Step 801, reading pre-processed data from the central database through the data layer; including real-time environmental data, pest and disease data and growth prediction data;

[0048] Step 802: extract key features of the data through the feature extraction layer; the feature extraction layer extracts features through a principal component analysis method, and stores the extracted feature data in a feature database;

[0049] Step 803: The context perception layer reads the processed feature data from the feature database and identifies the current context state, including drought, rainstorm, high density of people, high incidence of pests and diseases, and vigorous growth of plants; the context feature perception layer divides the data into different categories through k cluster analysis to identify the current context state; the k cluster analysis selects the best clustering result through multiple initialization method, and the formula expression is:

[0050]

[0051] In formula (1), c i represents the i-th cluster center; X represents the set of all data points; ||Xc j || represents the data point X and the selected cluster center c j The Euclidean distance between j represents the variance of the jth cluster center, which is used to adjust the influence of distance;

[0052] Step 804: setting optimization goals through the goal setting layer, including plant growth conditions, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects; the goal setting layer sets an initial weight value for each optimization goal through an expert system combined with historical data, and the goal setting layer dynamically adjusts the weight value according to actual environmental changes and management needs through a feedback mechanism based on reinforcement learning, and continuously optimizes the weight setting;

[0053] Step 805: construct a mixed integer linear programming model including the objectives of maximizing plant growth efficiency, minimizing the incidence of pests and diseases, and optimizing resource use costs through the optimization modeling layer. The mixed integer linear programming model dynamically adjusts the weights between different objectives based on real-time situations through a dynamic adjustment function. The formula expression of the dynamic adjustment function is:

[0054]

[0055] In formula (2), f(x) represents the comprehensive objective function; w i (t) represents the weight of the i-th target at time t; g i (x) represents the value of the i-th objective function; γ i represents the penalty coefficient of the i-th objective, which is used to adjust the nonlinear effect of the objective function; θ i represents the threshold of the i-th target, which is used to define the benchmark value of the objective function; β i represents the smoothing parameter of the i-th objective, which is used to control the smoothness of the objective function;

[0056] Step 806: Generate the optimal management strategy through the algorithm solution layer, including irrigation duration, fertilizer type, fertilizer quantity, fertilizer frequency, and plant location and type; the optimization solution layer solves the optimization model through a genetic algorithm, and the genetic algorithm gradually optimizes the individuals in the population through selection, crossover and mutation operations, and handles conflicts between targets through non-dominated sorting and crowding distance calculation; the algorithm solution layer also continuously monitors the optimization effect through a feedback mechanism based on online learning, and dynamically adjusts the optimization strategy according to actual conditions.

[0057] As a further technical solution of the present invention, the adaptive control algorithm realizes dynamic adjustment of the irrigation system and fertilization equipment through PID control algorithm and fuzzy control, and ensures that the system maintains optimal performance in a dynamic environment through an online learning algorithm; wherein, the PID control algorithm dynamically adjusts control parameters through proportional control, integral control and differential control; the fuzzy control converts complex nonlinear relationships into fuzzy rules through a fuzzy inference system, and makes decisions based on the fuzzy rule base; the online learning algorithm updates the control parameters in real time through a recursive least squares method.

[0058] As a further technical solution of the present invention, an urban greening system based on big data includes a multi-source data acquisition module for collecting multimodal data of urban greening areas, including temperature and humidity, light intensity, soil moisture, air quality, plant species, growth cycle, vegetation growth status and human density data;

[0059] A data integration module, used for cleaning, aligning, converting and fusing the multimodal data collected by the multi-source data collection module through a deep heterogeneous fusion algorithm;

[0060] The real-time data analysis module is used to analyze the integrated data in real time, including vegetation growth trend prediction, pest and disease warning, and human activity pattern recognition, and output real-time analysis results, including irrigation demand prediction, fertilization recommendations, pest and disease warning information, and human activity hotspots;

[0061] Adaptive decision-making module, which is used to build a situation-based multi-objective optimization model based on real-time analysis results through mixed integer linear programming, situation awareness and genetic algorithm, dynamically adjust irrigation plans, fertilization strategies, pest control measures and greening area layout, and realize data-driven intelligent management;

[0062] User interaction feedback module, used to provide user interaction functions, including user query, feedback, activity participation, and display real-time data and results of greening management;

[0063] The output end of the multi-source data acquisition module is connected to the input end of the data integration module; the output end of the data integration module is connected to the input end of the real-time data analysis module; the output end of the real-time data analysis module is connected to the input end of the adaptive decision module; the user interaction feedback module is connected to the above modules.

[0064] The positive beneficial effects of the present invention are:

[0065] 1. Collect data from multiple sources through API interfaces and IoT wireless communication technology, not only limited to environmental sensors, but also including plant growth cycle databases, public data sources, drone inspection systems, etc. These data are centrally stored in a temporary buffer. Subsequently, all data are imported into a distributed real-time data stream processing platform, which uses a deep heterogeneous fusion algorithm to clean, align, transform and fuse the data to ensure data consistency and availability. This process solves the problem of inconsistent data formats, units, and sampling frequencies caused by diverse data sources, laying a solid foundation for subsequent advanced analysis.

[0066] 2. Based on the image data of green areas from the drone inspection system, combined with the machine learning algorithm of image recognition, it is possible to efficiently identify the symptoms of pests and diseases. In addition, when the system detects abnormal environmental data, it can automatically trigger emergency response measures, such as increasing the monitoring frequency or adjusting the irrigation strategy in real time. This mechanism significantly enhances the system's response speed and accuracy to environmental changes. At the same time, by training historical data through the long short-term memory network (LSTM), an accurate growth prediction model is constructed, enabling the system to make reasonable adjustments in advance when faced with unforeseen situations such as extreme weather, further improving the system's dynamic adaptability.

[0067] 3. In order to overcome the limitations of the existing system in multi-objective optimization, this solution adopts a deep Q-network algorithm to dynamically adjust the weight value of the optimization target according to environmental changes, ensuring the balance between different targets. Specifically, the system defines the optimization target and initial weight value based on multiple dimensions such as plant growth conditions, pest and disease symptoms, and water resource utilization efficiency. In addition, a situation-based multi-objective optimization model is constructed through mixed integer linear programming, situational awareness, and genetic algorithms. The model can comprehensively consider factors such as real-time environmental data, pest and disease data, and growth prediction data to generate the most optimized comprehensive management strategy. In the end, these strategies were applied in actual operations, such as adjusting the duration of irrigation, selecting appropriate types and quantities of fertilizers, planning the location and types of plants, etc., thereby achieving effective allocation of resources and improving overall management efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1A step diagram of an urban greening optimization method based on big data according to the present invention;

[0069] Figure 2 It is a structural framework step diagram of the growth prediction model of the present invention;

[0070] Figure 3 This is a schematic diagram of the urban greening system based on big data of the present invention;

[0071] Figure 4 This is a step diagram of the dynamic adjustment method of the deep Q network algorithm of the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] In the present invention, if Figure 1-Figure 4 As shown: A method for optimizing urban greening based on big data includes the following steps:

[0074] S1. Obtain temperature and humidity, light intensity, soil moisture and air quality data from environmental sensors through API interfaces and IoT wireless communication methods, obtain plant species and growth cycle data from plant growth cycle databases, obtain human density and activity area data from public data sources, and regularly obtain image data of green areas from drone inspection systems, and store them in temporary buffers; S2. Input the data into a distributed real-time data stream processing platform to clean, align, convert and fuse the collected data; the distributed real-time data stream processing platform integrates data sets in a unified format through a deep heterogeneous fusion algorithm, including environmental parameters, vegetation growth status, and human activity patterns in a unified format; S3. Based on the image data of green areas of the drone inspection system, identify pest and disease symptoms through an image recognition-based machine learning algorithm, and automatically trigger corresponding emergency response measures when abnormal environmental data is detected through an event-driven mechanism, including increasing monitoring frequency and adjusting irrigation strategies; S4. Train historical environmental data, plant growth data, and pest and disease symptoms through a long short-term memory network to build a growth prediction Model; S5, input the real-time collected environmental data into the growth prediction model, obtain the growth prediction results of future environmental changes, and generate optimization strategies including irrigation plans, fertilization strategies, pest and disease control measures and greening area layout based on the prediction results; S6, define optimization goals and initial weight values, including plant growth status, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects, and dynamically adjust the weight value of each goal according to environmental changes through a deep Q network algorithm; S7, construct a situation-based multi-objective optimization model through mixed integer linear programming, situational awareness and genetic algorithms, input real-time environmental data, pest and disease data, growth prediction data and optimization strategies into the situation-based multi-objective optimization model, obtain the optimal comprehensive management strategy, including irrigation time, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type, and store the output results in a database; S8, send the output results of the situation-based multi-objective optimization model to the control system, and the control system dynamically adjusts the operating parameters of the irrigation system and fertilization equipment and facilities through an adaptive control algorithm according to the optimization strategy.

