Fertilizer demand analysis method based on AI planting plan

Abstract

The invention belongs to the technical field of intelligent agriculture, and discloses a fertilizer demand analysis method based on an AI planting plan. Comprising the steps of collecting crop whole growth cycle data, performing preprocessing analysis to obtain a multi-source structured data set, and inputting a planting analysis model to obtain an AI planting plan; combining agronomic knowledge to establish a dynamic agricultural knowledge graph, and fusing data to obtain a nutrient demand matrix; the method comprises the following steps: collecting soil nutrient space heating power and historical fertilization time sequence data, and analyzing to obtain space-time nutrient demand feature vectors; an initial matrix is generated through attention mechanism fusion, a fertilizer demand matrix is obtained through reinforcement learning optimization, and dynamic calibration is performed based on leachate monitoring after the fertilizer demand matrix is converted into fertilization operation; according to the method, nutrient waste or insufficient supply caused by the fact that genotype difference is not considered in a general fertilization mode is avoided, the crop yield is remarkably increased, and meanwhile the nutrient leaching loss rate and the fertilizer cost are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart agriculture technology, more particularly, the present application relates to a fertilizer demand analysis method based on AI planting plan. BACKGROUND

[0002] In current agricultural fertilization management, the traditional mode relies on the experience of farmers or general fertilization schemes, which has the limitation of "one size fits all" and does not consider the differences in crop genotypes (such as the different nutrient absorption characteristics of nitrogen-efficient / phosphorus-efficient genotypes), which can easily lead to mismatch between nutrient supply and crop demand, resulting in problems such as fertilizer waste, nutrient leaching, and soil degradation.

[0003] A method and system for evaluating the effect of fertilization in tobacco fields based on big data are disclosed in Chinese patent CN118798491B, which includes: acquiring multi-source data of the tobacco field, including soil physical and chemical properties (pH, available nitrogen, phosphorus, and potassium, etc.), tobacco plant phenotype data (plant height, leaf area, tobacco grade), meteorological environmental data (temperature, rainfall), historical fertilization data (fertilization amount, ratio, time), and combining remote sensing images to obtain macroscopic growth information of the tobacco field; cleaning, standardizing, and integrating the collected data to construct a tobacco field fertilization database, and simultaneously using knowledge graph to correlate static indicators such as soil, plants, and fertilization; establishing an evaluation model to correlate fertilization data with indicators such as tobacco yield, tobacco quality, and soil nutrient changes, and calculating parameters such as fertilizer utilization rate and nutrient surplus rate to judge the pros and cons of the fertilization scheme. The invention collects, processes, and analyzes multi-source data such as soil physical samples, plant physical samples, real-time soil monitoring data, hyperspectral remote sensing image data, and unmanned aerial vehicle image data, comprehensively evaluates the effect of fertilization, and uses more abundant data to improve the comprehensiveness of the evaluation of fertilization effect. The use of knowledge graph technology to evaluate and analyze soil properties improves the processing efficiency of a large number of data indicators, and the use of knowledge graph to mine the correlation between different indicators improves the accuracy of the evaluation of fertilization effect.

[0004] Although the above method can meet most scenarios, research and practical application of the above method and prior art have found that the above method and prior art at least have the following defects:

[0005] It is unable to customize fertilization strategies for different genotypes (such as japonica rice and indica rice, low potassium tolerant varieties and sensitive varieties), and only a "one size fits all" general evaluation standard can be used.

[0006] In view of this, the present application proposes a fertilizer demand analysis method based on AI planting plan to solve the above problems. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purpose, the present application provides the following technical solution: a fertilizer demand analysis method based on AI planting plan, comprising the following steps:

[0008] Crop data of the whole growth cycle of the crop is collected, preprocessed and analyzed to obtain a multi-source structured data set;

[0009] The multi-source structured data set is taken as an input of a planting analysis model to obtain an AI planting plan;

[0010] A dynamic agricultural knowledge graph is constructed according to agricultural knowledge and the AI planting plan, and the dynamic agricultural knowledge graph is fused with the multi-source structured data set to analyze and obtain a nutrient demand matrix;

[0011] Soil nutrient spatial thermal data and historical fertilization timing data are collected to analyze and obtain a spatio-temporal nutrient demand feature vector;

[0012] The nutrient demand matrix and the spatio-temporal nutrient demand feature vector are fused through an attention mechanism to generate an initial nutrient matrix, which is adjusted through a reinforcement learning optimizer, a reward function is designed according to a predicted crop yield and a predicted nutrient leaching rate mixed prediction of fertilizer cost, and a fertilizer demand matrix is obtained;

[0013] The fertilizer demand matrix is converted into actual fertilization operations, and dynamic calibration is performed based on leachate monitoring.

[0014] Further, the multi-source structured data set includes rhizosphere data, crown layer data, environmental data and gene phenotype data;

[0015] The rhizosphere data includes pH value, ion concentration, specifically nitrogen ion, phosphorus ion and potassium ion, and conductivity of the rhizosphere; the crown layer data includes red edge normalized vegetation index, chlorophyll fluorescence parameter and leaf yellowing area proportion; the environmental data includes photosynthetically active radiation, precipitation intensity, accumulated temperature and wind speed; and the gene phenotype data includes genotype markers and corresponding expression characteristics;

[0016] The method for obtaining the crown layer data includes:

[0017] The multi-spectral data of the crop is converted into digital quantitative values, which are calibrated into actual reflectivity through a preset standard gray board;

[0018] The digital quantitative values of the pure sky light area are obtained, and the average value and the standard deviation of the digital quantitative values of the pure sky light area are counted to obtain the average quantitative value and the average standard deviation;

[0019] The dark field digital quantitative value in the absence of light is obtained, and the average value is counted as the dark digital quantitative value, and the corrected sky light quantitative value is calculated according to the average quantitative value and the dark digital quantitative value;

[0020] The sky light radiation brightness is calculated according to the radiation scaling gain coefficient, the radiation scaling offset coefficient and the corrected sky light quantitative value;

[0021] Extract the average quantization value of the center of the gray board image, calculate the DN value after dark field correction, calculate the corresponding gray board radiation brightness, and combine the sky light radiation brightness, the gray board reflectivity of the preset standard gray board to obtain the sky light reflectivity;

[0022] Based on the sky light reflectivity, the actual reflectivity is corrected to obtain the corrected reflectivity;

[0023] According to the corrected reflectivity and the infrared reflectivity, the red edge normalized vegetation index is calculated and obtained;

[0024] The initial fluorescence and the maximum fluorescence are collected, the variable fluorescence is calculated according to the initial fluorescence and the maximum fluorescence, and the ratio of the variable fluorescence and the maximum fluorescence is calculated to obtain the chlorophyll fluorescence parameter;

[0025] The crop canopy image is collected, the Gaussian filter is used to remove the crop canopy image noise, and the crop canopy image is converted from the RGB color space to the HSV color space; the HSV range of the yellowing area is counted, and the threshold method is used to generate a binary image; the total number of yellowing area pixels and the total number of leaf pixels are counted, and the yellowing area ratio is calculated.

