Method, system and equipment for generating special fertilizer formula for crops

Abstract

The invention provides a method, a system and equipment for generating a special fertilizer formula for crops. The method comprises the following steps: acquiring crop growth data of any region; constructing a nitrogen fertilizer recommendation model according to the crop growth data, and obtaining a nitrogen application amount prediction value and a yield prediction value; constructing a phosphate fertilizer recommendation model based on a phosphate fertilizer grading standard according to the yield prediction value and the actually measured data of the content of the rapidly available phosphorus in the soil, and obtaining a phosphorus application amount prediction value; establishing a potash fertilizer recommendation model based on a potash fertilizer grading standard, and obtaining a potassium application amount prediction value; respectively calculating the base fertilizer amount and the topdressing amount of nitrogen, phosphorus and potassium according to the predicted value of the nitrogen application amount, the predicted value of the phosphorus application amount, the predicted value of the potassium application amount and the base-topdressing ratio; constructing a base fertilizer formula model according to the base fertilizer amount of nitrogen, phosphorus and potassium; and constructing a dynamic adjustment model according to the topdressing amount of nitrogen, phosphorus and potassium, and calculating a topdressing correction value. The method can accurately and efficiently formulate a fertilization strategy which is suitable for crops, lands and conditions, obviously improves the utilization efficiency of the fertilizer, and reduces excessive application.

Description

Technical Field

[0001] This invention belongs to the field of smart planting technology, specifically relating to a method, system, and equipment for generating crop-specific fertilizer formulas. Background Technology

[0002] Currently, my country's agricultural production is at a critical stage of transformation from "high input, high consumption" to "high efficiency, green development." Chemical fertilizers, as one of the main means of increasing crop yields, have long supported national food security. However, over-reliance on experience-based fertilization and static formulas not only leads to low fertilizer utilization rates but also brings significant ecological and economic problems. These are specifically manifested in the following aspects: I. Currently, most regions still rely primarily on paper maps or administrative region-based recommendations, failing to dynamically update fertilization strategies based on differences in crops, plots, and years. The creation of fertilization prescription maps and formula recommendations largely depend on human experience or static mean models, lacking data-driven and feedback optimization mechanisms. This leads to significant deviations, low accuracy, and high difficulty in widespread adoption during practical application.

[0003] Second, at the fertilizer product level, many "general-purpose compound fertilizer" formulas have failed to be optimized for specific crops, regions, and yield targets, resulting in "inaccurate formulas, unclear dosages, and unstable effects." Farmers lack clear references for fertilizer formulation, while enterprises lack a data-driven system platform to support formula optimization and product iteration, creating a break in the "production-supply-use" chain.

[0004] Third, current explorations in intelligent fertilization mostly focus on single-point breakthroughs (such as variable-rate fertilization equipment or remote sensing monitoring), lacking a closed-loop system from data perception and algorithm reasoning to formula generation, prescription implementation, and result feedback. Especially at the county and field scale, there is a lack of intelligent systems capable of efficiently integrating multi-source information such as soil, crops, and meteorology to output highly reliable fertilization prescriptions.

[0005] Therefore, there is an urgent need to develop an intelligent system that integrates "data acquisition, fertilization decision-making, formula optimization, prescription output, and equipment integration" to achieve automation, personalization, and high efficiency in fertilization of major crops, and truly support the development of modern agriculture towards precision, high yield, high efficiency, and environmental protection. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method, system and equipment for generating crop-specific fertilizer formulas, which can dynamically adjust the fertilization strategy according to the differences of different crops, different regions and different years, thereby improving the reliability of the fertilizer formula.

[0007] This invention provides a method for generating a crop-specific fertilizer formula, comprising the following steps: S1. Obtain crop growth data for any region, including measured climate data, measured soil data, measured management data, measured soil available phosphorus content, measured soil available potassium content, measured yield data, and NO3. - Actual data from rinsing, N2O emissions, and NH3 volatilization; S2. Construct a nitrogen fertilizer recommendation model based on machine learning based on the crop growth data, and obtain the predicted value of nitrogen application rate and the predicted value of yield; S3. Based on the predicted yield and the measured data of available phosphorus content in the soil, construct a phosphorus fertilizer recommendation model based on the phosphorus fertilizer grading standard, and obtain the predicted phosphorus application rate. S4. Based on the predicted yield and the measured data of available potassium content in the soil, construct a potassium fertilizer recommendation model based on the potassium fertilizer grading standard, and obtain the predicted potassium application rate. S5. Calculate the amount of nitrogen, phosphorus, and potassium fertilizer as basal fertilizer and topdressing fertilizer based on the predicted nitrogen application rate, the predicted phosphorus application rate, the predicted potassium application rate, and the basal-topdressing ratio. S6. Construct a base fertilizer formula model based on the amount of nitrogen, phosphorus, and potassium in the base fertilizer, and obtain the base fertilizer formula; S7. Construct a dynamic adjustment model based on the nutrient deficiency-dynamic adjustment factor during the key topdressing period according to the amount of nitrogen, phosphorus and potassium topdressing, and calculate the topdressing correction value of nitrogen, phosphorus and potassium to generate the topdressing formula.

[0008] Furthermore, step S2 includes the following steps: S201. Based on the measured climate data, measured soil data, measured management data, DNDC model, and APSIM model, obtain the predicted yield and NO3. - Predicted values ​​for leaching, N2O emissions, and NH3 volatilization; S202, Based on the crop growth data, yield forecast, and NO3... - A nitrogen fertilizer recommendation model based on an attention mechanism recurrent neural network was constructed using leaching predictions, N2O emission predictions, and NH3 volatilization predictions. S203. Based on the measured climate data, measured soil data, measured management data, and nitrogen fertilizer recommendation model, calculate the predicted nitrogen application rate and yield under optimal conditions.

