Delta15N-based wolfberry soil nitrogen circulation system evaluation method

Abstract

The invention provides a delta15N-based wolfberry soil nitrogen circulation system evaluation method, and belongs to the technical field of intelligent agriculture and precise fertilization. The method aims at solving the problems that traditional Chinese wolfberry fertilization depends on experience and static soil testing, and accurate fertilization decision based on a nitrogen cycle mechanism is difficult to achieve. According to the method, different nitrogen source treatment tests are set, the content and delta 15N values of soil ammonium nitrogen and nitrate nitrogen, soil culture transformation data and delta 15N values of Chinese wolfberry leaves are measured synchronously, and fruit quality indexes are obtained; establishing a soil-plant delta15N response relationship based on the data and calculating a nitrogen conversion rate; integrating the parameters to construct a comprehensive evaluation model, wherein the model outputs a soil nitrogen plant effectiveness score, a nitrogen cycle strength score and a fruit quality prediction interval according to an input soil delta 15N value; and generating a fertilization decision scheme containing the specific nitrogen source type and the application amount based on the evaluation result. The method is mainly used for nitrogen nutrition diagnosis and precise fertilization management in wolfberry planting.

Description

Technical Field

[0001] This invention belongs to the field of smart agriculture and precision fertilization technology, specifically relating to a method based on δ 15 Evaluation method of soil nitrogen cycle system for wolfberry (N). Background Technology

[0002] In goji berry cultivation, scientific fertilization is a crucial step in ensuring yield, improving quality, and maintaining soil health. Traditional goji berry fertilization management mainly relies on growers' experience and routine testing of basic soil nutrient content. This model generally suffers from insufficient attention to the dynamic transformation process of soil nitrogen, especially lacking real-time and accurate quantitative assessment methods for the migration, transformation, and absorption of nitrogen in the soil-plant system.

[0003] To optimize fertilization, agricultural research and practice have attempted to introduce more refined soil testing and plant nutrition diagnostic techniques. However, conventional soil nitrogen speciation analysis only provides static content information at a specific point in time, failing to reveal the nitrogen conversion rate, source contribution, and dynamic processes of plant uptake. This leaves the assessment of soil nitrogen supply capacity at a superficial level, making it difficult to explain the differentiated effects of different fertilization measures mechanistically, and unable to predict long-term effects after fertilization. Furthermore, the nitrogen nutrition status of wolfberry plants has traditionally been indirectly assessed mainly through leaf nitrogen content or chlorophyll measurement, but these indicators are easily affected by environmental factors and cannot effectively distinguish whether nitrogen originates from current season fertilizers or the soil's existing reserves, limiting the timeliness and specificity for guiding precise topdressing. Summary of the Invention

[0004] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0005] One objective of this invention is to address the problem that traditional wolfberry fertilization relies on experience and static soil testing, which cannot systematically quantify the dynamic transformation process of soil nitrogen, the effectiveness of plant absorption, or correlate these process indicators with the final fruit quality, thus making it difficult to achieve precise fertilization decisions based on the nitrogen cycle mechanism.

[0006] One object of the present invention is to provide a δ-based 15 Evaluation methods for the soil nitrogen cycle system of wolfberry (N) include: Field trials with different nitrogen source treatments were conducted in the wolfberry planting area. Each treatment corresponded to a fertilization program that included a specific nitrogen source type and application rate. During the main growth stages of wolfberry, the ammonium nitrogen and nitrate nitrogen content and their δ-values ​​in the soil under each treatment were measured simultaneously. 15 N value, changes in inorganic nitrogen content in soil before and after incubation to reflect nitrogen transformation dynamics, and δ0.05 of functional leaves of wolfberry plants. 15The N value was obtained, and the quality indicators of wolfberry fruit in the corresponding experimental plots were acquired. The fruit quality indicators included soluble solids content, total sugar content, protein content, and amino acid content. Based on the measured soil inorganic nitrogen δ 15 N value and δ of wolfberry leaves 15 N value, established by regression analysis for soil inorganic nitrogen δ 15 N-mean and δ of wolfberry leaves 15 Quantitative functional relationship between N values; Based on the changes in inorganic nitrogen content before and after soil incubation, the apparent net mineralization rate and apparent net nitrification rate of soil nitrogen were calculated. The regression coefficients in the quantitative functional relationship, the values ​​of the apparent net mineralization rate and net nitrification rate, and the fruit quality indicators are used to construct a δ-value for soil ammonium nitrogen and nitrate nitrogen. 15 The N-value is the input, and the comprehensive evaluation model outputs a quantitative score of soil nitrogen plant availability, a quantitative score of nitrogen cycle transformation intensity, and a fruit quality prediction range. The fruit quality prediction range includes the prediction range of soluble solids content, total sugar content, protein content, and amino acid content. The target wolfberry field was evaluated based on the comprehensive evaluation model. The evaluation included: detecting the current δ¹⁸O values ​​of ammonium nitrogen and nitrate nitrogen in the soil of the field. 15 The N value is input into the model to obtain the soil nitrogen plant availability quantitative score, nitrogen cycle transformation intensity quantitative score, and fruit quality prediction range for the field. Based on the comparison results between the fruit quality prediction range and the preset quality target, as well as the values ​​of the soil nitrogen plant availability quantitative score and nitrogen cycle transformation intensity quantitative score, a fertilization decision scheme including nitrogen source type and application rate recommendations is generated. The preset quality target includes the baseline values ​​of soluble solids content, total sugar content, protein content, and amino acid content.

