Method, system, and equipment for generating the optimal fertilizer blend ratio for agricultural crops.

Abstract

This invention provides a method, system, and equipment for generating the optimal blending ratio of fertilizers specifically for agricultural crops. [Solution] The method includes acquiring crop growth data, constructing a nitrogen fertilizer recommendation model based on the crop growth data and acquiring predicted nitrogen application rates and yield predictions, constructing a phosphorus fertilizer recommendation model based on phosphorus fertilizer grading standards based on yield predictions and actual soil fast-acting phosphorus content data and acquiring predicted phosphorus application rates, constructing a potassium fertilizer recommendation model based on potassium fertilizer grading standards and acquiring predicted potassium application rates, calculating the base fertilizer and top dressing amounts of nitrogen, phosphorus, and potassium based on the predicted nitrogen application rates, predicted phosphorus application rates, predicted potassium application rates, and base fertilizer and top dressing ratios, constructing a base fertilizer blending ratio model based on the base fertilizer amounts of nitrogen, phosphorus, and potassium, and constructing a dynamic adjustment model based on the top dressing amounts of nitrogen, phosphorus, and potassium and calculating top dressing correction values.

Description

[Technical Field]

[0001] This invention belongs to the field of smart cultivation technology, and more specifically, relates to a method, system, and equipment for generating the blending ratio of fertilizers specifically for agricultural crops. [Background technology]

[0002] Currently, Chinese agricultural production is undergoing a critical phase of transformation from "high input, high consumption" to "high efficiency, greening." Chemical fertilizers have long supported China's food security as one of the main means of improving crop yields. However, excessive reliance on empirical fertilization and static mixing ratios not only reduces fertilizer utilization but also leads to significant ecological and economic problems. Specifically, these include the following:

[0003] 1. Currently, many regions still rely primarily on paper maps or administrative district recommendations, and are unable to dynamically update fertilization strategies according to differences in different crops, different plots, and different years. The drawing of fertilizer mixture ratio diagrams and recommendation of mixture ratios often rely on human experience or static average models, and lack data-driven and feedback optimization mechanisms. As a result, deviations in the actual application process are relatively large, accuracy is low, and widespread adoption is difficult.

[0004] 2. At the fertilizer product level, the mixing ratios of large quantities of "general-purpose compound fertilizers" cannot be optimized for specific crops, regions, and yield targets, leading to "mixing ratio errors, unknown usage amounts, and unstable effects." Farmers lack clear reference standards for fertilizer mixing, and companies lack data to support a system platform for optimizing mixing ratios and iterating on products, resulting in a break in the "production-supply-application" chain.

[0005] Thirdly, traditional intelligent exploration of fertilization often remains at a breakthrough point (for example, variable fertilization machines or remote sensing monitoring), and there is no closed-loop system from data sensing, algorithm inference to formulation ratio generation, formulation ratio use, and result feedback. In particular, at the county and plot scales, there is no smart system that can efficiently integrate multi-source information such as soil, crops, and meteorology to output a reliable fertilization formulation ratio.

[0006] Therefore, in order to achieve the automation, personalization, and efficiency of fertilization for major crops and reliably support the development of modern agriculture in the direction of accuracy, high production, high efficiency, and environmental protection, it is necessary to develop a smart system that integrates "data collection - fertilization decision-making - formulation ratio optimization - formulation ratio output - equipment adaptation".

Summary of the Invention

Problems to be Solved by the Invention

[0007] Provided are a method, a system, and equipment for generating a formulation ratio of a special fertilizer for crops, which can dynamically adjust the fertilization strategy according to differences in crops, regions, and years, thereby improving the reliability of the fertilization formulation ratio.

Means for Solving the Problems

[0008] The present invention provides a method for generating a formulation ratio of a special fertilizer for crops, which is step S1 of obtaining crop growth data in any region, where the crop growth data includes climate measured data, soil measured data, management measure measured data, soil available phosphorus content measured data, soil available potassium content measured data, yield measured data, NO3 - leaching measured data, N2O emission measured data, and NH3 volatilization measured data, step S1, and step S2 of constructing a nitrogen fertilizer recommendation model based on machine learning based on the crop growth data and obtaining a nitrogen application rate prediction value and a yield prediction value. Step S3 of constructing a phosphate fertilizer recommendation model based on the phosphate fertilizer grading criteria and obtaining a predicted phosphate application rate based on the predicted yield value and the measured data of the available phosphorus content in the soil; Step S4 of constructing a potassium fertilizer recommendation model based on the potassium fertilizer grading criteria and obtaining a predicted potassium application rate based on the predicted yield value and the measured data of the available potassium content in the soil; Step S5 of calculating the base fertilizer amount and top dressing amount of nitrogen, phosphorus, and potassium respectively based on the predicted nitrogen application rate, predicted phosphate application rate, predicted potassium application rate, and base fertilizer to top dressing ratio; Step S6 of constructing a base fertilizer blending ratio model based on the base fertilizer amount of nitrogen, phosphorus, and potassium and obtaining the blending ratio of the base fertilizer; Step S7 of constructing a dynamic adjustment model based on the lack element-dynamic adjustment factor during the important top dressing period based on the top dressing amount of nitrogen, phosphorus, and potassium, calculating the top dressing correction value of nitrogen, phosphorus, and potassium, and generating the blending ratio of the top dressing, including.