[0009] Furthermore, step S3 includes the following steps: S301. The recommended phosphorus application rate is calculated based on the predicted yield value and the recommended phosphorus fertilizer application rate formula. S302. Determine the current phosphorus application rate for the region based on the measured data of available phosphorus content in the soil and the phosphate fertilizer grading standards. S303. Multiply the recommended phosphorus application rate by the phosphorus application rate ratio to obtain the predicted phosphorus application rate.

[0010] Furthermore, step S4 includes the following steps: S401. The recommended potassium application rate is calculated based on the predicted yield and the recommended potassium fertilizer application rate formula. S402. Determine the current potassium application rate ratio based on the measured data of available potassium content in the soil and the potassium fertilizer grading standards. S403. Multiply the recommended potassium application rate by the potassium application rate ratio to obtain the predicted potassium application rate.

[0011] Furthermore, the expression for the base fertilizer formulation model is as follows: ; in, Represents the largest integer not exceeding A. Represents the largest integer not exceeding B. Represents the largest integer not exceeding C. , , ω represents the fertilizer application ratio, W N W is the predicted value for nitrogen application. P W is the predicted value for phosphorus application rate. K A represents the predicted potassium application rate, B represents the nitrogen application rate coefficient, C represents the phosphorus application rate coefficient, and C represents the potassium application rate coefficient.

[0012] Furthermore, step S7 includes the following steps: S701. Obtain remote sensing data of the current area within the previous sampling period before the topdressing period, and extract its NDVI value; S702. Determine the deficiency judgment conditions and dynamic adjustment factors based on the NDVI value and the preset NDVI threshold range; S703. Multiply the dynamic adjustment factor by the amount of nitrogen, phosphorus and potassium topdressing to obtain the nitrogen, phosphorus and potassium topdressing correction value.

[0013] Furthermore, the deficiency criterion and dynamic adjustment factor in S702 are determined in the following manner: ; Where m is the dynamic adjustment factor, σ1, σ2, and σ3 are adjustment coefficients, σ1∈[-7%,-2%], σ2∈[2%,7%], σ3∈[σ2+5%,σ2+10%], ε is the floating coefficient, 0<ε<1, NDVI max This is the maximum value within the preset threshold range.

[0014] Furthermore, step S7 includes the following steps: S701. Obtain leaf images of crops in the current region; S702. Input the blade image into the Qwen2.5-VL-7B multimodal model to determine the deficiency nutrient judgment conditions and dynamic adjustment factors; S703. Multiply the dynamic adjustment factor by the amount of nitrogen, phosphorus and potassium topdressing to obtain the nitrogen, phosphorus and potassium topdressing correction value.

[0015] This invention also provides a crop-specific fertilizer formulation generation system, comprising: The acquisition module is used to acquire crop growth data for any region. This crop growth data includes measured climate data, measured soil data, measured management data, measured soil available phosphorus content, measured soil available potassium content, measured yield data, and NO3. - Actual data from rinsing, N2O emissions, and NH3 volatilization; The nitrogen fertilizer recommendation model calculation module is used to construct a machine learning-based nitrogen fertilizer recommendation model based on the crop growth data, and obtain the predicted nitrogen application rate and the predicted yield. The phosphate fertilizer recommendation model calculation module is used to construct a phosphate fertilizer recommendation model based on the phosphate fertilizer grading standard based on the predicted yield value and the measured data of available phosphorus content in the soil, and to obtain the predicted value of phosphorus application rate. The potassium fertilizer recommendation model calculation module is used to construct a potassium fertilizer recommendation model based on the potassium fertilizer grading standard based on the predicted yield value and the measured data of available potassium content in the soil, and to obtain the predicted potassium application rate. The base fertilizer and topdressing calculation module is used to calculate the amount of nitrogen, phosphorus and potassium base fertilizer and topdressing fertilizer based on the predicted nitrogen application rate, the predicted phosphorus application rate, the predicted potassium application rate and the base fertilizer and topdressing ratio, respectively. The base fertilizer formula generation module is used to construct a base fertilizer formula model based on the amount of nitrogen, phosphorus, and potassium in the base fertilizer and obtain the base fertilizer formula. The topdressing formula generation module is used to construct a dynamic adjustment model based on the nutrient deficiency-dynamic adjustment factor during the key topdressing period according to the amount of nitrogen, phosphorus and potassium topdressing, and to calculate the topdressing correction value of nitrogen, phosphorus and potassium to generate the topdressing formula.

[0016] The present invention also provides a crop-specific fertilizer formula generation device, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the above method.

[0017] The beneficial effects of this invention are: (1) The present invention can realize the precise and efficient formulation of fertilization strategies that are "tailored to crops, local conditions and time", significantly improve fertilizer utilization efficiency and reduce excessive application.

[0018] (2) The present invention dynamically adjusts the fertilization strategy based on real-time environment and crop status, avoids the risk of static formula mismatch, and enhances the system adaptability. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method in this invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0021] As shown in Figure 1, this invention provides a method for generating a crop-specific fertilizer formula, comprising the following steps: S1. Obtain crop growth data for any region. This data includes measured climate data, measured soil data, measured data on management practices, measured soil available phosphorus content, measured soil available potassium content, measured yield data, and NO3. - Data includes leaching measurements, N2O emission measurements, and NH3 volatilization measurements. Climate data includes annual average temperature and annual average precipitation. Soil data includes soil organic carbon, total nitrogen, pH, bulk density, and clay content. Management data includes crop type (rice, wheat, and corn), fertilizer type (urea, mineral nitrogen fertilizer, urea with added inhibitors), nitrogen application method (top application, deep application), soil tillage method (conventional tillage and no-till), number of nitrogen fertilizer applications (one, two, or three), and nitrogen application rate.