[0007] Preferably, in the method, during the same period of the field trial and the corresponding period of the evaluation of the target field, at least one type of environmental factor data is simultaneously monitored and acquired: Meteorological data, including average temperature and rainfall; Soil physical data, including soil volumetric water content; When constructing the comprehensive evaluation model, the acquired environmental factor data are compared with the δ values ​​of soil ammonium nitrogen and nitrate nitrogen. 15 The N values ​​are combined to form the input feature set of the model.

[0008] Preferably, in the method, the comprehensive evaluation model is constructed using a gradient boosting decision tree algorithm; The input features of the gradient boosting decision tree model include: the soil ammonium nitrogen δ15 N value, nitrate nitrogen δ 15 The N value, the regression coefficient in the quantitative function relationship, the numerical values ​​of the apparent net mineralization rate and net nitrification rate, and the key environmental factor data; The model is configured to simultaneously predict the output soil nitrogen plant availability quantification score, nitrogen cycle transformation intensity quantification score, and fruit quality prediction range based on the input features.

[0009] Preferably, in the method, the gradient boosting decision tree model is trained by minimizing a multi-task loss function; The multi-task loss function is a linear weighted sum of the prediction error terms corresponding to the soil nitrogen plant availability quantification score, the nitrogen cycle transformation intensity quantification score, and the fruit quality prediction interval, respectively. The weight coefficients in the linear weighted sum are determined by iteratively optimizing the weight coefficients using validation set performance feedback during model training.

[0010] Preferably, in the method, the step of iteratively optimizing the weight coefficients using validation set performance feedback is achieved by executing a process including the following steps: Assign a set of preset initial values ​​to the weight coefficients, and divide the field test dataset into a non-overlapping training subset and a validation subset; Multiple optimization loops are performed. In each loop, the gradient boosting decision tree model is trained on the training subset using the current weight coefficients. The trained model is used to make predictions on the validation subset and the prediction errors of the soil nitrogen plant availability quantification score, nitrogen cycle transformation intensity quantification score, and fruit quality prediction interval are calculated respectively. Then, a set of new weight coefficient candidate values ​​is calculated according to the prediction error and the preset weight update algorithm. Substitute the new candidate values ​​of the weight coefficients into the model and re-evaluate the overall prediction performance on the validation subset. If the overall performance is better than the performance using the current weight coefficients, then update the weight coefficients to the candidate values. Repeat the optimization loop and evaluation update process until the preset optimization termination condition is reached, and use the final obtained weight coefficient as the determined value in the linear weighted sum.

[0011] Preferably, in the method, the generation of a fertilization decision scheme containing nitrogen source type and application rate recommendations is accomplished by querying a preset decision mapping knowledge base; The decision mapping knowledge base uses the quantitative score of soil nitrogen plant availability, the quantitative score of nitrogen cycle transformation intensity, and the comparison results of the fruit quality prediction interval with the preset quality target as joint input conditions. The knowledge base pre-stores the mapping relationship between multiple sets of input condition combinations and corresponding output decisions, where each set of output decisions includes a specific recommended nitrogen source type and application amount.

[0012] Preferably, in the method, the generation of a fertilization decision scheme including nitrogen source type and application amount suggestions is achieved through a trained fertilization scheme recommendation model; The fertilization program recommendation model uses the quantitative score of soil nitrogen availability to plants, the quantitative score of nitrogen cycle transformation intensity, and the comparison results between the fruit quality prediction interval and the preset quality target as input features. The fertilization scheme recommendation model uses fertilization scheme data that have been verified as preferred in historical agronomic trials and correspond to the input features as training labels; The model outputs a recommended nitrogen source type and application rate that matches the system state represented by the input features.

[0013] Preferably, in the method, the fertilization scheme recommendation model is a multilayer perceptron neural network; The multilayer perceptron neural network includes an input layer, at least one hidden layer, and an output layer; wherein the output layer is configured to simultaneously output discrete classification results representing different nitrogen source types and continuous numerical values ​​representing recommended application amounts.

[0014] Preferably, in the method described, the preset weight update algorithm used in the iterative optimization of the weight coefficients using validation set performance feedback is as follows: Based on the latest prediction error of each task on the validation subset, dynamically adjust its weight coefficient in the loss function; Specifically, an error benchmark threshold is set; for tasks whose prediction error exceeds the error benchmark threshold, their loss weight is increased in the next iteration according to the excess ratio, and the increase is proportional to the excess ratio; for tasks whose prediction error is below the error benchmark threshold, their loss weight remains unchanged.

[0015] The present invention has at least the following beneficial effects: This invention provides a method based on δ 15 A systematic evaluation method for nitrogen (N) is used to establish a full-chain correlation model from soil nitrogen form dynamics and plant uptake response to fruit quality. This method can be based on readily available soil δ¹⁸O3 data. 15 The N signal enables quantitative evaluation of soil nitrogen availability to plants, nitrogen cycle transformation intensity, and prediction of fruit quality, thus providing a clear decision-making basis for generating scientific fertilization programs and helping to reduce the reliance on experience and blind spots in traditional fertilization.