[0026] Further, the corresponding expression characteristics include relative expression amount, nitrogen transport rate, potassium transport rate and enzyme activity;

[0027] The method for obtaining the gene phenotype data comprises:

[0028] The genomic DNA of the crop is collected, the target gene is screened, the target gene can select a nutrient utilization related gene with a clear function, and specific primers are designed for the SNP site of the target gene;

[0029] The SNP alleles are distinguished by competitive allele-specific PCR to obtain the typing results, the target gene and the SNP site genotype combination are formed to form a genotype marker;

[0030] The relative expression amount of the target gene in the root system is detected by real-time quantitative PCR;

[0031] The 15N and 42K isotopes are labeled by 15N and 42K isotope labeling method to detect the 15N and 42K abundance, and the nitrogen transport rate and the potassium transport rate are calculated;

[0032] The protein expression amount is detected by specific antibody;

[0033] The pNP standard solution is configured, the absorbance values corresponding to different pNP standard solution concentrations are detected by the sodium p-nitrophenyl phosphate colorimetric method, a standard curve with the pNP concentration as the horizontal coordinate and the absorbance value as the vertical coordinate is constructed, a regression equation is fitted and obtained, the pNP solution concentration corresponding to the current absorbance value is calculated according to the regression equation, and the enzyme activity is calculated according to the total volume of the reaction system, the enzyme liquid dilution multiple, the protein mass in the reaction system, the incubation time and the enzyme liquid volume.

[0034] Further, the method for obtaining the nitrogen transport rate comprises:

[0035] The crop after the starvation treatment is moved into the 15N-labeled nutrient solution, and is cultured at 25 DEG C for 2 hours under the light for 12 hours to obtain a labeled crop;

[0036] The root system of the labeled crop is washed with the nitrogen-free nutrient solution for 3 times to remove the residual 15N on the surface, the root system and the above-ground part are separated, are killed at 105 DEG C for 30 minutes, are dried at 70 DEG C until the weight is constant, are crushed and are passed through a 100-mesh sieve to obtain a powder sample;

[0037] 10 mg of the powder sample is taken, and the 15N abundance is detected by an elemental analyzer-isotope ratio mass spectrometer;

[0038] The nitrogen transport rate is calculated according to the 15N abundance, the culture time and the dry weight of the root system.

[0039] Further, the method for constructing the dynamic agricultural knowledge graph comprises:

[0040] The knowledge components are defined, including entities, relationships, attributes and gene regulation intensities;

[0041] The entities include gene phenotype entities, crop entities, environment entities, nutrient entities and agronomic entities;

[0042] The relationships include crop-gene phenotype relationships, crop-environment relationships, crop-nutrient relationships, crop-agronomic relationships, gene phenotype-nutrient relationships, environment-gene phenotype relationships, environment-nutrient relationships and agronomic-nutrient relationships;

[0043] The attributes include gene attributes, environment attributes and agronomic attributes;

[0044] The gene regulation intensity is obtained by weighted calculation according to the corresponding expression characteristics;

[0045] Whether the entities pointing to the same real object in different data sources are the same is judged based on the similarity calculation, whether the data of the same entity in different data sources conflicts is detected, and the conflicting data is resolved;

[0046] After resolving conflicts, knowledge components are used to build a knowledge graph. The knowledge graph is updated based on real-time data to obtain a dynamic agricultural knowledge graph, which is then stored in a graph database.

[0047] Furthermore, methods for obtaining the nutrient requirement matrix include:

[0048] Multi-source structured datasets are converted into knowledge components using predefined knowledge extraction rules, including mapping rules, natural language processing techniques, and convolutional neural networks. Entity nodes are constructed based on the knowledge components, and the relationship chain is activated through a message passing algorithm to output a knowledge-driven nutrient demand feature vector. A nutrient demand matrix is ​​then constructed based on the nutrient demand feature vector.

[0049] Furthermore, methods for obtaining nutrient requirement feature vectors include:

[0050] The correlation between entities is calculated based on the latest data using the Pearson correlation coefficient, and the weights are updated according to the correlation.

[0051] For each entity node, the top K neighbors are collected in descending order of relation weight, and the neighbor node information is merged using the mean aggregation method;

[0052] When new data input causes entity attributes to be updated, incremental updates to the relationship chain are triggered, and the affected subgraphs are recalculated.

[0053] Embedding vectors of crops, gene phenotypes, environment, and agronomic entities are extracted from the dynamic agricultural knowledge graph after message transmission.

[0054] Relational attention is calculated based on the entity's attention weight, cosine similarity, and relation weight. The original features are then weighted according to relational attention to obtain the nutrient demand feature vector.

[0055] Furthermore, methods for obtaining spatiotemporal nutrient demand feature vectors include:

[0056] Soil nutrient data is collected and used as input to a convolutional neural network to obtain spatial distribution features. Historical fertilization time series data is obtained and used as input to a time series analysis model to obtain time-dependent features. Spatial distribution features and time-dependent features are concatenated to obtain a spatiotemporal nutrient demand feature vector.

[0057] Furthermore, methods for obtaining the fertilizer demand matrix include:

[0058] By fusing dual-channel features through an attention mechanism, an initial nutrient matrix is ​​generated. This matrix is ​​then adjusted using a reinforcement learning optimizer. The multi-source structured dataset, soil nutrient spatial thermodynamic data, and historical fertilization time series data are used as state vectors, while the fertilizer ratio adjustment dimension and fertilization time series adjustment dimension are used as action vectors. A reward function is designed based on the predicted crop yield, predicted nutrient leaching rate, and predicted fertilizer cost to obtain the fertilizer demand matrix.

[0059] Furthermore, methods for achieving dynamic calibration based on leachate monitoring include:

[0060] Collect the ion concentration of the root zone leachate F hours after fertilization, and calculate the adjusted fertilizer utilization coefficient based on the initial fertilizer utilization rate, the target leachate nutrient concentration, and the measured leachate nutrient concentration.

[0061] If the target leachate nutrient concentration is greater than the measured leachate nutrient concentration, the amount of fertilizer applied in the next round will be adjusted according to the adjusted fertilizer utilization rate coefficient.