[0022] To ensure the comprehensiveness and representativeness of the database, 10,000 grid scenarios were randomly selected from each of the three crops for pre-training. Specifically, each sampling randomly selected one grid and one scenario from approximately 72,805 grids and 72 management combination scenarios. The selected 10,000 samples are almost evenly distributed across the country, fully representing all scenario combinations and covering agricultural production scenarios under different crops, climates, soils, and management practices, thus ensuring the synthetic database's rapid response to diverse agricultural production conditions. All the above data underwent feature standardization and one-hot encoding preprocessing to ensure consistent data format and readability.

[0023] S2. Construct a machine learning-based nitrogen fertilizer recommendation model based on crop growth data, and obtain predicted nitrogen application rates and yields. Specifically, S2 includes the following steps: S201. Based on measured climate data, measured soil data, measured management data, DNDC model, and APSIM model, yield forecasts and NO3 are obtained. - Predicted values ​​for leaching, N2O emissions, and NH3 volatilization. Due to the influence of measured data on yield and NO3... -Since the sample sizes for leaching values, N2O emissions, and NH3 volatilization values ​​are relatively small, this step introduces the DNDC and APSIM models to simulate these four data points to meet training requirements. Climate data, soil data, and management data are input into the APSIM model to obtain yield predictions. Climate data, soil data, and management data are input into the DNDC model to obtain NO3 volatilization data. - The predicted values ​​for rinsing, N2O emissions, and NH3 volatilization were obtained. This provided a sufficient sample size of data, which was then used as the dataset for model training.

[0024] S202, based on crop growth data, yield forecasts, and NO3... - A nitrogen fertilizer recommendation model based on an attention mechanism recurrent neural network was constructed using predicted values ​​of leaching, N2O emissions, and NH3 volatilization. The knowledge-guided machine learning model was trained using synthetic data generated from the crop growth model APSIM and the nitrogen management model DNDC, as well as observational data collected from literature. The training model network used was a multi-task deep feedforward neural network (MLP), whose overall structure can be summarized as: Input Projection → Backbone MLBlock stacked → Task-specific Output Headers. The network design aims to simultaneously predict multiple target variables, including yield, N2O emissions, NO3⁻ leaching, and NH3 emissions, and supports rapid fine-tuning by freezing the backbone module.

[0025] I. Model Input Layer: 1. Input Features: Includes continuous numerical features and categorical embeddings. The input dimension is the sum of the number of continuous features and the embedding dimensions of all categorical features.

[0026] 2. Linear Projection: A single-layer fully connected network (Linearlayer) projects the original input into a 128-dimensional hidden space. Function: Unifies feature dimensions and generates a latent vector representation suitable for subsequent MLPBlock processing.

[0027] II. Backbone Feature Extraction Module (BackboneMLPBlock): 1. Module stacking: The backbone consists of 6 MLB Blocks with the same structure stacked sequentially.

[0028] Each MLPBlock has an input dimension of 128 and an output dimension of 128.

[0029] 2. Single MLPBlock structure: The first fully connected layer (FC) is 128 to 512 (the hidden layer dimension is 4 times the input dimension), followed by the GELU activation function (GELU).

[0030] Dropout: Used to prevent overfitting.

[0031] Second layer fully connected (FC): 512→128.

[0032] Residual Connection: The input is added to the output of the second layer to enhance gradient flow and training stability.

[0033] Layer Normalization (LayerNorm): Normalizes the residual output to improve training stability and convergence speed.

[0034] Each block structure is as follows: Input 128 -> Fully connected 512 -> GELU activation -> Dropout -> Fully connected 128 -> Residual connection -> LayerNorm.

[0035] 3. Module Functions: The backbone is responsible for extracting general, high-dimensional feature representations from the input features, providing a shared foundation for multi-task prediction.

[0036] III. Multi-task output branches (Task-specificHeads): 1. Structural Design: Five independent mini-MLPs are derived from the 128-dimensional feature vector output from the main branch, each used to predict a target variable (including production, PFP, N2O emissions, NH3 volatilization, and NO3). - (Swimming). Each small MLP consists of two fully connected network layers: 128→32→1, with the intermediate layer activated using GELU and supplemented by Dropout.

[0037] 2. Functional Description: The small output network focuses on task-related feature mapping to achieve independent prediction for different tasks.

[0038] Backbone Freezing: 1. Definition and Principle: Freeze refers to fixing the network parameters during training and not updating gradients. During training, the parameter update formula is: ; When param.requires_grad=False, the gradient of the corresponding parameter θ is not calculated, the optimizer will not modify its value, and the forward propagation only depends on the existing parameters.

[0039] 2. Implementation method: Freeze the network by traversing the network parameters and identifying the name of the backbone module: for name, param in model.named_parameters(): if "MLPBlock" in name: # or backbone module name param.requires_grad = False 3. Training effect after freezing: The parameters of the backbone feature extraction module remain unchanged, and only the parameters of the output branches (task-specific heads) are updated.

[0040] Advantages: It retains existing knowledge in the core framework, quickly adapts to new tasks, and reduces the risk of overfitting.

[0041] 4. Comparison between fine-tuning and full-scale training: Backbone: A deep, general feature extraction network that can be frozen.

[0042] OutputHead: A task-dependent network whose parameters can be updated to adapt to new tasks.