[0075] In the specific implementation, S2 fills missing values ​​through mean interpolation, median interpolation and K nearest neighbor interpolation. Mean interpolation is applicable to numerical data, and missing values ​​are filled by calculating the mean of the same feature; median interpolation is applicable to numerical data, and missing values ​​are filled by calculating the median of the same feature; K nearest neighbor interpolation is applicable to numerical and categorical data, and missing values ​​are filled by finding the values ​​of similar samples. Then the Z-Score method is used to identify and process outliers. The Z-Score method calculates the standardized score of each data point, and regards the data points whose standardized scores exceed a certain threshold (such as 3σ) as outliers, and corrects or deletes them. Then linear interpolation and spline interpolation are used to align timestamps. Linear interpolation is applicable to data with a small time interval, and the missing values ​​within the time interval are filled by the linear interpolation method; spline interpolation is applicable to data with a large time interval, and the time series curve is fitted by the spline interpolation method to fill the missing values. Spatial coordinate alignment is performed through geocoding and coordinate transformation. Geocoding converts address information into longitude and latitude coordinates, and coordinate transformation unifies data in different coordinate systems into the same coordinate system. Based on the aligned data, the features of image and time series data are extracted through convolutional neural network (CNN) and recurrent neural network (RNN), and the mutual information method, principal component analysis (PCA) and feature engineering are used to generate feature vectors through dynamic weight adjustment mechanism. The fused feature vectors are used as the input of the deep neural network, and the nonlinear mapping relationship between input data and output data is learned through multi-layer perceptron (MLP) and autoencoder (AE) to generate a data set in a unified format. The nonlinear mapping relationship between input data and output data is learned through forward propagation and back propagation algorithms, and nonlinear characteristics are introduced through the activation function ReLU to improve the expression ability of the model. The low-dimensional representation of the data is learned through the encoder and decoder, and the model parameters are optimized through the mean square error (MSE).

[0076] In specific implementation, the comprehensive data set generated by the deep heterogeneous fusion algorithm covers a wide range of information such as environmental parameters, vegetation growth status, and human activity patterns, providing comprehensive decision-making support for urban greening management. Based on these data, managers can more accurately evaluate greening effects, optimize resource allocation, and improve urban greening levels. S2's real-time processing and high reliability ensure the timeliness and accuracy of data, enabling urban greening management to quickly respond to environmental changes and emergencies. This not only improves management efficiency, but also reduces management costs.

[0077] In specific implementation, the hardware operating environment of the deep heterogeneous fusion algorithm includes: high-performance computing servers equipped with multi-core CPUs and high-performance GPUs. Large-capacity, high-read and high-write speed storage devices. Distributed real-time data stream processing platform for real-time data collection, cleaning, alignment, conversion and fusion. Stable network communication equipment and power cooling system.

[0078] Based on the above hardware equipment, a comparative experiment was designed to verify the advantages of the deep heterogeneous fusion algorithm in processing big data and performing data fusion. The experiment included the deep heterogeneous fusion algorithm (Group A) and the rule-based data fusion algorithm (Group B). The rule-based data fusion algorithm relies on predefined rules and logic to integrate data from different sources. The same data set was selected and experiments were conducted on Group A and Group B respectively. Each group of experiments was conducted five times to ensure the stability and reliability of the results. The running time, data fusion accuracy, data fusion integrity, CPU occupancy, memory occupancy, and disk I / O parameters of each experiment were recorded. The experimental data record table is shown in Table 1:

[0079] Table 1 Experimental data record table

[0080]

[0081] The results of the comparative experiment show that the deep heterogeneous fusion algorithm (group A) is superior to the traditional algorithm (group B) in terms of running time, data fusion accuracy and data fusion integrity. The average running time of group A is only about 120 seconds, while group B takes about 240 seconds, indicating that the deep heterogeneous fusion algorithm has higher efficiency in processing big data. In terms of data fusion accuracy, the average accuracy of group A reached 98.6%, which is much higher than the 90.0% of group B, indicating that the deep heterogeneous fusion algorithm has higher accuracy in data fusion. At the same time, the data fusion integrity of group A is also better than that of group B, indicating that the deep heterogeneous fusion algorithm can more completely retain the characteristics and information of the original data. In addition, group A also shows better performance in terms of resource occupation, with lower CPU occupation, memory occupation and disk I / O than group B. In summary, the deep heterogeneous fusion algorithm has significant advantages in processing big data and performing data fusion, and provides strong technical support for the intelligent management of urban greening systems.

[0082] Specifically, the working principle of the deep heterogeneous fusion algorithm is as follows: first, fill in missing values, including mean interpolation: when the data distribution is relatively uniform, use the mean to fill in missing values. Median interpolation: for skewed distributed data, the median can better represent the central trend of the data. K nearest neighbor interpolation: fill in missing values ​​based on the values ​​of the K nearest neighbors, and use the Euclidean distance to calculate the proximity; after standardizing the data, use the Z-Score method to identify and process outliers that exceed the threshold. The expression is Where u is the original data, k is the mean, a kis the standard deviation. The z value indicates the degree to which the data point deviates from the mean. If the absolute value of the z value exceeds the set threshold, the data point is treated as an outlier and processed. Next, for data with uniform time intervals, linear interpolation is used to align timestamps. For data with uneven time intervals, spline interpolation is used to align timestamps. Spline interpolation is a more flexible and accurate interpolation method that approximates the original data by constructing a piecewise polynomial function. For data with spatial attributes, we need to convert the address into longitude and latitude coordinates through geocoding, and convert data in different coordinate systems into a unified coordinate system through coordinate conversion.

[0083] In the feature extraction stage, the convolution operation formula of the convolutional neural network is expressed as r = h + f (Wx + b), where x is the input image, W is the convolution kernel, b is the bias term, f is the convolution operation, and h is the activation function used to introduce nonlinear characteristics. The convolution kernel slides on the input image and calculates the weighted sum to extract features. Next, the time dependency in the time series data is captured through the loop structure. The time series data contains dynamic information in time. The formula is expressed as h t =1+ε(W h h t-1 +W x x t +b), h c is the hidden state at time step t, W h and W x is the weight matrix, x t is the input data at time step t, b is the bias term, and ε is the activation function; the temporal dependency in the time series data is captured by transferring the hidden state between time steps.

[0084] After the feature extraction stage, these feature vectors are fused into a feature vector of a unified format, in which the weights are dynamically adjusted according to the importance of the feature vectors through a dynamic weight adjustment mechanism. The dynamic weight adjustment mechanism fuses the feature vectors from different sources into a feature vector of a unified format by weighted summation, that is, by using ρ, τ and These three dynamic weights dynamically adjust the feature vectors from CNN, RNN and other sources; in order to learn the nonlinear mapping relationship between input data and output data, the deep learning model is trained through multi-layer perceptron (MLP) and autoencoder. Multi-layer perceptron (MLP) learns the nonlinear mapping relationship between input data and output data through a multi-layer neural network structure. First, the weighted sum is calculated and the nonlinear characteristics are introduced through the activation function to obtain the output, followed by the encoding and decoding process of the autoencoder, and the low-dimensional representation of the data is learned by minimizing the reconstruction error.