[0009] Furthermore, step S2 includes: Step S201 of obtaining the predicted yield value, NO3 - leaching prediction value, N2O emission prediction value, and NH3 volatilization prediction value based on the measured climate data, measured soil data, measured management measure data, DNDC (Denitrification-Decomposition) model, and APSIM (Agricultural Production Systems Simulator) model; Step S202 of constructing a nitrogen fertilizer recommendation model of a recurrent neural network based on the attention mechanism based on the crop growth data, predicted yield value, NO3 - leaching prediction value, N2O emission prediction value, and NH3 volatilization prediction value; Step S203 of calculating the predicted nitrogen application rate and predicted yield value under the optimal conditions based on the measured climate data, measured soil data, measured management measure data, and nitrogen fertilizer recommendation model, including.

[0010] Furthermore, step S3 includes: Step S301 calculates the recommended phosphorus application rate based on the yield prediction value and the recommended phosphorus fertilizer application rate formula, Step S302 involves determining the current regional phosphate application rate ratio based on the measured soil rapid-acting phosphate content data and phosphate fertilizer grading standards. The method includes step S303, which involves multiplying the recommended phosphate application rate by the phosphate application rate ratio to obtain a predicted phosphate application rate.

[0011] Furthermore, step S4 is, Step S401 involves calculating the recommended potassium application rate based on the yield prediction value and the recommended potassium fertilizer application rate formula, Step S402 involves determining the current regional potassium application rate ratio based on the measured soil rapid-acting potassium content data and potassium fertilizer grading standards. The method includes step S403, which involves multiplying the recommended potassium application rate by the potassium application rate to obtain a predicted potassium application rate.

[0012] Furthermore, the expression for the base fertilizer mixing ratio model is ω = [A]% + [B]% + [C]%, However, [A] is the largest integer not exceeding A, [B] is the largest integer not exceeding B, and [C] is the largest integer not exceeding C.

number

[0013] Furthermore, step S7 is, Step S701 involves acquiring remote sensing data for the current region that falls within the sampling cycle before the top dressing period, and extracting the NDVI (Normalized Difference Vegetation Index) value. Step S702 determines the element deficiency judgment conditions and dynamic adjustment factors based on the NDVI value and a predetermined NDVI threshold range, The method includes step S703, which involves multiplying the aforementioned dynamic adjusting factor by the amount of nitrogen, phosphorus, and potassium added as top dressing to obtain a correction value for nitrogen, phosphorus, and potassium added as top dressing.

[0014] Furthermore, the element deficiency determination conditions and dynamic adjustment factors in step S702 are determined as follows:

number

[0015] Furthermore, step S7 is, Step S701 involves obtaining leaf images of crops in the current region, Step S702 involves inputting the aforementioned leaf images into the Qwen2.5-VL-7B multimodal model to determine the element deficiency judgment conditions and dynamic adjustment factors. The method includes step S703, which involves multiplying the aforementioned dynamic adjusting factor by the amount of nitrogen, phosphorus, and potassium added as top dressing to obtain a correction value for nitrogen, phosphorus, and potassium added as top dressing.

[0016] The present invention further provides a system for generating the blending ratio of fertilizers specifically for agricultural crops. An acquisition module used to obtain crop growth data for any region, wherein the crop growth data includes actual climate data, actual soil data, actual management measures data, actual soil rapid-acting phosphorus content data, actual soil rapid-acting potassium content data, actual yield data, NO3- An acquisition module including measured leaching data, measured N2O emissions data, and measured NH3 volatilization data, A nitrogen fertilizer recommendation model calculation module is used to construct a machine learning-based nitrogen fertilizer recommendation model based on the aforementioned crop growth data, and to obtain predicted nitrogen application rates and yields. A phosphate fertilizer recommendation model calculation module is used to construct a phosphate fertilizer recommendation model based on phosphate fertilizer grading standards and to obtain a phosphate application rate prediction value, based on the aforementioned yield prediction value and actual soil rapid-acting phosphate content data. A potassium fertilizer recommendation model calculation module is used to construct a potassium fertilizer recommendation model based on potassium fertilizer grading standards, using the aforementioned yield prediction values ​​and measured soil rapid-acting potassium content data, and to obtain potassium application rate prediction values. A base fertilizer and top dressing calculation module used to calculate the base fertilizer and top dressing amounts of nitrogen, phosphorus, and potassium, respectively, based on predicted nitrogen application rates, predicted phosphorus application rates, predicted potassium application rates, and base fertilizer and top dressing ratios. A base fertilizer mixing ratio generation module is used to construct a base fertilizer mixing ratio model based on the amounts of nitrogen, phosphorus, and potassium in the base fertilizer, and to obtain the base fertilizer mixing ratio. This includes a top dressing ratio generation module used to construct a dynamic adjustment model based on the amount of nitrogen, phosphorus, and potassium applied as top dressing, based on deficiency element-dynamic adjustment factors during critical top dressing periods, and to calculate top dressing correction values ​​for nitrogen, phosphorus, and potassium, and to generate top dressing ratios.

[0017] The present invention also provides equipment for generating the blending ratio of fertilizers specifically for agricultural crops, comprising a processor and memory, wherein the memory is used to store a computer program, and the processor is configured to execute the computer program to realize the steps of the above method. [Effects of the Invention]

[0018] (1) The present invention can accurately and efficiently create a fertilization strategy of "fertilizing according to the type, location, and time of crops", significantly improve the utilization efficiency of fertilizers, and reduce over-application.

[0019] (2) The present invention dynamically adjusts the fertilization strategy based on real-time environment and crop conditions, avoids the risk of inconsistency in static mixing ratios, and improves the adaptability of the system.

Brief Description of the Drawings

[0020] [Figure 1] It is a flowchart of the method in the present invention.

