Road bougainvillea speetabilis intelligent fertilization supervision method and system

Abstract

The invention relates to a road bougainvillea speetabilis intelligent fertilization supervision method and system, and the method comprises the steps: enabling collected multispectral image data and thermal imaging image data to be precisely matched with GPS coordinates, enabling soil component parameters to be aligned with the corresponding multispectral image data and thermal imaging image data, and forming a spatiotemporal data set, performing feature extraction on the multispectral image by adopting a convolutional neural network, constructing a growth attitude comprehensive analysis model, comparing a result output by the growth attitude comprehensive analysis model, namely a current growth state quantized value, with a preset standard based on historical growth data and a bougainvillea speetabilis growth rule, and calculating the missing amount of related fertilizers, a precise fertilization formula and a fertilization time node are determined by combining the current fertility condition of the soil, a control instruction is generated according to the fertilization formula and the time node, and the fertilization robot executes fertilization operation based on the control instruction and a preset path. The method has the effects of improving the fertilizer utilization rate, reducing the labor cost and improving the maintenance efficiency.

Description

Technical Field

[0001] This application relates to the technical field of smart agriculture and precision fertilization, and in particular to a method and system for intelligent fertilization monitoring of road bougainvillea. Background Technology

[0002] With the acceleration of urbanization, green plants (such as bougainvillea) along urban roadsides and overpasses play an important role in beautifying the environment and improving the ecology. Autumn and winter are critical periods for bougainvillea growth, and its growth status and nutrient requirements differ significantly from those in summer. However, traditional fertilization management methods often lack precision and fail to meet the plant's needs at different growth stages and under different environmental conditions. Therefore, developing an intelligent fertilization monitoring method is of great significance for improving the growth quality and maintenance efficiency of bougainvillea.

[0003] In existing plant care techniques, fertilization management mainly relies on manual experience or simple soil testing methods. For example, soil fertility levels are determined by periodically collecting soil samples for laboratory analysis; or fertilization needs are judged based on plant appearance characteristics (such as leaf color and growth rate). In addition, some more advanced technologies are beginning to incorporate sensor networks to monitor parameters such as soil moisture and temperature in real time to assist in fertilization decisions. However, these methods still have many limitations when applied to plants along roadsides and on elevated structures.

[0004] A major drawback of existing fertilization management techniques is the lack of a comprehensive assessment of plant growth status and soil fertility. Traditional fertilization methods typically focus only on a single indicator of soil fertility, neglecting the actual growth needs of plants and the influence of environmental factors. This imprecise fertilization approach may not only lead to fertilizer waste but also negatively impact plant growth and the ecological environment. Summary of the Invention

[0005] To achieve precise fertilization, this application provides an intelligent fertilization monitoring method and system for roadside bougainvillea.

[0006] Firstly, this application provides an intelligent fertilization monitoring method for roadside bougainvillea, employing the following technical solution: A method for intelligent monitoring of fertilization application for roadside bougainvillea includes: In the bougainvillea planting area, soil composition parameters, multispectral image data and thermal imaging image data were collected simultaneously, and the corresponding GPS coordinates, timestamps and growth stage information were recorded. The collected multispectral and thermal image data are precisely matched with GPS coordinates. Based on the timestamp, the soil composition parameters are aligned with the corresponding multispectral and thermal image data to form a spatiotemporal dataset. Convolutional neural networks were used to extract features from multispectral images. Combined with the temperature field distribution of thermal imaging data, the health status, leaf density and distribution of bougainvillea canopy were identified. Long short-term memory networks were used to analyze the time series of soil composition parameters and predict the trend of soil fertility change. The image analysis results and soil parameter prediction results were fused through an attention mechanism to construct a comprehensive analysis model of growth posture. Based on historical growth data and the growth pattern of bougainvillea, nutritional requirement standards corresponding to different growth postures are set. The results output by the growth posture comprehensive analysis model, i.e. the quantitative value of the current growth status, are compared with the preset standards to calculate the amount of fertilizer deficiency. Combined with the current soil fertility status, the precise fertilizer formula and fertilization time are determined. Based on the fertilizer formula and time schedule, control instructions are generated, and the fertilization robot performs fertilization operations based on the control instructions and preset paths.

[0007] By employing the above technical solution, this method simultaneously collects soil, multispectral images, and thermal imaging data, and combines this with GPS coordinates and timestamps to form a spatiotemporal dataset. Convolutional neural networks and long short-term memory networks are used to analyze plant growth status and soil fertility changes, constructing a comprehensive analysis model. Based on this model, precise fertilizer formulations and timing points are determined, and the fertilization operation is performed by a fertilization robot, achieving intelligent and precise fertilization, improving fertilizer utilization, ensuring the healthy growth of bougainvillea, reducing labor costs, and improving maintenance efficiency.

[0008] Secondly, this application provides an intelligent fertilization monitoring system for roadside bougainvillea, employing the following technical solution: A roadside bougainvillea intelligent fertilization monitoring system includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is loaded and executed by the processor, it implements the roadside bougainvillea intelligent fertilization monitoring method as described in the first aspect. Attached Figure Description

[0009] Figure 1 This is a flowchart of an intelligent fertilization monitoring method for road bougainvillea according to an embodiment of this application.

[0010] Figure 2 This is a flowchart of another embodiment of the present application that fuses image analysis results with soil parameter prediction results through an attention mechanism to construct a comprehensive analysis model of growth posture. Detailed Implementation

[0011] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-2 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0012] Reference Figure 1 This application discloses an intelligent fertilization monitoring method for roadside bougainvillea, comprising the following steps: In step S100, soil composition parameters, multispectral image data, and thermal imaging image data are collected simultaneously in the bougainvillea planting area, and the corresponding GPS coordinates, timestamps, and growth stage information are recorded.

[0013] Among them, soil composition parameters refer to the content of nutrients such as nitrogen (N), phosphorus (P), and potassium (K) in the soil, which are obtained through a portable soil analyzer. Multispectral image data consists of images captured by a multispectral camera, containing spectral information in different bands (such as red and near-infrared light), used to analyze the physiological state of plants. Thermal imaging image data reflects the surface temperature distribution of plants, used to determine the health status of plants.

[0014] GPS coordinates: Geographical location information obtained through GPS devices, used to accurately locate the data collection point. Timestamp: Records the specific time of data collection, generated by the built-in clock module of the data collection device. Growth stage information: Records the growth stage of the plant based on its morphological characteristics (such as the number of flower buds, leaf color, etc.), observed and recorded manually.

[0015] The necessary and complete process is described as follows: In the bougainvillea planting area, a portable soil analyzer was used to collect soil composition parameters, recording the content of key components such as nitrogen, phosphorus, and potassium. Simultaneously, a multispectral camera was used to photograph the bougainvillea plants, setting the red light band (650-680nm) and near-infrared band (750-850nm) to acquire multispectral images. Next, a thermal imager was used to scan the plants, acquiring thermal images. During the data collection process, the coordinates of the collection points were recorded using a GPS device, and the timestamp was recorded by the device's built-in clock module. At the same time, the morphological characteristics of the bougainvillea were manually observed, and their growth stage information (such as budding stage, full bloom stage, etc.) was recorded. After all data collection was completed, the collected soil composition parameters, multispectral image data, thermal image data, GPS coordinates, timestamps, and growth stage information were stored in the data acquisition terminal to provide basic data for subsequent analysis.

[0016] Step S200: Accurately match the collected multispectral image data and thermal imaging image data with GPS coordinates. Using timestamps as a reference, align the soil composition parameters with the corresponding multispectral image data and thermal imaging image data to form a spatiotemporal dataset.

[0017] The necessary and complete process is described as follows: First, the acquired multispectral and thermal imaging data are precisely matched with GPS coordinates. Using GPS coordinates as a spatial reference, each image corresponds one-to-one with its acquisition location. Next, soil composition parameters are aligned with the corresponding multispectral and thermal imaging data using timestamps. For example, a timestamp matching algorithm is used to associate soil composition data acquired at the same time point with image data. Specifically, if the timestamp of a multispectral image is "2025-07-07 10:00:00", then thermal imaging images and soil composition parameter data with the same timestamp are found and combined into a single data unit. Finally, all aligned data units are integrated into a spatiotemporal dataset, providing structured data support for subsequent analysis.

[0018] In step S300, a convolutional neural network is used to extract features from the multispectral image. Combined with the temperature field distribution of the thermal imaging data, the health status, leaf density and distribution of the bougainvillea canopy are identified. A long short-term memory network is used to analyze the time series of soil composition parameters and predict the trend of soil fertility change. The image analysis results and the soil parameter prediction results are fused through an attention mechanism to construct a comprehensive analysis model of growth posture.

[0019] Convolutional Neural Networks (CNNs) are deep learning models used to extract features from images. By processing images through convolutional and pooling layers, they can identify local features within the image. Thermal imaging data, specifically the temperature field distribution reflected in thermal images, is used to analyze the health status of plants.

[0020] Long Short-Term Memory Network (LSTM): A variant of Recurrent Neural Network (RNN) used to process time-series data, capable of capturing long-term dependencies in the data. Attention Mechanism: A modeling mechanism used to identify and emphasize important features or information in data, helping the model focus on key parts. Comprehensive Growth Posture Analysis Model: A model built by fusing image analysis results and soil parameter prediction results, used for comprehensive analysis of plant growth status.

[0021] The complete process of this step can be found in steps S310 to S340, and will not be repeated here.

[0022] Step S400: Based on historical growth data and the growth pattern of bougainvillea, set the nutritional requirement standards corresponding to different growth postures. Compare the results output by the growth posture comprehensive analysis model, i.e., the quantitative value of the current growth status, with the preset standards to calculate the amount of fertilizer deficiency. Combined with the current soil fertility status, determine the precise fertilizer formula and fertilization time.

[0023] The data includes: Historical growth data: Recorded data on the growth status of bougainvillea in the past, including soil composition and leaf health status, used to analyze growth patterns. Nutritional requirement standards: Nutritional requirement thresholds set according to the growth stage and season of bougainvillea, used to assess the current nutritional status. Quantitative growth status values: Numerical expressions of plant growth status obtained through a comprehensive growth posture analysis model, used for comparison with standards.

[0024] Calculate the required fertilizer deficiency: By comparing the current growth status with the standard nutrient requirements, calculate the type and amount of fertilizer that needs to be supplemented. Precision fertilizer formulation: A fertilization plan developed based on the fertilizer deficiency and soil fertility status, including the type and amount of fertilizer. Fertilization timing: The optimal fertilization time is determined based on the bougainvillea's growth patterns and soil fertility trends.

[0025] The complete process of this step can be found in steps S410 to S460, and will not be repeated here.

