Intelligent honey pomelo fertilization decision-making system and method based on multi-source data fusion

Abstract

The invention relates to the technical field of agricultural resource management and optimization decision making, in particular to a honey pomelo intelligent fertilization decision making system and method based on multi-source data fusion, and the system comprises a multi-source data sensing module, a space-time fusion and feature generation module, a dynamic partition decision making engine module and a decision making execution and interaction module. The method comprises the following steps: acquiring soil physicochemical properties with geographic coordinates, a crown scale remote sensing spectrum and meteorological environment multi-source spatio-temporal data, mapping the data to a gridding geographic information system unit through spatio-temporal alignment, generating a collaborative diagnosis feature vector by combining agronomic knowledge model fusion analysis, and inputting the collaborative diagnosis feature vector into a prediction model to obtain an initial nutrient demand quantity; and then integrating multi-model and constraint condition multi-target optimization, generating a partition variable fertilization prescription, finally converting the partition variable fertilization prescription into an intelligent fertilization machinery instruction and a visual farming scheme, and correcting model parameters according to fertilization effect feedback data. According to the invention, deep fusion of multi-source data is realized, and the accuracy and scientificity of fertilization decision making are improved.

Description

Technical Field

[0001] This invention relates to the field of agricultural resource management and optimization decision-making technology, specifically to an intelligent fertilization decision-making system and method for pomelos based on multi-source data fusion. Background Technology

[0002] Existing models and systems related to pomelo fertilization decision-making lack dynamic self-optimization capabilities. The parameters of nutrient requirement prediction models are mostly fixed and not dynamically corrected based on the actual effects after fertilization. Deviations between model predictions and actual planting conditions cannot be identified and corrected in a timely manner. As the soil environment and tree growth status in the planting area dynamically change, the model's prediction accuracy gradually decreases, reducing the reliability of subsequent fertilization decisions. Furthermore, existing fertilization decisions fail to adequately consider the multi-objective synergy of fertilization costs, fruit quality, yield improvement, and soil nutrient cycling protection. Some decisions prioritize yield increase while neglecting cost control, while others focus excessively on cost at the expense of fruit quality. Moreover, soil nutrient balance and environmental maintenance are rarely incorporated as rigid constraints into the decision-making system, easily leading to an imbalance between economic and ecological benefits.

[0003] The current implementation of fertilization decisions for pomelos suffers from inconsistencies. Some decision-making systems generate abstract fertilization plans that lack compatibility with intelligent fertilization machinery, failing to directly translate them into executable machine instructions. This necessitates manual interpretation and conversion, reducing efficiency and increasing the risk of parameter deviations during conversion, thus hindering the accurate implementation of fertilization decisions. Furthermore, the low level of visualization in these plans fails to provide intuitive and easy-to-understand agricultural guidance for growers unfamiliar with intelligent equipment. This results in a high barrier to entry for precision fertilization technology, hindering its widespread adoption among pomelo growers and limiting its large-scale application in the pomelo cultivation sector.

[0004] In summary, the current field of pomelo fertilization management suffers from numerous problems, including extensive fertilization methods, simplistic data application, fixed decision-making models, lack of multi-objective consideration, and inefficient implementation. Existing technologies are no longer sufficient to meet the demands of pomelo cultivation towards precision, intelligence, and green development. There is an urgent need to develop an intelligent pomelo fertilization decision-making system and method that can integrate multi-source spatiotemporal data, achieve collaborative analysis of soil and tree environment, consider multi-objective optimization, possess dynamic self-correction capabilities, and efficiently connect decision-making and execution. This would address the various pain points in current pomelo fertilization management and promote the high-quality development of the pomelo cultivation industry. Summary of the Invention

[0005] The purpose of this invention is to provide a smart fertilization decision system and method for pomelos based on multi-source data fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A smart fertilization decision-making system for pomelos based on multi-source data fusion includes:

[0008] The multi-source data sensing module is used to acquire spatiotemporal data of pomelo planting areas, including soil physicochemical property data with geographic coordinate information, canopy-scale remote sensing spectral data, and meteorological environment data.

[0009] The spatiotemporal fusion and feature generation module, which is connected to the multi-source data sensing module, is used for:

[0010] Soil physicochemical property data, canopy-scale remote sensing spectral data, and meteorological environmental data are spatiotemporally aligned and mapped to a unified gridded geographic information system unit.

[0011] Based on the agronomic knowledge model, the aligned data within the same geographic information system unit are fused and analyzed to generate a collaborative diagnostic feature vector that reflects at least the soil's basic fertilization capacity, the tree's real-time nutritional status, and the influence coefficient of environmental factors.

[0012] The dynamic partitioning decision engine, whose communication connection is with the spatiotemporal fusion and feature generation module, is used for:

[0013] Receive the collaborative diagnostic feature vector and input it into the pre-trained pomelo fertilizer requirement prediction model to obtain the initial nutrient requirements of each unit;

[0014] By integrating the fertilization cost model, quality and yield objective function and nutrient cycling constraints, and using gridded geographic information system units as the basic decision-making units, the initial nutrient demand is optimized in multiple objectives to generate a fertilization prescription with zonal variables including fertilizer type, precise dosage and recommended operation time.

[0015] Among them, the pomelo fertilizer requirement prediction model performs parameter self-correction based on the effect feedback data corresponding to historical fertilizer prescriptions and the updated collaborative diagnostic feature vector;

[0016] The decision execution and interaction module, which is connected to the dynamic partition decision engine, is used to convert the partition variable fertilizer prescription into an instruction set or a visualized agricultural guidance scheme that can be parsed and executed by intelligent fertilizer machinery and then output.

[0017] A smart fertilization decision-making method for pomelos based on multi-source data fusion, the method being executed by a smart fertilization decision-making system for pomelos based on multi-source data fusion, includes the following steps:

[0018] Acquire spatiotemporal data of pomelo planting areas, including soil physicochemical property data with geographic coordinate information, tree canopy-scale remote sensing spectral data, and meteorological environmental data;

[0019] The acquired soil physicochemical property data, canopy-scale remote sensing spectral data, and meteorological environmental data are spatiotemporally aligned and mapped to a unified gridded geographic information system unit.

[0020] Based on the agronomic knowledge model, the aligned data within the same gridded geographic information system unit are fused and analyzed to generate a collaborative diagnostic feature vector that reflects at least the soil's basic fertilization capacity, the tree's real-time nutritional status, and the influence coefficient of environmental factors.

[0021] The collaborative diagnostic feature vector is input into the pre-trained pomelo fertilizer requirement prediction model to obtain the initial nutrient requirements of each gridded geographic information system unit.

[0022] By integrating the fertilization cost model, quality and yield objective function and nutrient cycling constraints, and using gridded geographic information system units as the basic decision-making units, the initial nutrient demand is optimized in multiple objectives to generate a fertilization prescription with zonal variables including fertilizer type, precise dosage and recommended operation time.

[0023] The partitioned variable fertilizer prescription is converted into an instruction set or a visualized agricultural guidance scheme that can be parsed and executed by intelligent fertilizer application machinery and then output.

[0024] In particular, after the fertilization operation is completed, the parameters of the pomelo fertilizer requirement prediction model are self-calibrated based on the feedback data of the effect of historical fertilization prescriptions and the updated collaborative diagnostic feature vector.

[0025] As can be seen from the technical solution provided by the present invention above, the beneficial effects of the intelligent fertilization decision-making system and method for pomelos based on multi-source data fusion provided by the present invention are:

[0026] This invention integrates soil physicochemical property data with geographic coordinate information, tree canopy-scale remote sensing spectral data, and meteorological environmental data. Through spatiotemporal alignment, it maps various types of data to a unified gridded geographic information system unit. Combined with agronomic knowledge models, it completes the fusion analysis of multi-dimensional data, accurately explores the soil's basic fertilization capacity, the tree's real-time nutritional status, and the influence of environmental factors within different grid units. This breaks through the limitations of traditional fertilization relying on single data and uniform fertilization across the entire area, enabling fertilization decisions to accurately match the spatial heterogeneity of pomelo planting areas and meet the actual growth needs of pomelos in different regions.

[0027] This invention uses a gridded geographic information system unit as the decision-making unit, integrates a fertilization cost model, a quality and yield objective function, and nutrient cycling constraints to carry out multi-objective optimization solutions. Under the premise of satisfying soil nutrient balance, nutrient loss control, and agronomical safe application, it balances the need to minimize the total fertilization cost and maximize the comprehensive score of fruit quality and yield through Pareto optimal front analysis. This not only effectively reduces the fertilization input cost in the pomelo planting process, but also specifically improves the quality and yield of pomelos, thereby optimizing and improving the economic benefits of pomelo planting.

[0028] The pomelo nutrient requirement prediction model in this invention can perform parameter self-correction based on feedback data of the effect after fertilization and updated collaborative diagnostic feature vectors. Combined with the actual state of the tree, soil and environment after fertilization, it quantifies the model prediction deviation and iteratively updates the internal parameters to continuously improve the accuracy of the model's nutrient requirement prediction. This design enables the fertilization decision system to form a complete closed loop of prediction, decision, execution, feedback and correction, which can adapt to the dynamic changes of soil, tree and environment in the pomelo planting area and ensure the long-term stability of the fertilization decision effect.

[0029] This invention converts zonal variable fertilization prescriptions into a set of drive instructions that can be directly parsed and executed by intelligent fertilization machinery through a decision execution and interaction module. This automates and standardizes pomelo fertilization operations, replacing the traditional manual experience-based fertilization mode and significantly improving agricultural efficiency. At the same time, the module generates a visualization scheme that includes a differentiated fertilizer dosage distribution map and agricultural guidance cards. This transforms professional grid-based fertilization decisions into intuitive and easy-to-understand operation guidelines, adapting to the needs of growers with different planting skill levels and effectively expanding the practical application coverage of precision fertilization technology in pomelo cultivation.

[0030] This invention enables precise fertilization based on the soil's basic fertilization capacity and the tree's real-time nutrient balance, determining the fertilizer type and application rate as needed, thus avoiding fertilizer waste caused by indiscriminate fertilization from the source. At the same time, by setting fertilization boundaries through nutrient cycling constraints, it balances soil nutrient balance and fertilizer effectiveness, reducing nutrient loss and soil pollution caused by excessive fertilization. While improving the economic benefits of pomelo cultivation, it also protects the soil ecological environment of the planting area, promoting the green, ecological, and sustainable development of pomelo cultivation.