Embodiments for Carrying out the Invention

[0021] Hereinafter, the present invention will be further described with reference to the drawings. The following examples are only for more clearly explaining the technical solutions of the present invention and do not limit the protection scope of the present invention.

[0022] As shown in FIG. 1, the present invention provides a method for generating a mixing ratio of fertilizers for crops, including the following steps S1 to S7.

[0023] S1. Obtain the crop growth data of any region. The crop growth data includes climate measured data, soil measured data, management measure measured data, soil available phosphorus content measured data, soil available potassium content measured data, yield measured data, NO3 - leaching measured data, N2O emission measured data, and NH3 volatilization measured data. The climate measured data includes annual average temperature and annual average precipitation. The soil measured data includes soil organic carbon, total nitrogen, pH, unit weight, and clay content. The management measure measured data includes the type of crops (rice, wheat, corn), the type of fertilizers (urea, mineral nitrogen fertilizers, urea with inhibitors), the nitrogen application method (surface application, deep application), the soil tillage method (conventional tillage and no tillage), the number of nitrogen fertilizer applications (1 time, 2 times, 3 times), and the nitrogen application rate.

[0024] To ensure the comprehensiveness and representativeness of the database, 10,000 grid scenes were randomly selected from each of the three crops for pre-training. Specifically, for each sample, one grid and one scene were randomly selected from approximately 72,805 grids and 72 types of management combination scenes. The selected 10,000 samples are distributed fairly evenly across all regions of China and can adequately represent all scene combinations, covering agricultural production conditions under different crops, climates, soils, and management measures, thereby ensuring the rapid reflection of diversified agricultural production conditions in the synthetic database. All of the above data underwent feature standardization and one-hot coding preprocessing to ensure consistency and readability of the data format input.

[0025] S2 constructs a machine learning-based nitrogen fertilizer recommendation model based on crop growth data and obtains predicted nitrogen application rates and yields. Specifically, S2 includes the following steps S201 to S203.

[0026] S201, actual climate data, actual soil data, actual management measures data, yield prediction values ​​based on DNDC model and APSIM model, NO3 - Predicted leaching values, N2O emission values, and NH3 volatilization values ​​are obtained. Yield and NO3 are measured using actual data. - Because the sample sizes for the four data points—leaching values, N2O emissions, and NH3 volatility—were relatively small, this step involved simulating these four data points using the DNDC and APSIM models to meet training requirements. Climate measurement data, soil measurement data, and control measure measurement data were input into the APSIM model to obtain yield predictions. Climate measurement data, soil measurement data, and control measure measurement data were input into the DNDC model to simulate NO3 - Predicted values ​​for leaching, N2O emissions, and NH3 volatilization were obtained. In this way, a sufficient amount of sample data was obtained, and this data was used to train the model.

[0027] S202, Crop growth data, Yield forecast, NO3- A recurrent neural network nitrogen fertilizer recommendation model based on attention mechanisms is constructed based on predicted leaching, predicted N2O emissions, and predicted NH3 volatilization. A knowledge-guided machine learning model is trained based on synthetic data generated by the crop growth model APSIM and the nitrogen management model DNDC, as well as observational data collected from the literature. The training model network used is a multi-task deep feedforward neural network (Multilayer Perceptron, MLP), and its overall structure can be generalized as Input Projection → Principal Feature Extraction Module (BackboneMLPBlock stack) → Task-specific Output Heads. The network design considers yield, N2O emissions, and NO3. - The aim is to simultaneously predict multiple target variables, including leaching and NH3 emission, and to support high-speed fine-tuning by freezing the main module.

[0028] 1. Model Input Layer: 1. Input Features: Embedding vectors containing continuous numerical features and classification variables (categorical embeddings). The input dimension is the sum of the number of continuous features plus the dimensions of all classification feature embeddings.

[0029] 2. Linear Projection: A single-layer fully connected network (Linearlayer) is used to project the original input into a 128-dimensional hidden space. Function: Unifies the feature dimensions and generates a hidden vector representation suitable for subsequent MLPBlock processing.

[0030] 2. Main Feature Extraction Module (BackboneMLPBlock): 1. Module Stack: The main unit is composed of six MLPBlocks with the same structure stacked in sequence.

[0031] Each MLPBlock has an input dimension of 128 and an output dimension of 128.

[0032] 2. Single MLPBlock structure: The first layer has a fully connected (FC) architecture: 128 → 512 (the dimension of the hidden layer is four times the input dimension), and later the GELU activation function (Gaussian Error Linear Unit, GELU) is used.

[0033] Dropout: Used to prevent overfitting.

[0034] Fully connected (FC) layers in the second layer: 512 → 128.

[0035] Residual Connection: The input is added to the output of the second layer, enhancing the gradient stream and training stability.

[0036] Layer Normalization (LayerNorm): This process normalizes the residual output to improve training stability and convergence speed.

[0037] Each Block structure is Input 128->Full Connection 512->GELU Activation->Dropout->Full Connection 128->Residual Connection->LayerNorm.

[0038] 3. Module Function: The core extracts general-purpose, high-dimensional feature representations from input features, providing a shared foundation for multi-task prediction.

[0039] 3. Multitask output branches (Task-specificHeads): 1. Structural Design: From the 128-dimensional feature vector output from the main unit, five independent small MLPs were separated, each with one target variable (yield, PFP (Partial Factor Productivity), N2O emissions, NH3 volatilization, and NO3). -It is used to predict (including elution). Each small MLP consists of a two-layer fully connected network of 128→32→1, with the hidden layer activated using GELU and aided by Dropout.