[0026] In step S500, control instructions are generated based on the fertilizer formula and time nodes, and the fertilizer robot performs the fertilization operation based on the control instructions and the preset path.

[0027] Among them, the fertilization robot is an automated robot equipped with a navigation system and fertilization equipment, used to perform fertilization tasks according to preset paths and instructions. Control instructions are generated based on precise fertilizer formulas and time points to guide the operation of the fertilization robot. The preset path is the planned route for the fertilization robot within the planting area to ensure even fertilization coverage.

[0028] The necessary complete process is described as follows: Based on the precise fertilization formula and fertilization time node generated in step S400, control instructions are generated. These instructions include the amount of fertilizer, the type of fertilizer, and the specific time and location of fertilization. For example, the control instruction might be "Fertilize bougainvillea planting area A at 9:00 AM on October 15, 2025, with 3 grams of nitrogen fertilizer per square meter and 1 gram of phosphorus fertilizer per square meter." The fertilization robot executes the fertilization operation according to the control instructions and the preset path. The preset path is generated through a path planning algorithm to ensure that the fertilization robot can evenly cover the entire planting area. For example, a grid-based path planning is used to divide the planting area into multiple small blocks, and the robot fertilizes each block sequentially. At the same time, the fertilization robot is equipped with a fertilization feedback system to monitor the amount and location of fertilizer in real time, ensuring the accuracy and uniformity of the fertilization operation. If a deviation in the amount of fertilizer is detected, the feedback system will automatically adjust the output of the fertilization equipment to ensure that the fertilization task is completed as planned.

[0029] Reference Figure 2 The steps for constructing a comprehensive growth posture analysis model by fusing image analysis results with soil parameter prediction results through an attention mechanism include: Step S310: Establish the joint decision matrix of environment and growth stage, using external environmental parameters and growth stage status as input variables.

[0030] The environment-growth stage joint decision matrix is ​​a decision model that combines external environmental parameters (such as light, temperature, and humidity) with plant growth stages (such as germination, growth, and flowering) to assess plant growth under different environmental conditions. External environmental parameters include light intensity, temperature, humidity, and wind speed, which are collected in real time by sensors.

[0031] The necessary and complete process is described below: First, establish an environment-growth stage joint decision matrix. This matrix takes external environmental parameters (such as light intensity, temperature, and humidity) and growth stage states (such as germination, growth, and flowering) as input variables. For example, light intensity is divided into three levels: "low," "medium," and "high"; temperature is divided into three intervals: "low temperature," "suitable temperature," and "high temperature"; and growth stages are divided into three phases: "germination," "growth," and "flowering." Using these input variables, a three-dimensional matrix is ​​constructed, where each cell represents a specific combination of environment and growth stage. For example, a cell in the matrix might represent the combination of "medium light intensity, suitable temperature, and growth stage (growth phase)."

[0032] Step S320: Using a self-attention mechanism, based on the combined state output by the joint decision matrix, calculate the fusion coefficient of multispectral, thermal imaging spatial features and soil parameter time series features, and dynamically adjust the weights of each feature based on different combinations of environment and growth stage.

[0033] Among them, the self-attention mechanism is a model mechanism that automatically learns the importance weights of different parts of the input data, making the model focus more on key information. The combined state output by the joint decision matrix is ​​the result of a specific combination of environment and growth stage obtained from the environment-growth stage joint decision matrix. Multispectral and thermal imaging spatial features are spatial features related to plant growth status extracted from multispectral and thermal images.

[0034] Soil parameter time series features: Features extracted from time series data of soil component parameters to reflect the changing trends of soil fertility. Fusion coefficient: Used to measure the importance weight of different features under the current environmental and growth stage combination. Dynamic weight adjustment: The weights of each feature are adjusted in real time according to the current environmental and growth stage combination to adapt to different analytical needs.

[0035] The necessary and complete process is described as follows: Using a self-attention mechanism, the fusion coefficients of multispectral and thermal imaging spatial features and soil parameter time-series features are calculated based on the combined states output by the joint decision matrix. First, multispectral images, thermal imaging images, and soil parameter time-series data are input into the self-attention model. The self-attention mechanism dynamically adjusts the weights of each feature by calculating the similarity and correlation between these features. For example, in the combination of "high temperature, growing season," the temperature distribution features in the thermal imaging image may be more critical for assessing plant growth status, and the self-attention mechanism will assign it a higher weight; while in the combination of "low temperature, germination period," the nutrient content features in the soil parameter time series may be more important, and their weight will be increased accordingly.

[0036] Step S330: Link the growth stage gating unit and the environmental parameter response module to trigger feature enhancement or suppression logic for a specific environment-growth stage combination.

[0037] The Growth Stage Gating Unit (GTU) is a module used to control the flow of features according to the plant's growth stages, determining which features are important at the current stage. The Environmental Parameter Response Module (EPR) is a module used to adjust feature weights based on environmental conditions, dynamically adjusting feature importance according to environmental changes.

[0038] Feature Enhancement: Enhancing the influence of certain features through algorithms or models to make them more representative in subsequent analyses. Feature Suppression: Reducing the influence of certain features through algorithms or models to prevent them from misleading the analysis results. Specific Environment-Growth Stage Combinations: Specific combinations of conditions obtained from the environment-growth stage joint decision matrix, used to trigger feature enhancement or suppression logic.

[0039] The necessary complete process is described as follows: The growth stage gating unit and the environmental parameter response module work together to trigger feature enhancement or suppression logic for a specific environment-growth stage combination. Specifically, the growth stage gating unit filters features related to the current growth stage (e.g., flowering period, vegetative growth period). For example, during the flowering period, the number of flower buds and the health status of the flowers may be more important, and the gating unit will increase the weight of these features. Simultaneously, the environmental parameter response module adjusts the feature weights based on the current environmental conditions (e.g., high temperature, drought). For example, in a high-temperature environment, the temperature distribution features in the thermal imaging image may need to be enhanced, while soil moisture features may need to be suppressed.

[0040] Step S340: The image and soil parameter features after collaborative fusion, weight adjustment and feature enhancement and suppression are input into the model architecture. After integration by the fully connected layer and operation by the ReLU function, a comprehensive analysis model of growth posture is constructed, and the output is a quantitative vector of growth status including health index, nutrient stress level and water stress level.

[0041] The structure includes: **Fully Connected Layer:** A layer in a neural network used to integrate input features into the output, where each neuron is connected to all neurons in the previous layer. **ReLU Function (Rectified LinearUnit):** An activation function that introduces non-linearity, its formula is f(x) = max(0,x), which sets negative values ​​to 0 and retains positive values. **Growth Status Quantization Vector:** A vector output by the model containing information such as health index, nutrient stress level, and water stress level, used to quantify the plant's growth status.

[0042] The necessary and complete process is described as follows: The image and soil parameter features, after co-fusion, weight adjustment, and feature enhancement and suppression processing, are input into the comprehensive analysis model of growth posture. The model uses a fully connected layer to integrate the input features, introduces nonlinear factors through the ReLU activation function, and finally outputs a quantitative vector of growth status containing health index, nutrient stress level, and water stress level.

[0043] The specific process is as follows: First, the processed features are input into the fully connected layer of the model. For example, the input features include multispectral image features, thermal imaging image features, and soil parameter time series features. The fully connected layer integrates these features into an intermediate feature vector through weighted summation. Next, the intermediate feature vector is processed by the ReLU activation function to remove negative values ​​and retain positive values, further extracting useful information. Finally, the model outputs a growth status quantification vector, for example, a health index of 0.8 (indicating good health), a nutrient stress level of 1 (indicating mild stress), and a water stress level of 2 (indicating moderate stress).

[0044] The steps to determine the precise fertilizer formula and timing include: Step S410: Obtain microclimate data, chlorophyll fluorescence parameters, soil parameters, growth stage information, and historical fertilization data for the bougainvillea planting area.

[0045] The data includes microclimate data such as temperature, humidity, light intensity, and wind speed, collected in real time by miniature weather stations installed in the planting area. Chlorophyll fluorescence parameters, an indicator of plant photosynthetic efficiency, are measured using a chlorophyll fluorometer. Historical fertilization data includes the time, type, and amount of previous fertilizations, obtained from fertilization records.

[0046] The necessary complete process is described as follows: In step S410, firstly, microclimate data of the bougainvillea planting area, including temperature, humidity, light intensity, and wind speed, is acquired in real time using a micro-weather station. Next, chlorophyll fluorescence parameters are measured using a chlorophyll fluorometer to assess the plant's photosynthetic efficiency. Simultaneously, soil parameters such as soil texture, fertility, and pH are acquired using soil sensors, or relevant data are extracted from laboratory analysis reports. Furthermore, information on the bougainvillea's growth stage is recorded; for example, the number of leaves and flower buds is observed manually to determine whether it is in the budding, growth, or flowering stage. Finally, historical fertilization data, including fertilization time, fertilizer type, and amount, is compiled from fertilization records.

[0047] Step S420: Input the above data into the pre-constructed four-dimensional coupled model of environment-physiology-soil-growth stage, fuse the data by improving the GBDT algorithm, and output the nitrogen, phosphorus and potassium absorption efficiency correction coefficient.

[0048] The four-dimensional coupled model of environment-physiology-soil-growth stage is used to analyze the efficiency of plant nutrient absorption, comprehensively considering environmental conditions, plant physiological state, soil characteristics, and growth stage. The improved GBDT algorithm is an improved algorithm based on gradient boosting decision tree (GBDT), which improves the model's prediction accuracy by optimizing the tree's growth strategy and feature selection. Nitrogen, phosphorus, and potassium absorption efficiency correction coefficients are used to adjust the absorption efficiency of nitrogen, phosphorus, and potassium nutrients, dynamically calculated based on the plant's physiological state and environmental conditions.

[0049] The necessary and complete process is described as follows: Microclimate data (such as temperature and humidity), chlorophyll fluorescence parameters (reflecting photosynthetic efficiency), soil parameters (such as fertility and pH), and growth stage information (such as germination period and growth period) are input into a pre-constructed four-dimensional coupled model of environment-physiology-soil-growth stage. This model integrates these multi-dimensional data through an improved GBDT algorithm to calculate the correction coefficients for nitrogen, phosphorus, and potassium absorption efficiency.

[0050] The specific process is as follows: First, the improved GBDT algorithm automatically identifies the feature combinations that have the greatest impact on nutrient uptake efficiency through feature selection and tree growth strategy optimization. For example, the model may find that temperature and soil pH have a significant impact on nitrogen fertilizer uptake efficiency. Next, the model calculates correction coefficients based on these features. For example, under high temperature and acidic soil conditions, nitrogen fertilizer uptake efficiency may decrease, and the correction coefficients will be adjusted accordingly to values ​​less than 1.