[0016] This invention enhances the adaptability of the comprehensive evaluation model to different climate and soil moisture conditions by requiring the simultaneous inclusion of key environmental factor data in the evaluation process. This enables the model's predictive output to reflect the influence of environmental covariates, improving the stability and reliability of evaluation results and fertilization decisions under variable field conditions.

[0017] This invention effectively addresses soil delta-gravity by using a gradient boosting decision tree algorithm to construct a comprehensive evaluation model. 15 The algorithm incorporates multi-dimensional and nonlinear characteristics such as N-value, mechanistic parameters, and environmental factors. It possesses strong predictive capabilities and feature importance assessment functions, facilitating the simultaneous and accurate prediction of three objectives: soil nitrogen plant availability score, nitrogen cycle intensity score, and quality prediction interval.

[0018] This invention introduces an optimizable multi-task loss function to train the model, providing a mechanism for balancing the learning of different prediction tasks. This helps coordinate the optimization objectives of each task during model training, alleviates conflicts between tasks, and promotes more balanced and stable performance of the model across multiple evaluation metrics.

[0019] This invention systematizes the complex weight tuning process by establishing an automated process that includes initialization, iterative training, evaluation, and updating to determine multi-task loss weights. This method allows for feedback and adjustments based on the model's actual performance on the validation set, helping to more efficiently find the optimal combination of task weights that maximizes the model's overall performance.

[0020] This invention introduces a pre-defined decision mapping knowledge base, providing clear rules for converting the system state score output by the model into specific fertilization operation recommendations. This approach allows the abstract "evaluation result" to be directly mapped to an "agronomic prescription" containing specific nitrogen source types and application rates, achieving a seamless transition from analysis to execution.

[0021] This invention provides an alternative approach based on historical experience learning by proposing a data-driven fertilization recommendation model to generate decisions. This method can learn the mapping relationship between states and decisions from a large number of historical optimal solutions, making the generated fertilization recommendations more data-supported, adaptable, and flexible.

[0022] This invention, by specifically defining the fertilization program recommendation model as a multilayer perceptron neural network, clarifies a network architecture capable of realizing complex nonlinear mappings. This architecture can simultaneously handle classification (nitrogen source type) and regression (application rate) tasks, providing a stable and feasible computational framework for generating integrated fertilization decisions.

[0023] This invention quantifies and regularizes the adjustment process of multi-task loss weights by setting a weight update algorithm based on an error benchmark threshold. This rule adjusts the weights proportionally according to the deviation of each task's prediction error from the benchmark, providing a clear and repeatable automated adjustment logic for task balancing during model training.

[0024] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0025] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0026] This invention provides a method based on δ 15 The evaluation method for the nitrogen cycle system in wolfberry soil includes: setting up field trials in wolfberry planting areas with different nitrogen source treatments, each treatment corresponding to a fertilization program containing a specific nitrogen source type and application rate; and simultaneously measuring the ammonium nitrogen and nitrate nitrogen content and their δ-values ​​in the soil under each treatment during the main growth stages of wolfberry. 15 N value, changes in inorganic nitrogen content in soil before and after incubation to reflect nitrogen transformation dynamics, and δ0.05 of functional leaves of wolfberry plants. 15 The N value was used to obtain the quality indicators of wolfberry fruit in the corresponding experimental plots. These fruit quality indicators included soluble solids content, total sugar content, protein content, and amino acid content. Based on the measured soil inorganic nitrogen δ... 15 N value and δ of wolfberry leaves 15 N value, established by regression analysis for soil inorganic nitrogen δ 15 N-mean and δ of wolfberry leaves 15 A quantitative functional relationship between N values; based on the changes in inorganic nitrogen content before and after soil cultivation, the apparent net mineralization rate and apparent net nitrification rate of soil nitrogen are calculated; the regression coefficients in the quantitative functional relationship, the values ​​of the apparent net mineralization rate and net nitrification rate, and the fruit quality indicators are used to construct a δ-value relationship between soil ammonium nitrogen and nitrate nitrogen. 15 The model uses N as input and outputs a comprehensive evaluation of soil nitrogen plant availability, nitrogen cycle transformation intensity, and fruit quality prediction range. The fruit quality prediction range includes the prediction ranges for soluble solids content, total sugar content, protein content, and amino acid content. The model is used to evaluate a target wolfberry field, including detecting the δ¹⁸O values ​​of current ammonium nitrogen and nitrate nitrogen in the soil. 15The N value is input into the model to obtain the soil nitrogen plant availability quantitative score, nitrogen cycle transformation intensity quantitative score, and fruit quality prediction range for the field. Based on the comparison results between the fruit quality prediction range and the preset quality target, as well as the values ​​of the soil nitrogen plant availability quantitative score and nitrogen cycle transformation intensity quantitative score, a fertilization decision scheme including nitrogen source type and application rate recommendations is generated. The preset quality target includes the baseline values ​​of soluble solids content, total sugar content, protein content, and amino acid content.

[0027] The implementation of this invention mainly includes the following five stages: field trials and data collection, key relationship analysis, comprehensive evaluation model construction, model application and decision generation.