[0031] This invention comprehensively records and stores various types of raw data, processing results, and operational information throughout the entire process of multi-source data perception, spatiotemporal fusion analysis, fertilizer requirement prediction, fertilizer prescription generation, and decision implementation. All fertilization decisions, from their formulation to their execution, can be fully traced. When problems arise during fertilization, staff can quickly pinpoint the source and cause of the problem based on the stored data, providing detailed data support for system maintenance optimization and fertilization effect analysis, thus achieving refined and standardized management of pomelo fertilization.

[0032] This invention deeply integrates geographic information systems, remote sensing technology, machine learning, multi-objective optimization, and other technologies with agronomical knowledge of pomelo cultivation to construct a full-process intelligent fertilization decision-making system from data collection and analysis to decision generation and implementation. This system automates and digitizes pomelo fertilization management. It replaces the cumbersome process of manual sampling, analysis, and plan formulation in traditional fertilization, significantly improving the overall management efficiency of pomelo cultivation. It provides practical reference for agricultural resource management and optimization decision-making, and promotes the intelligent upgrading of the fruit tree planting field. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the structure of a pomelo intelligent fertilization decision system based on multi-source data fusion according to the present invention;

[0034] Figure 2 This is a schematic diagram of the steps of a smart fertilization decision-making method for pomelos based on multi-source data fusion according to the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0036] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0037] like Figure 1-2 As shown, this embodiment of the invention provides a smart fertilization decision-making system for pomelos based on multi-source data fusion, comprising:

[0038] The multi-source data sensing module is used to acquire spatiotemporal data of pomelo planting areas, including soil physicochemical property data with geographic coordinate information, canopy-scale remote sensing spectral data, and meteorological environment data.

[0039] The spatiotemporal fusion and feature generation module, which is connected to the multi-source data sensing module, is used for:

[0040] Soil physicochemical property data, canopy-scale remote sensing spectral data, and meteorological environmental data are spatiotemporally aligned and mapped to a unified gridded geographic information system unit.

[0041] Based on the agronomic knowledge model, the aligned data within the same geographic information system unit are fused and analyzed to generate a collaborative diagnostic feature vector that reflects at least the soil's basic fertilization capacity, the tree's real-time nutritional status, and the influence coefficient of environmental factors.

[0042] The dynamic partitioning decision engine, whose communication connection is with the spatiotemporal fusion and feature generation module, is used for:

[0043] Receive the collaborative diagnostic feature vector and input it into the pre-trained pomelo fertilizer requirement prediction model to obtain the initial nutrient requirements of each unit;

[0044] By integrating the fertilization cost model, quality and yield objective function and nutrient cycling constraints, and using gridded geographic information system units as the basic decision-making units, the initial nutrient demand is optimized in multiple objectives to generate a fertilization prescription with zonal variables including fertilizer type, precise dosage and recommended operation time.

[0045] Among them, the pomelo fertilizer requirement prediction model performs parameter self-correction based on the effect feedback data corresponding to historical fertilizer prescriptions and the updated collaborative diagnostic feature vector;

[0046] The decision execution and interaction module, which is connected to the dynamic partition decision engine, is used to convert the partition variable fertilizer prescription into an instruction set or a visualized agricultural guidance scheme that can be parsed and executed by intelligent fertilizer machinery and then output.

[0047] In this embodiment, the multi-source data perception module is the core foundation for data acquisition in the pomelo intelligent fertilization decision-making system based on multi-source data fusion. It provides a standardized data source with accurate and comprehensive geographic coordinate information for the system's subsequent data analysis and fertilization decision-making through multi-dimensional spatiotemporal data collection and professional processing.

[0048] The multi-source data sensing module is primarily responsible for acquiring spatiotemporal data of the pomelo planting area. Its core components include three types of data: soil physicochemical properties data with geographic coordinates, tree-canopy-scale remote sensing spectral data, and meteorological environmental data. The module dynamically analyzes the basic geographic and production data of the pomelo planting area to scientifically determine the set of geographic coordinates for soil sampling points. Using this as the core benchmark, it conducts laboratory physicochemical analysis of soil samples, extracts remote sensing spectral information at the pomelo tree canopy scale, and performs meteorological data fusion and interpolation processing for the planting area. This ensures accurate matching of all collected data with geographic coordinates and completes the gridding of meteorological environmental data. Finally, it outputs three types of standardized spatiotemporal data with geographic coordinate information, providing unified and standardized input data for subsequent spatiotemporal fusion and feature generation modules. This ensures the scientific, accurate, and spatially adaptable nature of pomelo fertilization decisions from the data source. The multi-source data sensing module is used to acquire spatiotemporal data of the pomelo planting area, specifically including:

[0049] Unit for determining the coordinates of sampling points:

[0050] Basic data retrieval and analysis: Real-time retrieval of historical yield maps and topographic maps of pomelo planting areas, while reading the system's preset soil sampling density parameters, to comprehensively analyze the historical yield distribution characteristics and topographic features, clarify the spatial difference patterns of soil and tree growth within the planting area, and provide data support for the layout of sampling points;

[0051] Sampling point coordinate calculation: Based on the spatial difference patterns of the planting area obtained from the analysis, combined with the preset sampling density, the geographic coordinates of the soil sampling points that can accurately reflect the regional characteristics are calculated through the geospatial analysis algorithm, taking into account the representativeness and uniformity of the sampling points, and avoiding data deviation caused by unreasonable layout of sampling points.

[0052] Coordinate set output: The calculated geographic coordinates of all soil sampling points are integrated to generate a standardized set of geographic coordinates of soil sampling points and output it. This set of coordinates becomes the geographic benchmark for subsequent data collection, ensuring the consistency of various types of data in geographic space.

[0053] Soil physicochemical property data acquisition unit:

[0054] Soil sample collection: Based on the coordinates of the sampling points, the set of geographic coordinates output by the unit is determined. Soil samples are collected in the corresponding geographic locations of the pomelo planting area. The agronomic specifications for soil sampling are strictly followed to ensure the representativeness and effectiveness of the collected samples and to accurately match the geographic coordinates of the sampling points.

[0055] Laboratory physicochemical analysis: The collected soil samples are sent to a professional laboratory to conduct comprehensive testing and analysis of the soil's physicochemical properties. The testing indicators include core indicators such as soil organic matter content, pH value, available nitrogen content, available phosphorus content, and available potassium content, generating corresponding soil physicochemical property test data.

[0056] Geographic coordinate association: The soil physicochemical property test data obtained from laboratory testing are accurately associated with the corresponding sampling point geographic coordinates, and a unique geographic coordinate information is matched for each set of physicochemical property data, ultimately generating soil physicochemical property data with geographic coordinate information;

[0057] Tree canopy-scale remote sensing spectral data acquisition unit:

[0058] Remote sensing image retrieval: Based on the sampling point coordinates, determine the set of geographic coordinates output by the unit, retrieve high-resolution remote sensing images that can completely cover the pomelo planting area from a professional remote sensing platform, and ensure that the shooting time and resolution of the remote sensing images meet the technical requirements for tree canopy spectral information extraction.

[0059] Spectral information extraction: Based on the coordinates of the sampling points in the geographic coordinate set, precise coordinate positioning is performed in the high-resolution remote sensing image, and the pixel spectral information of the pomelo tree canopy at the corresponding location of each sampling point is extracted to ensure that the extracted spectral information is accurately matched with the spatial location of the pomelo tree.

[0060] Spectral data generation: The spectral information of all extracted pomelo tree canopy pixels is integrated and organized, and the corresponding sampling point geographic coordinate information is associated to generate canopy-scale remote sensing spectral data with geographic coordinate information, which fully reflects the spatial distribution of the spectral characteristics of pomelo trees in the planting area.

[0061] Meteorological and environmental data acquisition unit:

[0062] Meteorological data acquisition: Based on the geographical range covered by the set of sampling point coordinates, multi-source meteorological observation data within this range are retrieved from a professional meteorological data service, covering core meteorological indicators such as temperature, precipitation, sunshine, and humidity, to ensure that the monitoring dimensions and temporal resolution of the meteorological data meet the analytical needs of fertilization decision-making.

[0063] Data fusion and interpolation: The retrieved multi-source meteorological observation data are fused to eliminate the bias and redundancy between different data sources. Then, a spatial interpolation algorithm is used to perform spatial interpolation on the fused meteorological data to realize the spatial distribution calculation of meteorological data in the planting area.

[0064] Gridded meteorological data generation: The interpolated meteorological data is precisely matched with the geographic coordinates of the sampling points to generate corresponding meteorological and environmental data for each geographic coordinate point. Finally, gridded meteorological and environmental data matching each geographic coordinate point is generated to achieve a precise correspondence between meteorological data and geographic space.

[0065] Furthermore, the dynamic sampling point layout technology, based on geographic information system analysis theory and agronomic sampling standards, integrates historical yield data, topographic data, and preset sampling density of pomelo planting areas for multi-dimensional analysis to uncover the spatial heterogeneity characteristics of soil and tree growth within the planting area. The heterogeneity characteristics are quantified using geospatial interpolation algorithms, and the geographic coordinates of the sampling points are calculated accordingly. This ensures that the layout of the sampling points accurately reflects the differences in soil and tree growth within the area, guaranteeing the representativeness of the sampling points while dynamically adjusting the sampling density based on regional characteristics, thus avoiding resource waste or insufficient sampling caused by uniform sampling.

[0066] The precise detection technology for soil physicochemical properties is based on the physicochemical analysis methods of soil science. It conducts standardized laboratory tests on collected soil samples and uses professional testing instruments and reagents to quantitatively analyze key indicators such as soil organic matter, pH value, available nitrogen, available phosphorus, and available potassium. At the same time, it combines geographic information association technology to match the detected physicochemical data with the geographic coordinates of the sampling points, realizing the spatial labeling of soil physicochemical property data. This allows the soil data to accurately reflect the spatial distribution characteristics of the planting area and provides precise physicochemical indicators to support subsequent analysis of soil fertility capacity.

[0067] The canopy-scale spectral information extraction technology is based on remote sensing spectroscopy and geographic coordinate positioning technology. It utilizes the pixel features of high-resolution remote sensing images and, through precise matching of geographic coordinates, locates the location of the pomelo tree canopy corresponding to each sampling point in the remote sensing image. It then extracts core spectral information such as pixel spectral reflectance at that location. This spectral information can indirectly reflect the growth characteristics of the pomelo tree, such as chlorophyll content, nutritional status, and water status. By linking spectral information with geographic coordinates, it realizes the spatial expression of tree growth characteristics, providing spectral data support for subsequent diagnosis of tree nutritional status.