[0051] Step S430: Based on the correction coefficient, the balance equation containing the reinforcement learning module is used to calculate the nitrogen, phosphorus and potassium nutrient deficiencies by combining the growth stage sensitivity coefficient, environmental correction factor and soil texture correction term.

[0052] The reinforcement learning module is an algorithm based on trial-and-error learning that dynamically adjusts model parameters through interaction with the environment to optimize the decision-making process. The balance equation is a mathematical model used to calculate the deficiencies of nitrogen, phosphorus, and potassium nutrients, which dynamically adjusts nutrient requirements by incorporating plant growth needs, soil fertility, and absorption efficiency correction coefficients.

[0053] Growth stage sensitivity coefficient: A coefficient set according to the plant's sensitivity to nutrient requirements at different growth stages. For example, the flowering period is more sensitive to phosphorus fertilizer requirements. Environmental correction factor: A coefficient that adjusts nutrient requirements and absorption efficiency based on environmental conditions (such as temperature and humidity). Soil texture correction item: A correction item that adjusts nutrient retention and supply capacity based on soil texture (such as sandy soil, loam, and clay).

[0054] The necessary complete process is described below: Based on the nitrogen, phosphorus, and potassium absorption efficiency correction coefficients output in step S420, the nitrogen, phosphorus, and potassium nutrient deficiencies are calculated using a balance equation containing a reinforcement learning module. The specific process is as follows: 1. Input Data: Input the nitrogen, phosphorus, and potassium absorption efficiency correction coefficients, growth stage sensitivity coefficients, environmental correction factors, and soil texture correction terms into the balance equation. For example, for bougainvillea in the flowering period, the growth stage sensitivity coefficient may be set to a higher value for phosphorus fertilizer (e.g., 1.2), while it may be set to 1.0 for nitrogen and potassium fertilizers.

[0055] 2. Equilibrium Equation Calculation: Based on the input correction coefficients and sensitivity coefficients, the equilibrium equation formula can be expressed as: ; Here, it is assumed that the theoretical requirements for nitrogen, phosphorus, and potassium are N0, N1, N2, N3, N4, N5, N6, N7, N8, N9, N10, N11 需求 P 需求 and K 需求 The current nutrient content in the soil is N 土壤 P 土壤 and K 土壤 The absorption efficiency correction factors are respectively The growth stage sensitivity coefficient, environmental correction factor, and soil texture correction term are S, respectively. 生长 E 环境 and T 土壤 .

[0056] Dynamically adjust nutrient requirements. For example, assuming the current nitrogen fertilizer content in the soil is 100 mg / kg, and the theoretical requirement of the plant at the current growth stage is 150 mg / kg, with a nitrogen fertilizer absorption efficiency correction coefficient of 0.8, then the actual nitrogen fertilizer requirement is 150 / 0.8 = 187.5 mg / kg. Therefore, the nitrogen fertilizer deficiency is 187.5 - 100 = 87.5 mg / kg.

[0057] 3. Reinforcement Learning Optimization: The reinforcement learning module dynamically adjusts the parameters in the equilibrium equation based on historical fertilization data and plant growth feedback. For example, if the plant grows well after fertilization, the reinforcement learning module will maintain the current parameters; if the plant grows poorly, it will adjust the parameters to optimize subsequent fertilization decisions.

[0058] Step S440: Using nutrient deficiency as input, a multi-objective optimization model is constructed by improving the NSGA-Ⅲ algorithm. Combined with soil urease activity and rhizosphere microbial diversity constraints, the optimal fertilization formula is generated.

[0059] Among them, nutrient deficiency: the deficient amounts of nutrients such as nitrogen, phosphorus, and potassium calculated through the balance equation, used to guide fertilization. Improved NSGA-Ⅲ algorithm: a multi-objective optimization algorithm based on non-dominated sorting, which improves solution diversity and convergence by improving selection and mutation strategies. Soil urease activity: an indicator reflecting urease activity in the soil, obtained through soil testing, used to assess soil fertility and nutrient transformation capacity.

[0060] Rhizosphere microbial diversity: An indicator obtained through soil microbial community analysis, reflecting the richness and diversity of rhizosphere microorganisms and influencing plant nutrient uptake. Multi-objective optimization model: A model that comprehensively considers multiple objectives (such as nutrient requirements, soil fertility, and environmental impact) to generate optimal fertilization formulas. Optimal fertilization formula: A fertilization plan generated based on the multi-objective optimization model, including fertilizer type, application rate, and application ratio.

[0061] The complete process of this step can be found in steps S441 to S444, and will not be repeated here.

[0062] Step S450: Use the GraphLSTM model to predict soil moisture content, daily average temperature and cold wave probability, and combine fertilizer characteristics with growth stage time windows to determine the fertilization time nodes through a dynamic time window algorithm.

[0063] The GraphLSTM model is a hybrid model combining a graph neural network and a long short-term memory (LSTM) network to process spatiotemporal series data, capturing the spatiotemporal dependencies of variables such as soil moisture content and temperature. Soil moisture content is the soil moisture content monitored in real time by a soil moisture sensor. Average daily temperature is the daily average temperature calculated from weather station data or historical meteorological data. Cold wave probability is the probability of a cold wave occurring, obtained through weather forecasting models or analysis of historical meteorological data.

[0064] Fertilizer characteristics: including the fertilizer's dissolution rate and release cycle, obtained from fertilizer instructions or experimental data. Growth stage time window: the appropriate time range for fertilization determined based on the plant's growth stage, for example, one week before flowering. Dynamic time window algorithm: an algorithm that dynamically adjusts the fertilization timing based on real-time data, considering plant growth stage, fertilizer characteristics, and environmental conditions.

[0065] The complete process of this step can be found in steps S451 to S453, and will not be repeated here.

[0066] Step S460: After fertilization, collect data on the growth status of bougainvillea and feed it back to the reinforcement learning module to optimize model parameters, forming a closed-loop regulation.

[0067] Among them, the post-fertilization growth status data includes plant growth status data obtained through monitoring methods such as multispectral images, thermal imaging images, and chlorophyll fluorescence parameters, reflecting the fertilization effect. Closed-loop regulation involves feeding the post-fertilization plant growth status back into the model, adjusting model parameters through a reinforcement learning module to form a dynamically optimized closed-loop system.

[0068] The necessary and complete process is described as follows: After fertilization, growth status data of the bougainvillea is collected, including multispectral images, thermal images, chlorophyll fluorescence parameters, etc. This data is acquired through sensors or monitoring equipment. This data is fed back to the reinforcement learning module to optimize model parameters and form a closed-loop regulation.

[0069] The specific process is as follows: First, image data of the bougainvillea after fertilization is collected using a multispectral camera and thermal imager, and chlorophyll fluorescence parameters are measured using a chlorophyll fluorescence meter. These data reflect the physiological state and health status of the plant. Next, these data are input into the reinforcement learning module. The reinforcement learning module dynamically adjusts the model parameters according to the plant's growth status. For example, if the plant's growth status is good after fertilization (e.g., normal leaf color, no pests or diseases), the reinforcement learning module will maintain the current fertilization strategy; if the growth status is poor (e.g., yellowing leaves, slow growth), the fertilization formula or fertilization timing will be adjusted.

[0070] The steps to generate the optimal fertilizer formula include: Step S441: Collect soil urease activity and rhizosphere microbial diversity index and set constraint thresholds to normalize the nitrogen, phosphorus and potassium nutrient deficiencies. Among them, the constraint threshold is the upper or lower limit set for soil urease activity and rhizosphere microbial diversity index, used to assess soil health. Normalization treatment converts nitrogen, phosphorus, and potassium nutrient deficiencies to the same dimensional range (e.g., 0 to 1) for subsequent optimization calculations.

[0071] The necessary and complete process is described below: First, soil urease activity and rhizosphere microbial diversity index are collected. Soil urease activity is obtained through soil enzyme activity detection methods, such as colorimetric determination. Rhizosphere microbial diversity index is obtained through high-throughput sequencing or biosensors, such as analyzing microbial community structure through 16S rRNA gene sequencing. Next, constraint thresholds are set according to soil health standards, for example, a lower limit threshold of 50 U / g for soil urease activity and a lower limit threshold of 3.0 for rhizosphere microbial diversity index.

[0072] Subsequently, the nitrogen, phosphorus, and potassium nutrient deficiencies were normalized. Assuming a nitrogen deficiency of 100 kg / ha, a phosphorus deficiency of 50 kg / ha, and a potassium deficiency of 80 kg / ha, these deficiencies were normalized to a range of 0 to 1. For example, using the maximum-minimum normalization method, the nitrogen deficiency was normalized to 0.8, the phosphorus deficiency to 0.4, and the potassium deficiency to 0.6. These normalized deficiencies will serve as input to the subsequent optimization model, providing data support for generating the optimal fertilization formula.

[0073] Step S442: If the detected value exceeds the constraint threshold, the dynamic reference point adjustment mechanism of the improved NSGA-Ⅲ algorithm is triggered to dynamically increase the density of reference points in the feasible solution region according to the proportion of nutrient deficiency.

[0074] The dynamic reference point adjustment mechanism is a mechanism used to improve the NSGA-III algorithm. It optimizes the distribution of solutions by dynamically adjusting reference points, ensuring both diversity and uniformity of solutions. The feasible solution region is the set of solutions that satisfy all constraints in the optimization problem. The reference point density, in the NSGA-III algorithm, refers to the number and distribution density of reference points used to guide the population distribution.

[0075] The complete process of this step can be found in steps S4421 to S4424, and will not be repeated here.

[0076] Step S443: Construct an optimization model with the objectives of minimizing fertilization costs, maximizing nutrient replenishment efficiency, and minimizing soil ecological impact. Input the normalized missing values ​​into the model, and generate the Pareto optimal solution set by introducing a non-dominated ranking based on constraint priority and an elite strategy that retains the optimal individuals with a pre-set proportion.

[0077] Among them, the multi-objective optimization model is an optimization model that considers multiple objectives simultaneously to find the optimal solution among conflicting objectives. Minimizing fertilization costs: One of the optimization objectives is to reduce fertilization costs by decreasing fertilizer application or selecting low-cost fertilizers. Maximizing nutrient replenishment efficiency: One of the optimization objectives is to ensure maximum nutrient absorption efficiency by plants through the rational allocation of fertilizer proportions. Minimizing soil ecological impact: One of the optimization objectives is to reduce negative impacts on the soil ecosystem, such as preventing nutrient loss and soil acidification, by optimizing fertilization programs.

[0078] Non-dominated sorting: An algorithm for multi-objective optimization that ranks solutions by their relative merits to find the Pareto optimal set. Elite strategy: An optimization strategy that ensures the inheritance of high-quality solutions in the population by retaining a predetermined proportion of the best individuals. Pareto optimal set: In multi-objective optimization problems, the set of solutions that cannot improve one objective without harming other objectives.