[0085] In the specific implementation of S3, this step first uses drones to regularly inspect green areas according to the predetermined flight route and schedule to capture high-definition images. The drones are equipped with high-resolution cameras and GPS positioning systems to ensure the clarity of the images and the accuracy of the geographic location information. The images captured by the drones are filtered to remove noise and reduce environmental interference. The Gaussian filter smoothes the image through the convolution kernel to retain the main features of the image. Then the details of the image are enhanced through histogram equalization and contrast adjustment. Histogram equalization makes the brightness and contrast of the image more uniform by redistributing pixel values. Contrast adjustment makes the details in the image more prominent by adjusting the range of pixel values. The preprocessed images are extracted through convolutional neural networks. The convolution layer uses multiple filters to perform convolution operations on the image to generate feature maps. The pooling layer performs maximum pooling operations on the feature maps to reduce computational complexity. The activation function (such as ReLU) performs nonlinear transformations on the output of the convolution layer to increase the expression ability of the model. Next, candidate regions for pest and disease symptoms are generated. The sliding window method slides at different positions and scales of the image to generate candidate regions. RPN generates more accurate candidate regions through convolutional neural networks to improve detection efficiency. Feature extraction is then performed on the candidate regions to generate feature vectors. This process usually uses deep feature maps of convolutional neural networks. Feature vectors are classified and bounding box regressed through fully connected layers and cross entropy loss functions to determine the location and category of pest and disease symptoms. The fully connected layer maps feature vectors to category labels and bounding box coordinates. The cross entropy loss function is used to optimize the classification task, and the smooth L1 loss function is used to optimize the bounding box regression task. Pest and disease symptoms in images are annotated through manual annotation or semi-supervised learning to generate training data sets. Manual annotation ensures the accuracy of the data, and semi-supervised learning improves the generalization ability of the model through a small amount of annotated data and a large amount of unlabeled data. The model is trained through backpropagation algorithms and optimization algorithms (such as SGD) to optimize model parameters and improve the accuracy and reliability of the model. Finally, the identified pest and disease symptoms are transmitted to the central server in real time through an event-driven mechanism. The central server automatically triggers corresponding emergency response measures based on the detection results, such as increasing the monitoring frequency and adjusting the irrigation strategy.

[0086] Among them, the event-driven mechanism reads the pest symptom data from the message queue through the event processor, generates response measure events, and pushes the response measure events to the message queue. The emergency response measures are triggered by the rule engine and decision tree algorithm. The operation process of the event processor is as follows:

[0087] Initialize the event processor: Set the connection parameters of the message queue and initialize the event processor.

[0088] Listening to the message queue: The event processor continuously listens to the message queue, waiting to receive pest and disease symptom data.

[0089] Read data: When there is new pest and disease symptom data in the message queue, the event processor reads the data.

[0090] Generate response measure events: Generate corresponding response measure events based on the read pest and disease symptom data.

[0091] Push events: Push the generated response action events to the message queue for subsequent processing.

[0092] In the specific implementation, the drone inspection system and image recognition technology can quickly and accurately detect the symptoms of pests and diseases in green areas and discover potential problems in time. This helps to take measures in advance to prevent the spread of pests and diseases and protect the results of greening. In addition, the highly automated detection process reduces the dependence on manual labor and reduces labor costs. The drone inspection system can work automatically at the scheduled time and route, reducing the burden on staff. Secondly, through convolutional neural networks and target detection algorithms, the location and category of pest and disease symptoms can be accurately located, providing a scientific basis for urban greening management. Managers can formulate more accurate management strategies based on the test results to improve the level of greening management. At the same time, timely detection and treatment of pest and disease problems can help maintain ecological balance and promote the sustainable development of the city. By adjusting irrigation strategies and increasing monitoring frequency, water waste can be effectively reduced and resource utilization efficiency can be improved. Finally, through the event-driven mechanism, when abnormal environmental data is detected, the corresponding emergency response measures can be automatically triggered. This helps to quickly respond to emergencies, reduce losses, and ensure the safety and stability of urban greening.

[0093] In the specific implementation, the hardware working environment of the machine learning algorithm based on image recognition in actual work includes: UAV inspection system: model: DJI Mavic 2Enterprise; camera: 4K high-definition camera, 12 million pixels; battery life: 31 minutes; flight control: GPS positioning, automatic flight path planning; central server: processor: Intel XeonGold 6248R (2.4GHz, 24Cores); memory: 128GB DDR4 RAM; storage: 2TB NVMe SSD; operating system: Ubuntu 20.04LTS; GPU: NVIDIA Tesla V100 (32GB); communication equipment: wireless communication module: 4G / 5G wireless communication module; data transmission protocol: MQTT protocol; image processing software: development environment: Python 3.8; deep learning framework: TensorFlow 2.4; image processing library: OpenCV 4.5;

[0094] Based on the above hardware, an image dataset containing symptoms of pests and diseases was prepared. The dataset contains 1,000 images, of which 500 are normal images and 500 are images containing symptoms of pests and diseases. The dataset is divided into a training set (70%) and a test set (30%). By comparing the performance of the machine learning algorithm based on image recognition (Group A) and the traditional algorithm (Group B) in pest and disease detection, the positive and beneficial effects of the machine learning algorithm based on image recognition are verified. Among them, Group A: Machine learning algorithm based on image recognition, using convolutional neural network (CNN) for feature extraction and target detection. Group B: Traditional algorithm, using rule-based methods (such as color segmentation, edge detection, etc.) for pest and disease detection. The experimental record table is shown in Table 2:

[0095] Table 2 Experimental record form

[0096] Number of experiments Group Detection time (seconds) Detection accuracy (%) False alarm rate (%) 1 Group A 12.3 95.6 2.1 1 Group B 18.5 82.4 12.3 2 Group A 12.1 95.8 2.0 2 Group B 18.7 82.2 12.5 3 Group A 12.4 95.5 2.2 3 Group B 18.6 82.3 12.4 4 Group A 12.2 95.7 2.1 4 Group B 18.8 82.1 12.6 5 Group A 12.3 95.6 2.1 5 Group B 18.5 82.4 12.3

[0097] As shown in Table 2, the machine learning algorithm based on image recognition has shown higher detection accuracy, lower false alarm rate and faster processing speed in pest and disease detection, which has significant positive and beneficial effects. This provides more efficient and accurate technical support for urban greening management.

[0098] In the specific implementation of S4 and S5, the construction of the growth prediction model first collects historical environmental data (such as temperature, humidity, light), plant growth data (such as growth rate, health status) and pest and disease symptom data through various sensors and devices. These data are cleaned, aligned and standardized to ensure the quality and consistency of the data. Data preprocessing is the basis of model training, which improves the availability of data by removing noise and filling missing values. The model includes an input layer, an embedding layer, a long short-term memory layer, an attention mechanism layer, a fully connected layer and an output layer: the input layer receives the preprocessed data, which includes time series environmental parameters, plant growth status and pest and disease symptoms. The input layer converts the raw data into a format that the model can process. The data is usually input in the form of a two-dimensional array, where each row represents a time step and each column represents a feature. The input layer is also responsible for passing the data to the next layer for further processing.

[0099] The embedding layer converts the input data into a fixed-length vector representation for subsequent processing. The embedding layer captures the intrinsic characteristics of the data by learning a low-dimensional representation of the input data. The embedding layer captures the intrinsic characteristics of the data by learning a low-dimensional representation of the input data. The embedding layer is usually used to process discrete features (such as categorical variables) and convert them into a continuous vector form. In this model, the embedding layer learns the best vector representation through the training process, so that similar input data are close in the vector space;

[0100] Long Short-Term Memory Layer (LSTM) LSTM is a special recurrent neural network (RNN) that can capture long-term dependencies in time series data and extract temporal features that have an important impact on plant growth. Through multiple LSTM units, the model can effectively extract temporal features and understand the impact of environmental changes on plant growth. LSTM is a special recurrent neural network (RNN) that can effectively handle long-term dependency problems in time series data. LSTM controls the flow of information by introducing gating mechanisms (input gate, forget gate, and output gate) to avoid gradient vanishing and gradient exploding problems. At each time step, the LSTM unit receives the input data of the current time step and the hidden state of the previous time step. The LSTM unit includes: Forget gate: decides which information to discard. The forget gate calculates a value between 0 and 1 through a sigmoid activation function, indicating the degree of retention of each cell state. Input gate: determines which new information will be stored in the cell state. The input gate calculates new candidate cell states through a sigmoid activation function and a tanh activation function. Cell state update: combines the outputs of the forget gate and the input gate to update the cell state. Output gate: decides which information to output. The output gate calculates a value between 0 and 1 through a sigmoid activation function, indicating the output degree of each cell state. Finally, a tanh activation function is used to convert the cell state into an output value.