[0062] The technical effects and advantages of the fertilizer demand analysis method based on AI planting plan of this invention are as follows:

[0063] This invention collects multi-source structured datasets covering crop genotypes, canopy data, etc., and inputs them into an LSTM planting analysis model whose network parameters are optimized using a nature-inspired optimization algorithm. This trains a model capable of accurately learning the association patterns between different genotypes and planting plans to generate AI planting plans tailored to specific genotypes. Simultaneously, it constructs a dynamic agricultural knowledge graph containing genotype entities, gene regulation strength, and planting plan attributes. Through weighted fusion of gene regulation strength, it resolves conflicts between multi-source data, clearly depicting the specific association patterns between different genotypes and nutrient requirements and agronomic measures. Finally, it fuses knowledge-driven nutrient requirement feature vectors and spatiotemporal nutrient requirement feature vectors using an attention mechanism, combining them to predict crop growth. A reinforcement learning optimizer, targeting crop yield, nutrient leaching rate, and fertilizer cost, generates an initial nutrient matrix and adjusts it into a specific nutrient requirement matrix. This matrix is ​​then transformed into actual fertilization operations and dynamically calibrated through leachate monitoring. This effectively breaks away from the traditional "one-size-fits-all" general fertilization assessment standard, enabling customized fertilization for different crop genotypes. It avoids nutrient waste or insufficient supply caused by the failure to consider genotype differences under the general fertilization model, and significantly increases crop yield while reducing nutrient leaching rate and fertilizer cost. Furthermore, it can continuously adapt to the nutrient absorption and utilization characteristics of different crop genotypes through dynamic calibration, ultimately achieving the technical effect of precision fertilization, coordinated agricultural production, and ecological environmental protection. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the fertilizer demand analysis method based on AI planting plan of the present invention;

[0065] Figure 2 This is a schematic diagram of the data flow in this invention;

[0066] Figure 3 This is a schematic diagram of the method for obtaining gene phenotype data according to the present invention;

[0067] Figure 4 This is a schematic diagram of the method for constructing a dynamic agricultural knowledge graph according to the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example 1

[0070] Please see Figure 1 - Figure 2 As shown, this embodiment provides a fertilizer demand analysis method based on AI planting plans, including the following steps:

[0071] Collect crop data throughout the entire crop growth cycle, and perform preprocessing and analysis to obtain a multi-source structured dataset;

[0072] Multi-source structured datasets include rhizosphere data, canopy data, environmental data, and genotype data;

[0073] Rhizosphere data include rhizosphere pH, ion concentration (specifically nitrogen, phosphorus, and potassium ions), and electrical conductivity; canopy data include red-edge normalized vegetation index, chlorophyll fluorescence parameters, and the percentage of leaf yellowing area; environmental data include photosynthetically active radiation, precipitation intensity, accumulated temperature, and wind speed; and gene phenotypic data include genotype markers and their corresponding expression characteristics.

[0074] Methods for obtaining canopy data include:

[0075] The crop's multispectral data is converted into digital quantized values, which are then calibrated to actual reflectance using a preset standard gray card; for example, the actual reflectance... ,in, The digital quantization value is obtained by converting the analog electrical signal into a digital-to-analog converter. The calibration slope can be obtained by fitting a standard gray board. The calibration intercept can be obtained by fitting a standard gray board.

[0076] The digital quantization value of the pure sky light region is obtained, and the mean quantization value and mean standard deviation of the digital quantization value of the pure sky light region are calculated. The pure sky light region is an image region that does not contain any non-sky light interference elements such as crops, soil, shadows, clouds, and birds, and is used solely to capture and record natural sky light signals.

[0077] Obtain the dark field digital quantization value under no light conditions, and calculate the average value as the dark digital quantization value. Calculate the corrected skylight quantization value based on the average quantization value and the dark digital quantization value.

[0078] Sky radiance is calculated based on the radiometric calibration gain coefficient, radiometric calibration offset coefficient, and corrected skylight quantization value; for example, skylight radiance... ,in, The radiometric calibration gain coefficient is a calibration coefficient that is matched to a specific band (such as the multispectral band required for calculating skylight) and must be obtained through a standard radiometric calibration procedure. To correct the sky light quantization value; The radiation calibration offset coefficient is... The calibration intercept for band matching is also obtained through the standard radiometric calibration procedure.

[0079] The average quantized value at the center of the grayboard image is extracted, the DN value after dark field correction is calculated, and the corresponding grayboard radiance is obtained. The skylight reflectance is then calculated by combining the skylight radiance and the grayboard reflectance of a preset standard grayboard. (The grayboard radiance is then used to calculate the skylight reflectance.) ,in, The radiation calibration gain coefficient can be obtained by fitting a standard gray plate. This is the DN value after dark field correction. , The average quantization value at the center of the grayscale image. This refers to the DN value for the dark field. The radiometric calibration offset coefficient can be obtained by fitting a standard gray card; sky reflectance. ,in, Let be the reflectance of the gray panel, and be a known quantity. The brightness of the sky's radiant light;

[0080] The actual reflectance is corrected based on the skylight reflectance to obtain the corrected reflectance; for example, the corrected reflectance... ;

[0081] The red-edge normalized vegetation index is calculated based on the corrected reflectance and infrared reflectance; for example, the red-edge normalized vegetation index. , The corrected infrared reflectance; The red-edge band reflectivity is the corrected value.

[0082] Initial fluorescence and maximum fluorescence were collected. Variable fluorescence was calculated based on these two values. The ratio of variable fluorescence to maximum fluorescence was then calculated to obtain chlorophyll fluorescence parameters. ,in, Maximum fluorescence; Initial fluorescence; It exhibits variable fluorescence;

[0083] Crop canopy images were acquired, and Gaussian filtering was used to remove noise. The crop canopy images were then converted from the RGB color space to the HSV color space. The HSV range of the yellowed areas was statistically analyzed, and a binary image was generated using a thresholding method. The total number of pixels in the yellowed areas and the total number of pixels in the leaves were counted to calculate the proportion of yellowed area. The proportion of yellowed leaf area was then calculated, as shown below. ,in, This represents the total number of pixels in the yellowed region. This represents the total number of pixels on the blades.