[0068] Meteorological data fusion and spatial interpolation technology is based on multi-source data fusion theory and spatial interpolation algorithms. First, a data assimilation algorithm is used to fuse multi-source meteorological observation data, eliminating data biases between different monitoring stations and equipment, thus achieving standardized integration of meteorological data. Then, the Kriging interpolation algorithm is used to perform spatial interpolation processing on the fused meteorological data. Its core formula is: ,in, The meteorological element values ​​for the points to be interpolated. The number of known meteorological monitoring points participating in the interpolation. For the first Weight coefficients for known monitoring points For the first Measured values ​​of meteorological elements at known monitoring points;

[0069] This algorithm calculates the meteorological element values ​​for each geographical coordinate point within the pomelo planting area, realizing the gridded transformation of meteorological data from discrete monitoring points to continuous spatial regions. This allows the meteorological data to accurately match the geographical coordinates of the planting area, providing spatialized meteorological data support for subsequent environmental factor impact analysis.

[0070] In this embodiment, the spatiotemporal fusion and feature generation module is the core hub of the pomelo intelligent fertilization decision system based on multi-source data fusion, which realizes multi-source spatiotemporal data collaborative analysis and precise extraction of agronomic features. It performs standardized processing and deep fusion of various spatiotemporal data collected by the multi-source data perception module, and combines agronomic knowledge models to mine the agronomic rules behind the data, generating feature vectors that can accurately reflect the synergistic state of soil, tree and environment in pomelo planting areas. This provides scientific and accurate data analysis results for the dynamic partitioning decision engine, and is a key link connecting data collection and fertilization decision, ensuring the scientificity and accuracy of fertilization decisions.

[0071] The spatiotemporal fusion and feature generation module is primarily responsible for spatiotemporal alignment of soil physicochemical property data, canopy-scale remote sensing spectral data, and meteorological environmental data output from the multi-source data perception module. It maps various data types to standardized gridded geographic information system units, and then performs multi-dimensional fusion analysis on the aligned data within the same grid unit based on an agronomic knowledge model. This extracts core agronomic indicators related to the soil and tree environment, generating a collaborative diagnostic feature vector that comprehensively reflects the soil's basic fertilization capacity, the tree's real-time nutritional status, and the influence coefficients of environmental factors. This provides core data support for the dynamic zoning decision engine's nutrient requirement prediction and fertilizer prescription generation. Simultaneously, it enables the spatialization of data from discrete sampling points to continuous spatial regions, improving the spatial accuracy of fertilization decisions. The spatiotemporal fusion and feature generation module includes:

[0072] Spatiotemporal alignment and mesh mapping unit:

[0073] Data time reference normalization: Receive soil physicochemical property data, canopy-scale remote sensing spectral data, and meteorological environmental data output from the multi-source data sensing module, extract observation time information for each type of data, determine a unified observation period for the system, perform time reference normalization processing on each type of data from different acquisition times, synchronize the time information of all data to this unified observation period, eliminate analytical bias caused by asynchronous data acquisition times, and ensure that various types of data are comparable and compatible in the time dimension;

[0074] Unified geographic coordinate transformation: Extract geographic coordinate information of various types of data after time-base normalization, call the system's preset geographic coordinate system parameters, and perform unified geographic coordinate transformation on soil physicochemical property data, tree canopy-scale remote sensing spectral data, and meteorological environment data in different coordinate systems. This ensures that the spatial location information of all data is matched to the preset geographic coordinate system, eliminating spatial location deviations caused by differences in coordinate systems and achieving standardization of various types of data in spatial dimensions.

[0075] Planting area grid division: Based on the unified geographic coordinate system, the geographic boundary information of the pomelo planting area is read. According to the grid division scale preset by the system, the entire pomelo planting area is divided into multiple grid units of the same size. All grid units together constitute a unified gridded geographic information system unit. Each grid unit corresponds to a unique geographic spatial location, which becomes the basic spatial unit for subsequent data fusion and fertilization decision-making.

[0076] Spatial data interpolation and resampling: For each gridded geographic information system unit, a spatial interpolation algorithm is used to interpolate soil physicochemical property data and meteorological environmental data at discrete sampling point locations to the center of the grid unit, realizing the spatial transformation of discrete data into continuous grid units; at the same time, pixel resampling processing is performed on the canopy-scale remote sensing spectral data to accurately match the spectral pixel information to the corresponding grid unit, so that the spatial scale of the spectral data is consistent with the grid unit, ensuring the spatial matching of the three types of data within the grid unit;

[0077] Grid cell data integration: Soil physicochemical property data, meteorological environment data, and tree canopy-scale remote sensing spectral data after spatial interpolation and pixel resampling are classified and integrated according to gridded geographic information system units. Independent and complete soil physicochemical property dataset, tree canopy-scale remote sensing spectral dataset, and meteorological environment dataset are generated for each grid cell. All datasets are associated with the geospatial information of the corresponding grid cell, realizing centralized storage and unified management of multi-source data within the grid cell.

[0078] Collaborative diagnostic feature generation unit:

[0079] Soil basic fertilization feature extraction: Read the soil physicochemical property dataset in each grid cell, and extract the soil organic matter content, pH value, available nitrogen content, available phosphorus content and available potassium content as basic indicators for soil fertilization capacity assessment. Input these basic indicators into the system's preset soil nutrient supply capacity assessment model, and output the soil basic fertilization feature vector that represents the basic fertilization capacity of the corresponding grid cell through model calculation. This vector can accurately quantify the soil's own nutrient supply level.

[0080] Canopy spectral derivation parameter analysis: Retrieve the canopy-scale remote sensing spectral dataset in each grid cell, input it into the system's preset pomelo canopy spectral analysis model, and extract canopy spectral derivation parameters such as the relative chlorophyll content, carotenoid index and water stress index that can reflect the growth status of pomelo trees through the analysis and calculation of spectral information by the model. These parameters are important basis for characterizing the real-time growth and nutritional status of the tree.

[0081] Tree nutrient status diagnosis: The soil basic fertilization feature vector is fused with the corresponding tree canopy spectral derived parameters. The fused dataset is input into the system's preset tree nutrient balance diagnosis model. Through the model's comprehensive analysis and calculation, a tree nutrient status vector that reflects the real-time balance of nitrogen, phosphorus, and potassium elements in the tree is generated. This vector can accurately reflect the matching relationship between the tree's current nutrient requirements and soil fertilization.

[0082] Environmental impact coefficient calculation: Extract the meteorological environment dataset in each grid cell, input it into the system's preset environmental nutrient migration and absorption impact model, combine it with the agronomic laws of pomelo growth, and calculate the environmental impact coefficient vector through model operation to represent the degree of influence of temperature, precipitation and light conditions on fertilizer effectiveness and tree absorption efficiency. This vector can quantify the degree of influence of environmental factors on fertilization effect.

[0083] Collaborative diagnostic feature vector integration: Read the soil basic fertilization feature vector, tree nutrient status vector, and environmental influence coefficient vector corresponding to each grid unit, associate the age and phenological information of pomelo trees in each grid unit, and integrate the above vectors and information according to the feature fusion rules preset by the system to finally generate a collaborative diagnostic feature vector that can comprehensively reflect the soil basic fertilization capacity, the real-time nutrient status of the tree, and the influence coefficient of environmental factors. This vector is the core input data for the dynamic partition decision engine to predict fertilizer requirements.

[0084] Furthermore, the spatiotemporal data standardization and fusion technology, based on spatiotemporal data fusion theory and geographic information system technology, standardizes multi-source heterogeneous data from both temporal and spatial dimensions. In the temporal dimension, a unified observation period is established to normalize the time reference, eliminating the asynchronous nature of data acquisition. In the spatial dimension, a unified geographic coordinate system is used for coordinate transformation, combined with grid partitioning to achieve spatial discretization of planting areas. Then, spatial interpolation and pixel resampling techniques are used to match the discrete multi-source data to a unified grid cell, achieving dual standardization of multi-source data in both spatiotemporal dimensions. The core of this technology is to transform heterogeneous data of different types, acquisition times, and spatial locations into standardized data with the same spatiotemporal reference, laying a data foundation for subsequent agronomic feature analysis. The spatial interpolation uses the Kriging interpolation algorithm, whose core formula is: ,in, The meteorological element values ​​for the points to be interpolated. The number of known meteorological monitoring points participating in the interpolation. For the first Weight coefficients for known monitoring points For the first Measured values ​​of meteorological elements at known monitoring points;

[0085] Soil nutrient supply capacity assessment technology is based on soil science and agronomy theories, with soil organic matter content, pH value, available nitrogen content, available phosphorus content, and available potassium content as core assessment indicators. These indicators directly determine the soil's nutrient reserves and supply capacity. By constructing a soil nutrient supply capacity assessment model, the core indicators are quantitatively weighted and calculated, transforming multi-dimensional physicochemical indicators into feature vectors that comprehensively reflect the soil's nutrient supply capacity. The model's weight settings are based on the soil nutrient requirements of pomelo growth, and the weights of different indicators are adjusted according to the differences in nutrient requirements at different growth stages of pomelo to ensure that the assessment results match the actual growth needs of pomelo.

[0086] Canopy spectral analysis technology is based on the correlation between remote sensing spectroscopy and plant physiology. Physiological characteristics of pomelo trees, such as chlorophyll content, carotenoid content, and water status, will form specific spectral features in spectral reflectance. By constructing a pomelo canopy spectral analysis model, feature extraction and calculation are performed on remote sensing spectral information at the canopy scale, and spectral reflectance is converted into spectral derivative parameters that can directly reflect the physiological state of the tree. The model is trained on spectral samples of pomelo trees under different nutritional states to establish a quantitative mapping relationship between spectral features and tree physiological indicators, ensuring that the analyzed parameters can accurately reflect the real-time nutritional status of the tree.

[0087] The tree nutrient surplus and deficit diagnostic technology is based on the soil-tree nutrient supply and demand balance theory. It integrates the soil basic nutrient supply characteristic vector and the tree canopy spectral derived parameters to construct a tree nutrient surplus and deficit diagnostic model. The model first analyzes the soil's nutrient supply capacity, and then combines the tree canopy spectral derived parameters to reflect the actual growth status of the tree. By comparing the nutrient requirements of pomelo trees at different growth stages, it quantitatively analyzes the actual surplus and deficit of nitrogen, phosphorus, and potassium elements in the tree, and finally generates a tree nutrient status vector. The core of this technology is to achieve synergistic analysis of soil nutrient supply capacity and the actual nutrient requirements of the tree, avoiding the bias caused by relying solely on soil or tree data for diagnosis.