[0079] The necessary complete process is described below: the normalized nitrogen, phosphorus, and potassium nutrient deficiencies from step S441 are input into a multi-objective optimization model. This model aims to minimize fertilization costs, maximize nutrient replenishment efficiency, and minimize soil ecological impact. It solves for the Pareto optimal solution set through non-dominated ordination and an elite strategy.

[0080] The specific process is as follows: First, optimization objectives are set. Fertilization costs are calculated based on fertilizer prices and dosages; nutrient replenishment efficiency is assessed based on a plant uptake efficiency correction coefficient; soil ecological impact is measured by combining soil urease activity and rhizosphere microbial diversity indices. Next, individuals in the population are non-dominated and ranked to select the Pareto optimal solution set. For example, if one fertilization scheme is superior to another in terms of fertilization cost but slightly inferior in terms of nutrient replenishment efficiency, these two schemes are not mutually dominant and are both included in the Pareto optimal solution set. Then, the optimal individuals of a pre-set proportion (e.g., the top 10%) are retained to ensure the inheritance of high-quality solutions. For example, the 10% of individuals with the highest nutrient replenishment efficiency and lowest fertilization cost are selected from the current population as the basis for the next generation. Finally, a Pareto optimal solution set containing multiple optimization objectives is generated. Each solution represents a fertilization scheme that achieves a balance between fertilization cost, nutrient replenishment efficiency, and soil ecological impact, providing a scientific basis for precision fertilization.

[0081] Step S444: Use the TOPSIS method to calculate the closeness between each solution set individual and the ideal solution, select the solution with the highest closeness as the optimal fertilization formula, and output the nitrogen, phosphorus and potassium ratio and fertilization amount.

[0082] Among them, TOPSIS (Technique for Order Preference by Similarity to IdealSolution) is a multi-attribute decision-making method that selects the optimal solution by calculating the similarity between each alternative solution and the ideal solution and the negative ideal solution.

[0083] The necessary complete process is described as follows: In step S444, the optimal fertilizer formula is selected from the Pareto optimal solution set generated in step S443. The TOPSIS method is used to select the optimal solution by calculating the closeness of each solution to the ideal solution and the negative ideal solution.

[0084] The specific process is as follows: First, determine the ideal solution and negative ideal solution for each optimization objective. For example, for the objective of minimizing fertilization costs, the ideal solution is the lowest cost value, and the negative ideal solution is the highest cost value; for the objective of maximizing nutrient supplementation efficiency, the ideal solution is the highest efficiency value, and the negative ideal solution is the lowest efficiency value. Next, calculate the distance between each solution and the ideal and negative ideal solutions. The proximity formula is: ; Where D i Let D be the distance between the i-th solution and the ideal solution. i maxD represents the minimum distance between the solution and the ideal solution. i This represents the maximum distance between the solution and the negative ideal solution. The closer the proximity value is to 1, the closer the solution is to the ideal solution. Finally, the solution with the highest proximity value is selected from the Pareto optimal solution set as the optimal fertilization formula. For example, if three solutions have proximity values ​​of 0.85, 0.90, and 0.88 respectively, the solution with a proximity value of 0.90 is selected. The final output is the optimal fertilization formula, including the ratio and application rate of nitrogen, phosphorus, and potassium, such as 100 kg / ha of nitrogen, 50 kg / ha of phosphorus, and 80 kg / ha of potassium. Using the TOPSIS method, the fertilization scheme with the best overall performance can be selected from multiple optimization objectives, providing a scientific basis for precision fertilization.

[0085] The steps for dynamically increasing the density of reference points in the feasible solution region according to the proportion of nutrient deficiency include: Step S4421: Calculate the proportion of nitrogen, phosphorus, and potassium nutrient deficiencies in the total amount, and identify the dominant nutrient with the largest current deficiency.

[0086] The necessary and complete process is described below: calculate the proportion of nitrogen, phosphorus, and potassium nutrient deficiencies in the total amount, and identify the dominant nutrient with the largest current deficiency. The specific process is as follows: First, obtain the nutrient deficiencies of nitrogen, phosphorus, and potassium, for example, 100 kg / ha, 50 kg / ha, and 80 kg / ha, respectively. Next, calculate the total deficiency: 100 + 50 + 80 = 230 kg / ha. Then, calculate the percentage of each nutrient deficiency: nitrogen 100 / 230 ≈ 0.435 (43.5%), phosphorus 50 / 230 ≈ 0.217 (21.7%), and potassium 80 / 230 ≈ 0.348 (34.8%). By comparing the percentages, it is determined that the dominant nutrient with the largest deficiency is nitrogen (43.5%). This process is completed through simple ratio calculations, without the need for complex algorithms. Identifying the dominant nutrient provides a basis for subsequent optimization strategies, such as prioritizing nitrogen supplementation in fertilization plans.

[0087] Step S4422: When the soil urease activity is lower than the preset threshold or the rhizosphere microbial diversity index does not meet the standard, the dynamic densification mechanism is activated; according to the proportion of each nutrient deficiency, the number of new reference points is allocated to the sub-regions of the corresponding nutrient dimension in the three-dimensional target space.

[0088] The algorithm includes the following components: Dynamic Encryption Mechanism: A mechanism used to optimize the algorithm by dynamically adjusting the distribution of reference points, thereby improving the search accuracy in key areas. Soil Urease Activity: An indicator obtained through soil enzyme activity detection equipment, reflecting the activity level of urease in the soil. Preset Thresholds: Pre-set upper and lower limits for soil urease activity and rhizosphere microbial diversity index, used to assess soil health. Three-Dimensional Target Space: A target space with nitrogen, phosphorus, and potassium nutrient deficiency as dimensions, used for the distribution of reference points in the optimization algorithm. Number of Reference Points: The number of reference points used in the optimization algorithm to guide population distribution.

[0089] The necessary complete process is described as follows: When soil urease activity is detected to be below a preset threshold (e.g., 50 U / g) or the rhizosphere microbial diversity index is not up to standard (e.g., below 3.0), a dynamic densification mechanism is activated. At this time, based on the proportion of nitrogen, phosphorus, and potassium nutrient deficiencies in the total amount calculated in step S4421 (e.g., nitrogen 43.5%, phosphorus 21.7%, potassium 34.8%), the number of newly added reference points is allocated to the sub-regions of the corresponding nutrient dimensions in the three-dimensional target space. For example, assuming the initial number of reference points is 100, based on the proportion of nutrient deficiencies, 43.5 new reference points are added to the nitrogen dimension, 21.7 to the phosphorus dimension, and 34.8 to the potassium dimension.

[0090] Step S4423: In the target dimension corresponding to the dominant missing nutrient, the density of reference points is increased, and the increase is proportional to the proportion of the missing nutrient. For the non-dominant nutrient dimension, the density of reference points is decreased accordingly, and the decrease is inversely proportional to the proportion of the missing nutrient.

[0091] Among them, reference point distribution density: In the optimization algorithm, the density of reference points distributed in the target space affects the search accuracy of the algorithm. Dominant missing nutrient: The nutrient with the largest current missing amount, determined by step S4421, is the focus of attention during the optimization process.

[0092] The necessary complete process is described as follows: In step S4423, based on the percentage of nitrogen, phosphorus, and potassium nutrient deficiencies calculated in step S4421 (e.g., nitrogen 43.5%, phosphorus 21.7%, potassium 34.8%), the distribution density of reference points in the three-dimensional target space is adjusted. For dominant deficient nutrients (such as nitrogen), the distribution density of reference points is increased in the corresponding target dimension, with the increase proportional to the percentage of nutrient deficiency. For example, nitrogen has the highest deficiency percentage (43.5%), so the density of reference points is increased in the nitrogen dimension. For non-dominant nutrients (such as phosphorus and potassium), the distribution density of reference points is appropriately reduced, with the reduction inversely proportional to the percentage of nutrient deficiency. For example, phosphorus has the lowest deficiency percentage (21.7%), so the density of reference points is reduced in the phosphorus dimension. Through this dynamic adjustment, the optimization algorithm can perform a more refined search in the key nutrient dimensions while avoiding wasting computational resources in non-key dimensions, thereby improving optimization efficiency and solution quality.

[0093] Step S4424 finally generates an encrypted set of reference points, so that the distribution of reference points in the feasible solution region matches the proportion of nutrient deficiency.

[0094] The encrypted reference point set, which is a set of reference points with dynamically adjusted density, is used to guide the search direction of the optimization algorithm.

[0095] The necessary complete process is described below: Finally, an encrypted set of reference points is generated, ensuring that the distribution of these reference points in the feasible solution region matches the proportion of nutrient deficiency. This process builds upon the previous steps (S4421 to S4423), ensuring that the optimization algorithm can perform a more refined search in critical regions.

[0096] Specifically, through the dynamic adjustments made in the preceding steps, the distribution of reference points has been optimized based on the proportions of nitrogen, phosphorus, and potassium nutrient deficiencies. For example, nitrogen fertilizer deficiency accounts for the highest proportion (43.5%), so the density of reference points in the nitrogen fertilizer dimension has been increased; while phosphorus fertilizer deficiency accounts for the lowest proportion (21.7%), so the density of reference points in the phosphorus fertilizer dimension has been appropriately reduced. Ultimately, the generated densified set of reference points better reflects the actual needs of the current soil and plants.

[0097] Using the GraphLSTM model to predict soil moisture content, daily average temperature, and cold wave probability, and combining fertilizer characteristics with growth stage time windows, the steps to determine fertilization timing using a dynamic time window algorithm include: Step S451: Real-time collection of microclimate, soil parameters and severe weather remote sensing data; identification of weather types and extraction of occurrence probability and duration through convolutional neural network; spatiotemporal alignment with bougainvillea growth stage information and historical fertilization data.

[0098] Among them, severe weather remote sensing data includes weather information acquired through satellite remote sensing technology, such as cloud images and precipitation distribution. Spatiotemporal alignment involves matching data from different sources according to time and space to ensure data consistency and comparability.

[0099] The necessary and complete process is described below: Real-time acquisition of microclimate data (such as temperature, humidity, and light intensity), soil parameters (such as moisture content, fertility, and pH value), and remote sensing data on severe weather. A convolutional neural network (CNN) is used to identify weather types and extract their probability of occurrence and duration. For example, a CNN model can identify a cold wave that may occur within the next week, predicting a 30% probability of occurrence and a duration of 2 days.

[0100] Next, the identified weather types and their characteristics are spatiotemporally aligned with bougainvillea growth stage information (such as budding stage, growth stage, and flowering stage) and historical fertilization data. For example, assuming the bougainvillea is currently in its growth stage, historical fertilization data shows that the last fertilization was two weeks ago. Through spatiotemporal alignment, this data is integrated into a unified spatiotemporal framework, providing a foundation for subsequent environmental parameter prediction and determination of fertilization timing.