[0084] The method for obtaining canopy data involves multispectral data calibration, sky reflectance correction, and calculation of key indicators. The resulting corrected red-edge normalized vegetation index (NMR), chlorophyll fluorescence parameters, and leaf yellowing area percentage can accurately capture the phenotypic differences in nutrient response among different crop genotypes. For example, nitrogen-efficient genotypes exhibit a higher corrected near-infrared to red-edge reflectance ratio, more stable photosynthetic efficiency, and a lower leaf yellowing area percentage under the same nitrogen conditions. In contrast, ordinary or nutrient-sensitive genotypes show a decrease in the corrected NMR, a reduction in chlorophyll fluorescence parameters, and an increase in yellowing area under the same nutrient supply. These differential phenotypic data can serve as direct phenotypic evidence for the specific nutrient requirements of different crop genotypes, breaking the limitations of traditional "one-size-fits-all" assessment standards that rely solely on general soil or plant mean data. This allows subsequent fertilization strategies to be precisely matched to the nutrient absorption and utilization characteristics of different genotypes, effectively solving the technical problem of not being able to customize fertilization strategies for different genotypes.

[0085] Reference Figure 3 Methods for obtaining gene phenotype data include:

[0086] Collect genomic DNA from crops, screen for target genes, and select nutrient utilization-related genes with known functions. Design specific primers for the SNP sites of the target genes.

[0087] Methods for collecting crop genomic DNA include:

[0088] Take 50 mg of pretreated crop sample, put it into a centrifuge tube containing 2 steel balls, add 500 μL of CTAB buffer, and grind it at 60 Hz for 2 min using a tissue homogenizer to break the cell wall, release nucleic acid, and obtain homogenate.

[0089] The CTAB extraction buffer was prepared as follows: 2g CTAB, cetyltrimethylammonium bromide, for lysing cell membranes; 10mL 1mol / L Tris-HCl, pH 8.0, for maintaining pH stability; 4mL 0.5mol / L EDTA, pH 8.0, for inhibiting nucleases; 28mL 5mol / L NaCl, for dissolving DNA; and sterile water was added to a final volume of 100mL. Before use, 0.2mL of β-mercaptoethanol was added to prevent phenol oxidation and protect the DNA.

[0090] Transfer the homogenate to a new centrifuge tube and incubate at 65°C for 30 minutes. Gently invert the tube once every 10 minutes to mix and promote cell lysis.

[0091] Add 500 μL of chloroform-isoamyl alcohol, gently invert 10 times, centrifuge at 4℃ and 12000 rpm for 10 min, and transfer the upper aqueous phase containing DNA to a new tube.

[0092] Add 400 μL of pre-cooled isopropanol, gently invert until a white flocculent precipitate appears, and let stand at -20°C for 30 min to promote DNA precipitation;

[0093] Centrifuge the new tube at 4℃ and 10000rpm for 5min, discard the supernatant, add 500μL of 75% ethanol to wash the precipitate twice, discard the ethanol, and air dry at room temperature to avoid complete drying which would make the DNA difficult to dissolve. Dissolve the DNA in 50μL of sterile water to obtain the crop's genomic DNA, which can be temporarily stored at 4℃.

[0094] It can perform purity and integrity testing on genomic DNA to ensure the validity of subsequent experiments;

[0095] Purity test: The ratio of OD260 / OD280 is measured by ultraviolet spectrophotometer. If the ratio is 1.8 to 2.0, it is acceptable. A ratio lower than 1.8 indicates protein contamination, and a ratio greater than 2.0 indicates RNA contamination. The two contaminants need to be treated with proteinase K or RNase A, respectively.

[0096] Integrity test: Take 2 μL of genomic DNA and electrophoresis it on a 1% agarose gel containing EB staining agent at 120V for 20 min. Observe it under a UV gel imaging system. If a single bright band without diffusion appears, it indicates that the genomic DNA is in good integrity.

[0097] The following principles should be followed when designing specific primers for SNP sites of a target gene:

[0098] The amplified region covers the SNP site of the target gene. If the SNP site is located in the middle 1 / 3 of the primer amplification fragment, it will facilitate subsequent sequencing or genotyping.

[0099] The primer parameters should be 18-25 bp in length, 55-62℃ in temperature, and the difference in temperature between the upstream and downstream primers should not exceed 2℃. The GC content should be 40%-60% to avoid hairpin structures and primer dimers.

[0100] Primers can be validated using NCBI BLAST to ensure that they only amplify the target gene and do not bind nonspecifically.

[0101] SNP alleles are distinguished by competitive allele-specific PCR to obtain genotyping results. The target gene is combined with the SNP locus genotype to form a genotype marker.

[0102] Methods for obtaining typing results include:

[0103] The reaction mixture consisted of 5 μL KASP Master Mix, 0.15 μL upstream primer, 0.15 μL downstream primer, 2.0 μL crop genomic DNA, and 2.7 μL sterile water.

[0104] The reaction system was pre-denatured at 94℃ for 15 min to activate Taq enzyme and denature DNA;

[0105] Perform touchdown cycles: denature the reaction system at 94℃ for 20s, anneal and extend it at 60℃-65℃ for 60s, and repeat the touchdown cycle ten times. The annealing temperature can be reduced by 0.5℃ in each cycle to improve primer specificity.

[0106] Fluorescence detection cycles were performed: the reaction system was denatured at 94℃ for 20s, annealed and extended at 60℃ for 60s, and the fluorescence detection cycles were repeated 26 times. The FAM / HEX fluorescence signal was detected after each cycle.

[0107] After the fluorescence detection cycle is completed, FAM / HEX fluorescence signals are acquired and analyzed to obtain the genotyping results; if only FAM signal is positive: the genotype is TT; if only HEX signal is positive: the genotype is CC; if both FAM and HEX signals are positive: the genotype is TC, i.e., heterozygous.

[0108] The relative expression level of the target gene in the root system was detected by real-time quantitative PCR.

[0109] RNA was extracted from the crop, and its purity and integrity were tested.

[0110] 1 μg of crop RNA was extracted and cDNA was synthesized using a reverse transcription kit. The cDNA synthesis can be carried out through the following reaction system: 1 μg of crop RNA, 1 μL of primer, 1 μL of enzyme, 4 μL of buffer, and water to a final volume of 20 μL; after reacting at 37°C for 15 min, the reaction was carried out at 85°C for 5 s to obtain cDNA.

[0111] qPCR reaction: A 20 μL qPCR reaction system was incubated at 95℃ for 30 s, and then the reaction was repeated 40 times at 95℃ for 5 s and at 60℃ for 30 s. This was used to construct a melting curve with a temperature range of 65℃ to 95℃, with a 5 s pause at each 0.5℃ increment. The qPCR reaction system consisted of 10 μL qPCR Mix, 0.4 μL each of forward and reverse primers, 2 μL cDNA, and 7.2 μL water.

[0112] The circulation thresholds of the target gene and the control gene were obtained based on melting curve analysis. The relative circulation threshold of the target gene was calculated, and the relative expression level was calculated based on this relative circulation threshold. (The last sentence appears to be incomplete and possibly refers to a different topic.) ,in, The cycle threshold of the target gene; The cycle threshold for the control gene;

[0113] The abundance of 15N was detected by 15N isotope labeling, and the transport activity was calculated to obtain the nitrogen transport rate. The method for obtaining the potassium transport rate is similar to that for obtaining the nitrogen transport rate, and will not be described in detail here.