[0028] Phase 1: Field Trials and Systematic Data Collection. Representative goji berry planting areas were selected, and experimental plots were established containing at least three typical nitrogen source treatments, such as: urea-only treatment, well-rotted sheep manure-only treatment, and a combination of urea and sheep manure treatment. Each treatment had a clearly defined and fixed nitrogen source type and application rate per unit area. During key phenological stages of goji berries, such as bud break and leaf expansion and fruit enlargement, simultaneous sampling and measurement were conducted in each experimental plot. Soil samples were collected, and the contents of ammonium nitrogen and nitrate nitrogen, as well as their respective deltaic values, were determined. 15 N value. Simultaneously, a standard soil incubation experiment was conducted to obtain basic data for calculating the nitrogen conversion rate by measuring changes in soil inorganic nitrogen content before and after incubation. Functional leaves of corresponding wolfberry plants were collected, and their δ0.05 values ​​were measured. 15 N value. During the fruit ripening period, fruits from each plot were harvested, and key quality indicators such as soluble solids, total sugar, protein, and amino acids were measured. Protein content was determined using the Kjeldahl method, and amino acid content was determined using the ninhydrin colorimetric method or an automated amino acid analyzer. All data were recorded one-to-one according to the experimental plots.

[0029] Phase II: Quantification of soil-plant nitrogen response and transformation processes. Based on the data from Phase I, the δ¹⁸O₂ value of inorganic nitrogen in soil for each treatment was first calculated. 15 The weighted average of N. The weighted average is calculated as follows: the measured soil ammonium nitrogen content and nitrate nitrogen content are used as weights, and the corresponding δ values ​​are calculated. 15 The N value is used for weighted averaging. Specifically, the weighted average is equal to "ammonium nitrogen δ". 15 "N value multiplied by ammonium nitrogen content" and "nitrate nitrogen δ" 15 The sum of N value multiplied by nitrate nitrogen content, then divided by the sum of ammonium nitrogen and nitrate nitrogen content. This average value is used as the independent variable, corresponding to the δ value of wolfberry leaves. 15With N as the dependent variable, linear regression analysis was performed to establish a quantitative relationship function between the two, and the regression coefficients were recorded. Simultaneously, using soil incubation experimental data, the apparent net mineralization rate and apparent net nitrification rate of soil nitrogen in each experimental plot were calculated. These two rate parameters were used to quantify the intensity of key transformation processes in the soil nitrogen cycle.

[0030] Phase 3: Comprehensive Evaluation Model Construction. The regression coefficients, apparent net mineralization rate, and nitrification rate values ​​obtained in Phase 2, along with measured fruit quality indicators, are integrated. Multivariate statistical analysis or machine learning algorithms, such as gradient boosting decision trees, are used to train this data. The goal of the training is to construct a predictive model based on the ammonium nitrogen and nitrate nitrogen δ values ​​of new, unplanted soil samples. 15 Using the N value as input, the system can simultaneously output predicted values: a quantitative score characterizing the availability of soil nitrogen to wolfberry, a quantitative score characterizing the intensity of soil nitrogen cycle transformation, and prediction ranges for key quality indicators of wolfberry fruit. Specifically, the quantitative score for soil nitrogen availability to plants is a value between 0 and 100, with higher scores indicating easier absorption and utilization of soil nitrogen by wolfberry; the quantitative score for nitrogen cycle transformation intensity is also a value between 0 and 100, with higher scores indicating more active mineralization and nitrification processes of soil nitrogen. The prediction range can be provided separately for indicators such as soluble solids, total sugar, protein, and amino acids, or one or more key indicators can be selected for output based on preset targets.

[0031] Phase Four: Model Application and Fertilization Decision Generation. For target wolfberry fields requiring fertilization guidance, soil samples were collected during their critical growth stages using the methods employed in Phase One, and the current ammonium nitrogen and nitrate nitrogen δ-values ​​in the soil were measured. 15 N value. This pair of δ 15 The N-value is input into the comprehensive evaluation model constructed in the third stage, and the model can then output the three evaluation results for the field. The decision-making system compares the obtained fruit quality prediction range with the preset high-quality production target, and combines the specific values ​​of soil nitrogen availability score and cycle intensity score. Through preset decision rules or by querying the knowledge base, it finally generates a fertilization plan that includes recommended nitrogen source types and their suggested application rates. The preset high-quality production target is, for example, soluble solids content not less than 20%, total sugar content not less than 25 mg / g, protein content not less than 10%, and total amino acid content not less than 5 mg / g.

[0032] Phase 5: Solution Validation and Iteration. The fertilization effects implemented according to the above decision-making scheme can be used as new data points to supplement the historical database for future optimization and updates of the comprehensive evaluation model and decision-making rules.

[0033] The closest existing technology is the conventional soil testing and fertilizer recommendation method for wolfberry cultivation. This method is typically based on the chemical determination of the content of readily available nutrients such as ammonium nitrogen and nitrate nitrogen in the soil, combined with the target yield and a broad fertilizer utilization coefficient, to calculate the approximate total nitrogen fertilizer requirement. Its recommendations mainly focus on the total nutrient content, and the fertilizer formula is relatively fixed with a long update cycle.

[0034] The main difference between this invention and the prior art described above is that: Existing technologies primarily evaluate the "static stock" of nitrogen in soil; this invention further utilizes δ0.05... 15 N and culture experiments were conducted to evaluate the "dynamic transformation process" of nitrogen (such as mineralization and nitrification intensity) and "plant uptake efficiency", providing richer system state information from a mechanistic perspective.