[0101] Step S452: Construct severe weather nodes in the graph structure of the GraphLSTM model, establish the association edges between severe weather nodes and soil moisture content and daily average temperature nodes, learn the influence weights of severe weather on soil moisture content, daily average temperature and cold wave probability through attention mechanism, and output the predicted values ​​of environmental parameters after pre-training the model with historical data.

[0102] Among them, severe weather nodes: In the graph structure of the GraphLSTM model, these represent nodes indicating severe weather events, and are associated with nodes such as soil moisture content and average daily temperature. Attention mechanism: A model mechanism used to dynamically adjust the connection weights between different nodes, enabling the model to focus on key information. Association edges: In the graph structure, the edges connecting different nodes represent the mutual influence between them. Pre-training: Training the model using historical data allows it to learn patterns and regularities in the data.

[0103] The necessary complete process is described as follows: Based on the data collected and aligned in step S451, environmental parameters are predicted using a GraphLSTM model. First, severe weather nodes are constructed in the graph structure of the GraphLSTM model, and their association edges with soil moisture content and daily average temperature nodes are established. For example, the association edge between the cold wave node and the soil moisture content node represents the impact of the cold wave on the soil moisture content.

[0104] Next, an attention mechanism is used to learn the weights of severe weather on soil moisture content, daily average temperature, and the probability of a cold wave. This attention mechanism dynamically adjusts the weights of associated edges, allowing the model to focus on key influencing factors. For example, the model might find that a cold wave has a higher weight on soil moisture content and a lower weight on daily average temperature.

[0105] The GraphLSTM model is pre-trained using historical data, enabling it to learn the relationship between severe weather and environmental parameters. For example, historical data shows that cold waves typically lead to decreased soil moisture content and lower daily average temperatures. Through pre-training, the model can output predicted values ​​of environmental parameters for a future period, such as soil moisture content, daily average temperature, and the probability of a cold wave.

[0106] Step S453: Introduce severe weather risk assessment factors into the dynamic time window algorithm. Based on the predicted weather intensity and duration, combined with fertilizer dissolution rate, nutrient release cycle and key growth period time window, dynamically adjust the start and end boundaries of the fertilization time window to determine the fertilization time node that meets the fertilizer characteristics and bougainvillea growth requirements.

[0107] The algorithm includes the following components: Dynamic Time Window Algorithm: This algorithm dynamically adjusts fertilization timing based on real-time data, considering plant growth stages, fertilizer characteristics, and environmental conditions. Severe Weather Risk Assessment Factor: This indicator assesses the impact of severe weather on fertilization timing, calculated based on predicted weather intensity and duration. Fertilizer Dissolution Rate: The rate at which fertilizer dissolves in the soil, obtained from fertilizer instructions or experimental data. Nutrient Release Cycle: The duration of nutrient release from the fertilizer, obtained from fertilizer instructions or experimental data. Critical Growth Stage Time Window: The appropriate time range for fertilization determined based on the plant's growth stage, for example, one week before flowering.

[0108] The necessary complete process is described as follows: Based on the environmental parameter prediction results of step S452, the fertilization time node is determined using a dynamic time window algorithm. First, a severe weather risk assessment factor is introduced into the dynamic time window algorithm, and the fertilization time window is adjusted according to the predicted weather intensity and duration. For example, if a cold wave is predicted to occur within the next 3 days and last for 2 days, the risk assessment factor will suggest completing fertilization before the cold wave arrives.

[0109] Next, by combining fertilizer characteristics (such as dissolution rate and nutrient release cycle) with the critical growth stage time window of bougainvillea, the start and end boundaries of the fertilization time window are dynamically adjusted. For example, if the fertilizer dissolves in 1 day, the nutrient release cycle is 30 days, and the bougainvillea is one week before its critical flowering period, the algorithm will suggest starting fertilization 2 days before the cold wave to ensure that the fertilizer dissolves and begins to release nutrients before the cold wave arrives, while meeting the nutrient requirements during the flowering period.

[0110] Ultimately, the algorithm outputs fertilization time points that meet the fertilizer characteristics and bougainvillea growth requirements, ensuring that fertilization is carried out under optimal environmental conditions, thereby improving fertilizer utilization efficiency and plant growth.

[0111] Based on the fertilizer formula and time schedule, control instructions are generated. The steps by which the fertilization robot performs the fertilization operation based on the control instructions and the preset path include: Step S4521: Combine the high-precision three-dimensional electronic map of the planting area, the distribution coordinates of bougainvillea along the road, and the terrain slope data to generate a path initialization command that includes the start point, end point, and fertilization stop points of the operation.

[0112] The system includes: a high-precision 3D electronic map containing detailed information on the planting area's terrain, roads, and vegetation, acquired using Geographic Information System (GIS) technology; Bougainvillea distribution coordinates along roads, the geographic coordinates of the bougainvillea planting locations, acquired using GPS devices or GIS technology; terrain slope data, slope information acquired through topographic surveying or GIS technology; and path initialization instructions, including path planning instructions for the start point, end point, and fertilization stops, used to guide the fertilization robot in performing its tasks.

[0113] The necessary and complete process is described below: Combining a high-precision 3D electronic map of the planting area, the coordinates of the bougainvillea distribution along the road, and terrain slope data, a path initialization command is generated, including the start point, end point, and fertilization stop points. The specific process is as follows: First, a high-precision 3D electronic map, bougainvillea distribution coordinates, and terrain slope data are integrated into a unified Geographic Information System (GIS) platform. For example, GIS software is used to match the GPS coordinates of the bougainvillea with the road network on the map, while considering the impact of terrain slope on path planning. Next, a path planning algorithm (such as the A* algorithm or Dijkstra's algorithm) is used to generate the fertilization robot's operating path. During path planning, terrain slope and bougainvillea distribution are considered to ensure the feasibility and efficiency of the path. For example, paths through steep slopes are avoided to reduce the risk of the robot climbing steep inclines. Finally, based on the path planning results, path initialization instructions containing the start point, end point, and stopping fertilization points are generated. These instructions will serve as the basis for the fertilization robot to execute its tasks, ensuring that it can complete the fertilization operation according to the predetermined path.

[0114] Step S4522: Obtain the severe weather prediction data output by the GraphLSTM model and the hardware parameters of the fertilizing robot, and analyze whether the operational constraints are met. The hardware parameters include the driving range and the maximum climbing angle. If the constraints are met, proceed to step S4523; otherwise, proceed to step S4524.

[0115] The data includes: severe weather prediction data: future weather conditions output by the GraphLSTM model, including the probability, intensity, and duration of severe weather such as cold waves and heavy rain. Fertilizer robot hardware parameters: technical parameters such as the robot's range and maximum climbing angle, obtained from technical manuals provided by the robot manufacturer or through experimental testing. Operational constraints: conditions that ensure the fertilizer robot can complete its tasks safely and efficiently, such as whether the range meets operational requirements and whether the maximum climbing angle is suitable for the terrain slope.

[0116] The necessary and complete process is described below: Obtain severe weather prediction data output by the GraphLSTM model and the hardware parameters of the fertilizing robot, and analyze whether the operational constraints are met. The specific process is as follows: First, severe weather forecast data for a future period is obtained from the GraphLSTM model, including the probability, intensity, and duration of cold waves, heavy rain, etc. For example, the model predicts a moderate cold wave that will last for two days within the next three days.

[0117] Next, the hardware parameters of the fertilizing robot were obtained, such as a range of 10 kilometers and a maximum climbing angle of 30 degrees. These parameters were obtained through technical manuals or experimental tests provided by the robot manufacturer.

[0118] Next, analyze whether these data meet the operational constraints. For example, if the total length of the operational path is 8 kilometers and the maximum slope on the path is 25 degrees, then the robot's current range and climbing ability both meet the operational requirements. However, if the predicted cold wave intensity is too high, it may affect the robot's operational safety, in which case further evaluation is needed.

[0119] If the constraints are met, step S4523 is executed to optimize the path and operation parameters using preset rules; if the constraints are not met, step S4524 is executed to adjust the task according to a preset strategy. In this way, step S4522 ensures that the fertilizing robot can complete the fertilizing task safely and efficiently.

[0120] Step S4523: The path and operation parameters are optimized using preset rules. The fertilization robot regenerates the path based on the optimized path and operation parameters and performs the fertilization operation.

[0121] The system includes: Preset rules: Optimization rules formulated based on the fertilization task requirements and robot hardware characteristics, used to adjust the path and operational parameters. Path optimization: By adjusting path planning, it ensures that the fertilization robot can complete the task efficiently while avoiding the impact of terrain obstacles and inclement weather. Operational parameters: Including parameters such as fertilization speed and fertilizer amount, which are adjusted according to plant needs and environmental conditions. Optimization algorithms: Such as genetic algorithms and simulated annealing algorithms, used to find the optimal path and operational parameters.

[0122] The complete process of this step can be found in steps S4523-1 to S4523-2, and will not be repeated here.

[0123] Step S4524: Adjust the task according to the preset strategy.

[0124] Task adjustment strategies are rules for replanning or adjusting fertilization tasks based on current conditions and constraints. These strategies are typically based on preset priorities and conditional judgments. Priority rules determine which tasks or parameters need priority adjustment when constraints are not met. For example, prioritizing path adjustment over fertilizer application. Conditional judgments are logical judgments based on real-time data (such as severe weather forecasts or robot hardware status) to decide whether to adjust the task. Dynamic adjustments dynamically modify the task plan based on real-time data to adapt to the current environment and equipment status.

[0125] The complete process of this step can be found in steps S4524-1 to S4524-3, and will not be repeated here.

[0126] The steps for setting up rules include: Step S4523-1: When the probability of a cold wave is greater than or equal to the preset risk threshold, prioritize routes that are sheltered from the wind and have sufficient sunlight in the route planning, and lock the operation time during the period with the highest average daily temperature within the fertilization time window.

[0127] Among them, the cold wave probability is predicted by the GraphLSTM model, reflecting the likelihood of a cold wave occurring in the near future. The preset risk threshold is a pre-set cold wave probability threshold used to determine whether operational strategies need adjustment. The sheltered and well-lit route is prioritized in route planning to reduce the impact of cold waves on operations. The fertilization time window is the suitable time range for fertilization determined based on plant growth needs and environmental conditions. The period with the highest average daily temperature is the period with the highest average daily temperature within the fertilization time window, typically the warmest part of the day.