[0088] The environmental factor impact quantification technology is based on agricultural meteorology and nutrient science theories. Meteorological environmental factors such as temperature, precipitation, and sunlight directly affect the decomposition, transformation, and migration of fertilizers and the tree's nutrient absorption efficiency. By constructing an environmental nutrient migration and absorption impact model, meteorological environmental data is combined with agronomic parameters of pomelo nutrient absorption to quantify the impact coefficients of fertilizer effectiveness and tree absorption efficiency under different meteorological conditions, generating an environmental impact coefficient vector. The model's calculation process incorporates the growth characteristics of pomelo; for example, a suitable temperature range can improve tree absorption efficiency, while excessive precipitation can lead to nutrient loss and reduce fertilizer effectiveness, ensuring that the coefficients accurately reflect the actual impact of environmental factors.

[0089] The multi-feature fusion generation technology is based on feature fusion theory and agronomic knowledge model. It integrates soil basic fertilization feature vector, tree nutrient status vector, and environmental influence coefficient vector, and associates them with pomelo tree age and phenological stage information. It performs calculations according to preset feature fusion rules. The fusion rules are based on the growth patterns and fertilization requirements of pomelo trees at different ages and phenological stages. Different weights are assigned to each feature vector, and the multi-dimensional feature vectors are integrated into a single collaborative diagnostic feature vector. This achieves the collaborative representation of multiple factors in the soil, tree, and environment, and allows the feature vector to comprehensively and accurately reflect the overall status of pomelo planting within the grid unit.

[0090] In this embodiment, the dynamic partitioning decision engine is the core hub of the pomelo intelligent fertilization decision system based on multi-source data fusion to achieve precise fertilization decision-making. It generates differentiated fertilization prescriptions adapted to the grid units of the pomelo planting area through in-depth analysis and multi-objective optimization of collaborative diagnostic feature vectors. At the same time, it realizes the dynamic self-correction of the pomelo fertilizer requirement prediction model to ensure the scientific accuracy and dynamic adaptability of fertilization decisions.

[0091] The dynamic zoning decision engine is primarily responsible for receiving the collaborative diagnostic feature vector output by the spatiotemporal fusion and feature generation module. It calculates the initial nutrient requirements of each gridded geographic information system unit using a pre-trained pomelo fertilizer requirement prediction model. Then, it integrates the fertilization cost model, quality and yield objective functions, and nutrient cycle constraints to perform multi-objective optimization, generating a zoning-specific fertilization prescription that includes precise application rates for fertilizer varieties and recommended operating times. Simultaneously, based on fertilization effect feedback data and updated collaborative diagnostic feature vectors, it performs parameter self-calibration of the pomelo fertilizer requirement prediction model, providing precise fertilization decision-making basis for the decision execution and interaction module. This enables differentiated and precise fertilization decisions for gridded units in the pomelo planting area. The dynamic zoning decision engine includes:

[0092] Fertilizer requirement prediction unit:

[0093] Feature vector reception and preprocessing: Real-time reception of collaborative diagnostic feature vectors from various gridded geographic information system units output by the spatiotemporal fusion and feature generation module; data standardization processing of feature vectors to eliminate the dimensional differences of feature data of different dimensions; ensuring that the data input into the pomelo fertilizer requirement prediction model has consistency and validity; laying the data foundation for subsequent fertilizer requirement prediction calculations.

[0094] Multi-output prediction network operation: The preprocessed collaborative diagnostic feature vector is input into the multi-output prediction network of the pomelo fertilizer requirement prediction model. First, the collaborative diagnostic feature vector is decoupled, and sub-feature sets strongly correlated with nitrogen, phosphorus and potassium requirements are extracted respectively. Then, through the nitrogen requirement prediction sub-network, phosphorus requirement prediction sub-network and potassium requirement prediction sub-network, the corresponding sub-feature sets are fused with the soil basic fertilizer supply feature vector, tree nutrient status vector and environmental impact coefficient vector in the collaborative diagnostic feature vector respectively, to complete the independent prediction operation of the initial nitrogen, phosphorus and potassium requirements of each gridded geographic information system unit.

[0095] Initial nutrient requirements aggregation: The initial nitrogen fertilizer requirements output by the nitrogen demand prediction subnetwork, the initial phosphate fertilizer requirements output by the phosphorus demand prediction subnetwork, and the initial potassium fertilizer requirements output by the potassium demand prediction subnetwork are classified and integrated. The data is aggregated according to the gridded geographic information system units to generate the initial nutrient requirements corresponding to each unit, providing benchmark data for subsequent multi-objective optimization solutions.

[0096] Multi-objective optimization solution unit:

[0097] Fertilization cost model establishment: Using gridded geographic information system units as decision-making units, a fertilization cost model is established with the goal of minimizing the total fertilization cost. Based on the system's built-in library of selectable fertilizer varieties and fertilizer nutrient content parameters, combined with the correlation between the unit cost and application rate of different fertilizer varieties, the model is constructed and parameters are assigned, quantifying the correspondence between fertilization cost and fertilizer variety and application rate.

[0098] Quality and yield objective function establishment: With the goal of maximizing the comprehensive score of expected fruit quality and yield, a quality and yield objective function is constructed. The initial nutrient requirements of each gridded geographic information system unit are used as the baseline input. The age and phenological information of the pomelo tree corresponding to the grid unit are associated with the information. The nutritional requirements of pomelo at different growth stages are incorporated to complete the construction of the objective function and realize the quantitative expression of quality and yield objectives.

[0099] Nutrient cycling constraint setting: Combining agronomic knowledge and environmental maintenance requirements for pomelo cultivation, nutrient cycling constraints are set. Nutrient loss coefficient constraints are determined based on the environmental impact coefficient vector. Soil nutrient balance constraints are set based on the soil basic fertilizer supply characteristic vector. Thresholds for safe application of fertilizer per application are set based on agronomic knowledge, forming a multi-dimensional nutrient cycling constraint system and defining the feasible range of fertilization decisions.

[0100] Multi-objective optimization solution operation: The initial nutrient requirements of each gridded geographic information system unit, along with the diagnostic feature vector, optional fertilizer variety library, and fertilizer nutrient content parameters, are input into the multi-objective optimization solver. Under the premise of satisfying the nutrient cycle constraint, the fertilizer cost model and the quality and yield objective function are solved in parallel. Pareto optimal front analysis is used to explore the balance relationship between cost and quality and yield objectives, and to obtain an optimal solution set covering different optimization tendencies.

[0101] Zonal variable fertilization prescription generation: For each gridded geographic information system unit, a recommended solution that is suitable for the actual needs of pomelo planting is selected from the set of optimized solutions. Based on the recommended solution, the specific fertilizer variety combination and the precise dosage of each variety are determined. The recommended operation time is determined by combining the pomelo phenological period information corresponding to the grid unit. All information is integrated to generate the zonal variable fertilization prescription corresponding to each gridded geographic information system unit.

[0102] Model parameter self-calibration unit:

[0103] Acquisition of effect feedback data: After the fertilization operation is completed, the yield data and quality test data of the corresponding areas of the historical fertilization prescription after fruit harvest are collected. The data are standardized and verified, invalid data is removed, and effect feedback data that meets the model calibration requirements are generated to provide a practical basis for model deviation analysis.

[0104] Updated collaborative diagnostic feature vector generation: The new round of canopy-scale remote sensing spectral data and meteorological environment data of the pomelo planting area are obtained through the multi-source data perception module. Combined with historical soil physicochemical property data, the spatiotemporal alignment and fusion analysis of the data are completed through the spatiotemporal fusion and feature generation module to generate updated collaborative diagnostic feature vectors corresponding to each gridded geographic information system unit, reflecting the actual state of the tree soil and environment after fertilization.

[0105] Prediction deviation signal calculation: The effect feedback data is compared with the quality and yield prediction targets set before the implementation of historical fertilizer prescriptions. The deviation value is calculated based on the comparison results. The prediction deviation signal is generated based on the sign and magnitude of the deviation value to quantify the prediction error of the pomelo fertilizer requirement prediction model.

[0106] Parameter adjustment calculation: The prediction deviation signal and the updated co-diagnostic feature vector are input into the model parameter correction algorithm. Based on the direction and amplitude of the prediction deviation signal, combined with the actual response state of the tree reflected by the updated co-diagnostic feature vector, the parameter adjustment of the multi-output prediction network in the pomelo fertilizer requirement prediction model is calculated by the algorithm to clarify the update amplitude and direction of the model parameters.

[0107] Model parameter iterative update: The internal parameters of the pomelo fertilizer requirement prediction model are iteratively updated layer by layer using the parameter adjustment amount obtained from the solution, and the parameter replacement and calibration are completed to realize the parameter self-correction of the pomelo fertilizer requirement prediction model and improve the subsequent prediction accuracy of the model.

[0108] Furthermore, the multi-output feature decoupling prediction technology, based on feature engineering and deep learning network theory, decouples the high-dimensional collaborative diagnostic feature vectors and mines sub-feature sets strongly correlated with nitrogen, phosphorus, and potassium requirements through feature extraction algorithms, thereby achieving effective separation of multiple nutrient requirement features. Relying on the structural design of the multi-output prediction network, independent prediction sub-networks are constructed for nitrogen, phosphorus, and potassium requirements. Each sub-network integrates the corresponding sub-feature set with the soil basic fertilization feature vector, tree nutrient status vector, and environmental influence coefficient vector. Combined with the agronomic rules of pomelo nutrient requirements, the network operation logic is constructed to achieve accurate and independent prediction of the initial nutrient requirements, while ensuring the spatial matching of the prediction results with the gridded geographic information system units.

[0109] The multi-objective collaborative optimization solution technology is based on multi-objective optimization theory and Pareto optimality criterion. It uses gridded geographic information system units as the basic decision-making units to construct a dual-objective optimization system with fertilization cost model and quality and yield objective functions. At the same time, it incorporates nutrient cycling constraints to define the decision boundary. The two objective functions are solved in parallel by a multi-objective optimization solver. The Pareto optimality front analysis method is used to screen out non-dominated solutions to form an optimal solution set. Each solution can achieve a balance between cost and quality objectives to different degrees. Combined with the actual needs of pomelo planting, a recommended solution is selected to achieve a synergistic balance between cost control, quality improvement and yield increase in fertilization decisions, while ensuring that the fertilization plan meets the requirements of soil nutrient cycling and agronomic safety.