[0035] The core input of existing technologies is nutrient content; the core input of this invention is the nitrogen stable isotope δ¹⁸. 15 N value. δ 15 The N value not only reflects the amount of nitrogen, but also contains information about the source and transformation pathway of nitrogen, which cannot be provided by conventional chemical determination.

[0036] Existing technologies rely on direct calculations of "content-total," lacking a correlation between process and final quality. This invention's decision-making is based on a correlation model of "process signal-system state-quality prediction." It first utilizes δ... 15 The N-signal diagnoses the system's internal state (effectiveness and conversion strength) and predicts the potential output quality. Then, decisions are generated based on the state diagnosis and quality objectives, resulting in a longer logical chain and a more systematic consideration of factors.

[0037] In existing technologies, formulas or coefficient tables are fixed; this invention introduces a predictive model trained on historical data, which can learn complex nonlinear relationships, giving its evaluation and prediction capabilities the potential to improve itself through data accumulation.

[0038] In a preferred embodiment, the method involves simultaneously monitoring and acquiring at least one type of environmental factor data during the same period of the field experiment and the corresponding period for evaluating the target field: meteorological data, including average temperature and rainfall; soil physical data, including soil volumetric water content; and when constructing the comprehensive evaluation model, the acquired environmental factor data is compared with the δ¹⁸O values ​​of soil ammonium nitrogen and nitrate nitrogen. 15 The N values ​​are combined to form the input feature set of the model.

[0039] This specific implementation method integrates environmental factors as covariates into the construction and application of the comprehensive evaluation model.

[0040] Phase 1: Synchronous Monitoring and Acquisition of Environmental Data. During field trials, a small-scale weather station and soil moisture monitoring points must be established within the experimental area. Throughout the entire experimental period, especially during periods coinciding with soil and plant sample collection, environmental data must be continuously recorded and acquired. Two types of data are collected: meteorological data, including the cumulative daily average temperature and cumulative rainfall during the sampling period; and soil physical data, primarily the direct measurement of soil volumetric water content in each experimental plot using sensors during sampling. All of this environmental data must be correlated with the corresponding soil δ¹⁸O values. 15 N measurement value, leaf δ 15 A strict temporal and spatial correspondence should be established between N-value and fruit quality data. Similarly, when evaluating the application of the model to the target field, it is also necessary to obtain data on temperature, rainfall, and soil volumetric moisture content in the area where the field is located within the corresponding evaluation period.

[0041] Phase Two: Integrating Environmental Data as Features into the Model. When constructing the comprehensive evaluation model, environmental variables need to be added to the data preparation steps. An extended feature vector is constructed for each training sample (i.e., each experimental plot). This vector includes not only soil δ... 15 In addition to indicators such as N-value, regression coefficient, and transformation rate, environmental factor data (such as average temperature, rainfall, and soil moisture content) acquired simultaneously from the community also need to be added as a separate component. Subsequently, this extended feature vector, integrating mechanistic indicators and environmental covariates, along with the corresponding system evaluation target values ​​(effectiveness score, intensity score, and quality interval), is used to train a machine learning model, such as a gradient boosting decision tree. During training, the model will simultaneously learn soil δ0... 15 The complex pattern in which the N signal and environmental conditions jointly influence the system state.

[0042] The third stage: Evaluation based on integrated environmental information during application. When applying a trained model to a specific target field, in addition to inputting the δ¹⁴ values ​​of soil ammonium nitrogen and nitrate nitrogen... 15 In addition to the N value, the average temperature, rainfall, and soil moisture content of the field during the current evaluation period, either measured or obtained from nearby weather stations, must be input simultaneously. The model will integrate these soil signals with environmental background information to output more accurate system state evaluation results and quality predictions for this specific environmental condition.

[0043] In some embodiments, the evaluation method may be based solely on soil-plant δ 15 This method uses N-relationships to construct a fertilization decision model, focusing on isotopic signal transduction within the soil and plants, but failing to systematically integrate external environmental conditions as independent input variables. This method assumes or expects δ... 15 The correlation between the N signal and the system state is sufficiently robust under different environmental conditions.

[0044] The main difference between the specific embodiments of the present invention described above and the method is as follows: The dataset for this method mainly revolves around isotopes and chemical indicators of soil and plants themselves; the implementation of this invention systematically incorporates key driving factors affecting nitrogen cycle rate and pathways—meteorological conditions such as temperature and rainfall, and soil physical conditions such as volumetric water content—into the necessary data collection and modeling system, thereby constructing a comprehensive data framework of "environment-driven-soil response-plant absorption" that is closer to the real field scenario.

[0045] The input variables of this method focus on intrinsic system indicators, implicitly assuming that these indicators are sufficient to explain the main variations in the system state. The implementation of this invention relaxes this assumption by adding environmental factors as explicit covariates to the model input, enabling the model to learn and quantify the same soil δ0. 15 The different system states corresponding to the N signal under different environmental backgrounds enhance the explanatory dimension of the model.

[0046] Because this method does not explicitly consider environmental covariates, the model's predicted output may be closer to an estimate under the average environmental conditions of the training data. When applied to new locations or new years with significantly different training environments, the prediction accuracy may be affected. The embodiments of this invention, by introducing environmental variables, endow the model with a certain degree of "contextual awareness," enabling its predictions to adjust to specific input environmental conditions. This is expected to improve the model's extrapolation applicability and prediction stability under different climate and soil moisture conditions.