[0114] The starved crops were transferred into a nutrient solution labeled with 15N and cultured for 2 hours at 25°C under 12 hours of light to obtain labeled crops.

[0115] The roots of the marked crop were rinsed three times with nitrogen-free nutrient solution to remove residual 15N from the surface. The roots and aboveground parts were separated, blanched at 105℃ for 30 min, dried at 70℃ to constant weight, pulverized and passed through a 100-mesh sieve to obtain powder samples.

[0116] Take 10 mg of powder sample and use an elemental analyzer-isotope ratio mass spectrometer to detect the 15N abundance.

[0117] The nitrogen translocation rate was calculated based on 15N abundance, culture time, and root dry weight. (Example: Nitrogen translocation rate) ,in, The abundance of 15N in crops; The abundance of 15N is the control crop; Nitrogen content of the control crop; Dry weight of control crop; For training time; The dry weight of the root system;

[0118] Protein expression levels were detected using specific antibodies.

[0119] Take the roots of the crop, add 500 μL of protein extraction buffer, grind in an ice bath for 30 min, centrifuge at 4℃ and 12000 rpm for 15 min, and take the supernatant.

[0120] The protein concentration in the supernatant was determined using a BCA kit;

[0121] Take 20 μg of protein sample, add 5× loading buffer, denature at 95℃ for 5 min, load onto SDS-PAGE gel, perform electrophoresis on a stacking gel at 80V for 30 min, and then perform electrophoresis on a separating gel at 120V for 90 min; where 5× indicates that the concentration of the loading buffer is 5 times the working concentration required for the experiment, and it needs to be diluted proportionally in actual use.

[0122] The protein on the gel was transferred to a PVDF membrane and blocked with 5% skim milk at room temperature for 2 hours to block non-specific binding sites.

[0123] Primary antibody (1:1000 dilution) was prepared using the blocked protein and incubated overnight at 4°C. The membrane was washed three times with TBST buffer for 10 min each time. Secondary antibody (1:5000 dilution) was prepared using the blocked protein and incubated at room temperature for 1 h. The membrane was washed three times with TBST.

[0124] ECL kit development, image capture, and ImageJ software for quantitative analysis of band gray values;

[0125] If the protein gray value of the GmPAP12-AA genotype is ≥1.3 times that of the GG genotype, it indicates that its GmPAP12 protein content is significantly higher.

[0126] A pNP standard solution was prepared, and the absorbance values ​​corresponding to different pNP standard solution concentrations were detected by the sodium nitrophenyl phosphate colorimetric method. A standard curve was constructed with pNP concentration as the x-axis and absorbance value as the y-axis. A regression equation was obtained by fitting the curve, and the pNP solution concentration corresponding to the current absorbance value was calculated based on the regression equation. The enzyme activity was calculated based on the total volume of the reaction system, the enzyme dilution factor, the protein content in the reaction system, the incubation time, and the enzyme volume. (Enzyme activity...) ,in, The pNP concentration is the value in the regression equation. The total volume of the reaction system; This refers to the dilution factor of the enzyme solution; Protein content; Incubation time; This represents the volume of the enzyme solution.

[0127] The method for obtaining gene phenotypic data involves screening target genes with clearly defined nutrient utilization functions, designing specific primers for competitive allele-specific PCR typing to obtain genotype markers, and combining this with real-time quantitative PCR to detect the relative expression levels of target genes, 15N isotope labeling to determine nitrogen and potassium transport rates, specific antibody detection of protein expression levels, and sodium p-nitrophenyl phosphate colorimetric method to calculate enzyme activity. This approach precisely reveals the molecular and phenotypic differences in key physiological processes of nutrient absorption and utilization among different crop genotypes from multiple dimensions, including genotyping, expression levels, nutrient transport activity, functional protein content, and enzyme catalytic efficiency. This differentiated gene phenotypic information provides precise genotype-level evidence for subsequent fertilization strategy development, breaking the limitations of a "one-size-fits-all" universal evaluation standard. It allows fertilization strategies to be customized according to the nutrient utilization characteristics of different crop genotypes, effectively solving the technical problem of not being able to tailor fertilization strategies for different genotypes.

[0128] By using multi-source structured datasets as input to the planting analysis model, an AI-powered planting plan can be obtained.

[0129] Training methods for planting analysis models include:

[0130] Pre-collect S groups of planting training data, which includes multi-source structured datasets and AI planting plans.

[0131] The planting training data is used as the input to the planting analysis model, and the AI ​​planting plan is used as the output of the planting analysis model. With the goal of minimizing the error between the output AI planting plan and the actual AI planting plan, the network parameters of the LSTM model are optimized through a natural heuristic optimization algorithm to obtain the network parameters that minimize the error between the control parameters output by the planting analysis model and the actual control parameters. The planting analysis model constructed with the corresponding network parameters is used as the trained planting analysis model.

[0132] The training method of the planting analysis model involves pre-collecting S sets of planting training data, including multi-source structured datasets and corresponding AI planting plans. The goal is to minimize the error between the AI ​​planting plan output by the model and the actual AI planting plan. Natural heuristic optimization algorithms are used to optimize the LSTM model network parameters, enabling the trained planting analysis model to accurately learn the correlation between multi-source structured data of different crop genotypes and AI planting plans adapted to their nutrient utilization characteristics. This allows the model to output customized AI planting plans when receiving multi-source structured datasets of different crop genotypes as input, effectively breaking the limitations of a "one-size-fits-all" general evaluation standard. This provides core model-level support for solving the technical problem of not being able to customize fertilization strategies for different genotypes.

[0133] A dynamic agricultural knowledge graph is constructed based on agronomic knowledge and AI planting plans. The dynamic agricultural knowledge graph is then integrated with multi-source structured datasets to analyze and obtain a nutrient requirement matrix.

[0134] Reference Figure 4 Methods for constructing dynamic agricultural knowledge graphs include:

[0135] Define knowledge components, including entities, relationships, attributes, and gene regulatory strength;

[0136] Entities include genetic phenotype entities, crop entities, environmental entities, nutrient entities, and agronomic entities;

[0137] The relationships include crop-gene phenotype relationships, crop-environment relationships, crop-nutrient relationships, crop-agronomic relationships, gene phenotype-nutrient relationships, environment-gene phenotype relationships, environment-nutrient relationships, and agronomic-nutrient relationships;

[0138] Attributes include genetic attributes, environmental attributes, agronomic attributes, and planting plan attributes;

[0139] The intensity of gene regulation is obtained by weighting the corresponding expression characteristics; the weighting can be obtained by optimizing a natural-inspired optimization algorithm.