[0110] The calculation formula for the fertilizer cost model is as follows: ,in, The total cost of fertilization, The number of fertilizer varieties, For the first The unit cost of fertilizer varieties, For the first (Application amount of fertilizer type);

[0111] The formula for calculating the objective function of quality and output is as follows: ,in, The overall score is based on quality and quantity. This refers to the amount of nitrogen fertilizer applied. This refers to the amount of phosphate fertilizer applied. This refers to the amount of potassium fertilizer applied. The age of the pomelo tree This is the phenological period for pomelos. This is the production forecast function. For quality prediction function, This is the production weighting coefficient. This is the quality weighting coefficient. ;

[0112] The model dynamic parameter self-calibration technology is based on closed-loop feedback control theory and the iterative optimization principle of machine learning models. It uses the actual effect feedback data after fertilization as the basis for closed-loop feedback. By comparing the actual effect with the prediction target, a prediction deviation signal is generated to quantify the prediction error of the model. Combined with the updated collaborative diagnostic feature vector generated by the actual state of the tree, soil and environment after fertilization, the actual causes of the error are explored. The deviation signal and the actual state features are fused and analyzed through the model parameter calibration algorithm to solve for the appropriate parameter adjustment amount. The parameter adjustment amount is used to iteratively update the multi-output prediction network parameters of the pomelo fertilizer requirement prediction model to achieve dynamic calibration of the model parameters. This enables the model to continuously adapt to the changes in soil, tree and environment in the pomelo planting area, forming a closed-loop optimization system of prediction decision feedback calibration, and continuously improving the prediction accuracy of the model.

[0113] The workflow of the dynamic partitioning decision engine is as follows:

[0114] Initialization phase:

[0115] After the dynamic partitioning decision engine starts, it completes the self-check of the hardware devices, including the data processing unit, model calculation engine, optimization solver, data storage unit, etc., to ensure that the hardware devices are operating normally.

[0116] The initial parameters of the pomelo fertilizer requirement prediction model, the basic parameters of the fertilizer cost model, the weight coefficients of the quality and yield objective function, and the threshold parameters of the nutrient cycle constraint conditions are loaded. Communication connections are established with the spatiotemporal fusion and feature generation module, the multi-source data perception module, and the decision execution and interaction module to prepare for receiving collaborative diagnostic feature vectors.

[0117] Feature vector reception and preprocessing stage:

[0118] The system receives collaborative diagnostic feature vectors from each gridded geographic information system unit in real time from the spatiotemporal fusion and feature generation module, and performs integrity verification on the received feature vectors to ensure that the data is free of missing or abnormal data.

[0119] The verified collaborative diagnostic feature vectors are subjected to data standardization processing to eliminate the dimensional differences of feature data in different dimensions, generate standardized feature vectors that meet the requirements of model operation, and complete the data preprocessing.

[0120] Initial nutrient requirement forecasting phase:

[0121] The preprocessed standardized feature vectors are input into the multi-output prediction network of the pomelo fertilizer requirement prediction model. Feature decoupling is performed on the feature vectors, and sub-feature sets that are strongly correlated with nitrogen, phosphorus and potassium requirements are extracted respectively.

[0122] Each nutrient prediction subnetwork integrates the corresponding sub-feature set and the soil basic fertilization feature vector, tree nutrient status vector, and environmental influence coefficient vector from the collaborative diagnostic feature vector to complete the prediction calculation of the initial demand for each nutrient.

[0123] The initial nutrient requirements are collected according to the gridded geographic information system units, and the initial nutrient requirements corresponding to each unit are generated, which then enters the subsequent multi-objective optimization solution stage.

[0124] Multi-objective optimization solution stage:

[0125] Using gridded geographic information system units as decision-making units, a fertilization cost model and quality and yield objective functions are established, nutrient cycling constraints are set, and a multi-objective optimization system is constructed.

[0126] The initial nutrient requirement, the collaborative diagnostic feature vector, the selectable fertilizer variety library, and the fertilizer nutrient content parameters are input into the multi-objective optimization solver. Under the premise of satisfying the nutrient cycle constraint, the fertilizer cost model and the quality and yield objective function are solved in parallel.

[0127] The optimal solution set is obtained through Pareto optimal front analysis. Recommended solutions are selected for each gridded geographic information system unit. Combined with pomelo phenological information, a zonal variable fertilization prescription is generated, which includes the precise amount of fertilizer for the fertilizer variety and the recommended operation time.

[0128] Fertilizer prescription output stage:

[0129] The generated fertilization prescriptions for regional variables are standardized to ensure that the format and content of the prescriptions meet the input requirements of the decision execution and interaction module.

[0130] The standardized zonal variable fertilization prescriptions are synchronously output to the decision execution and interaction module, providing a decision basis for fertilization instruction conversion and agricultural guidance program generation. At the same time, the fertilization prescription data is stored in the local data unit to provide historical data for subsequent model parameter self-calibration.

[0131] Model parameter self-calibration stage:

[0132] After the fertilization operation is completed, a new round of canopy-scale remote sensing spectral data and meteorological environment data are acquired through the multi-source data sensing module. Combined with historical soil physicochemical property data, updated collaborative diagnostic feature vectors are generated. At the same time, yield data and quality test data after fruit harvest are collected to generate effect feedback data.

[0133] The effect feedback data is compared with the historical quality and yield prediction targets to generate a prediction deviation signal. The prediction deviation signal and the updated collaborative diagnostic feature vector are input into the model parameter correction algorithm to solve for the parameter adjustment amount.

[0134] The internal parameters of the fertilizer requirement prediction model for pomelos are iteratively updated using parameter adjustment to complete the self-calibration of model parameters. The updated model will be applied to the next round of fertilizer requirement prediction calculation.

[0135] End phase:

[0136] Once the dynamic partitioning decision engine completes the output of the current fertilization prescription and the self-calibration of model parameters, it automatically closes the temporary communication connections with each external module, saves all the original data and processing results of this decision, releases the module's temporary system resources, and awaits the next round of fertilization decision instructions.

[0137] In this embodiment, the decision execution and interaction module is the core hub for implementing fertilization decisions based on the multi-source data fusion-based intelligent fertilization decision system for pomelos. It builds a bridge between system decision-making and intelligent fertilization machinery operations by professionally analyzing and transforming the fertilization prescriptions of regional variables. At the same time, it provides growers with intuitive and easy-to-understand agricultural guidance solutions to ensure that precise fertilization decisions are efficiently transformed into actual planting operations.

[0138] The decision execution and interaction module is primarily responsible for parsing the fertilization prescriptions for regional variables output by the dynamic regional decision engine, accurately extracting the core operational elements, constructing standardized structured instruction sequences, and converting them into a set of drive instructions that intelligent fertilization machinery can directly parse and execute. Simultaneously, it generates visualized agricultural guidance schemes based on the fertilization prescriptions. This module distributes the drive instruction sets to the intelligent fertilization machinery control terminal and pushes the visualized agricultural guidance schemes to the user interface, achieving a dual implementation of mechanized automatic execution of fertilization decisions and manual visual reference. It ensures the accuracy and integrity of instructions and schemes throughout the entire process, connecting the system's decision-making with actual agricultural operations. The decision execution and interaction module includes:

[0139] Prescription analysis and element extraction unit:

[0140] Prescription Reception and Verification: The system receives fertilization prescriptions for partitioned variables from the dynamic partitioning decision engine in real time, verifies the integrity and format of the received prescription data, checks for missing parameters, format errors, etc., marks prescriptions that fail verification and synchronously feeds them back to the dynamic partitioning decision engine, and initially stores prescriptions that pass verification.

[0141] Core element extraction: According to the preset element extraction rules, the qualified zoning variable fertilizer prescriptions are analyzed in a structured manner. The core operational elements such as the recommended application time and geographical coordinates of fertilizer varieties corresponding to each grid unit are identified and extracted one by one to ensure the comprehensiveness and accuracy of element extraction.

[0142] Element collection and organization: All extracted operational elements are classified and collected according to the geographic coordinates of the grid units. The element data is standardized and organized to unify the data format and units, and a standardized element dataset corresponding to each grid unit is generated, providing a unified and standardized data source for subsequent instruction conversion and scheme production.

[0143] Machine instruction conversion and generation unit:

[0144] Structured instruction sequence construction: Based on a standardized feature dataset, with the geographic coordinates of grid units as the core identifier, and integrating information such as fertilizer category codes, target application rates, and timestamps, a coherent structured instruction sequence is constructed according to the geographic spatial order of the pomelo planting area to ensure that the spatial logic of the instruction sequence matches the actual layout of the planting area;

[0145] Drive instruction set conversion: Read the system's preset intelligent fertilizer applicator model and communication protocol parameters, and convert the structured instruction sequence into a drive instruction set that the intelligent fertilizer applicator's control system can directly parse and execute according to the corresponding protocol rules, thereby achieving a precise mapping of decision parameters to mechanical execution parameters;

[0146] Instruction set verification and optimization: The converted drive instruction set is verified by both logic and parameters to check for logical conflicts between instructions and whether parameters exceed the mechanical execution threshold. Problems found during verification are corrected in real time. At the same time, the execution order of the instruction set is optimized to improve the smoothness of intelligent fertilization machinery during operation.

[0147] Visualized agricultural solutions creation unit:

[0148] Differentiated fertilization distribution map drawing: retrieve the geographic information system parameters of the pomelo planting area, spatially fuse the fertilizer variety and application data in the standardized element dataset with geographic coordinates, and use spatial rendering technology to draw a differentiated fertilization distribution map of each grid unit in the planting area, intuitively presenting the fertilization differences in different areas.

[0149] Agricultural guidance card generation: Based on the zoning rules of pomelo planting areas, the agricultural guidance cards with associated recommended operation times are generated by integrating elements such as fertilizer varieties, precise application rates, and operation times for each grid unit, clarifying the specific fertilization operation requirements for each area and simplifying the operation reference dimensions for growers.

[0150] Standardization of the plan: The format of the completed differentiated fertilizer application distribution map and the generated agricultural guidance card is standardized to adapt to the display requirements of the user interface and the offline printing needs. At the same time, the content of the plan is simplified and optimized to ensure the simplicity and clarity of information transmission.

[0151] Instruction Issuance and Solution Push Unit:

[0152] Mechanical terminal command issuance: Establish an encrypted communication connection with the intelligent fertilizer application machinery control terminal, and issue the verified and optimized drive command set to the corresponding control terminal through a dedicated communication link. Record the time node of command issuance and the amount of data transmitted in real time to ensure the security and integrity of command transmission.