[0101] The attention mechanism layer improves the model's explanatory power and accuracy by dynamically adjusting the importance of different time steps. The attention mechanism layer assigns different weights to the features of different time steps according to the input of the current time step, thereby focusing on key information. The attention mechanism dynamically adjusts the importance of different time steps to improve the model's explanatory power and accuracy. The attention mechanism enables the model to focus on key information by calculating the weight of each time step. Specifically, it includes: Calculating the attention score: Calculating the attention score of each time step through a feedforward neural network (usually a fully connected layer). The attention score reflects the importance of the time step to the current prediction. Normalization: The attention score is normalized by the softmax function so that its sum is 1 to form the attention weight. Weighted summation: Multiply the feature vector of each time step by its corresponding attention weight, and then perform weighted summation to obtain the final feature representation.

[0102] The fully connected layer performs nonlinear transformation on the extracted features to generate the final prediction results. The fully connected layer learns complex nonlinear relationships through multiple layers of neurons to improve the expressiveness of the model. The fully connected layer learns complex nonlinear relationships through multiple layers of neurons to convert the features extracted by the LSTM layer and the attention mechanism layer into the final prediction results. The fully connected layer usually contains multiple hidden layers, each of which performs nonlinear transformation through an activation function (such as ReLU) to increase the expressiveness of the model. The last layer is usually a linear layer that outputs the prediction results.

[0103] The output layer outputs the predicted results of plant growth under future environmental changes, including growth rate, health status, and pest and disease risk. The output layer converts the output of the fully connected layer into specific prediction results. The output layer usually uses an appropriate activation function (such as sigmoid or softmax) for the final nonlinear transformation to generate the prediction results. For example, for the prediction of growth rate, a linear activation function can be used; for the prediction of pest and disease risk, a sigmoid activation function can be used to generate a probability value.

[0104] The steps for S5 to generate an optimized management strategy are as follows: Real-time data collection: Real-time environmental data (such as temperature, humidity, and light) and plant growth data (such as growth rate and health status) are collected through sensors. These real-time data are transmitted to the central server through the Internet of Things technology to ensure the real-time and accuracy of the data.

[0105] Data input: The real-time collected data is input into the growth prediction model. The model generates growth prediction results for future environmental changes based on historical data and real-time data.

[0106] Irrigation plan: Based on the vegetation growth requirements and soil moisture data, an expert system generates an irrigation plan. The expert system determines the time, frequency and amount of irrigation based on the water requirements of the plants and soil moisture. The irrigation plan ensures the rational use of water resources and reduces waste through optimization algorithms.

[0107] Fertilization strategy: Based on the vegetation growth status and soil nutrient data, a fertilization plan is generated through a nutrition model. The nutrition model determines the type, amount and frequency of fertilizer application based on the nutritional needs of the plant and the soil nutrient status. The fertilization strategy ensures the rational application of fertilizers, improves fertilizer utilization and reduces environmental pollution through optimization algorithms. The nutrition model first substantially completes the integration and cleaning of individual or group nutritional data through data collection and preprocessing technology. This step uses a database management system and a data cleaning algorithm to ensure the accuracy and consistency of the input data, laying the foundation for subsequent analysis. Then, the model uses a personalized nutrition assessment algorithm to substantially analyze the nutritional needs of different individuals. Based on factors such as the individual's age, gender, weight, height, body fat percentage, health status, and eating habits, the algorithm calculates nutritional recommendations and diet plans that meet individual needs through regression analysis and machine learning models (such as decision trees, random forests, etc.). In addition, the model also realizes dynamic adjustment of nutrition management through real-time monitoring and feedback mechanisms (such as smart devices monitoring dietary intake and physical condition) to ensure the timeliness and effectiveness of the plan. This mechanism relies on Internet of Things technology and mobile application development technology, enabling users to easily obtain personalized nutritional advice and continuously optimize nutritional plans based on feedback.

[0108] Pest and disease control measures: According to the early warning information of pests and diseases and the growth status of vegetation, control measures are formulated through the expert system. According to the severity of pests and diseases and the health status of plants, the expert system determines measures such as spraying pesticides and pruning diseased branches. Pest and disease control measures ensure the effectiveness and timeliness of control measures through optimization algorithms, and reduce the impact of pests and diseases on plants. The expert system first substantially completes the extraction and structured storage of domain expert knowledge through knowledge acquisition and representation technology. This step uses data mining technology and natural language processing technology to extract valuable knowledge from a large number of documents, cases and expert interviews, and represents it in the form of ontology and rule base. Then, the system substantially realizes the application of knowledge and problem solving through the reasoning engine. Based on forward reasoning and rule matching algorithms, the reasoning engine retrieves relevant facts and rules from the knowledge base according to the questions or requests entered by the user, and gradually derives new conclusions or suggestions. In addition, the expert system also substantially improves the availability and accuracy of the system through user interaction and feedback mechanisms. This mechanism uses user interface design technology and machine learning algorithms to enable users to easily input information and view results, while evaluating and adjusting the inference results based on user feedback to continuously optimize the performance of the system.

[0109] Green area layout: According to the human flow activity pattern and vegetation growth status, the spatial planning method is used to generate the layout of the green area and the types of vegetation. The spatial planning method takes into account the density of human flow, the activity area and the growth habits of plants, optimizes the layout of the green area, improves the greening effect and enhances the beauty of the city. The spatial planning method first substantially completes the collection and processing of basic data for spatial planning through data collection and integration technology. This step uses geographic information system (GIS) and remote sensing technology to obtain multi-source data such as topography, natural resources, and social economy, and realizes unified management and efficient access to data through data fusion algorithms and spatial database technology. Next, the planning method substantially evaluates the potential for spatial development and the carrying capacity of resources and environment through spatial analysis and evaluation technology. This step uses spatial overlay analysis, multi-criteria decision analysis (MCDA) and ecosystem service assessment models to comprehensively consider multiple factors such as resources and environment, economy and society, policies and regulations, and provide a scientific basis for spatial planning. In addition, the planning method also substantially formulates a planning scheme that meets regional development goals and constraints through planning scheme design and optimization technology. This step uses intelligent optimization algorithms (such as genetic algorithms, particle swarm algorithms) and planning simulation technology to compare and optimize multiple planning schemes to ensure the feasibility and sustainability of the schemes. Finally, through planning implementation and monitoring and evaluation technologies (such as planning implementation effect evaluation, environmental impact assessment, etc.), the planning schemes are implemented and dynamically adjusted to ensure the realization of planning goals.

[0110] In the specific implementation, through LSTM and attention mechanism, the model can capture the long-term dependencies in time series data and improve the accuracy of growth prediction. This helps to discover potential problems in advance, take preventive measures and reduce losses. Compared with traditional time series prediction methods, the growth prediction model based on deep learning has higher accuracy and robustness, and can better cope with complex and changing environmental conditions. In addition, the irrigation plan generated by the expert system can reasonably arrange the irrigation time and frequency according to the water demand of the plant and the soil moisture, reduce water resource waste and improve irrigation efficiency. The fertilization plan generated by the nutrition model can reasonably arrange the type, amount and frequency of fertilization according to the nutritional needs of the plant and the nutrient status of the soil, improve fertilizer utilization and reduce environmental pollution. The pest and disease control measures formulated by the expert system can take measures such as spraying pesticides and pruning diseased branches in time according to the severity of pests and diseases and the health status of the plants, reduce the impact of pests and diseases on plants, and protect the greening results. The greening area layout and vegetation types generated by the spatial planning method can optimize the layout of the greening area according to the activity pattern of human flow and the growth habits of plants, improve the greening effect and enhance the beauty of the city. Secondly, the system can respond to environmental changes and plant growth status in real time, generate optimized management strategies, and ensure the timeliness and effectiveness of management. Through expert systems and nutritional models, the system can generate intelligent management strategies based on plant growth needs and environmental conditions, reduce human intervention, and improve management efficiency. In addition, the system can coordinate multiple management links such as irrigation, fertilization, pest control, and greening area layout to achieve comprehensive optimization management and improve management efficiency and effectiveness.