[0128] The necessary complete process is described as follows: When the probability of a cold wave is greater than or equal to a preset risk threshold (e.g., 30%), the optimization mechanism is triggered. First, based on the cold wave probability output by the GraphLSTM model (e.g., 35%), it is determined that the operation strategy needs to be adjusted. Next, using high-precision 3D electronic maps and GIS technology, combined with real-time meteorological data, the path is prioritized to be sheltered from the wind and have sufficient sunlight, avoiding areas affected by the cold wave. At the same time, the operation time is locked during the period with the highest average daily temperature within the fertilization time window (e.g., 10:00 AM to 2:00 PM) to ensure a relatively warm operating environment. Finally, a composite control command containing path coordinates, fertilization parameters, and time window is generated to drive the fertilization robot to operate according to the optimized path and time.

[0129] Furthermore, step S4523-1 may also include the following steps: Step 1: Input the cold resistance physiological parameters of bougainvillea (including leaf thickness, degree of lignification of branches, etc.), combine the cold wave probability data output by the GraphLSTM model with the improved DQN algorithm, and call the historical data of fertilization effect in the same period of the near preset years as feedback, so that the risk threshold is dynamically adjusted with the growth stage of bougainvillea (the threshold in the seedling stage is lower than the preset proportion in the mature stage).

[0130] Among them, the cold resistance physiological parameters of bougainvillea include leaf thickness and the degree of lignification of branches, which are obtained through field measurements or historical data. An improved DQN algorithm, a deep reinforcement learning algorithm, is used to optimize the decision-making process. Historical fertilization effect data includes records of fertilization effects during the same period in the past, used for feedback and adjustment.

[0131] The necessary and complete process is described as follows: Input bougainvillea's cold resistance physiological parameters (such as leaf thickness and branch lignification degree), combine the cold wave probability data output by the GraphLSTM model with the improved DQN algorithm, and call historical data on fertilization effects during the same period in the past preset years as feedback to dynamically adjust the risk threshold according to the growth stage of bougainvillea. For example, the threshold in the seedling stage is lower than that in the mature stage by a preset proportion (such as 20%).

[0132] The specific process is as follows: First, the cold-resistance physiological parameters of bougainvillea were obtained through field measurements or historical data. For example, the leaf thickness was 0.5 mm and the lignification degree of the branches was 60%.

[0133] Next, the probability of a cold wave is predicted using a GraphLSTM model. For example, the model predicts a 40% probability of a cold wave within the next 3 days.

[0134] Then, historical data on fertilization effects from the same period over the past five years were used to analyze the effects of fertilization at different growth stages. For example, fertilization during the seedling stage was less effective, while fertilization during the mature plant stage was more effective.

[0135] Finally, the risk threshold was dynamically adjusted by improving the DQN algorithm and combining the above data. For example, the risk threshold was adjusted to 25% for seedlings and 30% for mature plants. In this way, the risk threshold was ensured to match the growth stage and cold resistance of bougainvillea, improving the accuracy of fertilization decisions.

[0136] Step 2: The mini weather station on the fertilizing robot monitors the ambient wind speed in real time. When the wind speed exceeds the preset level, the robot will immediately adjust the preset risk threshold ratio because strong winds will accelerate the loss of heat from the plants and increase the risk of frost damage. At the same time, the robot will control the wind deflector on the side of the robot to expand to 45° within a preset time to provide real-time environmental constraint parameters for subsequent path planning.

[0137] Among them, the mini weather station is a small meteorological monitoring device integrated into the fertilizing robot, which collects environmental parameters such as wind speed, wind direction, temperature, and humidity in real time. Wind speed rating: A rating based on meteorological standards (such as the Beaufort scale) used to determine wind intensity. Wind deflector: A foldable deflector installed on the side of the fertilizing robot to block strong winds and reduce heat loss from the plants. Risk threshold: A preset wind speed threshold used to trigger the wind deflector deployment mechanism.

[0138] The necessary and complete process is described as follows: The miniature weather station onboard the fertilizing robot monitors the ambient wind speed in real time. When the wind speed exceeds a preset level (e.g., level 6 wind), a linkage mechanism is immediately triggered. Strong winds accelerate heat loss from plants, exacerbating the risk of frost damage. Therefore, the system automatically raises the cold wave risk threshold (e.g., from 30% to 40%), while simultaneously controlling the wind deflectors on the side of the robot to expand to a 45° angle within 10 seconds, forming a windbreak. The deployment angle and timing of the wind deflectors are dynamically adjusted according to preset rules to ensure that the impact of wind damage is minimized without affecting the robot's movement. In addition, the wind deflector status is linked to the path planning system, providing real-time environmental constraint parameters for subsequent path optimization, ensuring that the robot can safely and efficiently complete the fertilization operation even in severe weather.

[0139] Step 3: Based on the risk threshold generated in Step 1 and the deployment status of the windbreak in Step 2, the fertilization robot calls high-precision three-dimensional road greening map data and uses the improved A* algorithm for path planning. The weight of the leeward coefficient in the algorithm cost function accounts for a preset proportion, giving priority to ensuring the cold resistance adaptability of the path.

[0140] The high-precision 3D road greening map data includes: a 3D map containing detailed information on roads, green belts, buildings, etc., acquired through LiDAR and remote sensing technologies. The leeward coefficient is an indicator reflecting the degree of wind influence on the route area, calculated through terrain analysis and meteorological data. The cost function is a function used to evaluate the quality of a route, incorporating factors such as distance, slope, and leeward coefficient.

[0141] The necessary complete process is described below: Based on the risk threshold generated in step 1 and the windbreak status in step 2, the fertilization robot calls high-precision 3D road greening map data and uses an improved A* algorithm for path planning. The leeward coefficient weight in the algorithm's cost function accounts for a preset proportion (e.g., 30%) to prioritize the path's cold-weather adaptability. The specific process is as follows: First, high-precision 3D map data is loaded to extract road network, green belt distribution, and terrain information. For example, the map shows that a certain road section is located on the leeward side of a mountain with a gentle slope.

[0142] Next, the A* algorithm is improved by introducing a leeward coefficient into the cost function to calculate the comprehensive cost of each path node. For example, paths with a low leeward coefficient (such as the leeward side of a mountain) have lower costs and are selected first.

[0143] Ultimately, the algorithm generates a path that balances the shortest distance with optimal cold resistance, ensuring the robot operates safely and efficiently under cold weather conditions.

[0144] Step 4: The fertilizing robot activates the lidar to scan the terrain of the work area and generate a centimeter-level 3D obstacle map. Simultaneously, it identifies the bougainvillea plant type (clump-forming / single-stem) through the visual navigation module. When the light intensity in the path area is detected to be lower than the preset value, the robot controls the retractable sunshade to slowly unfold within a preset distance. The onboard photovoltaic panel is activated simultaneously to reverse charge the robot's battery to store energy for subsequent operations.

[0145] The system includes: LiDAR (Light Detection and Ranging): A LiDAR sensor mounted on the fertilizing robot to scan the terrain of the work area in real time and generate high-precision 3D point cloud data. Centimeter-level 3D obstacle map: A detailed terrain model constructed from LiDAR point cloud data for path planning and obstacle avoidance. Visual navigation module: A computer vision-based navigation module that uses a camera to identify the bougainvillea plant type (e.g., clump-forming or single-stemmed). Retractable sunshade: An adjustable sunshade device mounted on the robot, deployed when sunlight is insufficient, and integrated with a photovoltaic panel for reverse charging. Preset light intensity: A minimum light threshold set according to the growth requirements of bougainvillea, monitored in real time by a light sensor.

[0146] The necessary complete process is described below. In step 4, the fertilizing robot activates its LiDAR to scan the work area, generating a centimeter-level 3D obstacle map. Simultaneously, it identifies the shape of the bougainvillea plants using a visual navigation module. When the light intensity in the path area is detected to be lower than a preset value (e.g., 5000 lux), the robot automatically deploys a retractable sunshade to a preset distance (e.g., 1 meter), and the onboard photovoltaic panels activate simultaneously to reverse charge the battery. For example, if the robot is working in a shady area with a light intensity of only 3000 lux, the system will trigger the sunshade to deploy and start the charging mode. Through real-time terrain scanning and adaptive light control, step 4 ensures that the robot can still operate efficiently under complex lighting conditions, while providing accurate terrain and light data support for subsequent path planning.

[0147] Step 5: Using the path data planned in Step 3 and the real-time illumination parameters collected in Step 4, combined with the hourly temperature and humidity prediction data pushed by the meteorological platform for the future preset duration, the fertilization robot uses the K-means time series clustering algorithm to lock the period with the highest daily average temperature. The clustering features include temperature fluctuation range (≤ preset temperature difference) and peak illumination intensity (≥ preset illumination value).

[0148] The algorithm includes: K-means time series clustering, an unsupervised machine learning algorithm used to divide time series data into multiple clusters based on similarity to identify the optimal application period; meteorological platform push data, which is real-time hourly temperature and humidity forecast data obtained from a meteorological service platform; clustering features, including temperature fluctuation range (≤ preset temperature difference, such as ±2℃) and peak light intensity (≥ preset light value, such as 8000 lux), used to screen suitable application periods; and the period with the highest average daily temperature, determined through cluster analysis, used to schedule fertilization operations.

[0149] The necessary complete process is described below: The fertilization robot uses the path data planned in step 3 and the real-time illumination parameters collected in step 4, combined with the hourly temperature and humidity forecast data pushed by the meteorological platform, to identify the period with the highest daily average temperature using the K-means time series clustering algorithm. The specific process is as follows: First, temperature and humidity data and light intensity data for the next 24 hours are input into the K-means model, with clustering characteristics set as temperature fluctuation ≤2℃ and peak light intensity ≥8000 lux. The model divides the data into multiple clusters, each representing a combination of environmental conditions.

[0150] Next, the clusters with the highest and lowest temperatures were selected as the optimal working hours. For example, the model identified 10 a.m. to 2 p.m. as the period with the most stable temperatures and the strongest sunlight.

[0151] Ultimately, this time period is designated as the fertilization window to ensure the robot operates under optimal environmental conditions, improving fertilization effectiveness and reducing the risk of cold waves. Step 5 utilizes K-means clustering analysis to achieve intelligent selection of the operation period.

[0152] Step 6: When the fertilizing robot predicts that the ambient temperature will drop to a preset critical value within a preset time, the built-in constant temperature heating device in the battery compartment will start 5 minutes in advance to ensure the battery life; the dual-channel timing module monitors the temperature change during the buffer period in real time. If the temperature drop exceeds the preset temperature difference, the nozzle recovery mechanism will be triggered to retract the fertilizer spraying component to the protective position and control the robot to stop in an emergency within a preset safe distance, providing an accurate time reference for subsequent differentiated operations.