[0153] User interface solution push: Standardized and visualized agricultural guidance solutions are pushed to the grower's user interface, supporting online viewing, local download and printing of the solutions to meet the grower's needs in different scenarios;

[0154] Transmission status monitoring and retransmission: Real-time monitoring of the transmission status of the driver instruction set and the agricultural guidance plan, accurately identifying transmission success and failure status results; if a transmission failure is detected, the retransmission mechanism is automatically triggered, and the retransmission operation is performed according to preset rules to ensure that all instructions and plans can be implemented.

[0155] Furthermore, the fertilizer prescription structured parsing technology, based on data semantic parsing and structured processing theory, pre-sets multi-dimensional element extraction rules for the fertilization prescription with partitioned variables output by the dynamic partitioning decision engine. Through rule matching, it achieves accurate identification and extraction of core operational elements in the prescription. The core technology lies in building a mapping relationship between prescription data and operational elements, which can dynamically adjust the extraction rules according to the updated prescription format to adapt to different prescription output formats. The extracted elements are collected and organized according to geographical coordinates, realizing the conversion of unstructured prescription data into structured element data and ensuring the consistency of data sources in subsequent processing stages.

[0156] The intelligent machinery instruction adaptation and conversion technology is based on equipment communication protocols and instruction encoding theory. The system pre-stores the communication protocols and instruction encoding rules for different models of intelligent fertilizing machinery. The core technology is to achieve accurate mapping between fertilization decision parameters and machinery execution parameters. The mapping process employs a parameter conversion algorithm, the expression of which is as follows: (in, For mechanical execution parameter values, For parameter matching coefficients, These are the fertilization decision parameter values. (For parameter correction values); This algorithm converts the decision parameters in the structured instruction sequence into execution parameters that are adapted to the mechanical control, and then encodes the instructions according to the coding rules of the corresponding machinery to generate a set of drive instructions that the mechanical control system can directly parse, ensuring that the mechanical execution actions are highly consistent with the fertilization decision requirements;

[0157] Geographic information fusion visualization technology, based on geographic information systems and data visualization theory, deeply integrates gridded fertilization data in fertilizer prescriptions with geospatial information of pomelo planting areas. Through spatial rendering algorithms, it uses color and numerical annotations to label fertilizer varieties and application rates in different grid units, creating differentiated fertilizer application distribution maps and achieving geospatial visualization of fertilization data. Simultaneously, combined with the division of planting areas into zones, it integrates and refines gridded fertilization elements to generate concise and intuitive agricultural guidance cards. This transforms professional gridded precision fertilization decisions into visual information that growers can directly understand, reducing the difficulty of operation and comprehension.

[0158] Multi-terminal communication and data transmission technology, based on network communication and data encryption theory, employs a dedicated encrypted transmission protocol to establish a two-way communication connection between the system and the grower's user interface at the intelligent fertilization machinery control terminal. The drive instruction set and agricultural guidance plan are encrypted before transmission, effectively ensuring data security and preventing data leakage and tampering. The technology also includes a real-time transmission status monitoring mechanism to instantly identify anomalies during transmission. The number of retransmissions triggered in case of transmission failure follows a formula. (in, The remaining number of retransmissions. The maximum number of retransmissions. The transmission time is calculated. (The base time for a single transmission) ensures both transmission reliability and transmission efficiency.

[0159] A smart fertilization decision-making method for pomelos based on multi-source data fusion, the method being executed by a smart fertilization decision-making system for pomelos based on multi-source data fusion, includes the following steps:

[0160] Step 1: Acquisition of multi-source spatiotemporal data:

[0161] This step is executed by the system's multi-source data sensing module. Its core objective is to acquire three types of spatiotemporal data: soil physicochemical properties, canopy-scale remote sensing spectra, and meteorological environment data related to the geographic coordinates of the pomelo planting area. This provides a precise and standardized raw data source for subsequent fertilization decisions. Specific operations include:

[0162] Dynamically determine the geographical coordinates of soil sampling points: retrieve historical yield maps and topographic maps of the pomelo planting area, combine them with the system's preset soil sampling density, comprehensively analyze the spatial difference patterns of soil and tree growth in the area, and calculate and output a set of representative and uniform geographical coordinates of soil sampling points through geospatial analysis algorithms, which serves as the geographical benchmark for all data collection.

[0163] Soil physicochemical property data collection: Soil samples were collected based on the above-mentioned set of geographic coordinates, strictly following agronomic sampling standards. The samples were sent to the laboratory for testing and analysis to extract core indicators such as soil organic matter content, pH value, available nitrogen content, available phosphorus content, and available potassium content. Corresponding geographic coordinates were matched to the test data to generate soil physicochemical property data with geographic coordinate information.

[0164] Collecting tree canopy-scale remote sensing spectral data: Based on the set of geographic coordinates of soil sampling points, high-resolution remote sensing images of the planting area are retrieved from a professional remote sensing platform. The pixel spectral information of the pomelo tree canopy corresponding to each sampling point is extracted by coordinate positioning. After associating with geographic coordinates, tree canopy-scale remote sensing spectral data is generated to reflect the spatial distribution of tree spectral characteristics.

[0165] Generate gridded meteorological environment data: Based on the geographical coordinate coverage of the sampling points, obtain multi-source meteorological observation data such as temperature, precipitation, sunshine, and humidity from the meteorological data server. First, the data is fused through a data assimilation algorithm to eliminate bias and redundancy. Then, the Kriging interpolation algorithm is used for spatial interpolation processing to generate gridded meteorological environment data that accurately matches each geographical coordinate point.

[0166] Step 2: Spatiotemporal alignment and grid mapping of multi-source data:

[0167] This step is performed by the system's spatiotemporal fusion and feature generation module. Its core objective is to eliminate the heterogeneity of multi-source data in the temporal and spatial dimensions, uniformly mapping all data to standardized gridded geographic information system units, thus achieving spatiotemporal standardization of the data. Specific operations include:

[0168] Time reference normalization: Extract the observation time information of three types of data: soil physicochemical properties, tree canopy-scale remote sensing spectrum, and meteorological environment, determine the unified observation period of the system, and perform time calibration on various types of data at different collection times to synchronize the time information of all data to the same period and eliminate the analysis bias caused by time asynchrony;

[0169] Unified geographic coordinate transformation: The system retrieves the preset geographic coordinate system parameters and transforms the geographic coordinates of the three types of data to the same coordinate system, eliminating spatial location deviations caused by different coordinate systems and achieving standardization of data spatial dimensions.

[0170] Planting area grid division: Based on the unified geographic coordinate system and combined with the geographic boundary information of the pomelo planting area, the entire planting area is divided into multiple grid units of the same size according to the preset scale, forming a unified gridded geographic information system unit. This unit is the basic spatial unit for all subsequent analysis and decision-making.

[0171] Spatial data interpolation and resampling: For each grid cell, the Kriging interpolation algorithm is used to interpolate the soil physicochemical properties data and meteorological environment data of discrete sampling points to the center of the grid cell; the canopy-scale remote sensing spectral data is resampled pixel to ensure that the spectral information is accurately matched to the corresponding grid cell, thus ensuring the spatial matching of the three types of data within the grid cell;

[0172] Grid cell data integration: The three types of data after interpolation and resampling are classified and organized according to grid cells. Independent and complete soil physicochemical properties dataset, canopy-scale remote sensing spectral dataset, and meteorological environment dataset are generated for each grid cell. All datasets are associated with the geospatial information of the corresponding grid cells, realizing centralized storage of multi-source data within the grid cells.

[0173] Step 3: Agronomic knowledge model fusion analysis to generate collaborative diagnostic feature vectors:

[0174] This step is performed by the system's spatiotemporal fusion and feature generation module. Its core objective is to perform deep fusion analysis on standardized data within the grid cells based on an agronomic knowledge model, extracting feature vectors that comprehensively reflect the synergistic state of soil, trees, and environment in the pomelo planting area. This provides core input data for subsequent fertilizer requirement prediction. Specific operations include:

[0175] Extracting the basic soil fertility supply feature vector: Organic matter, pH value, available nitrogen, available phosphorus, and available potassium are extracted as basic indicators from the soil physicochemical property dataset of each grid cell. These are then input into the soil nutrient supply capacity assessment model, and the basic soil fertility supply feature vector, which characterizes the basic soil fertility supply capacity of the grid cell, is output through quantitative weighted calculation.

[0176] Analysis of canopy spectral derived parameters: The remote sensing spectral dataset of the canopy scale of each grid unit is input into the pomelo canopy spectral analysis model to extract canopy spectral derived parameters that reflect the relative chlorophyll content, carotenoid index, and water stress index, which indirectly characterize the real-time growth and nutritional status of the tree.

[0177] Generate tree nutrient status vector: Integrate soil basic fertilization feature vector and tree canopy spectral derived parameters, input the tree nutrient surplus and deficit diagnostic model, compare the nutrient requirements of pomelo at different growth stages, quantify the real-time surplus and deficit status of nitrogen, phosphorus, potassium and magnesium elements in the tree, and generate tree nutrient status vector.

[0178] Calculate the environmental impact coefficient vector: Input the meteorological environment dataset of each grid cell into the environmental nutrient migration and absorption impact model, combine it with the agronomic laws of pomelo growth, quantify the impact of temperature, precipitation and light on fertilizer effectiveness and tree nutrient absorption efficiency, and generate the environmental impact coefficient vector.

[0179] Integrate and generate collaborative diagnostic feature vectors: collect the soil basic fertilization feature vectors, tree nutrition status vectors, and environmental influence coefficient vectors of each grid unit, associate the pomelo tree age and phenological period information corresponding to the grid unit, and perform integration calculations according to the system's preset feature fusion rules to finally generate collaborative diagnostic feature vectors that can comprehensively reflect the soil basic fertilization capacity, the real-time nutrition status of the tree, and the environmental factor influence coefficients.