[0043] Training strategy selection: Freeze the trunk + train the output branch: quick fine-tuning, suitable for scenarios with existing basic feature knowledge.

[0044] Full training: Learn new tasks from scratch; core parameters can be updated.

[0045] Based on the above model, specifically, S202 includes the following steps: S20201, Frozen NO3 - The leaching module, N2O emission module, and NH3 volatilization module will have their learning gradients set to zero. This will be based on crop growth data from 19 out of 22 years, yield forecasts, and NO3... - The predicted values ​​of leaching, N2O emissions, and NH3 volatilization were used as input data to train the yield and PFP modules in the machine learning model, obtaining the mapping relationship between the input data and yield and nitrogen fertilizer use efficiency. An adaptive learning method based on mean squared error was used to construct the training loss, starting with "easier" samples and gradually transitioning to "more difficult" samples.

[0046] The specific loss function is as follows: ; ; ; Among them, Loss step1 This is the loss function used in S20201, yj,i and These are the input data and predicted value for the j-th sample, respectively, v i,j It is an integer multiple used to determine whether to use the j-th sample to calculate the loss, N is the total number of samples in a simulation, λ is the threshold, λ0 is the initial value, and K growth This represents the growth rate.

[0047] S20202, Freeze the PFP module, utilize input data and combine knowledge-driven loss and reaction to train the output module in the machine learning model, NO3 - The scrubbing module, N2O emission module, and NH3 volatilization module obtain input data and output, NO3 - The relationship between leaching value, N2O emission value and NH3 volatilization value. In addition to the mean square error loss function Loss1, the loss function of S20202 also involves (1) the nitrogen balance loss function Loss2, in order to control the input nitrogen application rate and the predicted yield, N2O emission, NH3 volatilization and NO3. - The relationship between rinsing; (2) Control of N2O emissions, NH3 volatilization and NO3 - rinsing of SBN k The exponential response loss function, Loss3, is used. After two steps of pre-training, the knowledge-guided machine learning model can successfully predict yield, nitrogen fertilizer use efficiency (PFP), N2O emissions, NH3 volatilization, and NO3. - Washing. The specific loss function is as follows: ; ; ; ; ; Where α is the loss coefficient for each item, r is the Pearson correlation coefficient, b is the soil nitrogen residue coefficient, Nfer is the nitrogen application rate, and c k SBN is the crop nitrogen uptake coefficient. k The nitrogen surplus in the soil is represented by k, which represents the nitrogen content of wheat, corn, and rice, respectively. For the predicted yield of the target crop.

[0048] S20203, frozen backbone module and NO3 - The system includes a leaching module, an N2O emission module, and an NH3 volatilization module, with a learning rate set at 20% of the original learning rate. It utilizes measured climate data, soil data, management data, soil available phosphorus content data, soil available potassium content data, yield data, and NO3 data. -The production and PFP modules in the machine learning model were fine-tuned using actual rinsing data, actual N2O emission data, and actual NH3 volatilization data, combined with knowledge-driven loss. The loss function in S20203 includes the following parts: (1) the mean square error (MSE) loss function Loss1 between the observed and predicted values; (2) and The mean squared error loss function Loss2 between them; (3) constraints The loss function Loss3 is a quadratic function with a downward opening relationship between PFP and Nfer; (4) the loss function Loss4 is used to limit the negative response between PFP and Nfer, thereby enhancing the biological rationality and prediction accuracy of the model. The specific loss functions are as follows: ; ; ; ; Where a1, a2, and a3 are coefficients.

[0049] S20204, The main module and production module were frozen, NO3 - The leaching module and NH3 volatilization module were used, and the learning rate for the yield module and N2O emission module was set to 50% of the original learning rate to avoid overfitting and retain too much prior knowledge. Measured climate data, soil data, management data, soil available phosphorus content data, soil available potassium content data, yield data, and NO3 data were used. - The production and N2O emission modules in the machine learning model were fine-tuned using actual rinsing data, actual N2O emission data, and actual NH3 volatilization data, combined with knowledge-driven loss. The loss function in S20204 is similar to that in S20202, including conventional MSE loss and exponential response loss to SBN for controlling N2O, so it will not be elaborated further.

[0050] S20205, The main module and production module were frozen, NO3 - The leaching module and N2O emission module were implemented, and the learning rate for the yield module and NH3 volatilization module was set to 50% of the original learning rate to avoid overfitting and retain excessive prior knowledge. Measured climate data, soil data, management data, soil available phosphorus content data, soil available potassium content data, yield data, and NO3 data were used. -The production and NH3 volatilization modules in the machine learning model were fine-tuned using actual rinsing data, N2O emission data, and NH3 volatilization data, combined with knowledge-driven loss. The loss function in S20205 is similar to that in S20202, including conventional MSE loss and loss from controlling the exponential response of N2O to SBN, so it will not be elaborated further.

[0051] S20206, the main module, production module, NH3 volatilization module, and N2O emission module were frozen, and the production module and NO3 emission module were set. - The learning rate of the leaching module is 50% of the original learning rate to avoid overfitting and retain excessive prior knowledge. It utilizes measured climate data, soil data, management practice data, soil available phosphorus content data, soil available potassium content data, yield data, and NO3 data. - The measured data from rinsing, N2O emissions, and NH3 volatilization, combined with knowledge-driven loss analysis, were used to evaluate the yield module and NO3 in the machine learning model. - Fine-tuning of the rinsing module. The loss function in S20206 is similar to that in S20202, including the conventional MSE loss and the exponential response loss of N2O to SBN, so it will not be described in detail.