[0040] 2. Functional Description: The compact output network focuses on task-related feature mapping to achieve independent predictions for different tasks.

[0041] Backbone freezing: 1. Definition and Principle: Freezing refers to fixing network parameters during the training process and not performing gradient updates. During training, the parameter update formula is:

number

[0042] 2. Implementation method: Freezing is performed by traversing network parameters and identifying the name of the main module.

[0043] for name,param in model.named_parameters(): if “MLPBlock” in name:# or backbone module name param.requires_grad=False

[0044] 3. Training effect after freezing: The main feature extraction module parameters remain fixed, and only the output branch (task-specificheads) parameters are updated.

[0045] Advantages: It retains the principal's existing knowledge, allows for quick adaptation to new tasks, and reduces the risk of overfitting.

[0046] 4. Comparison of fine-tuning and full-volume training: Backbone: A deep, general-purpose feature extraction network that can be frozen.

[0047] Fine-tuning network device (OutputHead): This is a task-related network, and its parameters can be updated to adapt to new tasks.

[0048] Training strategy selection: Main branch freezing + output branch training: rapid fine-tuning, applying to existing foundational feature knowledge scenes.

[0049] Full training: Learns new tasks from scratch, and the main parameters can be updated.

[0050] Based on the above model, specifically, S202 includes steps S20201 to S20206.

[0051] S20201, NO3 - The leaching module, N2O emission module, and NH3 volatilization module are frozen, i.e., the learning gradient is set to zero. Crop growth data, yield predictions, and NO3 for 19 out of 22 years are collected. - Yield and PFP modules in a machine learning model were trained using predicted leaching, N2O emission, and NH3 volatilization values ​​as input data, and a mapping relationship between the input data, yield, and nitrogen fertilizer utilization efficiency was obtained. A training loss was constructed using an adaptive learning method based on mean squared error, and training was started with "relatively easy" samples and gradually progressed to "relatively difficult" samples.

[0052] The specific loss function is:

number

[0053] S20202, freeze the PFP module and use input data to combine loss and response of knowledge drive in the yield module of the machine learning model, NO3 - The leaching module, N2O emission module, and NH3 volatilization module were trained, and input data and yield, NO3 - The relationship between leaching value, N2O emission value, and NH3 volatilization value is obtained. In addition to the mean squared error loss function Loss1, the loss function of S20202 is further defined as (1) input nitrogen application rate and predicted yield, N2O emission, NH3 volatilization, and NO3 - (2) Nitrogen equilibrium loss function Loss2, (2) N2O emission, NH3 volatilization and NO3 - Leaching SBN k This relates to controlling the exponential response loss function Loss3 for the following. After two stages of pre-training, the knowledge-guided machine learning model controls yield, nitrogen fertilizer use efficiency (PFP), N2O emissions, NH3 volatilization, NO3 - Erosion can be successfully predicted. The specific loss function is:

number

[0054] S20203, main module and NO3 - The leaching module, N2O emission module, and NH3 volatilization module were frozen, and the learning rate was set to 20% of the initial learning rate. Climate measurement data, soil measurement data, management measure measurement data, soil rapid-acting phosphorus content measurement data, soil rapid-acting potassium content measurement data, yield measurement data, NO3 - The yield and PFP modules in the machine learning model are fine-tuned using measured leaching data, measured N2O emissions data, and measured NH3 volatilization data, combined with knowledge drive losses. The loss function in S20203 is (1) Mean-Square Error (MSE) loss function Loss1 between observed and predicted values, and (2) y' yield and Nfer×y' PFP The mean squared error loss function between (3)y' yield (4) The relationship between PFP and Nfer includes a quadratic relational loss function Loss3 that constrains the opening to be downward, and a negative response loss function Loss4 that restricts the relationship between PFP and Nfer, thereby enhancing the biological rationality and predictive accuracy of the model. The specific loss function is:

number

[0055] S20204, main module and yield module, NO3 - Overfitting and excessive retention of prior knowledge are avoided by freezing the leaching module and the NH3 volatilization module, and setting the learning rate of the yield module and the N2O emission module to 50% of the initial learning rate. Climate measurement data, soil measurement data, control measure measurement data, soil rapid-acting phosphorus content measurement data, soil rapid-acting potassium content measurement data, yield measurement data, NO3 -The yield module and N2O emission module in the machine learning model are fine-tuned using measured leaching data, measured N2O emission data, and measured NH3 volatilization data, combined with knowledge drive losses. The loss function in S20204 is similar to S20202 and includes control for the normal MSE loss and the exponential response loss of N2O to SBN, so its explanation is omitted.

[0056] S20205, main module and yield module, NO3 - Overfitting and excessive retention of prior knowledge are avoided by freezing the leaching module and N2O emission module, and by setting the learning rate of the yield module and NH3 volatilization module to 50% of the initial learning rate. Climate measurement data, soil measurement data, control measure measurement data, soil rapid-acting phosphorus content measurement data, soil rapid-acting potassium content measurement data, yield measurement data, NO3 - The yield module and NH3 volatilization module in the machine learning model are fine-tuned using measured leaching data, measured N2O emissions data, and measured NH3 volatilization data, combined with knowledge drive losses. The loss function in S20205 is similar to S20202 and includes control for the normal MSE loss and the exponential response loss of N2O to SBN, so its explanation is omitted.