[0180] Step 4: Predicting the fertilizer requirements of pomelos to obtain initial nutrient needs:

[0181] This step is executed by the system's dynamic partitioning decision engine. Its core objective is to input the collaborative diagnostic feature vector into a pre-trained pomelo fertilizer requirement prediction model, thereby achieving accurate prediction of the initial nitrogen, phosphorus, potassium, and magnesium requirements for each grid unit. Specific operations include:

[0182] Feature vector preprocessing: The received collaborative diagnostic feature vectors are subjected to integrity verification and data standardization to eliminate the dimensional differences of feature data of different dimensions and ensure that the data input to the model has consistency and validity;

[0183] Multi-output prediction network feature decoupling: The preprocessed collaborative diagnostic feature vector is input into the multi-output prediction network of the pomelo fertilizer requirement prediction model. The high-dimensional feature vector is decoupled through feature extraction algorithm, and sub-feature sets strongly related to nitrogen, phosphorus and potassium requirements are extracted respectively to achieve effective separation of multi-nutrient requirement features.

[0184] Initial nutrient requirements are predicted independently: The nitrogen, phosphorus, potassium, and magnesium requirement prediction sub-networks in the model are respectively integrated with the corresponding sub-feature sets and the soil basic fertilization, tree nutritional status, and environmental influence coefficient vectors in the collaborative diagnostic feature vector. Combined with the agronomic rules of pomelo nutrient requirements, the initial nitrogen, phosphorus, and potassium fertilizer requirements of the corresponding grid cells are calculated and output.

[0185] Initial nutrient requirements aggregation: The initial nitrogen, phosphorus, potassium and magnesium requirements of each grid unit are classified and integrated, and the data is collected according to the gridded geographic information system units to generate the initial nutrient requirements of each unit, which serve as the benchmark data for subsequent multi-objective optimization solutions.

[0186] Step 5: Multi-objective optimization solution to generate partitioned variable fertilization prescriptions:

[0187] This step is executed by the system's dynamic zoning decision engine. Its core objective is to integrate fertilization costs, quality and yield targets, and nutrient cycling constraints using gridded geographic information system units as the decision-making unit. This allows for multi-objective optimization of the initial nutrient requirements, generating a zoning-based variable fertilization prescription that is scientifically sound, economically viable, and practically applicable. Specific operations include:

[0188] Construct a multi-objective optimization model:

[0189] Establish a fertilization cost model: With the objective of minimizing total fertilization cost, a model is constructed based on the system's built-in library of selectable fertilizer varieties and nutrient content parameters, combined with the correlation between the unit cost and application rate of different fertilizer varieties. The calculation formula is as follows: (in, The total cost of fertilization, For the number of fertilizer varieties, For the first The unit cost of fertilizer For the first (Amount of fertilizer applied);

[0190] Establish a quality and yield objective function: With the goal of maximizing the combined score of expected fruit quality and yield, the initial nutrient requirement is used as the baseline input. The function is constructed by associating the pomelo tree age and phenological stage information of the grid cells. The calculation formula is as follows: (in, The overall score is based on quality and quantity. , , These refer to the application rates of nitrogen, phosphorus, potassium, and magnesium fertilizers, respectively. Tree age It is the phenological period. This is the production forecast function. For quality prediction function, , These are the weighting coefficients for yield and quality, and );

[0191] Nutrient cycling constraints are set: Combining pomelo planting agronomic knowledge and environmental maintenance requirements, a multi-dimensional constraint system is set, including nutrient loss coefficient constraints determined based on environmental impact coefficient vector, soil nutrient balance constraints set based on soil basic fertilizer supply characteristic vector, and safe application threshold for single fertilization set based on agronomic knowledge, thus defining the feasible range of fertilization decisions.

[0192] Multi-objective optimization solution: The initial nutrient requirements of each grid cell, the collaborative diagnostic feature vector, the optional fertilizer variety library and nutrient content parameters are input into the multi-objective optimization solver. Under the premise of satisfying the nutrient cycle constraint, the fertilizer cost model and the quality and yield objective function are solved in parallel. The non-dominated solution is screened out through Pareto optimal front analysis to form an optimal solution set that balances cost and yield and quality objectives.

[0193] Generate zonal variable fertilization prescriptions: For each grid cell, select a recommended solution from the optimized solution set that is suitable for the actual needs of pomelo planting, determine the specific fertilizer variety combination and the precise application amount of each variety, and determine the recommended operation time by combining the pomelo phenological information corresponding to the grid cell, and integrate all information to generate zonal variable fertilization prescriptions for each grid cell.

[0194] Step Six: Implementation and Visualization of Fertilization Decisions

[0195] This step is executed by the system's decision execution and interaction module. Its core objective is to transform the abstract fertilizer prescription into an executable instruction set for intelligent fertilization machinery, while simultaneously generating a visual agricultural guidance plan that growers can intuitively understand, thus enabling the practical implementation of fertilization decisions. Specific operations include:

[0196] Fertilizer prescription analysis and element extraction: Analyze the fertilization prescriptions generated by the dynamic zoning decision engine, extract the core operational elements such as fertilizer type, precise dosage, recommended operation time, and geographic coordinates for each grid unit, and classify and standardize them according to geographic coordinates;

[0197] Convert to intelligent fertilizer application machinery drive instruction set: Based on the extracted operation elements, construct a structured instruction sequence centered on the geographic coordinates of the grid unit, including fertilizer category code, target application amount and timestamp; according to the system's preset intelligent fertilizer application machinery model and communication protocol, convert the structured instruction sequence into a drive instruction set that the machinery control system can directly parse and execute, and complete the logic and parameter verification and optimization of the instruction set;

[0198] Generate visualized agricultural guidance schemes: Based on zonal variable fertilizer prescriptions and combined with geographic information system parameters of planting areas, draw distribution maps containing differentiated fertilizer application rates for each grid unit; integrate fertilizer elements by region to generate agricultural guidance cards with associated recommended operation times; complete the format standardization of the visualized schemes to adapt to online viewing, downloading and printing needs;

[0199] Command issuance and scheme push: Establish an encrypted communication connection with the intelligent fertilization machinery control terminal, issue the optimized drive command set to the control terminal; push the visualized agricultural guidance scheme to the grower's user interface, and record the transmission status of command issuance and scheme push in real time, and automatically trigger a retransmission mechanism for content that fails to be transmitted.

[0200] Step 7: Feedback on fertilization effects and self-calibration of model parameters:

[0201] This step is the closed-loop optimization stage of the entire decision-making method. It is executed by the system's dynamic partitioning decision engine in conjunction with multi-source data perception, spatiotemporal fusion, and feature generation modules. The core objective is to perform parameter self-calibration on the pomelo fertilizer requirement prediction model based on the actual feedback after fertilization, continuously improving the model's prediction accuracy, and ensuring that fertilization decisions always adapt to changes in the soil, trees, and environment of the planting area. Specific operations include:

[0202] Acquire a new round of multi-source data and generate updated collaborative diagnostic feature vectors: After fertilization is completed, the multi-source data sensing module acquires a new round of canopy-scale remote sensing spectral data and meteorological environmental data of the planting area. Combined with historical soil physicochemical property data, the spatiotemporal alignment and fusion analysis process of the spatiotemporal fusion and feature generation module generates updated collaborative diagnostic feature vectors for each grid unit, reflecting the actual state of the tree, soil and environment after fertilization.

[0203] Collect fertilization effect feedback data: Collect actual yield data and quality test data of the corresponding areas of historical fertilization prescriptions after fruit harvest, standardize and verify the data, remove invalid data, and generate effect feedback data that meets the model calibration requirements;

[0204] Generate prediction deviation signal: Compare the effect feedback data with the quality and yield prediction targets set before the implementation of historical fertilization prescriptions, calculate the deviation value, and generate a prediction deviation signal based on the sign and magnitude of the deviation value to quantify the prediction error of the pomelo fertilizer requirement prediction model.

[0205] Solving for model parameter adjustment: The prediction bias signal and the updated co-diagnostic feature vector are input into the model parameter correction algorithm. The algorithm calculates the parameter adjustment of the multi-output prediction network in the pomelo fertilizer requirement prediction model based on the direction and amplitude of the bias signal and the actual response state of the tree reflected by the updated feature vector.

[0206] Model parameter iterative update: Using the parameter adjustment amount obtained from the solution, the internal parameters of the pomelo fertilizer requirement prediction model are iteratively updated layer by layer to complete the self-correction of the model parameters. The updated model will be directly applied to the next round of pomelo fertilizer requirement prediction, forming a sustainable optimization closed loop of prediction-decision-execution-feedback-correction.

[0207] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source data fusion-based intelligent fertilizer application decision system for pomelos, characterized by: The application relates to a dynamic subarea variable fertilization prescription generation method based on multi-source data perception, which comprises the following steps: A multi-source data perception module is used to acquire the spatio-temporal data of a pomelo planting area, including soil physicochemical property data with geographic coordinate information, tree crown scale remote sensing spectral data and meteorological environment data; A spatio-temporal fusion and feature generation module is in communication connection with the multi-source data perception module and is used to: align the soil physicochemical property data, the tree crown scale remote sensing spectral data and the meteorological environment data in space and time, and map the data to a unified grid geographic information system unit; based on an agricultural knowledge model, fused analysis is conducted on the aligned data in the same geographic information system unit, and a collaborative diagnosis feature vector reflecting at least the soil basic fertilizer supply capacity, the real-time nutrition status of a tree body and the environmental factor influence coefficient is generated; a dynamic subarea decision engine is in communication connection with the spatio-temporal fusion and feature generation module and is used to: receive the collaborative diagnosis feature vector and input the feature vector into a pre-trained pomelo fertilizer requirement prediction model to obtain the initial nutrient requirement of each unit; an integrated fertilization cost model, a quality and yield target function and a nutrient circulation constraint condition are used to take the grid geographic information system unit as a basic decision unit, and the initial nutrient requirement is solved through multi-objective optimization to generate a subarea variable fertilization prescription containing a fertilization variety, a precise amount and a recommended operation time; wherein the pomelo fertilizer requirement prediction model is parameter self-corrected according to historical fertilization prescription corresponding effect feedback data and the updated collaborative diagnosis feature vector; a decision execution and interaction module is in communication connection with the dynamic subarea decision engine and is used to convert the subarea variable fertilization prescription into an instruction set or a visualized farming guidance scheme which can be parsed and executed by an intelligent fertilization machine and is output.

2. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 1, characterized in that: The multi-source data perception module is used to acquire the spatio-temporal data of a pomelo planting area, and specifically comprises the following steps: dynamically determining and outputting a geographic coordinate set of soil sampling points according to a historical yield map, a topographic map and a preset sampling density; collecting soil samples at corresponding positions according to the geographic coordinate set and sending the soil samples to a laboratory for physicochemical property analysis to generate soil physicochemical property data with geographic coordinate information; based on the geographic coordinate set, high-resolution remote sensing images covering the pomelo planting area are called from a remote sensing platform, and pixel spectral information of a pomelo tree crown corresponding to each sampling point is extracted according to coordinate positioning to generate tree crown scale remote sensing spectral data; according to the geographic range covered by the geographic coordinate set, multi-source meteorological observation data in the range are acquired from a meteorological data server, and the multi-source meteorological observation data are fused and interpolated to generate grid meteorological environment data matched with each geographic coordinate point.

3. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 1, characterized in that: The soil physicochemical property data, the tree crown scale remote sensing spectral data and the meteorological environment data are aligned in space and time and mapped to a unified grid geographic information system unit, which comprises the following steps: time reference normalization is conducted on the soil physicochemical property data, the tree crown scale remote sensing spectral data and the meteorological environment data respectively to synchronize the time information of all the data to the same observation period; the soil physicochemical property data, the tree crown scale remote sensing spectral data and the meteorological environment data after time normalization are converted through geographic coordinates and unified to a preset geographic coordinate system; Based on the geographic coordinate system, the pomelo planting area is divided into multiple grid units of the same size, forming a unified grid geographic information system unit; For each grid unit, the soil physical and chemical property data and meteorological environment data are interpolated from the sampling point position to the center of the grid unit using a spatial interpolation algorithm, and the tree crown scale remote sensing spectral data are matched to the corresponding grid unit through pixel resampling; The interpolated and resampled soil physical and chemical property data, tree crown scale remote sensing spectral data and meteorological environment data are integrated by grid unit to generate aligned soil physical and chemical property data set, tree crown scale remote sensing spectral data set and meteorological environment data set in each grid unit.

4. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 3, characterized in that: Based on the agronomic knowledge model, the aligned data in the same geographic information system unit are fused and analyzed to generate at least a collaborative diagnosis feature vector reflecting the soil basic fertilizer supply capacity, the real-time nutrient status of the tree and the influence coefficient of the environmental factors, including: According to the soil physical and chemical property data set, the soil organic matter content, pH value, available nitrogen content, available phosphorus content and available potassium content are extracted as basic indexes; The basic indexes are input into the preset soil nutrient supply capacity evaluation model to output the soil basic fertilizer supply feature vector representing the soil basic fertilizer supply capacity of the grid unit; The tree crown scale remote sensing spectral data set is input into the preset pomelo tree crown spectrum analysis model to extract the tree crown spectrum derived parameters reflecting the relative content of chlorophyll, carotenoid index and water stress index; The soil basic fertilizer supply feature vector and the tree crown spectrum derived parameters are fused and input into the preset tree nutrient surplus and deficiency diagnosis model to generate the tree nutrient status vector reflecting the real-time surplus and deficiency status of nitrogen, phosphorus, potassium and magnesium elements in the tree; The meteorological environment data set is input into the preset environmental nutrient migration and absorption influence model to calculate the environmental influence coefficient vector representing the influence degree of temperature, precipitation and light conditions on the fertilizer effectiveness and tree absorption efficiency; The soil basic fertilizer supply feature vector, the tree nutrient status vector and the environmental influence coefficient vector are integrated, and the information of pomelo tree age and phenology corresponding to the grid unit is associated, and the collaborative diagnosis feature vector is generated through the preset feature fusion rule.

5. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 1, characterized in that: The collaborative diagnosis feature vector is received and input into the pre-trained pomelo fertilizer requirement prediction model to obtain the initial nutrient requirement of each unit, including: The collaborative diagnosis feature vector is input into the pomelo fertilizer requirement prediction model, and the pomelo fertilizer requirement prediction model includes a multi-output prediction network; The multi-output prediction network first decouples the input collaborative diagnosis feature vector to extract a sub-feature set strongly related to nitrogen demand, phosphorus demand and potassium demand; The nitrogen demand prediction sub-network in the multi-output prediction network fuses the soil basic fertilizer supply feature vector, the tree nutrient status vector and the environmental influence coefficient vector in the collaborative diagnosis feature vector based on the sub-feature set related to nitrogen demand to output the initial nitrogen fertilizer requirement of the current grid unit; The phosphorus demand prediction sub-network in the multi-output prediction network fuses the soil basic fertilizer supply feature vector, the tree nutrition state vector, and the environmental influence coefficient vector in the collaborative diagnosis feature vector based on a set of sub-features related to phosphorus demand, and outputs the initial phosphorus fertilizer demand of the current grid cell; The potassium demand prediction sub-network in the multi-output prediction network fuses the soil basic fertilizer supply feature vector, the tree nutrition state vector, and the environmental influence coefficient vector in the collaborative diagnosis feature vector based on a set of sub-features related to potassium demand, and outputs the initial potassium fertilizer demand of the current grid cell; The initial nitrogen fertilizer demand, the initial phosphorus fertilizer demand, and the initial potassium fertilizer demand are collected to jointly constitute the initial nutrient demand of the corresponding grid cell.

6. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 5, characterized in that: The initial nutrient demand is multi-objective optimized by taking the fertilization cost model, the quality and yield objective function, and the nutrient cycling constraint condition as the basic decision unit of the grid-based geographic information system, to generate a variable partition fertilization prescription containing fertilization varieties, precise use amount, and recommended operation time, including: A fertilization cost model is established to minimize the total fertilization cost, which is calculated from the unit cost of different fertilizer varieties and their use amount; A quality and yield objective function is established to maximize the comprehensive score of expected fruit quality and yield, which takes the initial nutrient demand of each grid cell as the benchmark input, and is constructed in association with the tree age and phenology information in the collaborative diagnosis feature vector; Nutrient cycling constraint conditions are set, including nutrient loss coefficient constraints determined based on the environmental influence coefficient vector, soil nutrient balance constraints set based on the soil basic fertilizer supply feature vector, and single fertilization safe use threshold set based on agronomic knowledge; The initial nutrient demand, the collaborative diagnosis feature vector, the optional fertilizer variety library, and the fertilizer nutrient content parameters are input into the multi-objective optimization solver as independent decision units for each grid cell; The multi-objective optimization solver parallelly solves the fertilization cost model and the quality and yield objective function under the premise of meeting the nutrient cycling constraint conditions, and obtains an optimal solution set that balances cost and yield quality through Pareto optimal frontier analysis; For each grid cell, a recommended solution is selected from the optimal solution set, and a variable partition fertilization prescription containing specific fertilizer variety combination, precise use amount of each variety, and recommended operation time based on phenology is generated according to the recommended solution.

7. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 6, characterized in that: The honey pomelo fertilizer demand prediction model performs parameter self-correction based on the updated collaborative diagnosis feature vector and the effect feedback data corresponding to the historical fertilization prescription, including: After the fertilization operation is completed, the multi-source data perception module obtains new round of tree crown scale remote sensing spectral data and meteorological environment data of the honey pomelo planting area; Based on the new round of tree crown scale remote sensing spectral data and meteorological environment data, combined with historical soil physical and chemical property data, the updated collaborative diagnosis feature vector is generated through the spatio-temporal fusion and feature generation module; The yield data and quality detection data of the corresponding area after fruit harvesting are collected as effect feedback data; The effect feedback data and the quality and yield prediction target set before the execution of the historical fertilization prescription are compared to generate a prediction deviation signal; The prediction deviation signal and the updated collaborative diagnosis feature vector are input into a model parameter correction algorithm; The model parameter correction algorithm calculates a parameter adjustment amount of the multi-output prediction network in the citrus fertilizer requirement prediction model according to the direction and amplitude of the prediction deviation signal and in combination with the actual response state of the tree body reflected by the updated collaborative diagnosis feature vector; The internal parameters of the citrus fertilizer requirement prediction model are iteratively updated by using the parameter adjustment amount, and parameter self-correction is completed.

8. The multi-source data fusion based pomelo intelligent fertilization decision system according to claim 1, characterized in that: The decision execution and interaction module is used for converting the partition variable fertilization prescription into an instruction set that can be parsed and executed by an intelligent fertilization machine or a visualized farming guidance scheme and outputting, including: The partition variable fertilization prescription generated by the dynamic partition decision engine is parsed, and the operation elements corresponding to each grid unit in the partition variable fertilization prescription, such as the fertilization variety, the precise amount and the recommended operation time, are extracted; Based on the extracted operation elements, a structured instruction sequence is constructed, which takes the geographical coordinates of the grid unit as the center and contains the fertilizer variety code, the target application amount and the time stamp; According to the preset intelligent fertilization machine model and the communication protocol, the structured instruction sequence is converted into a driving instruction set that can be directly parsed and executed by the control system of the intelligent fertilization machine; At the same time, based on the partition variable fertilization prescription, a differentiated fertilization amount distribution map of each grid unit in the citrus planting area and a farming guidance card associated with the recommended operation time are generated as a visualized farming guidance scheme; The driving instruction set is sent to the control terminal of the intelligent fertilization machine, and the visualized farming guidance scheme is pushed to the user interaction interface.

9. A method for intelligent fertilization decision of pomelo based on multi-source data fusion, characterized in that: The method is executed by the citrus intelligent fertilization decision system based on multi-source data fusion according to any one of claims 1-8, and includes the following steps: Obtain the spatio-temporal data of the citrus planting area, including soil physical and chemical property data, tree crown scale remote sensing spectral data and meteorological environment data with geographical coordinate information; The obtained soil physical and chemical property data, tree crown scale remote sensing spectral data and meteorological environment data are spatio-temporally aligned and mapped to a unified grid geographic information system unit; Based on the agronomic knowledge model, the aligned data in the same grid geographic information system unit are fused and analyzed to generate a collaborative diagnosis feature vector reflecting at least the soil basic fertilization capacity, the real-time nutritional status of the tree body and the environmental factor influence coefficient; The collaborative diagnosis feature vector is input into a pre-trained citrus fertilizer requirement prediction model to obtain the initial nutrient demand of each grid geographic information system unit; Integrate the fertilization cost model, the quality and yield objective function and the nutrient recycling constraint condition, take the grid geographic information system unit as the basic decision unit, and perform multi-objective optimization on the initial nutrient demand to generate a partition variable fertilization prescription containing the fertilization variety, the precise amount and the recommended operation time; The partition variable fertilization prescription is converted into an instruction set that can be parsed and executed by an intelligent fertilization machine or a visualized farming guidance scheme and is outputted; Wherein, after the fertilization operation is completed, the citrus fertilizer requirement prediction model is parameter self-corrected according to the effect feedback data corresponding to the historical fertilization prescription and the updated collaborative diagnosis feature vector.

Images