[0140] Based on similarity calculation, it determines whether entities pointing to the same real-world object in different data sources are the same, detects whether data of the same entity from different data sources conflict, and resolves conflicts in the conflicting data.

[0141] When the genetic attributes of the same entity conflict with environmental, agronomic, or planting plan attributes, a weighted fusion based on gene regulatory intensity is used to resolve the conflict; such as the quantitative benchmark value obtained by weighted fusion based on gene regulatory intensity. ,in, In terms of gene regulation intensity; Values ​​representing the gene source; The values ​​can be environmental, agronomic, or planting task-related. In fertilizer demand analysis, if there is a conflict between nutrient requirements at the gene level and environmental constraints, the quantitative benchmark value obtained by weighted fusion of gene regulation intensity will be combined with gene regulation intensity to balance the influence of both and output an executable nutrient supply benchmark value.

[0142] After resolving conflicts, knowledge components are used to build a knowledge graph. The knowledge graph is updated based on real-time data to obtain a dynamic agricultural knowledge graph, which is then stored in a graph database.

[0143] The method for constructing a dynamic agricultural knowledge graph explicitly defines knowledge components including gene phenotype entities, gene-nutrient / environment-gene relationships, and gene regulatory strength. It structurally associates the gene attributes of different crop genotypes with suitable environmental, agronomic, and planting plan attributes. Furthermore, when handling conflicts between multiple data sources for the same entity, it uses gene regulatory strength as the core weight for weighted fusion and resolution, and continuously updates the graph content with real-time data. This allows the constructed dynamic agricultural knowledge graph to clearly depict the unique association patterns between different crop genotypes and specific nutrient requirements and suitable agronomic measures. It provides structured knowledge support based on genotype differences for fertilization strategy formulation, completely breaking the limitations of a "one-size-fits-all" universal evaluation standard, and effectively solving the technical problem of not being able to customize fertilization strategies for different genotypes.

[0144] Methods for obtaining a nutrient requirement matrix include:

[0145] Multi-source structured datasets are transformed into knowledge components using predefined knowledge extraction rules, including mapping rules, natural language processing techniques, and convolutional neural networks. Entity nodes are constructed based on the knowledge components, and the relationship chain is activated through a message passing algorithm to output a knowledge-driven nutrient demand feature vector. A nutrient demand matrix is ​​then constructed based on the nutrient demand feature vector.

[0146] Methods for obtaining nutrient requirement feature vectors include:

[0147] The correlation between entities is calculated based on the latest data using the Pearson correlation coefficient, and the weights are updated according to the correlation.

[0148] For each entity node, the top K neighbors are collected in descending order of relation weight, and the neighbor node information is merged using the mean aggregation method;

[0149] When new data input causes entity attributes to be updated, an incremental update of the relationship chain is triggered, and only the affected subgraphs are recalculated;

[0150] From the dynamic agricultural knowledge graph after message passing, extract the embedding vectors of crop, gene, environment and agronomic entities;

[0151] Relational attention is calculated based on the entity's attention weight, cosine similarity, and relation weight. The original features are then weighted according to this relational attention to obtain the nutrient demand feature vector. For example, entities... Attention weights ,in, Cosine similarity; The embedding vector of the nutrient entity; For embedding vectors of other entities related to nutrients; For entities Relationship weights.

[0152] The method for obtaining the nutrient requirement matrix and nutrient requirement feature vector first transforms a multi-source structured dataset containing crop genotype phenotypic data of different genotypes into knowledge components using predefined knowledge extraction rules that incorporate mapping rules, natural language processing techniques, and convolutional neural networks. Based on these components, entity nodes containing genotype entities are constructed, and then a message passing algorithm is used to activate the association chain. During this process, the correlation weights between entities are calculated and updated using the Pearson correlation coefficient to highlight the differences in association between genotypes and other entities. Then, the top K neighbors of each entity node are collected in descending order of relation weight, and the information is aggregated and fused using the mean. Key association information that is compatible with the genotype is prioritized, and new data leads to changes in entity genotypes. During the update process, only the affected subgraph is incrementally updated, ensuring that changes in the attributes of entities of different genotypes are reflected in the relationships in a timely manner. Subsequently, the embedding vectors of crop, gene, environment, and agronomic entities are extracted from the dynamic agricultural knowledge graph after message passing. Combined with cosine similarity and entity relationship weights, relationship attention is calculated and the original features are weighted. The resulting nutrient requirement feature vector accurately carries the exclusive nutrient requirement information of different genotype crops, avoiding the neglect of genotype differences by general evaluation standards. This provides a precise quantitative basis for nutrient requirements for customized fertilization strategies for different genotypes, thus effectively solving the technical problem of not being able to customize fertilization strategies for different genotypes and having to use a "one-size-fits-all" general evaluation standard.

[0153] Collect spatial thermodynamic data of soil nutrients and historical fertilization time series data, and analyze to obtain spatiotemporal nutrient demand feature vectors;

[0154] Methods for obtaining spatiotemporal nutrient demand feature vectors include:

[0155] The land was divided into plots using an E×E grid. Soil samples from the 0-20cm topsoil layer were collected at the center of each grid using a soil auger to obtain soil nutrient data. This data was then used as input to a convolutional neural network with a 3×3 spatial convolution kernel to obtain spatial distribution characteristics. Historical fertilization time-series data was also acquired and used as input to a time-series analysis model to obtain time-dependent characteristics. The spatial distribution and time-dependent characteristics were then concatenated to obtain a spatiotemporal nutrient demand feature vector. The training methods for the convolutional neural network and the time-series analysis model were the same as those for the planting analysis model.

[0156] This method divides land into plots using an E×E grid and collects topsoil samples to detect soil nutrient data. The data is then input into a convolutional neural network with a 3×3 spatial convolutional kernel to obtain the spatial distribution characteristics of soil nutrients. Simultaneously, historical fertilization time-series data is acquired and input into a time-series analysis model to obtain the time-dependent characteristics of fertilization. The spatial distribution characteristics and time-dependent characteristics are then concatenated to form a spatiotemporal nutrient demand feature vector. The convolutional neural network and time-series analysis model employ the same training method as the planting analysis model to ensure model effectiveness. This process accurately captures the spatial dimensional differences in soil nutrients across different plots and the fertilization impact patterns at different time periods. Furthermore, by combining the unique nutrient requirements of different crop genotypes, it can match precise nutrient supply criteria under corresponding spatiotemporal conditions, effectively breaking away from the "one-size-fits-all" universal fertilization assessment standard and facilitating the customization of personalized fertilization strategies for different crop genotypes.