[0057] S20206, freeze the main module and yield module, the NH3 volatilization module and the N2O emission module, and the yield module and NO3 - By setting the learning rate of the leaching module to 50% of the initial learning rate, overfitting and excessive retention of prior knowledge are avoided. Climate measurement data, soil measurement data, control measure measurement data, soil rapid-acting phosphorus content measurement data, soil rapid-acting potassium content measurement data, yield measurement data, NO3 - Using measured leaching data, measured N2O emissions data, and measured NH3 volatilization data, and combined with knowledge drive loss, the yield module and NO3 in the machine learning model are used. -Fine-tune the elution module. The loss function in S20206 is similar to that of S20202 and includes control for the normal MSE loss and the exponential response loss of N2O to SBN, so its explanation is omitted.

[0058] Based on S203, climate measurement data, soil measurement data, management measure measurement data, and nitrogen fertilizer recommendation models, the predicted nitrogen application rate and yield under optimal conditions are calculated. In this step, the Multi-Objective Optimization Evolutionary Algorithm (NSGA-III) is used to reveal the entire process solution for high-yield, high-efficiency, and emission-reduction cooperative multi-objective agricultural crops for China's three major food crops. The adjustment unit is a 1km x 1km grid with accuracy, and the objectives within a specific area are to maximize yield, maximize nitrogen fertilizer utilization efficiency, and simultaneously reduce NH3 volatilization, N2O emissions, and NO3 emissions. - The system is adjusted to minimize leaching, and the optimal combination of control measures is simulated within each grid using the following target equation.

[0059]

number

[0060] The specific steps of the Multi-Objective Optimization Evolutionary Algorithm (NSGA-III) are as follows: 1. Initialization: Calculate the Euclidean distance between weight vectors and determine the initial population x 1 , ..., x N Randomly generate and create one external population (EP) to remember high-quality individuals, with an initial empty set. 2. Population Update: For each optimization goal, two numbers k and t are randomly selected from the adjacency set B(i) to create a genetically modified genetic operator x k and xt Using this, a new solution is generated, repair and improvement heuristics based on the test problem are applied to y to generate y', high-quality individuals are selected to update the adjacent region solution B(i) and the outer population EP, and by analogy, this is repeated N times. 3. Condition completion: If the stop condition is met, the system stops and outputs EP. If the stop condition is not met, step 2 is repeated. This obtains the optimal nitrogen application rate after multi-objective optimization that maximizes yield, maximizes nitrogen fertilizer utilization efficiency, and minimizes environmental emissions, as well as the yield potential value that it achieves.

[0061] As can be seen from S2, conventional nitrogen fertilizer recommendation methods generally rely on empirical or static function models, making it difficult to comprehensively consider the interplay of soil attributes, climatic conditions, crop types, and management methods. This leads to applying nitrogen fertilizer "the same amount in the same place," resulting in low efficiency and high environmental risk. In this step, simulation data + observational data from the DNDC model and APSIM model are introduced to train a multi-task deep learning model, and yield, PFP, N2O, NO3 - It predicts multiple indicators, including NH3, possesses high generalization ability and interpretability, and improves prediction accuracy through multi-stage freezing and fine-tuning strategies.

[0062] Based on S3, yield predictions, and actual data on soil fast-acting phosphorus content, a phosphorus fertilizer recommendation model based on phosphorus fertilizer grading criteria is constructed, and phosphorus application rates are predicted. Quantitative monitoring of phosphorus fertilizers ensures that soil fast-acting phosphorus content remains stable within a certain range through regular monitoring and adjustment of fertilizer application to the soil fast-acting phosphorus content and its changing trends, thus preventing it from becoming a limiting factor in achieving target yields. Based on the phosphorus fertilizer monitoring method, recommended usage amounts of potassium phosphate fertilizers for different regions are determined. Specifically, S3 includes steps S301 to S303.

[0063] S301, calculate the recommended phosphate application rate based on the yield forecast and the recommended phosphate fertilizer application rate formula. The formula for calculating the recommended phosphate application rate is: The recommended phosphate application rate is calculated as: phosphate absorption per 100 kilograms of seeds × predicted yield.

[0064] Based on S302, measured data on soil fast-acting phosphorus content and phosphorus fertilizer grading standards, the current phosphorus application rate ratio for the region is determined. Tables 1 to 3 show the phosphorus fertilizer grading standards for three common crops in the main cultivation areas.

[0065] [Table 1]

[0066] [Table 2]

[0067] [Table 3]

[0068] S303. The recommended phosphate application rate is multiplied by the phosphate application rate ratio to obtain a predicted phosphate application rate. Taking a maize cultivation plot in the North China region as an example, if the soil's fast-acting phosphate content in that plot is 12.5 mg / kg, it falls into the low level category, and the corresponding recommended ratio is 130%. In this case, the predicted phosphate application rate for that plot = recommended phosphate application rate × 130%.

[0069] Based on S4, yield predictions, and actual measured data of soil fast-acting potassium content, a potassium fertilizer recommendation model is constructed based on potassium fertilizer grading standards, and a potassium application rate prediction is obtained. Potassium fertilizer quantitative monitoring ensures that the soil fast-acting potassium content remains stable within a certain range through regular monitoring and adjustment of fertilizer application to the soil fast-acting potassium content and its changing trends, thereby preventing it from becoming a limiting factor in achieving the target yield. Based on the potassium fertilizer monitoring method, the recommended usage amount of potassium fertilizer for different regions is determined. Specifically, S4 includes steps S401 to S403.

[0070] S401, calculate the recommended potassium application rate based on the yield forecast and the recommended potassium fertilizer application rate formula. The formula for calculating the recommended potassium application rate is: The recommended potassium application rate is calculated as: potassium absorption per 100 kilograms of seeds × predicted yield.