[0052] S203. Based on measured climate data, soil data, management data, and a nitrogen fertilizer recommendation model, the predicted nitrogen application rate and yield under optimal conditions are calculated. In this step, a multi-objective synergistic crop solution for high yield, high efficiency, and emission reduction in my country's three major grain crops is defined using the National Natural Science Foundation of China (NSGA-III) multi-objective optimization evolutionary algorithm. The adjustment unit is a 1km×1km grid, and the objectives within a specific area are to maximize yield, maximize nitrogen fertilizer use efficiency, and simultaneously reduce NH3 volatilization, N2O emissions, and NO3 emissions. - To minimize rinsing, the optimal combination of management measures is simulated within each grid using the following objective formula.

[0053] ; Where Z1 represents yield, Z2 represents nitrogen fertilizer utilization efficiency, and Y i,j Y represents the yield of crop j at grid i. FP E represents the yield potential that farmers can realize. i,j These represent the NH3 volatilization, N2O emissions, and NO3 emissions per unit area of ​​crop j in grid i, respectively. - Rinse.

[0054] The specific steps of the multi-objective optimization evolutionary algorithm (NSGA-III) are as follows: 1. Initialization: Calculate the Euclidean distance between the weight vectors and randomly generate the initial population x. 1 , ..., xN Create an external population (EP) to store high-quality individuals, initially empty; 2. Population Update: For each optimization objective, two random numbers k and t are selected from the neighborhood set B(i), and the genetic recombination operator x is used to update the population. k and x t To generate a new solution, y' is generated by applying the repair and improvement heuristic based on the test problem to y. High-quality individuals are selected to update the neighborhood solution B(i) and the external population EP. This process is repeated N times. 3. Termination Condition: If the termination condition is met, stop and output EP; otherwise, repeat step 2. This yields the optimal nitrogen application rate after multi-objective optimization, which maximizes yield, maximizes nitrogen fertilizer utilization efficiency, and minimizes environmental emissions, along with its potential yield value.

[0055] As shown in S2, traditional nitrogen fertilizer recommendation methods generally rely on experience or static function models, making it difficult to comprehensively consider the interactive effects of soil properties, climate conditions, crop type, and management methods. This results in nitrogen fertilizer application being applied in the same amount in the same location, leading to low efficiency and high environmental risks. In contrast, this step introduces simulated data from the DNDC model and APSIM model, combined with observational data, to train a multi-task deep learning model that predicts multiple indicators such as yield, PFP, N2O, NO3-, and NH3. This model has strong generalization ability and interpretability, and its prediction accuracy is improved through a multi-stage freezing and fine-tuning strategy.

[0056] S3. Based on the predicted yield and measured data of available phosphorus content in the soil, a phosphorus fertilizer recommendation model based on phosphorus fertilizer grading standards is constructed, and the predicted phosphorus application rate is obtained. Constant phosphorus fertilizer monitoring involves regularly monitoring the available phosphorus content and its changing trends in the soil and regulating fertilization to continuously stabilize the available phosphorus content within a certain range, ensuring it does not become a limiting factor for achieving the target yield. The recommended phosphorus and potassium fertilizer application rates for different regions are determined based on the constant phosphorus fertilizer monitoring method. Specifically, S3 includes the following steps: S301. The recommended phosphorus application rate is calculated based on the yield forecast and the recommended phosphorus fertilizer application rate formula. The formula for calculating the recommended phosphorus application rate is: Recommended phosphorus application rate = phosphorus uptake per 100 kg of grain × predicted yield.

[0057] S302. Determine the current phosphorus application rate for the region based on measured data of available phosphorus content in the soil and phosphate fertilizer grading standards. Phosphate fertilizer grading standards for the three common crops in the main planting areas are shown in Tables 1-3.

[0058] Table 1. Grading Standards for Available Phosphorus in Soils in Maize-Growing Areas Nationwide

[0059] Table 2. Grading Standards for Available Phosphorus in Soils in Wheat-Growing Areas Nationwide

[0060] Table 3. Grading Standards for Available Phosphorus in Soils in Rice-Growing Areas Nationwide

[0061] S303. Multiply the recommended phosphorus application rate by the phosphorus application rate ratio to obtain the predicted phosphorus application rate. Taking a corn-growing field in North China as an example, if the available phosphorus content in the soil of this field is 12.5 mg / kg, it belongs to the low level, and the corresponding recommended ratio is 130%. At this time, the predicted phosphorus application rate of this field = recommended phosphorus application rate × 130%.

[0062] S4. Based on the predicted yield and measured data of available potassium content in the soil, construct a potassium fertilizer recommendation model based on potassium fertilizer grading standards, and obtain the predicted potassium application rate. Constant potassium fertilizer monitoring, through regular monitoring of soil available potassium content and its changing trends and fertilization regulation, continuously stabilizes the soil available potassium content within a certain range, preventing it from becoming a limiting factor for achieving the target yield. The method based on constant potassium fertilizer monitoring determines the recommended potassium fertilizer application rate for different regions. Specifically, S4 includes the following steps: S401. The recommended potassium application rate is calculated based on the yield forecast and the recommended potassium fertilizer application rate formula. The formula for calculating the recommended potassium application rate is: Recommended potassium application rate = potassium uptake per 100 kg of grain × predicted yield.

[0063] S402. Determine the potassium application rate for the current region based on measured data of available potassium content in the soil and potassium fertilizer grading standards. See Table 4-6 for the phosphate fertilizer grading standards for the three common crops in their main planting areas.