[0153] The system includes: Ambient temperature prediction: based on real-time meteorological platform data and the future temperature trend output by the GraphLSTM model. Preset critical value: a set low-temperature threshold (e.g., 5℃) to trigger the protection mechanism. Constant temperature heating device: a heating module built into the battery compartment that automatically activates to maintain the battery's operating temperature. Dual-channel timing module: used to monitor temperature changes in real time during the buffer period to ensure timely response. Sprayer head recovery mechanism: automatically retracts the fertilizer spraying components to the protective position to prevent freezing damage. Emergency shutdown mechanism: automatically stops operation and evacuates to a safe distance in the event of a sudden temperature drop or equipment malfunction.

[0154] The necessary and complete process is described below: The fertilizing robot continuously monitors changes in ambient temperature. When it is predicted that the ambient temperature will drop below 5°C within 30 minutes, the system immediately activates the low-temperature warning mechanism. The constant-temperature heating device in the battery compartment starts 5 minutes in advance to ensure stable battery operation. Simultaneously, the dual-channel timing module monitors the rate of temperature change in real time. If a temperature drop exceeding 3°C / 10 minutes is detected, the nozzle recovery mechanism is immediately triggered to retract the fertilizer spraying components to the protective position, and the robot is controlled to stop urgently within a 10-meter safe distance. For example, if the temperature suddenly drops from 8°C to 4°C, the system will automatically execute the above protective measures.

[0155] Step 7: Based on the operating period and emergency stop mechanism determined in Step 6, the fertilization robot binds the path coordinates, time window and cold resistance level label of the bougainvillea variety, and performs graded control for weak cold resistance varieties: first, the traveling speed is reduced to a preset ratio of the rated speed, and after the machine body is stable, the height of the fertilizer spray nozzle is lowered to a preset distance (close to the root absorption area).

[0156] Among these features are: Cold Resistance Rating Labels: Grades are assigned based on the cold resistance of bougainvillea varieties, obtained through historical data or expert experience. Graded Control: Differentiated operational strategies are implemented for varieties with different cold resistance ratings. Travel Speed ​​Adjustment: The robot's speed is reduced to minimize mechanical damage to the plants. Fertilizer Nozzle Height Adjustment: The nozzle height is lowered to be close to the root absorption zone, improving nutrient utilization efficiency.

[0157] The necessary complete process is described below: Based on the operating period and emergency stop mechanism determined in step 6, the fertilization robot binds the path coordinates and time window to the cold resistance level label of the bougainvillea variety, and performs graded control for weakly cold-resistant varieties. First, the robot identifies the cold resistance level of the bougainvillea varieties in the current operating area. If it is a weakly cold-resistant variety (such as seedlings or tropical varieties), it automatically reduces its travel speed to 70% of the rated speed to reduce the impact of mechanical vibration on the plants. After the robot stabilizes, the height of the fertilizer nozzle is lowered to 10 cm above the ground, close to the root absorption area, to ensure rapid nutrient absorption. For example, for bougainvillea varieties with a "low" cold resistance level, the robot speed is reduced from 1.5 m / s to 1.0 m / s, and the fertilizer nozzle height is reduced from 30 cm to 10 cm.

[0158] Step 8: The fertilization robot uses an infrared thermal imager to monitor the plant surface temperature in real time. When the temperature is detected to be lower than the preset threshold, the detachable nozzle is automatically switched to a high flow mode because the plant needs extra nutrients to resist the cold wave in a low-temperature environment. This increases the amount of fertilizer sprayed by a preset ratio compared to conventional operations, thus achieving targeted nutrient replenishment. Among them, the temperature threshold is the set minimum surface temperature of the plant (e.g., 4℃). Below this value, a mechanism to increase the fertilizer application rate is triggered. The detachable nozzle is a modular nozzle mounted on the robot, supporting flow rate mode switching. The high flow rate mode operates with the nozzle at a high flow rate, increasing the fertilizer application rate by a preset percentage (e.g., 30%) compared to the regular mode.

[0159] The necessary complete process is described as follows: In step 8, the fertilization robot monitors the surface temperature of the bougainvillea plants in real time using an infrared thermal imager. When the detected temperature is below a preset threshold (e.g., 4°C), the system automatically determines that the plant is under low-temperature stress and requires additional nutrients to resist the cold wave. At this time, the robot controls the detachable nozzle to switch to a high-flow mode, increasing the amount of fertilizer sprayed by 30% compared to conventional operations. For example, the amount of fertilizer sprayed per plant is 50 ml in conventional mode, but increases to 65 ml in low temperatures. By dynamically adjusting the amount of fertilizer sprayed, step 8 ensures that the plants receive sufficient nutrients in low-temperature environments, enhancing their cold resistance and improving the accuracy and effectiveness of fertilization operations.

[0160] Step S4523-2: When the soil moisture content is greater than the preset ratio of field water holding capacity, the fertilization robot's travel speed is automatically reduced to the preset ratio of the rated speed, the fertilizer spraying pressure is reduced by the preset ratio, and a composite control command containing path coordinates, fertilization parameters, and time window is generated to drive the robot to work along the preset path.

[0161] Among them, soil moisture content: the soil moisture content monitored in real time by a soil moisture sensor. Field holding capacity preset ratio: a pre-set soil moisture content threshold used to determine whether the soil is too wet. Rated speed: the maximum travel speed of the fertilizing robot during normal operation. Fertilizer spraying pressure: the pressure setting when the fertilizing robot sprays fertilizer, affecting the distribution and coverage of the fertilizer. Composite control instructions: comprehensive instructions containing path coordinates, fertilization parameters, time windows, etc., used to drive the fertilizing robot to perform its tasks.

[0162] The necessary and complete process is described below: When the soil moisture content exceeds the preset proportion of field capacity, the fertilization robot automatically adjusts its travel speed and spraying pressure to adapt to the moist soil conditions. The specific process is as follows: First, soil moisture content is monitored in real time using soil moisture sensors. If the soil moisture content exceeds a preset proportion of field capacity (e.g., 80%), an operational parameter adjustment mechanism is triggered.

[0163] Next, the fertilization robot's travel speed is automatically reduced to a preset percentage of its rated speed (e.g., to 60%) to reduce the risk of slipping due to excessively wet soil. At the same time, the spraying pressure is reduced by a preset percentage (e.g., to 70%) to prevent excessive fertilizer concentration or runoff.

[0164] If multiple fertilizing robots are working collaboratively, the system will synchronously adjust the parameters of all robots to ensure consistency and efficiency in the operation. For example, through a wireless communication module, the main control system sends the adjusted parameters to each robot to ensure that they can fertilize at an optimized speed and pressure even under humid conditions.

[0165] Finally, a composite control command containing path coordinates, fertilization parameters, and a time window is generated to drive the fertilization robot to operate along the preset path and with adjusted parameters. In this way, step S4523-2 ensures that the fertilization robot can safely and efficiently complete the fertilization task under moist soil conditions, while reducing fertilizer waste and adverse effects on the soil.

[0166] The steps for setting up a strategy include: Step S4524-1: When the operation time exceeds the limit, the fertilization task will be split into at most 2 sub-tasks, with an interval of ≥1 hour between sub-tasks.

[0167] Among these, the following are considered errors: **Operating Time Exceeded:** The estimated completion time of the fertilization task exceeds the maximum allowable time for a single operation by the fertilization robot. **Subtasks:** The original fertilization task is broken down into multiple shorter subtasks so that the robot can complete them in stages. **Subtask Interval:** The time interval between two subtasks, used to ensure the robot has sufficient time for charging or other maintenance operations. **Preset Rules:** Splitting rules based on the robot's hardware characteristics and task requirements, used to determine the number and intervals of subtasks.

[0168] The necessary complete process is described below: When the estimated completion time of the fertilization task exceeds the robot's maximum allowable operating time, the fertilization task is split into at most two sub-tasks, ensuring that the interval between sub-tasks is ≥1 hour. The specific process is as follows: First, based on the fertilization robot's maximum allowed operating time (e.g., 8 hours) and the estimated total task time (e.g., 10 hours), determine whether the task needs to be split. If the task time exceeds the limit, the task splitting mechanism is triggered.

[0169] Next, the fertilization task is broken down into two sub-tasks according to preset rules. For example, the task can be divided into two 5-hour sub-tasks. Simultaneously, the interval between sub-tasks is set to be at least 1 hour to ensure the robot has sufficient time for charging or other maintenance operations. For instance, after completing the first sub-task, the robot returns to the charging point to charge for 1 hour before continuing with the second sub-task.

[0170] If multiple fertilizing robots are working collaboratively, the system will dynamically allocate sub-tasks based on the status and task requirements of each robot. For example, the main control system uses a wireless communication module to assign the split sub-tasks to different robots, ensuring that the tasks can be completed efficiently.

[0171] Finally, a composite control command containing subtask paths, fertilization parameters, and time windows is generated to drive the fertilization robot to execute subtasks according to preset paths and times.

[0172] Step S4524-2: When the path length exceeds the range, replan the path including temporary charging points based on the preset charging pile location data.

[0173] Among these, the path length exceeds the robot's single-charge range: The path length for the fertilization task exceeds the robot's single-charge range. Pre-set charging station location data: Pre-defined charging station location information, obtained through a Geographic Information System (GIS) or map data. Temporary charging points: Charging stations added to the path planning for the robot to recharge mid-journey. Path planning algorithm: Such as the A* algorithm or Dijkstra's algorithm, used to calculate the optimal path, taking into account the addition of charging points.

[0174] The necessary complete process is described below: When the path length of the fertilization task exceeds the single-charge range of the fertilization robot, the path, including temporary charging points, is replanned based on the preset charging station location data. The specific process is as follows: First, the location data of preset charging stations is obtained through a GIS system. This data, combined with the path length of the fertilization task and the robot's range, determines whether temporary charging points need to be added. For example, if the fertilization robot's range is 10 kilometers and the task path length is 15 kilometers, then charging points need to be added along the path.

[0175] Next, a path planning algorithm (such as the A* algorithm) is used to replan the route, ensuring that it includes temporary charging points. For example, the distance from the starting point to the first charging point is 5 kilometers, and after charging, the remaining 10 kilometers are traveled to the destination. The path planning algorithm considers terrain, road conditions, and charging point locations to generate the optimal route.

[0176] If multiple fertilizing robots are working together, the system will dynamically allocate paths based on the status and task requirements of each robot. For example, the main control system uses a wireless communication module to assign paths containing charging points to the robots, ensuring that the task can be completed efficiently.

[0177] Finally, a composite control command containing path coordinates, charging point locations, and fertilization parameters is generated to drive the fertilization robot to execute the task according to the preset path and time. In this way, step S542 ensures that the fertilization robot can safely and efficiently complete the fertilization task by planning a reasonable charging path, even when the path length exceeds its range.

[0178] Step S4524-3: When the terrain slope exceeds the limit, mark the area as an area to be supplemented and generate a geofence, skipping the immediate operation.