[0071] Based on measured data on the rapid-acting potassium content in soil (S402) and potassium fertilizer grading standards, the current regional potassium application ratio is determined. Tables 4-6 show the potassium fertilizer grading standards for three common crops in the main cultivation areas.

[0072] [Table 4]

[0073] [Table 5]

[0074] [Table 6]

[0075] S403. The predicted potassium application rate is obtained by multiplying the recommended potassium application rate by the potassium application rate ratio. Taking a maize cultivation plot in the North China region as an example, if the soil's fast-acting potassium content in that plot is 124.2 mg / kg, it falls into the high level category, and the corresponding recommended rate is 70%. In this case, the predicted potassium application rate for that plot = recommended potassium application rate × 70%.

[0076] Based on S5, the predicted nitrogen application rate, predicted phosphorus application rate, predicted potassium application rate, and the base fertilizer / top dressing ratio, the base fertilizer and top dressing amounts of nitrogen, phosphorus, and potassium are calculated, respectively. The base fertilizer / top dressing ratio is taken from the base fertilizer / top dressing ratio knowledge base; refer to Table 7.

[0077] [Table 7]

[0078] S6. A base fertilizer mixing ratio model is constructed based on the amounts of nitrogen, phosphorus, and potassium in the base fertilizer, and the mixing ratio of the base fertilizer is obtained. Specifically, the expression for the base fertilizer mixing ratio model is ω = [A]% + [B]% + [C]%. However, [A] is the largest integer not exceeding A, [B] is the largest integer not exceeding B, and [C] is the largest integer not exceeding C.

number

[0079] S7 constructs a dynamic adjustment model based on the amount of nitrogen, phosphorus, and potassium top dressing, using deficiency element-dynamic adjustment factors during critical top dressing periods, and calculates top dressing correction values ​​for nitrogen, phosphorus, and potassium to generate top dressing mixing ratios. This step aims to achieve inkjet-like fine-tuning of top dressing amounts based on real-time crop growth by combining remote sensing dynamic monitoring and target yield guidance. It is mainly applied to the "dynamic correction loop" for top dressing recommendations for nitrogen, phosphorus, and potassium, improving the efficiency of nutrient matching for crops during critical growing periods. Specifically, S7 includes steps S701 to S703.

[0080] Select the S701 MODIS NDVI product (MOD13Q1, 250m resolution, 16-day composite) to acquire remote sensing data for the current area in the sampling cycle prior to the topdressing period, and extract its NDVI value. This step further includes preprocessing operations using ArcGIS 10.2 to resample and shear up to the plot boundaries, extract the plot-average NDVI, and obtain the current NDVI value for each plot. The specific processes and principles of the preprocessing operations are all prior art and therefore will not be explained. The topdressing periods for different crops are determined based on the topdressing growth period knowledge base; refer to Table 8.

[0081] [Table 8]

[0082] In the case of corn, the top dressing period is during the large-trumpet stage; in the case of wheat, it is during the internode elongation stage; and in the case of paddy rice, it is during the greening stage and the panicle differentiation stage.

[0083] Based on S702, the NDVI value, and a predetermined NDVI threshold range, the criteria for determining elemental deficiency and the dynamic adjustment factors are determined. The predetermined NDVI threshold range is determined based on experience and is common knowledge; the specific numerical values ​​in this example are shown in Table 9.

[0084] [Table 9]

[0085] Specifically, the criteria for determining elemental deficiency and the dynamic adjustment factors in S702 were determined as follows:

number

number

[0086] S703. The dynamic adjustment factor is multiplied by the amount of nitrogen, phosphorus, and potassium added as top dressing to obtain the nitrogen, phosphorus, and potassium top dressing correction value.

[0087] As an alternative, S7 may further employ the following method, specifically S7 including steps S701 to S703.

[0088] S701 retrieves leaf images of crops in the current region. In this step, images of crops are uploaded by the user.

[0089] S702, leaf images are input into the Qwen2.5-VL-7B multimodal model to determine element deficiency criteria and dynamic modifiers. Core parameter configuration of the Qwen2.5-VL-7B multimodal model: Base 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) Number of training sessions: 30, Core processing flow: Data quality filtering: Blurred images are removed using the Peak signal-to-noise ratio (PSNR) method.

[0090] Cross-source data alignment: All images are resampled to unify their resolution.

[0091] Data annotation system: Construct a multidimensional metadata architecture.

[0092] Element deficiency types: Nitrogen / Phosphorus / Potassium Crop types: Wheat / Corn / Rice Growth stages: seedling stage / internode elongation stage / heading stage / grain filling stage, Symptom site: new leaves / old leaves / leaf veins / leaf edges, Supplementary annotation items: Data source (network / field), environmental parameters (illuminance / soil moisture content).

[0093] In this step, the Qwen2.5-VL-7B multimodal model outputs descriptions of crops, elemental deficiencies, and symptoms. The elemental deficiency judgment criteria and dynamic modifiers are determined as follows:

number

[0094] S703. The dynamic adjustment factor is multiplied by the amount of nitrogen, phosphorus, and potassium added as top dressing to obtain the nitrogen, phosphorus, and potassium top dressing correction value.

[0095] As can be seen from S7, conventional top dressing is time-fixed and applied in fixed amounts, ignoring real-time diagnosis of growth differences and vigor, resulting in phenomena such as "insufficient top dressing" and "excessive top dressing." This step provides two types of dynamic top dressing adjustment methods, one of which employs remote sensing technology to achieve dynamic top dressing control at the plot level of ±15% based on MODIS NDVI and elemental deficiency thresholds of crops.