[0064] Table 4. Grading Standards for Available Potassium in Soils in Maize-Growing Areas Nationwide

[0065] Table 5. Grading Standards for Available Potassium in Soils in Wheat-Growing Areas Nationwide

[0066] Table 6. Grading Standards for Available Potassium in Soils in Rice-Growing Areas Nationwide

[0067] S403. Multiply the recommended potassium application rate by the potassium application rate ratio to obtain the predicted potassium application rate. Taking a corn-growing field in North China as an example, if the available potassium content in the soil of this field is 124.2 mg / kg, it belongs to the high level, and the corresponding recommended ratio is 70%. At this time, the predicted potassium application rate of this field = recommended potassium application rate × 70%.

[0068] S5. Calculate the basal and topdressing amounts of nitrogen, phosphorus, and potassium fertilizers based on the predicted nitrogen, phosphorus, and potassium application rates and the basal-to-topdressing ratio. The basal-to-topdressing ratio is from the basal-to-topdressing ratio knowledge base, see Table 7.

[0069] Table 7. Baseline ratio of three crops in different regions (knowledge base)

[0070] S6. Construct a basal fertilizer formula model based on the nitrogen, phosphorus, and potassium basal fertilizer amounts, and obtain the basal fertilizer formula. Specifically, the expression of the basal fertilizer formula model is: ; in, Represents the largest integer not exceeding A. Represents the largest integer not exceeding B. Represents the largest integer not exceeding C. , , ; Where ω is the fertilization ratio, W N W is the predicted value for nitrogen application. P W is the predicted value for phosphorus application rate. K Here, A represents the predicted potassium application rate, B represents the nitrogen application rate coefficient, and C represents the phosphorus application rate coefficient. In this example... The number of solutions, n, in the basal fertilizer formulation model represents different feasible combinations. Each solution corresponds to a specific fertilizer formulation (e.g., 18-12-15, 20-10-15, etc.). These formulation combinations are then combined with the recommended total fertilizer application for each plot to form the final basal fertilizer formulation.

[0071] S7. Construct a dynamic adjustment model based on nutrient deficiency and dynamic adjustment factors during key topdressing periods, according to the topdressing amounts of nitrogen, phosphorus, and potassium, and calculate the correction values ​​for nitrogen, phosphorus, and potassium topdressing to generate a topdressing formula. This step aims to combine remote sensing dynamic monitoring with target yield guidance to achieve intelligent fine-tuning of topdressing amounts based on real-time crop growth. It is mainly applied to the "dynamic correction link" of nitrogen, phosphorus, and potassium topdressing recommendations to improve the nutrient matching efficiency of crops during key growth periods. Specifically, S7 includes the following steps: S701. Using the MODIS NDVI product (MOD13Q1, 250m resolution, 16-day composite), acquire remote sensing data for the current area within the sampling period preceding the topdressing period, and extract its NDVI values. This step also includes preprocessing: using ArcGIS 10.2 to resample and clip to plot boundaries, extract the plot mean NDVI, and obtain the current NDVI value for each plot. The specific process and principles of the preprocessing are existing technologies and will not be elaborated here. The topdressing period for different crops is determined according to the topdressing growth period knowledge base, as shown in Table 8.

[0072] Table 8. Knowledge Base of Sowing Dates and Topdressing Growth Periods for the Three Major Crops

[0073] For corn, the topdressing period is the large trumpet stage; for wheat, the topdressing period is the jointing stage; and for rice, the topdressing period is the greening stage and the young panicle differentiation stage.

[0074] S702. Determine the deficiency judgment conditions and dynamic adjustment factors based on the NDVI value and the preset NDVI threshold range. The preset NDVI threshold range is determined based on experience and is common knowledge. The specific values ​​in this example are shown in Table 9.

[0075] Table 9. Comparison of NDVI threshold ranges for three crops during the topdressing period.

[0076] Specifically, the nutrient deficiency judgment criteria and dynamic adjustment factors in S702 are determined in the following way: ; Where, m is the dynamic adjustment factor, σ1, σ2, and σ3 are adjustment coefficients, σ1∈[-7%,-2%], σ2∈[2%,7%], σ3∈[σ2+5%,σ2+10%], ε is the floating coefficient, 0<ε<1, NDVI max This represents the maximum value within the preset threshold range. The purpose of setting ε is to establish a buffer zone, preventing the difference in dynamic adjustment factors from being significantly greater than the difference in NDVI values ​​when the NDVI value is slightly greater or less than the extreme value, thus affecting adjustment accuracy. Therefore, users can determine the value of ε based on experience. In this example, taking rice as an example, generally, nitrogen has a greater impact on crop development, while phosphorus and potassium have relatively smaller impacts and are usually not deficient. Therefore, this example only considers nitrogen deficiency, and the values ​​of ε, σ1, σ2, and σ3 are 10%, -5%, 5%, and 10%, respectively. The deficiency judgment conditions and dynamic adjustment factors are as follows: If the NDVI value of the current region is 0.52, then m = 105%.

[0077] S703. Multiply the dynamic adjustment factor by the amount of nitrogen, phosphorus, and potassium topdressing to obtain the correction value for nitrogen, phosphorus, and potassium topdressing.

[0078] As an alternative example, S7 can also adopt the following method, specifically, S7 includes the following steps: S701. Obtain leaf images of crops in the current region. This step requires the user to upload crop images.

[0079] S702. Input the blade image into the Qwen2.5-VL-7B multimodal model to determine the missing nutrient judgment conditions and dynamic adjustment factors. Core parameter configuration of the Qwen2.5-VL-7B multimodal model: Basic model: Qwen2.5-VL-7B; Optimizer: AdamW; LoRA rank: 16; LoRA alpha: 32; Learning rate: 3e-4 (LoRA layer), 1e-6 (base model); Weight decay: 0.01; Batch size: 16 (batch_size=4, gradient_accumulation=4); Training rounds: 30; Core processing flow: Data quality filtering: Peak signal-to-noise ratio (PSNR) method is used to remove blurry images.