[0179] Among these, the following issues are identified: **Excessive Terrain Slope:** The terrain slope exceeds the fertilization robot's maximum climbing angle, preventing the robot from safely traversing the slope. **Areas Requiring Re-fertilization:** Areas where immediate fertilization is impossible due to excessively steep terrain slopes, requiring subsequent re-fertilization. **Geofencing:** Virtual boundaries defined using Geographic Information System (GIS) technology to mark specific areas. **Marking and Skipping Mechanism:** Areas requiring re-fertilization are identified and marked during path planning, allowing for skipping immediate operations.

[0180] The necessary complete process is described below: When there are areas in the fertilization task path where the terrain slope exceeds the fertilization robot's maximum climbing angle, these areas are marked as areas to be fertilized and geofences are generated, skipping the immediate task. The specific process is as follows: First, by using a high-precision 3D electronic map and terrain slope data, combined with the fertilization robot's maximum climbing angle (e.g., 30 degrees), areas with excessively steep slopes along the path are identified. For example, if an area has a slope of 35 degrees, exceeding the robot's climbing ability, that area is identified as an area requiring additional fertilization.

[0181] Next, geographic information system (GIS) technology is used to generate geofences for the areas to be supplemented, clearly defining their boundaries. Simultaneously, these areas are skipped during path planning, ensuring that the robot avoids excessively steep sections during real-time operations, thus mitigating operational risks caused by terrain.

[0182] For marked areas requiring fertilization, the system records their location and extent, and schedules dedicated fertilization tasks later. For example, fertilization can be carried out in the area after weather conditions improve or a suitable robot for climbing slopes is replaced.

[0183] Based on the same inventive concept, embodiments of the present invention provide an intelligent fertilization monitoring system for roadside bougainvillea, including a memory and a processor, wherein the memory stores information that can run on the processor to implement the following... Figures 1 to 2 The procedure for any method.

[0184] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for intelligent monitoring of fertilization application for roadside bougainvillea, characterized in that, include: In the bougainvillea planting area, soil composition parameters, multispectral image data and thermal imaging image data were collected simultaneously, and the corresponding GPS coordinates, timestamps and growth stage information were recorded. The collected multispectral and thermal image data are precisely matched with GPS coordinates. Based on the timestamp, the soil composition parameters are aligned with the corresponding multispectral and thermal image data to form a spatiotemporal dataset. Convolutional neural networks were used to extract features from multispectral images. Combined with the temperature field distribution of thermal imaging data, the health status, leaf density and distribution of bougainvillea canopy were identified. Long short-term memory networks were used to analyze the time series of soil composition parameters and predict the trend of soil fertility change. The image analysis results and soil parameter prediction results were fused through an attention mechanism to construct a comprehensive analysis model of growth posture. Based on historical growth data and the growth pattern of bougainvillea, nutritional requirement standards corresponding to different growth postures are set. The results output by the growth posture comprehensive analysis model, i.e. the quantitative value of the current growth status, are compared with the preset standards to calculate the amount of fertilizer deficiency. Combined with the current soil fertility status, the precise fertilizer formula and fertilization time are determined. Based on the fertilizer formula and time schedule, control instructions are generated, and the fertilization robot performs fertilization operations based on the control instructions and preset paths.

2. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 1, characterized in that, The image analysis results and soil parameter prediction results are fused with an attention mechanism to construct a comprehensive growth posture analysis model, including: Establish a joint decision matrix for the environment and growth stage, using external environmental parameters and growth stage status as input variables; By using a self-attention mechanism, based on the combined state output by the joint decision matrix, the fusion coefficient of multispectral and thermal imaging spatial features and soil parameter time series features is calculated, and the weights of each feature are dynamically adjusted based on different combinations of environment and growth stage. The linkage between the growth stage gating unit and the environmental parameter response module triggers feature enhancement or suppression logic for specific environment-growth stage combinations. The image and soil parameter features after collaborative fusion, weight adjustment and feature enhancement and suppression are input into the model architecture. After integration by fully connected layers and ReLU function operation, a comprehensive growth posture analysis model is constructed, and the output includes a quantitative vector of growth status containing health index, nutrient stress level and water stress level.

3. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 1, characterized in that, Determining the precise fertilizer formula and timing includes: Obtain microclimate data, chlorophyll fluorescence parameters, soil parameters, growth stage information, and historical fertilization data for the bougainvillea planting area; The above data is input into a pre-constructed four-dimensional coupled model of environment-physiology-soil-growth stage. The data is fused by improving the GBDT algorithm, and the nitrogen, phosphorus and potassium absorption efficiency correction coefficient is output. Based on the correction coefficient, the balance equation containing the reinforcement learning module is used, combined with the growth stage sensitivity coefficient, environmental correction factor and soil texture correction term, to calculate the nitrogen, phosphorus and potassium nutrient deficiency. Using nutrient deficiency as input, a multi-objective optimization model is constructed by improving the NSGA-Ⅲ algorithm, and the optimal fertilization formula is generated by combining soil urease activity and rhizosphere microbial diversity constraints. The GraphLSTM model was used to predict soil moisture content, daily average temperature and cold wave probability. Combined with fertilizer characteristics and growth stage time window, the dynamic time window algorithm was used to determine the fertilization time node. After fertilization, data on the growth status of bougainvillea are collected and fed back to the reinforcement learning module to optimize model parameters, forming a closed-loop regulation.

4. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 3, characterized in that, Generating optimal fertilizer formulations includes: Soil urease activity and rhizosphere microbial diversity index were collected and constraint thresholds were set to normalize the nitrogen, phosphorus and potassium nutrient deficiencies. If the detected value exceeds the constraint threshold, the dynamic reference point adjustment mechanism of the improved NSGA-Ⅲ algorithm is triggered, and the reference point density of the feasible solution region is dynamically increased according to the proportion of nutrient deficiency. An optimization model is constructed with the objectives of minimizing fertilization costs, maximizing nutrient replenishment efficiency, and minimizing soil ecological impact. The normalized missing values ​​are input into the model, and the Pareto optimal solution set is generated by introducing a non-dominated ranking based on constraint priority and an elite strategy that retains the optimal individuals with a pre-set proportion. The TOPSIS method is used to calculate the closeness between each solution set individual and the ideal solution. The solution with the highest closeness is selected as the optimal fertilization formula, and the nitrogen, phosphorus and potassium ratio and fertilization amount are output.

5. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 4, characterized in that, The density of reference points in the feasible solution region is dynamically increased according to the proportion of nutrient deficiency, including: Calculate the proportion of nitrogen, phosphorus, and potassium nutrient deficiencies in the total amount to identify the dominant nutrient with the largest current deficiency. When soil urease activity is below a preset threshold or the rhizosphere microbial diversity index fails to meet the standard, a dynamic densification mechanism is activated; according to the proportion of each nutrient deficiency, the number of new reference points is allocated to the sub-regions of the corresponding nutrient dimension in the three-dimensional target space. For the target dimension corresponding to the dominant missing nutrient, the density of reference points is increased, with the increase being proportional to the proportion of the missing nutrient; for the non-dominant nutrient dimension, the density of reference points is decreased accordingly, with the decrease being inversely proportional to the proportion of the missing nutrient. Finally, an encrypted set of reference points is generated, so that the distribution of reference points in the feasible solution region matches the proportion of nutrient deficiency.

6. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 3, characterized in that, Using the GraphLSTM model to predict soil moisture content, daily average temperature, and cold wave probability, and combining fertilizer characteristics with growth stage time windows, a dynamic time window algorithm is used to determine fertilization time nodes, including: Real-time collection of microclimate, soil parameters and severe weather remote sensing data; identification of weather types and extraction of occurrence probability and duration through convolutional neural network; spatiotemporal alignment with bougainvillea growth stage information and historical fertilization data. In the graph structure of the GraphLSTM model, a severe weather node is constructed, and its association edges with soil moisture content and daily average temperature nodes are established. The influence weights of severe weather on soil moisture content, daily average temperature and cold wave probability are learned through the attention mechanism. After the model is pre-trained using historical data, the predicted values ​​of environmental parameters are output. The algorithm incorporates a severe weather risk assessment factor into the dynamic time window algorithm. Based on the predicted weather intensity and duration, combined with fertilizer dissolution rate, nutrient release cycle and critical growth period time window, the start and end boundaries of the fertilization time window are dynamically adjusted to determine the fertilization time nodes that meet the fertilizer characteristics and bougainvillea growth requirements.

7. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 6, characterized in that, In the graph structure of the GraphLSTM model, a severe weather node is constructed, and its association edges with soil moisture content and daily average temperature nodes are established. An attention mechanism is used to learn the influence weights of severe weather on soil moisture content, daily average temperature, and cold wave probability. After pre-training the model using historical data, the predicted environmental parameters output include: By combining a high-precision 3D electronic map of the planting area, the distribution coordinates of bougainvillea along the road, and the terrain slope data, a path initialization command is generated that includes the start point, end point, and fertilization stop points of the operation. Obtain severe weather prediction data and fertilizer robot hardware parameters from the GraphLSTM model output, and analyze whether they meet the operational constraints. The hardware parameters include driving range and maximum climbing angle. If the constraints are met, the path and operation parameters are optimized using preset rules, and the fertilization robot regenerates the path based on the optimized path and operation parameters to perform the fertilization operation. If the constraints are not met, the task will be adjusted according to the preset strategy.

8. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 7, characterized in that, The preset rules include: When the probability of a cold wave is greater than or equal to the preset risk threshold, the route planning should prioritize routes that are sheltered from the wind and have sufficient sunlight, and the operation time should be locked during the period with the highest average daily temperature within the fertilization time window. When the soil moisture content is greater than the preset ratio of field water holding capacity, the fertilization robot's travel speed is automatically reduced to the preset ratio of the rated speed, the fertilizer spraying pressure is reduced by the preset ratio, and a composite control command containing path coordinates, fertilization parameters, and time window is generated to drive the robot to work along the preset path.

9. The intelligent fertilization monitoring method for roadside bougainvillea according to claim 7, characterized in that, Preset strategies include: If the work time exceeds the limit, the fertilization task will be split into a maximum of 2 sub-tasks, with an interval of ≥1 hour between sub-tasks; When the route length exceeds the range, the route including temporary charging points will be replanned based on the preset charging station location data. When the terrain slope exceeds the limit, mark the area as an area to be supplemented and generate a geofence to skip the immediate operation.

10. A smart fertilization monitoring system for roadside bougainvillea, characterized in that, It includes a memory, a processor, and a program stored in the memory and executable on the processor, which, when loaded and executed by the processor, implements the intelligent fertilization monitoring method for road bougainvillea as described in any one of claims 1 to 9.

Images