[0096] Another method employing leaf image recognition allows users to upload images of plants and combine them with the Qwen2.5-VL-7B multimodal model to perform accurate diagnosis of elemental deficiencies. The mixture ratio diagram is automatically updated after adjustment, realizing the true "time-varying" nature of the process.

[0097] The present invention further provides a system for generating the blending ratio of fertilizers specifically for agricultural crops. An acquisition module used to obtain crop growth data for any region, wherein the crop growth data includes actual climate data, actual soil data, actual management measures data, actual soil rapid-acting phosphorus content data, actual soil rapid-acting potassium content data, actual yield data, NO3 - An acquisition module including measured leaching data, measured N2O emissions data, and measured NH3 volatilization data, A nitrogen fertilizer recommendation model calculation module used to construct a machine learning-based nitrogen fertilizer recommendation model based on crop growth data, and to obtain predicted nitrogen application rates and yields, A phosphate fertilizer recommendation model calculation module is used to construct a phosphate fertilizer recommendation model based on phosphate fertilizer grading standards, and to obtain a phosphate application rate prediction value, based on yield prediction values ​​and actual soil rapid-acting phosphate content data. A potassium fertilizer recommendation model calculation module is used to construct a potassium fertilizer recommendation model based on potassium fertilizer grading standards, using yield prediction values ​​and actual soil rapid-acting potassium content data, and to obtain potassium application rate prediction values. A base fertilizer and top dressing calculation module used to calculate the base fertilizer and top dressing amounts of nitrogen, phosphorus, and potassium, respectively, based on predicted nitrogen application rates, predicted phosphorus application rates, predicted potassium application rates, and base fertilizer and top dressing ratios. A base fertilizer mixing ratio generation module is used to construct a base fertilizer mixing ratio model based on the amounts of nitrogen, phosphorus, and potassium in the base fertilizer, and to obtain the base fertilizer mixing ratio. This includes a top dressing ratio generation module used to construct a dynamic adjustment model based on the amount of nitrogen, phosphorus, and potassium applied as top dressing, based on deficiency element-dynamic adjustment factors during critical top dressing periods, and to calculate top dressing correction values ​​for nitrogen, phosphorus, and potassium, and to generate top dressing ratios.

[0098] Based on a similar inventive concept, the present invention further provides a device for generating the blending ratio of fertilizers specifically for agricultural crops, comprising a processor and memory, wherein the memory is used to store a computer program, and the processor is configured to execute the computer program to realize the steps of the above method.

[0099] The method for generating the blending ratio of crop-specific fertilizers, the system for generating the blending ratio of crop-specific fertilizers, and the equipment for generating the blending ratio of crop-specific fertilizers of the present invention can be applied to all regions of all countries, and is not limited to the above-mentioned countries and regions.

[0100] The above are merely preferred embodiments of the present invention, and it should be noted that those skilled in the art can make several improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered to fall within the scope of protection of the present invention.

Claims

1. Step S1 involves obtaining crop growth data for any region, wherein the crop growth data includes actual climate data, actual soil data, actual management measures data, actual soil rapid-acting phosphorus content data, actual soil rapid-acting potassium content data, actual yield data, NO ３ － Erosion measurement data, N ２ O emission measurement data and NH ３ Step S1, which includes volatile measurement data, Step S2 involves constructing a machine learning-based nitrogen fertilizer recommendation model based on the aforementioned crop growth data, and obtaining predicted nitrogen application rates and yields. Step S3 involves constructing a phosphate fertilizer recommendation model based on phosphate fertilizer grading standards and obtaining a phosphate application rate prediction value based on the yield prediction value and actual soil rapid-acting phosphate content data. Step S4 involves constructing a potassium fertilizer recommendation model based on potassium fertilizer grading standards, using the aforementioned yield prediction values ​​and measured soil rapid-acting potassium content data, and obtaining a predicted potassium application rate. Step S5 involves calculating the base fertilizer amounts and top dressing amounts for nitrogen, phosphorus, and potassium based on the predicted nitrogen application rate, predicted phosphorus application rate, predicted potassium application rate, and base fertilizer / top dressing ratio. Step S6 involves constructing a base fertilizer mixing ratio model based on the amounts of base fertilizers for nitrogen, phosphorus, and potassium, and obtaining the mixing ratio of the base fertilizers. A method for generating a blending ratio for a fertilizer specifically for agricultural crops, characterized by comprising step S7: constructing a dynamic adjustment model based on the amount of nitrogen, phosphorus, and potassium added as top dressing, based on the deficient elements during important top dressing periods and dynamic adjustment factors, and calculating the top dressing correction values ​​for nitrogen, phosphorus, and potassium to generate the blending ratio for top dressing.

2. The aforementioned step S2 is, Based on the aforementioned climate measurement data, soil measurement data, management measure measurement data, DNDC model and APSIM model, yield prediction values, NO ３ － Elution prediction value, N ２ O emission forecast values ​​and NH ３ Step S201 to obtain the predicted volatility value, The crop growth data, yield prediction value, NO ３ － leaching prediction value, N ２ O emission prediction value, and NH ３ Step S202 of constructing a nitrogen fertilizer recommendation model of a recurrent neural network based on an attention mechanism based on the volatilization prediction value, and A method for generating a blending ratio for crop-specific fertilizers according to claim 1, comprising step S203, which calculates a predicted nitrogen application rate and a predicted yield under optimal conditions based on the aforementioned climate measurement data, soil measurement data, management measure measurement data, and nitrogen fertilizer recommendation model.