[0111] In S6, the deep Q network defines multiple optimization objectives, including plant growth status, pest and disease symptoms, water resource utilization efficiency, and environmental protection effect. These objectives reflect the core needs of urban greening management. An initial weight value is assigned to each optimization objective, which reflects the importance of each objective in the overall optimization. The initial weight value can be determined by expert experience or historical data statistics. The DQN algorithm is a method that combines deep learning and reinforcement learning to solve complex decision-making problems. Its core idea is to evaluate the expected return of taking a certain action in a certain state through the Q function. In the urban greening system, the Q function represents the expected effect of taking a certain management measure (such as irrigation, fertilization, pest and disease control, etc.) under the current environmental state. The state represents the comprehensive information of the current environment, including plant growth status, pest and disease symptoms, soil moisture, light intensity, etc., usually expressed in vector form. The action represents the management measures that can be taken, such as increasing the irrigation frequency, adjusting the fertilization strategy, starting pest and disease control measures, etc., also expressed in vector form. The reward function is used to evaluate the effect of taking an action and is designed according to the changes in the optimization objectives, such as improved plant growth, reduced pests and diseases, and improved water resource utilization efficiency. The experience replay mechanism improves learning efficiency by storing and reusing past experiences. At each time step, the current state, action, reward, and next state are stored in the experience replay buffer. The target network is used to stabilize the learning process, and the parameters of the target network are regularly copied from the current network to ensure the stability of the target Q value. The ε-greedy strategy balances exploration (trying new actions) and exploitation (selecting the current optimal action) to ensure that the algorithm can gradually converge to the optimal strategy.

[0112] In practical applications, the working method of step S6 is as follows: First, define multiple optimization objectives, such as plant growth status, pest and disease symptoms, water resource utilization efficiency, and environmental protection effect, and assign initial weight values ​​to each objective. These weight values ​​can be determined by expert experience or historical data statistics to ensure that the importance of each objective in the overall optimization is reasonably reflected. Next, environmental data (such as temperature, humidity, and light), plant growth data (such as growth rate, health status), and pest and disease symptom data are collected in real time through various sensors and devices. The collected data is integrated into a state vector to represent the comprehensive information of the current environment. According to the current state, the optimal action is selected through the DQN algorithm. In the initial stage, the ε-greedy strategy is used to balance exploration and utilization to ensure that the algorithm can gradually converge to the optimal strategy. According to the selected action, the corresponding management measures are implemented, such as increasing the frequency of irrigation, adjusting the fertilization strategy, and starting pest and disease control measures. The reward value is calculated based on the changes in the actions taken and the optimization objectives. The reward value reflects the effect of taking a certain action, such as improved plant growth, reduced pests and diseases, and improved water resource utilization efficiency. The current state, action, reward, and next state are stored in the experience replay buffer, and training is performed through random sampling to improve learning efficiency. The parameters of the DQN network are updated through the back-propagation algorithm and optimization algorithm (such as Adam), and the weight value is optimized so that the model can better adapt to environmental changes. The parameters of the current network are regularly copied to the target network to ensure the stability of the target Q value. Through continuous iterative optimization, the model can dynamically adjust the weight value of each optimization target to achieve multi-target optimization.

[0113] Traditional urban greening management systems usually rely on fixed rules and manual adjustments, lacking real-time and adaptability. For example, rule-based irrigation systems can only irrigate according to a preset schedule and cannot be adjusted according to real-time environmental changes. The DQN algorithm can collect environmental data in real time and dynamically adjust management measures according to the current environmental status to ensure the timeliness and effectiveness of management. Through continuous iterative optimization, the DQN algorithm can adaptively adjust the weight value of each optimization target to improve the flexibility and adaptability of management. In addition, traditional urban greening management systems usually only focus on a single target, such as plant growth status or water resource utilization efficiency, and it is difficult to achieve multi-target optimization. For example, a rule-based pest control system may only focus on the reduction of pests and diseases, while ignoring the impact on plant growth status and water resource utilization efficiency. The DQN algorithm can consider multiple optimization targets at the same time and achieve multi-target optimization by dynamically adjusting the weight value of each target. This helps to comprehensively improve the effect of urban greening management and ensure the coordinated development of all aspects. Secondly, the traditional urban greening management system relies on expert experience and fixed rules, lacking learning ability and intelligent decision-making ability. For example, the rule-based irrigation system can only irrigate according to the preset rules and cannot make intelligent adjustments according to the actual situation. The DQN algorithm can learn the optimal management strategy from a large amount of historical data through deep learning technology, and has strong learning ability and intelligent decision-making ability. Through continuous iterative optimization, the DQN algorithm can gradually converge to the optimal strategy and improve the intelligent level of management. At the same time, the traditional urban greening management system has a large waste of resource utilization, such as over-irrigation and over-fertilization. For example, the rule-based irrigation system may irrigate when it is not necessary, resulting in water resource waste. The DQN algorithm can dynamically adjust management measures according to the current environmental status and optimization goals to improve resource utilization efficiency. Through precise irrigation and fertilization strategies, the DQN algorithm can reduce resource waste, improve resource utilization efficiency, and promote sustainable development. Finally, the traditional urban greening management system has shortcomings in environmental protection, such as excessive use of chemical pesticides. For example, the rule-based pest control system may overuse chemical pesticides, resulting in environmental pollution. The DQN algorithm can comprehensively consider the environmental protection effect, reduce the use of chemical pesticides by dynamically adjusting management measures, and improve the environmental protection effect. Through precise pest and disease control strategies, the DQN algorithm can reduce environmental pollution and protect the ecological environment.

[0114] The hardware working environment of the deep Q network algorithm in the urban greening system based on big data mainly includes: high-performance computing servers, data acquisition and processing equipment, network communication equipment and intelligent execution equipment. The high-performance computing server is responsible for running the deep Q network algorithm, performing complex data processing and model training; the data acquisition and processing equipment includes various sensors, environmental monitoring stations, etc., which are used to collect multi-dimensional data related to urban greening in real time; the network communication equipment ensures the efficient transmission of data between the server, data acquisition equipment and intelligent execution equipment; the intelligent execution equipment executes the corresponding greening management measures, such as irrigation, fertilization, pruning, etc., according to the optimal strategy output by the algorithm.

[0115] Based on the above hardware, the deep Q network algorithm (Group A) and the traditional algorithm (Group B) were used for comparative experiments. Both groups of experiments were conducted under the same experimental conditions, including the same urban greening area, the same data collection frequency, and the same greening management measures implementation standards. The traditional algorithm uses a rule-based algorithm. During the experiment, the plant growth status, pest and disease control effect, water resource utilization efficiency, and environmental protection effect indicators of each group of experiments were recorded under five repeated runs. The data record table is shown in Table 3:

[0116] Table 3 Experimental data record table

[0117]

[0118]

[0119] Through comparative experiments, it can be seen that the deep Q network algorithm (Group A) is superior to the traditional algorithm (Group B) in terms of plant growth conditions, pest control effects, water resource utilization efficiency and environmental protection effects. The average plant growth conditions of Group A increased by about 10%, the pest control effect increased by about 10%, the water resource utilization efficiency increased by about 12%, and the environmental protection effect increased by about 8%. This shows that the deep Q network algorithm can more accurately adjust the greening management strategy according to environmental changes and achieve more efficient and intelligent urban greening management.

[0120] In S7, mixed integer linear programming (MILP) is a mathematical optimization technique used to solve linear optimization problems involving continuous and discrete variables. In urban greening systems, MILP can be used to optimize decision variables such as irrigation duration, fertilizer type, fertilizer quantity, fertilizer frequency, and plant location and type. By defining objective functions and constraints, MILP can find the optimal solution that satisfies all constraints. In multi-objective optimization, MILP can be used to maximize plant growth, minimize pest and disease risks, and improve water resource utilization efficiency and environmental protection effects. By defining multiple objective functions, MILP can trade off between multiple objectives and find a comprehensive optimal solution. Context awareness refers to the ability of the system to automatically adjust its behavior and strategy based on the current environment and user needs. In urban greening systems, context awareness dynamically adjusts the input parameters of the optimization model by collecting environmental data (such as temperature, humidity, light), pest and disease data, and growth prediction data in real time. Context awareness technology can ensure that the input data of the optimization model is always up to date, improving the accuracy and timeliness of the optimization results. By monitoring environmental changes in real time, the system can adjust management strategies in a timely manner to respond to emergencies. Genetic algorithm is a global optimization algorithm based on natural selection and genetic mechanism. By simulating the biological evolution process, genetic algorithm can gradually optimize the quality of the solution through selection, crossover and mutation operations. In multi-objective optimization, genetic algorithm can find multiple Pareto optimal solutions and provide multiple optimization schemes for decision makers to choose. In urban greening systems, genetic algorithms can be used to optimize irrigation strategies, fertilization strategies and plant layout. By defining the fitness function, genetic algorithms can evaluate the effects of different management strategies and select the optimal management scheme.