[0157] An initial nutrient matrix is ​​generated by fusing dual-channel features through an attention mechanism. This matrix is ​​then adjusted using a reinforcement learning optimizer. Multi-source structured datasets, soil nutrient spatial-thermal data, and historical fertilization time-series data are used as state vectors, while fertilizer ratio adjustment and fertilization time-series adjustment dimensions are used as action vectors. A reward function is designed based on predicted crop yield, predicted nutrient leaching rate, and predicted fertilizer cost to obtain the fertilizer demand matrix. Time-based reward function ,in, To predict crop yield; To predict nutrient leaching rate; To predict fertilizer costs; , and These are all weighting coefficients, which can be obtained through natural heuristic optimization algorithms.

[0158] When generating the initial nutrient matrix by fusing dual-channel features through an attention mechanism, this mechanism can automatically focus on key features related to the nutrient requirements of different crop genotypes in multi-source data, avoiding the loss of genotype information caused by the fusion of generalized features. This ensures that the initial nutrient matrix already carries the unique nutrient requirements of different genotypes. Subsequently, when adjusting the matrix using a reinforcement learning optimizer, the weight coefficients of the designed reward function, which includes predictions of crop yield, nutrient leaching rate, and fertilizer cost, can be optimized using a naturalistic optimization algorithm that combines the characteristics of different genotypes. For example, for nitrogen-efficient genotypes, the optimizer tends to increase the yield weight and decrease the fertilizer cost weight; for nutrient-sensitive genotypes, it adjusts the weights to balance yield and nutrient supply. The resulting fertilizer requirement matrix can accurately match the nutrient utilization patterns and adaptation conditions of different crop genotypes, completely overcoming the limitations of a uniform nutrient matrix under a "one-size-fits-all" general evaluation standard. This provides a quantitative basis for customizing fertilization strategies for different genotypes, effectively solving the technical problem of not being able to customize fertilization strategies for different genotypes.

[0159] The fertilizer demand matrix is ​​translated into actual fertilization operations, and dynamic calibration is achieved based on leachate monitoring.

[0160] Methods for dynamic calibration based on leachate monitoring include:

[0161] Collect the ion concentration of the root zone leachate F hours after fertilization. Calculate the adjusted fertilizer utilization rate coefficient based on the initial fertilizer utilization rate, the target leachate nutrient concentration, and the measured leachate nutrient concentration. (The adjusted fertilizer utilization rate coefficient is then used as an example.) ,in, This refers to the initial utilization rate of fertilizer. The target leachate nutrient concentration; To measure the nutrient concentration in the leachate;

[0162] If the target leachate nutrient concentration is greater than the measured leachate nutrient concentration, the amount of fertilizer applied in the next round will be adjusted according to the adjusted fertilizer utilization rate coefficient.

[0163] After translating the fertilizer demand matrix into actual fertilization operations, a dynamic calibration method based on leachate monitoring collects the ion concentration of root zone leachate F hours after fertilization. Combining the initial fertilizer utilization rate, target, and measured leachate nutrient concentration, the adjusted fertilizer utilization rate coefficient is calculated using a formula. Then, the next round of fertilization is adjusted based on the comparison between the target and measured concentrations. This process can accurately capture the differences in leachate ion concentration caused by differences in nutrient absorption and utilization efficiency among different crop genotypes, and thus adjust the fertilization amount accordingly. This breaks the limitations of a uniform fertilization amount under a "one-size-fits-all" general evaluation standard, enabling fertilization operations to continuously adapt to the specific nutrient needs of different crop genotypes, thereby effectively solving the technical problem of not being able to customize fertilization strategies for different genotypes.

[0164] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection.