[0047] In a preferred embodiment, the comprehensive evaluation model is constructed using a gradient boosting decision tree algorithm; the input features of the gradient boosting decision tree model include: the soil ammonium nitrogen δ 15 N value, nitrate nitrogen δ 15 The model is configured to simultaneously predict, based on the input features, a quantitative score of soil nitrogen plant availability, a quantitative score of nitrogen cycle transformation intensity, and a fruit quality prediction range, including the N value, the regression coefficient in the quantitative function relationship, the numerical values ​​of the apparent net mineralization rate and net nitrification rate, and the key environmental factor data; the model is configured to simultaneously predict the output soil nitrogen plant availability quantitative score, nitrogen cycle transformation intensity quantitative score, and fruit quality prediction range.

[0048] To achieve a comprehensive evaluation of the soil nitrogen cycle system, a gradient boosting decision tree was selected as the core machine learning algorithm. Before building the model, a structured feature vector was prepared for each training sample (i.e., each field test plot). This feature vector contained the following types of data: the signal characteristics of the soil itself, namely the measured ammonium nitrogen δ... 15 N value and nitrate nitrogen δ 15The N-value; mechanistic characteristics derived from the soil-plant relationship, i.e., regression coefficients in the quantitative functional relationship obtained through regression analysis; characteristics characterizing the soil's internal transformation processes, i.e., the calculated apparent net mineralization rate and apparent net nitrification rate values; and covariate characteristics obtained from the external environment, i.e., key environmental factor data such as synchronously monitored average temperature, rainfall, and soil volumetric water content. These different types and sources of characteristics are combined into a unified input.

[0049] The gradient boosting decision tree algorithm was used to train the prepared feature vectors and their corresponding labels. The goal of the training process was to enable the model to simultaneously learn to predict three related but different target variables: the first is a quantitative score of soil nitrogen availability to plants, a continuous value reflecting the ease with which wolfberry plants can absorb nitrogen from the soil; the second is a quantitative score of nitrogen cycle transformation intensity, a continuous value reflecting the activity of nitrogen mineralization and nitrification processes within the soil; and the third is a predicted range for fruit quality, a range of key quality indicators for wolfberry fruits expressed as intervals. The model iteratively constructed multiple decision trees, learning the mapping rules from complex mixed features to these three targets.

[0050] Once trained, the model becomes a fixed prediction tool. When evaluating new target plots, simply measure their soil δ0 according to the same specifications. 15 The N value is obtained, along with the corresponding environmental factor data. These, along with the regression coefficients and conversion rates calculated from the field data (these calculation steps must follow the same standard methods as the training phase), together constitute a feature vector input model. The model will output the above three prediction results for the field in one go, completing a multi-dimensional diagnosis of the system state.

[0051] Single statistical models (such as multiple linear regression) or simpler machine learning models can be used to predict relevant indicators of soil nitrogen status, either separately or sequentially. However, these methods may face limitations in flexibility or efficiency when dealing with complex nonlinear interactions of multi-source heterogeneous data (such as isotope values, rates, and environmental factors) and coordinating multiple related but not entirely consistent prediction tasks.

[0052] The main difference between the specific embodiments of the present invention described above and this type of method is that: Gradient boosting decision tree algorithms are generally superior to traditional linear models or single decision trees in handling mixed-type features, automatically ranking feature importance, and capturing nonlinear relationships. This allows the model constructed in this implementation to more effectively extract features from soil δ... 15 Useful information is extracted from complex features composed of N-value, mechanistic parameters, environmental data, etc.

[0053] This implementation explicitly employs a single model architecture to simultaneously output three key evaluation metrics, rather than constructing multiple independent models. This multi-task learning framework allows models to share relevant information across different tasks during training, potentially improving data utilization efficiency and enabling the model to learn more robust feature representations that are beneficial for multiple tasks. This may be more advantageous than training multiple independent models, especially when training data is limited.

[0054] This implementation method can obtain evaluation results for all dimensions at once through forward propagation of a single model, making it more efficient and consistent in application. In contrast, separate modeling methods require running multiple models sequentially during application, and the prediction logic between each model is relatively independent. The output results may not be as consistent internally as those generated simultaneously from a single multi-task model.

[0055] In a preferred embodiment, the gradient boosting decision tree model is trained by minimizing a multi-task loss function; the multi-task loss function is a linear weighted sum of the prediction error terms corresponding to the soil nitrogen plant availability quantification score, the nitrogen cycle transformation intensity quantification score, and the fruit quality prediction interval; the weight coefficients in the linear weighted sum are determined by iteratively optimizing the weight coefficients using validation set performance feedback during model training.

[0056] When training a multi-task gradient boosting decision tree model, a specific loss function is defined to guide model optimization. This loss function consists of three parts, corresponding to the three objectives the model needs to predict: the prediction error term for the soil nitrogen plant availability quantification score, the prediction error term for the nitrogen cycle transformation intensity quantification score, and the prediction error term for the fruit quality prediction interval. These three error terms are combined into a total loss value through linear weighting; that is, the total loss equals the sum of each error term multiplied by its corresponding weight coefficient. These three weight coefficients are key adjustable hyperparameters in the model training process, and their magnitudes directly determine the degree of importance the model places on different prediction tasks during optimization.