[0080] Cross-source data alignment: Resample all images to unify resolution.

[0081] Data labeling system: Constructing a multi-dimensional metadata architecture.

[0082] Nutrient deficiency types: Nitrogen / Phosphorus / Potassium; Crop type: Wheat / Corn / Rice; Growth stages: seedling stage / jointing stage / heading stage / grain filling stage; Symptom location: new leaves / old leaves / leaf veins / leaf margins; Auxiliary annotation items: data source (online / field), environmental parameters (light intensity / soil moisture content).

[0083] In this step, the Qwen2.5-VL-7B multimodal model outputs crop information, nutrient deficiency type, and symptom description. Nutrient deficiency criteria and dynamic adjustment factors are determined as follows: ; Where σ1 < 0 and σ2 > 0.

[0084] S703. Multiply the dynamic adjustment factor by the amount of nitrogen, phosphorus, and potassium topdressing to obtain the correction value for nitrogen, phosphorus, and potassium topdressing.

[0085] As shown in S7, traditional topdressing uses fixed timing and amount, neglecting real-time differences in growth and diagnosis of crop vigor, resulting in "insufficient topdressing" and "over-topdressing." This step provides two methods for dynamic adjustment of topdressing. The method using remote sensing technology can achieve dynamic topdressing control at the plot level of ±15% based on MODIS NDVI and crop nutrient deficiency thresholds.

[0086] Another method, using leaf image recognition, allows users to upload plant photos and uses the Qwen2.5-VL-7B multimodal model for accurate nutrient deficiency diagnosis. The prescription diagram is automatically updated after adjustments, achieving true "time-dependent" results.

[0087] This invention also provides a crop-specific fertilizer formulation generation system, comprising: The acquisition module is used to acquire crop growth data for any region. This data includes measured climate data, measured soil data, measured data on management practices, measured soil available phosphorus content, measured soil available potassium content, measured yield data, and NO3 data. - Actual data from rinsing, N2O emissions, and NH3 volatilization.

[0088] The nitrogen fertilizer recommendation model calculation module is used to build a machine learning-based nitrogen fertilizer recommendation model based on crop growth data, and obtain the predicted value of nitrogen application rate and yield.

[0089] The phosphate fertilizer recommendation model calculation module is used to construct a phosphate fertilizer recommendation model based on phosphate fertilizer grading standards based on the predicted yield value and the measured data of available phosphorus content in the soil, and to obtain the predicted value of phosphorus application rate.

[0090] The potassium fertilizer recommendation model calculation module is used to construct a potassium fertilizer recommendation model based on potassium fertilizer classification standards based on the predicted yield value and the measured data of available potassium content in the soil, and to obtain the predicted value of potassium application rate.

[0091] The base fertilizer and topdressing calculation module is used to calculate the amount of nitrogen, phosphorus and potassium base fertilizer and topdressing fertilizer based on the predicted nitrogen application rate, the predicted phosphorus application rate, the predicted potassium application rate and the base fertilizer and topdressing ratio, respectively.

[0092] The base fertilizer formula generation module is used to construct a base fertilizer formula model based on the amount of nitrogen, phosphorus, and potassium in the base fertilizer and obtain the base fertilizer formula.

[0093] The topdressing formula generation module is used to construct a dynamic adjustment model based on the nutrient deficiency-dynamic adjustment factor during the key topdressing period according to the amount of nitrogen, phosphorus and potassium topdressing, and to calculate the topdressing correction value of nitrogen, phosphorus and potassium to generate the topdressing formula.

[0094] Based on the same inventive concept, the present invention also provides a crop-specific fertilizer formula generation device, including a processor and a memory, wherein the memory is used to store a computer program and the processor is used to execute the computer program to implement the steps of the above method.

[0095] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating a crop-specific fertilizer formula, characterized in that, Includes the following steps: S1. Obtain crop growth data for any region, including measured climate data, measured soil data, measured management data, measured soil available phosphorus content, measured soil available potassium content, measured yield data, and NO3. - Actual data from rinsing, N2O emissions, and NH3 volatilization; S2. Construct a nitrogen fertilizer recommendation model based on machine learning based on the crop growth data, and obtain the predicted value of nitrogen application rate and the predicted value of yield; S3. Based on the predicted yield and the measured data of available phosphorus content in the soil, construct a phosphorus fertilizer recommendation model based on the phosphorus fertilizer grading standard, and obtain the predicted phosphorus application rate. S4. Based on the predicted yield and the measured data of available potassium content in the soil, construct a potassium fertilizer recommendation model based on the potassium fertilizer grading standard, and obtain the predicted potassium application rate. S5. Calculate the amount of nitrogen, phosphorus, and potassium fertilizer as basal fertilizer and topdressing fertilizer based on the predicted nitrogen application rate, the predicted phosphorus application rate, the predicted potassium application rate, and the basal-topdressing ratio. S6. Construct a base fertilizer formula model based on the amount of nitrogen, phosphorus, and potassium in the base fertilizer, and obtain the base fertilizer formula; S7. Construct a dynamic adjustment model based on the nutrient deficiency-dynamic adjustment factor during the key topdressing period according to the amount of nitrogen, phosphorus and potassium topdressing, and calculate the topdressing correction value of nitrogen, phosphorus and potassium to generate the topdressing formula.