[0165] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing fertilizer requirements based on an AI planting plan, characterized by, The method comprises the following steps: Crop data in the whole growth period of crops is collected, preprocessed and analyzed to obtain a multi-source structured data set; The multi-source structured data set is taken as an input of a planting analysis model to obtain an AI planting plan; A dynamic agricultural knowledge graph is constructed according to agricultural knowledge and the AI planting plan, the dynamic agricultural knowledge graph is fused with the multi-source structured data set, and nutrient demand matrix is analyzed and obtained; Soil nutrient spatial thermal data and historical fertilization timing data are collected, and a space-time nutrient demand feature vector is analyzed and obtained; An initial nutrient matrix is generated by fusing the nutrient demand matrix and the space-time nutrient demand feature vector through an attention mechanism, the initial nutrient matrix is adjusted through a reinforcement learning optimizer, a reward function is designed according to a predicted crop yield and a predicted nutrient leaching rate mixed prediction of fertilizer cost, and a fertilizer demand matrix is obtained; The fertilizer demand matrix is converted into actual fertilization operations, and dynamic calibration is performed based on leachate monitoring. 2.The AI-based planting plan-based fertilizer requirement analysis method of claim 1, wherein, The multi-source structured data set comprises rhizosphere data, canopy data, environmental data and gene phenotype data; The rhizosphere data comprises pH value, ion concentration, specifically nitrogen ion, phosphorus ion and potassium ion, and conductivity of the rhizosphere; the canopy data comprises red edge normalized vegetation index, chlorophyll fluorescence parameter and yellowing area proportion of leaves; the environmental data comprises photosynthetically active radiation, precipitation intensity, accumulated temperature and wind speed; and the gene phenotype data comprises genotype marker and corresponding expression characteristics; The method for obtaining the canopy data comprises the following steps: The multi-spectral data of the crops is converted into digital quantitative values, and the actual reflectivity is calibrated through a preset standard gray board; The digital quantitative values of the pure sky light area are obtained, and the average value and the standard deviation of the digital quantitative values of the pure sky light area are counted to obtain the average quantitative value and the average standard deviation; The dark field digital quantitative values in the absence of light are obtained, and the average value is counted as the dark digital quantitative value, and the corrected sky light quantitative value is calculated according to the average quantitative value and the dark digital quantitative value; The sky light radiation brightness is calculated according to the radiation scaling gain coefficient, the radiation scaling offset coefficient and the corrected sky light quantitative value; The average quantitative value of the center of the gray board image is extracted, the DN value after dark field correction is calculated, the corresponding gray board radiation brightness is calculated, and the sky light reflectivity is calculated in combination with the sky light radiation brightness, the gray board reflectivity of the preset standard gray board; The actual reflectivity is corrected based on the sky light reflectivity to obtain the corrected reflectivity; The red edge normalized vegetation index is calculated according to the corrected reflectivity and the infrared reflectivity; The initial fluorescence and the maximum fluorescence are collected, the variable fluorescence is calculated according to the initial fluorescence and the maximum fluorescence, the ratio of the variable fluorescence to the maximum fluorescence is calculated, and the chlorophyll fluorescence parameter is obtained; The crop canopy image is collected, the crop canopy image noise is removed by using Gaussian filtering, and the crop canopy image is converted from the RGB color space to the HSV color space; the HSV range of the yellowing area is counted, and the threshold method is used to generate a binary image; the total number of yellowing area pixels and the total number of leaf pixels are counted, and the yellowing area proportion is calculated. 3.The AI-based planting plan-based fertilizer requirement analysis method of claim 2, wherein, The corresponding expression characteristics comprise relative expression amount, nitrogen transport rate, potassium transport rate and enzyme activity; The method for obtaining the gene phenotype data comprises the following steps: Collecting genomic DNA of crops, screening target genes, and designing specific primers for SNP sites of target genes; Differentiating SNP alleles by competitive allele-specific PCR to obtain typing results, combining target genes and SNP site genotypes to form genotype markers; Detecting the relative expression amount of target genes in root systems by real-time quantitative PCR; Detecting 15N and 42K abundance by 15N and 42K isotope labeling methods respectively, and calculating nitrogen transport rate and potassium transport rate; Detecting protein expression amount by specific antibodies; Configuring pNP standard solution, detecting absorbance values corresponding to different pNP standard solution concentrations by sodium p-nitrophenyl phosphate colorimetry, constructing a standard curve with pNP concentration as the abscissa and absorbance value as the ordinate, fitting to obtain a regression equation, calculating the pNP solution concentration corresponding to the current absorbance value according to the regression equation, and calculating the enzyme activity according to the total volume of the reaction system, enzyme liquid dilution multiple, protein mass in the reaction system, incubation time and enzyme liquid volume. 4.The AI-based planting plan-based fertilizer requirement analysis method of claim 3, wherein, The method for obtaining nitrogen transport rate comprises: Moving the crop after starvation treatment into 15N-labeled nutrient solution, culturing for 2 hours at 25°C under light for 12 hours to obtain a labeled crop; Rinsing the root system of the labeled crop with nitrogen-free nutrient solution for 3 times to remove the residual 15N on the surface, separating the root system and the aboveground part, killing green at 105°C for 30 minutes, drying at 70°C until constant weight, crushing and sieving through a 100-mesh sieve to obtain a powder sample; Taking 10 mg of the powder sample, detecting 15N abundance by an elemental analyzer-isotope ratio mass spectrometer; Calculating the nitrogen transport rate according to 15N abundance, culture time and root system dry weight. 5.The AI-based planting plan-based fertilizer requirement analysis method of claim 2, wherein, The method for constructing a dynamic agricultural knowledge graph comprises: Defining knowledge components, including entities, relationships, attributes and gene regulation intensity; Entities include gene phenotype entities, crop entities, environment entities, nutrient entities and agronomic entities; Relationships include crop-gene phenotype relationships, crop-environment relationships, crop-nutrient relationships, crop-agronomic relationships, gene phenotype-nutrient relationships, environment-gene phenotype relationships, environment-nutrient relationships and agronomic-nutrient relationships; Attributes include gene attributes, environment attributes and agronomic attributes; Gene regulation intensity is obtained by weighted calculation according to corresponding expression characteristics; Judging whether entities pointing to the same real object in different data sources are the same based on similarity calculation, detecting whether data of the same entity in different data sources conflict, and resolving the conflicting data; Building a knowledge graph by obtaining knowledge components after conflict resolution, updating the knowledge graph according to real-time data to obtain a dynamic agricultural knowledge graph, and storing the dynamic agricultural knowledge graph in a graph database. 6.The AI-based planting plan-based fertilizer requirement analysis method of claim 5, wherein, The method for obtaining a nutrient demand matrix comprises: Converting multi-source structured data sets into knowledge components by pre-defined knowledge extraction rules; constructing entity nodes according to the knowledge components, activating the associated relationship chain by a message passing algorithm, outputting a knowledge-driven nutrient demand feature vector, and constructing a nutrient demand matrix based on the nutrient demand feature vector. 7.The AI-based planting plan-based fertilizer requirement analysis method of claim 6, wherein, The method for obtaining a nutrient demand feature vector comprises: The correlation between the latest data of entities is calculated based on a Pearson correlation coefficient, and the weight is updated according to the correlation; For each entity node, the top K neighbors are collected in descending order of relationship weight, and the neighbor node information is fused in a mean aggregation manner; When new data input leads to entity attribute update, trigger relationship chain incremental update, and recalculate the affected subgraph; From the dynamic agricultural knowledge graph after message passing, the embedding vectors of crop, gene phenotype, environment and agronomic entities are extracted; According to the attention weight of the entity, the cosine similarity and the relationship weight, the relationship attention is calculated, the original features are weighted according to the relationship attention, and the nutrient demand feature vector is obtained. 8.The AI-based planting plan-based fertilizer requirement analysis method of claim 1, wherein, The method for obtaining the spatio-temporal nutrient demand feature vector comprises: Collecting soil nutrient data as the input of the convolutional neural network to obtain spatial distribution features, collecting historical fertilization time series data, and taking the historical fertilization time series data as the input of the time series analysis model to obtain time-dependent features, and splicing the spatial distribution features and the time-dependent features to obtain the spatio-temporal nutrient demand feature vector. 9.The AI-based planting plan-based fertilizer requirement analysis method of claim 1, wherein, The method for obtaining the fertilizer demand matrix comprises: Through the attention mechanism, the double-channel features are fused to generate an initial nutrient matrix, and through a reinforcement learning optimizer adjustment, the multi-source structured data set, the soil nutrient spatial thermal data and the historical fertilization time series data are taken as state vectors, the fertilizer ratio adjustment dimension and the fertilization time series adjustment dimension are taken as action vectors, a reward function is designed according to the predicted crop yield and the predicted nutrient leaching rate to mix the predicted fertilizer cost, and the fertilizer demand matrix is obtained. 10.The AI-based planting plan-based fertilizer requirement analysis method of claim 1, wherein, The method for realizing dynamic calibration based on leachate monitoring comprises: Collecting the ion concentration of the root zone leachate F hours after fertilization, calculating the adjusted fertilizer utilization rate coefficient according to the fertilizer initial utilization rate, the target leachate nutrient concentration and the measured leachate nutrient concentration; If the target leachate nutrient concentration is greater than the measured leachate nutrient concentration, the next round of fertilization amount is adjusted according to the adjusted fertilizer utilization rate coefficient.

Images