[0121] In practical applications, S7 first collects environmental data (such as temperature, humidity, light), pest and disease data, and growth prediction data in real time through various sensors and devices. The collected data is preprocessed, including data cleaning, alignment, and standardization, to ensure data quality and consistency. Next, an optimization model is built to define multiple optimization objectives, such as plant growth status, pest and disease risk, water resource utilization efficiency, and environmental protection effect. Decision variables are defined, including irrigation duration, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type. Constraints are defined, such as maximum irrigation water volume, maximum fertilizer amount, etc., based on actual conditions and restrictions. The optimization model is solved using a MILP solver (such as Gurobi or CPLEX) to find the optimal solution that meets all constraints. The input parameters of the optimization model are dynamically adjusted through the context-aware module to ensure the accuracy and timeliness of the optimization results. The optimization model is globally optimized using a genetic algorithm to find multiple Pareto optimal solutions and provide a variety of optimization schemes for decision makers to choose from. Based on the optimization results, the optimal comprehensive management strategy is generated, including irrigation duration, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type. The generated management strategy is stored in the database for subsequent management and decision-making. The generated management strategy is applied to actual management, such as adjusting the irrigation system, fertilization system, and plant layout. By real-time monitoring of environmental changes and management effects, feedback data is collected to further optimize the management strategy.

[0122] In the specific implementation, the operation process of the model is as follows: Initialize the parameters of the MILP model, including the objective function, decision variables and constraints. Initialize the parameters of the genetic algorithm, including population size, crossover probability, mutation probability and maximum number of iterations. Initialize the context-aware module and set the parameters for data collection and processing. Then collect environmental data, pest and disease data and growth prediction data in real time through sensors and equipment. Preprocess the collected data, including data cleaning, alignment and standardization, to ensure the quality and consistency of the data. Input the preprocessed data into the MILP model as the input parameters of the optimization model. Dynamically adjust the input parameters of the optimization model through the context-aware module to ensure the accuracy and timeliness of the optimization results. Use the MILP solver to solve the optimization model and find the optimal solution that meets all constraints. Use the genetic algorithm to globally optimize the optimization model, find multiple Pareto optimal solutions, and provide multiple optimization schemes for decision makers to choose. According to the optimization results, generate the optimal comprehensive management strategy, including irrigation time, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type. Store the generated management strategy in the database for subsequent management and decision-making. Finally, the generated management strategies are applied to actual management, such as adjusting the irrigation system, fertilization system and plant layout. By real-time monitoring of environmental changes and management effects, feedback data is collected to further optimize management strategies.

[0123] In specific implementation, traditional urban greening management systems usually only focus on a single goal, such as plant growth or water resource utilization efficiency, and it is difficult to achieve multi-objective optimization. For example, a rule-based pest control system may only focus on the reduction of pests and diseases, while ignoring the impact on plant growth and water resource utilization efficiency. The multi-objective optimization model can balance multiple goals by defining multiple optimization goals, such as plant growth, pest and disease risk, water resource utilization efficiency, and environmental protection effect, to find a comprehensive optimal solution and achieve multi-objective optimization. In addition, the multi-objective optimization model can collect environmental data, pest and disease data, and growth prediction data in real time through context perception technology, dynamically adjust the input parameters of the optimization model, ensure the accuracy and timeliness of the optimization results, and improve the real-time and adaptability of management. Secondly, the multi-objective optimization model can perform global optimization through genetic algorithms, find multiple Pareto optimal solutions, and provide multiple optimization schemes for decision makers to choose from, which improves the quality of optimization results and ensures the scientificity and rationality of management strategies. At the same time, the multi-objective optimization model can reduce resource waste and improve resource utilization efficiency by optimizing irrigation strategies and fertilization strategies. For example, through precise irrigation and fertilization strategies, the model can reduce the waste of water and fertilizer, improve resource utilization efficiency, and promote sustainable development. In addition, the multi-objective optimization model can reduce the use of chemical pesticides and improve environmental protection effects by comprehensively considering environmental protection effects. For example, through precise pest control strategies, the model can reduce environmental pollution and protect the ecological environment. Finally, the multi-objective optimization model can learn the optimal management strategy from a large amount of historical data through MILP and genetic algorithms, and has strong learning ability and intelligent decision-making ability. Through continuous iterative optimization, the model can gradually converge to the optimal strategy and improve the level of intelligent management.

[0124] In practical applications, the hardware working environment of the scenario-based multi-objective optimization model in the urban greening system based on big data mainly includes: high-performance computing servers, big data storage and processing equipment, Internet of Things sensor network, intelligent irrigation and fertilization system, and network communication equipment. High-performance computing servers are responsible for running complex optimization algorithms and data processing tasks; big data storage and processing equipment is used to store massive amounts of urban greening data and perform efficient data processing and analysis; the Internet of Things sensor network is responsible for real-time collection of environmental data, plant growth data, and pest and disease data; the intelligent irrigation and fertilization system executes specific greening management measures according to the strategy output by the optimization model; and network communication equipment ensures data communication and command transmission between devices.

[0125] Based on the above hardware environment, set up Group A: use a situation-based multi-objective optimization model (MILP + situational awareness + genetic algorithm). Group B: use a traditional rule-based optimization algorithm. Use Group A) and Group B for comparative experiments. Both groups of experiments were conducted under the same experimental conditions, including the same urban greening area, the same data collection frequency, and the same greening management measures implementation standards. During the experiment, the plant growth efficiency, resource consumption (such as irrigation water volume, fertilizer application), pest and disease control effect, and system response time of each group of experiments were recorded under five repeated runs. The experimental record data table is shown in Table 4:

[0126] Table 4 Comparative Experiment Record

[0127]

[0128] From the experimental data in Table 4, we can see that the scenario-based multi-objective optimization model (Group A) is superior to the traditional algorithm (Group B) in terms of plant growth efficiency, resource consumption, pest and disease control effect, and system response time. The average plant growth efficiency of Group A increased by about 7%, resource consumption decreased by about 10%, pest and disease control effect increased by about 7%, and system response time was shortened by about 50%. This shows that the scenario-based multi-objective optimization model can more accurately adjust the greening management strategy according to the real-time environment and plant growth conditions, and achieve more efficient and intelligent urban greening management.

[0129] In S8, the adaptive control algorithm is an algorithm that can dynamically adjust the control parameters according to the real-time data of the system operation. Its core idea is that during the operation of the system, through continuous learning and adjustment, the control system can adapt to environmental changes and achieve the optimal control effect. Adaptive control algorithms include model reference adaptive control (MRAC), self-tuning control (Self-Tuning Control) and fuzzy adaptive control. Model reference adaptive control (MRAC) defines an ideal reference model to make the output of the actual system as close as possible to the output of the reference model. By estimating the parameters of the system online, the controller gain is continuously adjusted to make the system output gradually approach the reference model. Self-tuning control (Self-Tuning Control) estimates the model parameters of the system online and dynamically adjusts the parameters of the controller to maintain good control performance in a changing environment. Fuzzy adaptive control uses a fuzzy logic system to dynamically adjust the control parameters according to the real-time status and historical data of the system, so that the system can adapt to complex and changing environments. The control system is responsible for dynamically adjusting the operating parameters of the irrigation system and fertilization equipment facilities according to the optimization strategy. The control system usually includes sensors, actuators, controllers and communication modules. Sensors are used to monitor the environment and plant status in real time, actuators are used to execute control commands, controllers are used to process data and generate control signals, and communication modules are used for data transmission. In the urban greening system, the control system dynamically adjusts the irrigation time of the irrigation system, the type, quantity and frequency of fertilizers applied by the fertilization equipment, and other parameters through adaptive control algorithms based on the output results of the scenario-based multi-objective optimization model to ensure the optimal realization of goals such as plant growth, pest and disease control, water resource utilization and environmental protection.