3. Step S3 is, Step S301: Calculate the recommended phosphorus application rate based on the yield prediction value and the recommended phosphorus fertilizer application rate formula. Step S302 involves determining the current regional phosphate application rate ratio based on the measured soil rapid-acting phosphate content data and phosphate fertilizer grading standards. A method for generating a blending ratio for a crop-specific fertilizer according to claim 1, comprising step S303, which involves multiplying the recommended phosphate application rate by the phosphate application rate ratio to obtain a predicted phosphate application rate.

4. Step S4 is, Step S401: Calculate the recommended potassium application rate based on the yield prediction value and the recommended potassium fertilizer application rate formula. Step S402 involves determining the current potassium application rate ratio for the region based on the measured data of the soil's rapid-acting potassium content and the potassium fertilizer grading standards. A method for generating a blending ratio for a fertilizer specifically for agricultural crops according to claim 1, comprising step S403, which involves multiplying the recommended potassium application rate by the potassium application rate ratio to obtain a predicted potassium application rate.

5. The expression for the base fertilizer mixing ratio model is ω = [A]% + [B]% + [C]%, However, [A] is the largest integer not exceeding A, [B] is the largest integer not exceeding B, and [C] is the largest integer not exceeding C. [Math 12] And ω is the fertilization ratio, W Ｎ This is the predicted nitrogen application rate, W Ｐ This is the predicted value for the amount of phosphoric acid applied, W Ｋ A is a predicted potassium application rate, B is a nitrogen application rate ratio coefficient, B is a phosphorus application rate ratio coefficient, and C is a potassium application rate ratio coefficient, as described in claim 1, for generating a blending ratio for a fertilizer specifically for agricultural crops.

6. Step S7 is, Step S701 involves acquiring remote sensing data for the current region that falls within the sampling cycle before the top dressing period, and extracting the NDVI value from that data. Step S702, in which element deficiency judgment conditions and dynamic adjustment factors are determined based on the NDVI value and a predetermined NDVI threshold range, A method for generating a blending ratio for a crop-specific fertilizer according to claim 1, comprising step S703, which involves multiplying the aforementioned dynamic adjusting factor by the amount of nitrogen, phosphorus, and potassium added as top dressing to obtain a correction value for nitrogen, phosphorus, and potassium added as top dressing.

7. The element deficiency determination conditions and dynamic adjustment factors in step S702 are determined as follows: [Number 13] However, m is a dynamic regulator, and σ １ , σ ２ , σ ３ σ is the adjustment coefficient, １ ∈[-7%, -2%], σ ２ ∈ [2%, 7%], σ ３ ∈[σ ２ +5%, σ ２ +10%], where ε is the floating coefficient, 0 < ε < 1, and NDVI ｍａｘ The method for generating the blending ratio of a fertilizer specifically for agricultural crops according to claim 6, characterized in that is the maximum value within a predetermined threshold range.

8. Step S7 is, Step S701 involves obtaining leaf images of crops in the current region, Step S702 involves inputting the aforementioned leaf image into the Qwen2.5-VL-7B multimodal model to determine the element deficiency judgment conditions and dynamic adjustment factors. A method for generating a blending ratio for a crop-specific fertilizer according to claim 1, comprising step S703, which involves multiplying the aforementioned dynamic adjusting factor by the amount of nitrogen, phosphorus, and potassium added as top dressing to obtain a correction value for nitrogen, phosphorus, and potassium added as top dressing.

9. A system for generating the mixing ratio of fertilizers specifically for agricultural crops, An acquisition module used to acquire crop growth data for any region, wherein the crop growth data includes actual climate data, actual soil data, actual management measures data, actual soil rapid-acting phosphorus content data, actual soil rapid-acting potassium content data, actual yield data, NO ３ － Erosion measurement data, N ２ O emission measurement data and NH ３ The acquisition module includes volatile measurement data, A nitrogen fertilizer recommendation model calculation module is used to construct a machine learning-based nitrogen fertilizer recommendation model based on the aforementioned crop growth data, and to obtain predicted nitrogen application rates and yields. A phosphate fertilizer recommendation model calculation module is used to construct a phosphate fertilizer recommendation model based on phosphate fertilizer grading standards and to obtain a phosphate application rate prediction value, based on the aforementioned yield prediction value and actual soil rapid-acting phosphate content data. A potassium fertilizer recommendation model calculation module is used to construct a potassium fertilizer recommendation model based on potassium fertilizer grading standards, using the aforementioned yield prediction values ​​and measured soil rapid-acting potassium content data, and to obtain potassium application rate prediction values. A base fertilizer and top dressing calculation module used to calculate the base fertilizer and top dressing amounts of nitrogen, phosphorus, and potassium, respectively, based on predicted nitrogen application rates, predicted phosphorus application rates, predicted potassium application rates, and base fertilizer and top dressing ratios. A base fertilizer mixing ratio generation module is used to construct a base fertilizer mixing ratio model based on the amounts of base fertilizers nitrogen, phosphorus, and potassium, and to obtain the base fertilizer mixing ratio. A system for generating the blending ratio of fertilizers for crops, characterized by comprising: a module for generating the blending ratio of top dressings, which is used to generate the blending ratio of top dressings by constructing a dynamic adjustment model based on the amount of top dressings of nitrogen, phosphorus, and potassium, which are deficient elements during important top dressing periods, and by calculating the top dressing correction values ​​for nitrogen, phosphorus, and potassium.

10. A device for generating the blending ratio of fertilizers for agricultural crops, comprising a processor and memory, wherein the memory is used to store a computer program, and the processor is configured to execute the computer program to realize the steps of the method according to claim 1.

Images