2. The method for generating a crop-specific fertilizer formula according to claim 1, characterized in that, S2 includes the following steps: S201. Based on the measured climate data, measured soil data, measured management data, DNDC model, and APSIM model, obtain the predicted yield and NO3. - Predicted values ​​for leaching, N2O emissions, and NH3 volatilization; S202, Based on the crop growth data, yield forecast, and NO3... - A nitrogen fertilizer recommendation model based on an attention mechanism recurrent neural network was constructed using leaching predictions, N2O emission predictions, and NH3 volatilization predictions. S203. Based on the measured climate data, measured soil data, measured management data, and nitrogen fertilizer recommendation model, calculate the predicted nitrogen application rate and yield under optimal conditions.

3. The method for generating a crop-specific fertilizer formula according to claim 1, characterized in that, S3 includes the following steps: S301. The recommended phosphorus application rate is calculated based on the predicted yield value and the recommended phosphorus fertilizer application rate formula. S302. Determine the current phosphorus application rate for the region based on the measured data of available phosphorus content in the soil and the phosphate fertilizer grading standards. S303. Multiply the recommended phosphorus application rate by the phosphorus application rate ratio to obtain the predicted phosphorus application rate.

4. The method for generating a crop-specific fertilizer formula according to claim 1, characterized in that, S4 includes the following steps: S401. The recommended potassium application rate is calculated based on the predicted yield and the recommended potassium fertilizer application rate formula. S402. Determine the current potassium application rate ratio based on the measured data of available potassium content in the soil and the potassium fertilizer grading standards. S403. Multiply the recommended potassium application rate by the potassium application rate ratio to obtain the predicted potassium application rate.

5. The method for generating a crop-specific fertilizer formula according to claim 1, characterized in that, The expression for the base fertilizer formulation model is: ; in, Represents the largest integer not exceeding A. Represents the largest integer not exceeding B. Represents the largest integer not exceeding C. , , ; ω represents the fertilizer application ratio, W N W is the predicted value for nitrogen application. P W is the predicted value for phosphorus application rate. K A represents the predicted potassium application rate, B represents the nitrogen application rate coefficient, C represents the phosphorus application rate coefficient, and C represents the potassium application rate coefficient.

6. The method for generating a crop-specific fertilizer formula according to claim 1, characterized in that, S7 includes the following steps: S701. Obtain remote sensing data of the current area within the previous sampling period before the topdressing period, and extract its NDVI value; S702. Determine the deficiency judgment conditions and dynamic adjustment factors based on the NDVI value and the preset NDVI threshold range; S703. Multiply the dynamic adjustment factor by the amount of nitrogen, phosphorus and potassium topdressing to obtain the nitrogen, phosphorus and potassium topdressing correction value.

7. The method for generating a crop-specific fertilizer formula according to claim 6, characterized in that, The deficiency criterion and dynamic adjustment factor in S702 are determined in the following manner: ； Where m is the dynamic adjustment factor, σ1, σ2, and σ3 are adjustment coefficients, σ1∈[-7%,-2%], σ2∈[2%,7%], σ3∈[σ2+5%,σ2+10%], ε is the floating coefficient, 0<ε<1, NDVI max This is the maximum value within the preset threshold range.

8. The method for generating a crop-specific fertilizer formula according to claim 6, characterized in that, S7 includes the following steps: S701. Obtain leaf images of crops in the current region; S702. Input the blade image into the Qwen2.5-VL-7B multimodal model to determine the deficiency nutrient judgment conditions and dynamic adjustment factors; S703. Multiply the dynamic adjustment factor by the amount of nitrogen, phosphorus and potassium topdressing to obtain the nitrogen, phosphorus and potassium topdressing correction value.

9. A crop-specific fertilizer formulation generation system, characterized in that, include: The acquisition module is used to acquire crop growth data for any region. This crop growth data includes measured climate data, measured soil data, measured management data, measured soil available phosphorus content, measured soil available potassium content, measured yield data, and NO3. - Actual data from rinsing, N2O emissions, and NH3 volatilization; The nitrogen fertilizer recommendation model calculation module is used to construct a machine learning-based nitrogen fertilizer recommendation model based on the crop growth data, and obtain the predicted nitrogen application rate and the predicted yield. The phosphate fertilizer recommendation model calculation module is used to construct a phosphate fertilizer recommendation model based on the phosphate fertilizer grading standard based on the predicted yield value and the measured data of available phosphorus content in the soil, and to obtain the predicted value of phosphorus application rate. The potassium fertilizer recommendation model calculation module is used to construct a potassium fertilizer recommendation model based on the potassium fertilizer grading standard based on the predicted yield value and the measured data of available potassium content in the soil, and to obtain the predicted potassium application rate. The base fertilizer and topdressing calculation module is used to calculate the amount of nitrogen, phosphorus and potassium base fertilizer and topdressing fertilizer based on the predicted nitrogen application rate, the predicted phosphorus application rate, the predicted potassium application rate and the base fertilizer and topdressing ratio, respectively. The base fertilizer formula generation module is used to construct a base fertilizer formula model based on the amount of nitrogen, phosphorus, and potassium in the base fertilizer and obtain the base fertilizer formula. The topdressing formula generation module is used to construct a dynamic adjustment model based on the nutrient deficiency-dynamic adjustment factor during the key topdressing period according to the amount of nitrogen, phosphorus and potassium topdressing, and to calculate the topdressing correction value of nitrogen, phosphorus and potassium to generate the topdressing formula.

10. A crop-specific fertilizer formulation generation device, characterized in that, It includes a processor and a memory, the memory being used to store a computer program, and the processor being used to execute the computer program to implement the steps of the method of claim 1.

Images