[0130] In practical applications, the working method of step S8 is as follows: First, environmental data (such as temperature, humidity, and light), plant growth data (such as growth rate, health status), and pest and disease data are collected in real time through various sensors and devices. The collected data is preprocessed, including data cleaning, alignment, and standardization, to ensure the quality and consistency of the data. Next, the preprocessed data is input into the scenario-based multi-objective optimization model to generate the optimal comprehensive management strategy, including irrigation duration, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type. The generated management strategy is stored in the database for use by the control system. The control system dynamically generates control signals through the adaptive control algorithm based on the optimization strategy. The adaptive control algorithm dynamically adjusts the control parameters based on real-time data to ensure the accuracy and timeliness of the control signal. For example, if the real-time data shows that the soil moisture is low, the adaptive control algorithm will increase the irrigation duration; if the plant growth condition is poor, the adaptive control algorithm will adjust the type and amount of fertilizer. The actuator executes the corresponding control command based on the generated control signal. For example, the automatic irrigation system adjusts the irrigation duration according to the control signal, and the automatic fertilization system adjusts the type and amount of fertilizer according to the control signal. By real-time monitoring of environmental changes and management effects, feedback data can be collected to further optimize control strategies.

[0131] In specific implementation, the control system based on the adaptive control algorithm can collect environmental data in real time, dynamically adjust the operating parameters of the irrigation system and fertilization equipment, and ensure the accuracy and timeliness of the control signal. This improves the real-time and adaptability of the system, enabling the system to better cope with environmental changes. In addition, the control system based on the adaptive control algorithm can dynamically adjust the control parameters according to real-time data, so that the system maintains good control performance in a changing environment. This improves the control effect and ensures the optimal realization of goals such as plant growth conditions, pest control, water resource utilization and environmental protection. Secondly, the control system based on the adaptive control algorithm can dynamically adjust the irrigation time and fertilization quantity according to real-time data, reduce resource waste and improve resource utilization efficiency. For example, through precise irrigation and fertilization strategies, the system can reduce the waste of water resources and fertilizers, improve resource utilization efficiency, and promote sustainable development. At the same time, the control system based on the adaptive control algorithm can dynamically adjust the pest control strategy according to real-time data, reduce the use of chemical pesticides, and improve the environmental protection effect. For example, through precise pest control strategies, the system can reduce environmental pollution and protect the ecological environment. Finally, the control system based on the adaptive control algorithm can learn the optimal control strategy from a large amount of historical data, and has strong learning ability and intelligent decision-making ability. Through continuous iterative optimization, the system can gradually converge to the optimal strategy and improve the level of intelligent management.

[0132] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A method for optimizing urban greening based on big data, characterized by: The following steps are involved: S1. Obtain temperature and humidity, light intensity, soil moisture and air quality data from environmental sensors through API interfaces and IoT wireless communication methods, obtain plant species and growth cycle data from plant growth cycle databases, obtain crowd density and activity area data from public data sources, and regularly obtain image data of green areas from drone inspection systems, and store them in temporary buffers; S2. Input the data into a distributed real-time data stream processing platform to clean, align, convert and fuse the collected data; the distributed real-time data stream processing platform integrates a unified format data set through a deep heterogeneous fusion algorithm, including unified format environmental parameters, vegetation growth status, and human activity patterns; S3, based on the image data of green areas of the drone inspection system, identifies the symptoms of pests and diseases through machine learning algorithms based on image recognition, and automatically triggers corresponding emergency response measures through event-driven mechanisms when abnormal environmental data is detected, including increasing the monitoring frequency and adjusting the irrigation strategy; S4. Train historical environmental data, plant growth data, and pest and disease symptoms through long short-term memory networks to build a growth prediction model; S5, inputting the real-time collected environmental data into the growth prediction model to obtain the growth prediction results of future environmental changes, and generating optimization strategies including irrigation plans, fertilization strategies, pest and disease control measures and greening area layout based on the prediction results; S6. Define optimization objectives and initial weight values, including plant growth conditions, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects, and dynamically adjust the weight value of each objective according to environmental changes through the deep Q network algorithm; S7, constructing a situation-based multi-objective optimization model through mixed integer linear programming, situational awareness and genetic algorithm, inputting real-time environmental data, pest and disease data, growth prediction data and optimization strategy into the situation-based multi-objective optimization model, obtaining the optimal comprehensive management strategy, including irrigation duration, fertilizer type, fertilizer quantity, fertilization frequency, and plant location and type, and storing the output results in a database; S8. Send the output results of the scenario-based multi-objective optimization model to the control system, and the control system dynamically adjusts the operating parameters of the irrigation system and fertilization equipment and facilities through an adaptive control algorithm according to the optimization strategy.

2. The urban greening optimization method based on big data according to claim 1 is characterized by: The working method of the deep heterogeneous fusion algorithm is as follows: fill missing values ​​through mean interpolation, median interpolation and K nearest neighbor interpolation, and identify and process outliers through the Z-Score method; align timestamps through linear interpolation and spline interpolation; align spatial coordinates through geocoding and coordinate conversion; Based on the aligned data, the deep heterogeneous fusion algorithm fuses the features of data from different sources through multimodal data; the multimodal data fusion extracts the features of image and time series data through convolutional neural networks and recurrent neural networks, and adopts mutual information method, principal component analysis and feature engineering, and generates feature vector F through dynamic weight adjustment mechanism. The fused feature vector F is used as the input of the deep neural network, and the nonlinear mapping relationship between input data and output data is learned through multilayer perceptron and autoencoder to generate a data set in a unified format; the multilayer perceptron learns the nonlinear mapping relationship between input data and output data through forward propagation and backpropagation algorithms, introduces nonlinear characteristics through activation function ReLU, and improves the expression ability of the model; the autoencoder learns the low-dimensional representation of data through encoder and decoder, and optimizes model parameters through mean square error MSE.

3. The urban greening optimization method based on big data according to claim 1 is characterized by: The working principle of the machine learning algorithm based on image recognition is: Step 301: pre-processing the image data collected by the drone through data cleaning and image enhancement, wherein the data cleaning removes noise in the image through Gaussian filtering, and the image enhancement enhances the details of the image through histogram equalization and contrast adjustment; Step 302: extract features from the image through a convolutional neural network; the convolutional neural network performs a convolution operation on the image through a convolution layer to extract local features, performs a maximum pooling operation on the feature map through a pooling layer to reduce computational complexity; and performs a nonlinear transformation on the output of the convolutional layer through an activation function; Step 303: locate and identify the symptoms of pests and diseases in the extracted image features through a target detection algorithm; the working method of the target detection algorithm is: r1. Generate candidate regions of pest and disease symptoms through sliding windows and region proposal networks. r2. Extract features of the candidate area through a convolutional neural network and generate a feature vector; r3. Classify the feature vector and perform bounding box regression through the fully connected layer and cross entropy loss function to determine the location and category of the pest and disease symptoms; r4. Ensure the accuracy and reliability of the machine learning algorithm through data labeling and model training; the data labeling is to label the pest and disease symptoms in the image through manual labeling or semi-supervised learning to generate a training data set; the model training is to train the model through the back propagation algorithm and the optimization algorithm SGD; r5. The identified pest and disease symptoms are transmitted to the central server in real time through an event-driven mechanism.

4. The urban greening optimization method based on big data according to claim 1 is characterized by: The working method of the event-driven mechanism is: read the pest and disease symptom data from the message queue through the event processor, generate response measure events, and push the response measure events to the message queue, and trigger emergency response measures through the rule engine and decision tree algorithm, including increasing the monitoring frequency, adjusting the irrigation strategy, and starting pest and disease control measures.

5. The urban greening optimization method based on big data according to claim 1 is characterized by: The growth prediction model includes an input layer, an embedding layer, a long short-term memory layer, an attention mechanism layer, a fully connected layer and an output layer; the operation process steps of the growth prediction model are as follows: Step 501: receiving and processing input data through the input layer, including historical environmental data, plant growth data, and pest and disease symptoms; Step 502: Convert the input data into a high-dimensional vector representation through the embedding layer to capture data features; The embedding layer converts plant species and pest types into continuous vector representations through a word embedding method; Step 503: receiving the high-dimensional vector output by the embedding layer through the long short-term memory layer, capturing the long-term dependency in the time series data, and extracting the time series features; the long short-term memory layer dynamically adjusts the retention and discarding of information through a gating mechanism; Step 504: The attention mechanism layer is used to enhance the model's attention to important time steps; the attention mechanism uses the attention mechanism to calculate the weight of each time step, and the weighted sum is used as the supplementary information output by the long short-term memory layer to enhance the model's sensitivity to key time steps; Step 505: The fully connected layer performs a nonlinear transformation on the extracted features to generate a prediction result; the fully connected layer performs a nonlinear transformation on the outputs of the long short-term memory layer and the attention mechanism layer through a ReLU activation function to map the features extracted by the long short-term memory layer and the attention mechanism layer to the output space; Step 506: Output the final growth prediction result through the output layer, including the plant growth status and pest and disease prediction under future environmental changes; the output layer generates a binary classification prediction result through the Sigmoid activation function and generates a multi-classification prediction result through the Softmax activation function.