[0057] The aforementioned weight coefficients are not fixed in advance, but are determined through an automated process based on validation set performance feedback. The specific steps are as follows: First, the complete field trial dataset is divided into a subset for model parameter training and a subset for performance evaluation. During training, multiple optimization loops are performed. In each loop, the model is trained using the current set of weight coefficients, and then the model's prediction errors for the three tasks are evaluated on the validation subset. Based on the calculation results of these three errors, a new set of potentially better candidate weight coefficients is automatically calculated according to a set of preset algorithm rules. Subsequently, the model is retrained using this new set of candidate values ​​and evaluated on the same validation set. If the overall performance of the model under the new weights (e.g., a weighted average of the errors of the three tasks) is better than the performance under the old weights, then this new set of weight coefficients is officially adopted for subsequent training. This process is repeated until the weight coefficients tend to stabilize or the preset number of loops is reached, ultimately determining a set of weight coefficients that result in better overall model performance.

[0058] Using the final set of weights, the model is trained once on the complete training data to obtain a fully trained multi-task prediction model.

[0059] When training multi-task prediction models, fixed, empirical weights can be used to combine the loss functions of different tasks, or multiple single-task models can be optimized independently. The fixed-weight method relies on prior knowledge or repeated trials and is difficult to adapt to different datasets or model structures; independent modeling cannot take advantage of any beneficial correlations that may exist between tasks.

[0060] The main difference between the specific embodiments of the present invention described above and these methods is that: This implementation abandons the fixed weight approach and introduces a closed-loop, iterative weight optimization mechanism based on validation set performance feedback. This means that the balance of the loss function is dynamically guided by the model's actual performance on a specific dataset, rather than relying entirely on preset experience, making weight adjustments more targeted and data-adaptive.

[0061] This implementation provides a standardized operating procedure, breaking down the weight optimization process into clearly defined steps such as initialization, iterative evaluation, rule updating, and performance comparison. This is more systematic and repeatable than parameter tuning methods that rely on manual trial and error.

[0062] The direct goal of weight optimization is to improve the model's "overall predictive performance" on the validation set. This forces the optimization process to consider the improvement of all tasks simultaneously, seeking an overall balance point, rather than pursuing the optimality of a single task in isolation. This helps alleviate potential conflicts arising from multiple tasks competing for model capacity, prompting the model to learn feature representations that are more beneficial to the overall prediction objective.

[0063] In a preferred embodiment, the method involves iteratively optimizing the weight coefficients using validation set performance feedback, which is achieved by executing a process comprising the following steps: assigning a set of preset initial values ​​to the weight coefficients and dividing the field trial dataset into non-overlapping training and validation subsets; performing multiple optimization loops, in each loop training the gradient boosting decision tree model on the training subset using the current weight coefficients, using the trained model to make predictions on the validation subset and calculating the prediction errors for the soil nitrogen plant availability quantification score, nitrogen cycle transformation intensity quantification score, and fruit quality prediction interval, respectively; then calculating a set of new candidate weight coefficient values ​​according to a preset weight update algorithm based on the prediction errors; substituting the new candidate weight coefficient values ​​into the model and re-evaluating the comprehensive prediction performance on the validation subset; if the comprehensive performance is better than the performance using the current weight coefficients, updating the weight coefficients to the candidate values; repeating the optimization loop and evaluation update process until a preset optimization termination condition is reached, and using the finally obtained weight coefficients as the determined values ​​in the linear weighted sum.

[0064] This specific implementation method achieves automated optimization of the weight coefficients of the loss function.

[0065] The optimization process begins with a set of manually set initial weight coefficients. The integrated complete dataset from the field trials is randomly divided into two non-overlapping subsets, ensuring equal representativeness across treatments: a training subset and a validation subset. The training subset is primarily used for learning and updating model parameters, while the validation subset is specifically used to evaluate the generalization performance of the model under different weight configurations and does not participate in the direct training of model parameters.

[0066] At the start of each loop, the loss function is configured using the current set of weight coefficients, and a full round of training is performed on the gradient boosting decision tree model on the training subset. After training, the model is used to predict all samples in the validation subset, and the prediction error values ​​of the model on the three target tasks are calculated respectively. Then, a pre-defined weight update algorithm is invoked. This algorithm takes the three prediction errors calculated above as input, performs calculations according to its internal logic (e.g., giving more attention to tasks with larger errors), and outputs a set of updated candidate weight coefficient values.

[0067] After obtaining new candidate weight values, their effectiveness needs to be verified. The loss function is reconfigured using this set of candidate weights, and a new model is trained from scratch on the training subset. Similarly, this new model is evaluated on the same validation subset, and a comprehensive performance metric (e.g., the weighted average or geometric mean of the errors from the three tasks) is calculated. This comprehensive performance metric is compared with the performance metric obtained using the weights from the previous round. If performance improves, the candidate weights are adopted as the "current weights" for the next optimization loop; if performance does not improve, the original weights are retained. This process is repeated until a preset termination condition is met. The termination condition can be set as reaching the maximum number of iterations, or the change in weight coefficients and comprehensive performance being less than a certain threshold over several consecutive iterations. Finally, the optimal weight coefficients determined after multiple iterations are used as the final values ​​in the linear weighted sum used to generate the final model.

[0068] In some embodiments, parameters can be manually tuned, that is, relying entirely on human experience to select weights by repeatedly trying different combinations of weights and observing the results; or a simple grid search can be used, that is, enumerating and trying out several predefined fixed weight combinations to select the set that performs best on the validation set.