6. The urban greening optimization method based on big data according to claim 1 is characterized by: The method for generating an irrigation plan, a fertilization strategy, pest and disease control measures, and an optimization strategy for the layout of green areas based on the prediction results in S5 is as follows: based on the vegetation growth requirements and soil moisture data, an irrigation plan is generated through an expert system, including the time, frequency, and amount of irrigation; based on the vegetation growth status and soil nutrient data, a fertilization plan is generated through a nutrition model, including the type, amount, and frequency of fertilizer; based on the pest and disease warning information and the vegetation growth status, control measures are formulated through an expert system, including spraying pesticides and pruning diseased branches; based on the human activity pattern and the vegetation growth status, the layout of the green area and the type of vegetation are generated through a spatial planning method.

7. The urban greening optimization method based on big data according to claim 1 is characterized by: The dynamic adjustment method of the deep Q network algorithm is: Step 701, defining optimization objectives, including plant growth conditions, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects, and setting initial weight values ​​for each optimization objective; Step 702: construct a deep neural network through a convolutional neural network and a fully connected layer; It includes an online network and a target network, wherein the online network is used to update and calculate the Q value of the current strategy in real time; the weight of the target network is regularly copied from the online network to calculate the target Q value to stabilize the training process; the convolutional neural network extracts the features of the input data through the convolution layer and the pooling layer to generate a feature vector; the fully connected layer maps the feature vector to the Q value, i.e., the expected reward value of each action, through a multi-layer perceptron; Step 703: collect environmental data in real time through the sensor network and the drone inspection system, including temperature, humidity, light intensity, and soil moisture, and pre-process the collected environmental data through normalization and standardization processing methods to construct an initialized environmental state space; Step 704: Combine the collected environmental data and the state information of the optimization target into a state vector as the input of the deep Q network; the state vector includes the current environmental data and the state information of the optimization target; Step 705: Select an action through the adaptive decay mechanism of the greedy strategy, wherein the greedy strategy randomly selects an action or selects the action with the highest Q value in the current state according to the exploration rate; the exploration rate is gradually reduced during the training process to balance exploration and utilization; Step 706: Execute the selected action, and collect environmental data and status information of the optimization target after the action is executed through the sensor network and the drone inspection system. If the plant growth condition improves, a positive reward is given, and if the pest and disease symptoms worsen, a negative reward is given; Step 707: Update the Q value through the Bellman equation, wherein the Bellman equation calculates the new Q value through the discount factor γ and the immediate reward r, calculates the difference between the current Q value and the target Q value through the mean square error loss function MSE, and updates the network parameters through the back propagation mechanism; Step 708: Periodically copy the weights of the online network to the target network through weight repetition to maintain the stability of the target value and reduce fluctuations during the training process; Step 709, continuously execute steps 704 to 708, interact with the environment, collect data, update network weights, and stop when the predetermined number of training rounds or convergence condition is met.

8. The urban greening optimization method based on big data according to claim 1 is characterized by: The scenario-based multi-objective optimization model includes a data reading layer, a feature extraction layer, a scenario perception layer, a target setting layer, an optimization modeling layer, and an algorithm solution layer; the working method based on scenario perception and multi-objective optimization model is: Step 801, reading pre-processed data from the central database through the data layer; including real-time environmental data, pest and disease data and growth prediction data; Step 802: extract key features of the data through the feature extraction layer; the feature extraction layer extracts features through a principal component analysis method, and stores the extracted feature data in a feature database; Step 803: The context perception layer reads the processed feature data from the feature database and identifies the current context state, including drought, rainstorm, high density of people, high incidence of pests and diseases, and vigorous growth of plants; the context feature perception layer divides the data into different categories through k cluster analysis to identify the current context state; the k cluster analysis selects the best clustering result through multiple initialization method, and the formula expression is: In formula (1), c i represents the i-th cluster center; X represents the set of all data points; ||Xc j || represents the data point X and the selected cluster center c j The Euclidean distance between j represents the variance of the jth cluster center, which is used to adjust the influence of distance; Step 804: setting optimization goals through the goal setting layer, including plant growth conditions, pest and disease symptoms, water resource utilization efficiency, and environmental protection effects; the goal setting layer sets an initial weight value for each optimization goal through an expert system combined with historical data, and the goal setting layer dynamically adjusts the weight value according to actual environmental changes and management needs through a feedback mechanism based on reinforcement learning, and continuously optimizes the weight setting; Step 805: construct a mixed integer linear programming model including the objectives of maximizing plant growth efficiency, minimizing the incidence of pests and diseases, and optimizing resource use costs through the optimization modeling layer. The mixed integer linear programming model dynamically adjusts the weights between different objectives based on real-time situations through a dynamic adjustment function. The formula expression of the dynamic adjustment function is: In formula (2), f(x) represents the comprehensive objective function; w i (t) represents the weight of the i-th target at time t; g i (x) represents the value of the i-th objective function; γ i Represents the penalty coefficient of the i-th objective, which is used to adjust the nonlinear impact of the objective function; θ i represents the threshold of the i-th target, which is used to define the benchmark value of the objective function; β i represents the smoothing parameter of the i-th objective, which is used to control the smoothness of the objective function; Step 806: Generate the optimal management strategy through the algorithm solution layer, including irrigation duration, fertilizer type, fertilizer quantity, fertilizer frequency, and plant position and type. The optimization solution layer solves the optimization model through a genetic algorithm, which gradually optimizes the individuals in the population through selection, crossover and mutation operations, and handles conflicts between targets through non-dominated sorting and crowding distance calculation. The algorithm solution layer also continuously monitors the optimization effect through a feedback mechanism based on online learning, and dynamically adjusts the optimization strategy according to actual conditions.

9. The urban greening optimization method based on big data according to claim 1 is characterized by: The adaptive control algorithm realizes dynamic adjustment of the irrigation system and fertilization equipment through PID control algorithm and fuzzy control, and ensures that the system maintains optimal performance in a dynamic environment through an online learning algorithm; wherein, the PID control algorithm dynamically adjusts control parameters through proportional control, integral control and differential control; the fuzzy control converts complex nonlinear relationships into fuzzy rules through a fuzzy inference system, and makes decisions based on the fuzzy rule base; the online learning algorithm updates control parameters in real time through a recursive least squares method.

10. An urban greening system based on big data, characterized by: An urban greening optimization method based on big data as applied to any one of claims 1 to 9, comprising: Multi-source data acquisition module, used to collect multi-modal data of urban greening areas, including temperature and humidity, light intensity, soil moisture, air quality, plant species, growth cycle, vegetation growth status and human density data; A data integration module, used for cleaning, aligning, converting and fusing the multimodal data collected by the multi-source data collection module through a deep heterogeneous fusion algorithm; The real-time data analysis module is used to analyze the integrated data in real time, including vegetation growth trend prediction, pest and disease warning, and human activity pattern recognition, and output real-time analysis results, including irrigation demand prediction, fertilization recommendations, pest and disease warning information, and human activity hotspots; Adaptive decision-making module, which is used to build a situation-based multi-objective optimization model based on real-time analysis results through mixed integer linear programming, situation awareness and genetic algorithm, dynamically adjust irrigation plans, fertilization strategies, pest control measures and greening area layout, and realize data-driven intelligent management; User interaction feedback module, used to provide user interaction functions, including user query, feedback, activity participation, and display real-time data and results of greening management; The output end of the multi-source data acquisition module is connected to the input end of the data integration module; the output end of the data integration module is connected to the input end of the real-time data analysis module; the output end of the real-time data analysis module is connected to the input end of the adaptive decision module; the user interaction feedback module is connected to the above modules.

Images