[0069] The main difference between the specific embodiments of the present invention described above and these methods is that: Compared to relying entirely on manual trial and error, this implementation defines a cyclical process that includes automatic evaluation, rule calculation, and decision feedback, minimizing human intervention (only initial values ​​and algorithm rules need to be set), and achieving a higher degree of automated optimization. Unlike simple grid search, it does not blindly enumerate all possible combinations, but rather "guides" the search direction through an algorithm that dynamically adjusts weights based on performance feedback, thus potentially making the exploration of high-dimensional weight spaces more efficient.

[0070] This implementation uses performance evaluation (on the validation set) directly as the basis for the next weight update, forming a closed-loop feedback system. The weight update algorithm utilizes the prediction error as a direct feedback signal, making the direction of weight adjustment traceable. Manual parameter tuning and grid search lack this integrated, model-performance-based specific feedback to guide each step of the adjustment.

[0071] This implementation breaks down a complex tuning task into a series of defined and repeatable steps, including initialization, iteration, evaluation, updating, comparison, and termination. This method is more systematic than unstructured trial and error, reduces the arbitrariness of the tuning process, and improves the reproducibility of the results.

[0072] In a preferred embodiment, the method involves generating a fertilization decision scheme that includes nitrogen source type and application rate recommendations by querying a pre-defined decision mapping knowledge base. The decision mapping knowledge base uses the soil nitrogen plant availability quantification score, nitrogen cycle transformation intensity quantification score, and the comparison results of the fruit quality prediction interval with the pre-defined quality target as joint input conditions. The knowledge base pre-stores the mapping relationship between multiple sets of input condition combinations and corresponding output decisions, wherein each set of output decisions includes specific recommended nitrogen source type and application rate values.

[0073] First, a structured decision rule base needs to be constructed. The construction of this knowledge base relies on historical agronomic experimental data, domain expert knowledge, or a combination of both. Specifically, the establishment of the knowledge base involves defining a series of "condition-action" mapping rules. Each rule contains two core parts: first, an "input condition combination," which consists of three elements: the numerical range of the soil nitrogen plant availability quantification score (e.g., low, medium, high), the numerical range of the nitrogen cycle transformation intensity quantification score, and the comparison result between the fruit quality prediction interval and the preset quality target (e.g., meeting the target, not meeting the target and with a gap). Second, a unique "output decision" corresponding to this condition combination, namely a specific fertilization operation recommendation, clearly indicating the recommended nitrogen source type (e.g., urea, organic fertilizer, or a specific ratio of compound fertilizer) and the recommended numerical range of application rate per unit area.

[0074] These defined mapping rules are encoded and organized to form a structured knowledge base, such as a decision table or rule engine, that can be queried by computer programs. During the model application phase, after the comprehensive evaluation model outputs three results for the target field, the system formats these three results into a joint input condition that conforms to the knowledge base query format. The system then searches the knowledge base for a rule that exactly matches or is closest to matching this input condition. Once a matching rule is found, the system extracts and outputs the corresponding fertilization decision plan, which includes the specific nitrogen source type and application rate.

[0075] The final decision-making solution will be presented to the user in a clear and unambiguous text or structured data format, such as "Recommended application of urea, with a suggested dosage of 15 to 18 kg per acre", thus completing the transformation from system status diagnosis to agronomic measure recommendations.

[0076] As an example, the decision mapping knowledge base includes the following simplified rule: When the soil nitrogen plant availability score and nitrogen cycle transformation intensity score output by the system evaluation are both below 40, it indicates that the soil's nitrogen supply capacity and nitrogen transformation activity are both at a low level. In this case, if the fruit quality prediction does not meet the target, the knowledge base recommends using a nitrogen source type of "urea and well-rotted organic fertilizer in combination" and suggests a pure nitrogen application rate of 8 to 10 kg per acre. Conversely, if the soil nitrogen plant availability score is above 70, it indicates that the soil currently has sufficient nitrogen supply, and regardless of the transformation intensity score, the knowledge base generally recommends not applying or only symbolically supplementing with a small amount of nitrogen fertilizer.

[0077] In some embodiments, decisions can be made based on a single indicator with a simple threshold, such as determining whether to fertilize and the approximate amount based solely on whether the soil nitrogen content test value is below a certain threshold.

[0078] The main difference between the specific embodiments of the present invention described above and these methods is that: The decision-making basis of this implementation method is a multi-dimensional, quantified comprehensive diagnosis of the system's state, encompassing soil nitrogen supply availability, nitrogen cycle intensity, and yield quality predictions. This approach is more systematic and comprehensive than decisions relying solely on a single nutrient content indicator. Compared to expert experience, it structures and explicitly codes experience, reducing the immediate dependence of decision-making on individual on-site subjective judgments.

[0079] This implementation method constructs a "decision mapping knowledge base," transforming complex decision-making logic that might otherwise be implicit in the minds of experts into a clear, storable, searchable, and reproducible set of rules. This makes the decision-making process transparent and traceable, whereas simple rules based on a single threshold or pure expert experience often lack such clear rule descriptions and consistency guarantees.

[0080] This implementation emphasizes precisely matching the quantitative score output by the model with a preset numerical range, thereby corresponding to specific agronomic operations. This is more accurate and actionable than rough suggestions based on a single indicator or broad expert opinions. It attempts to establish a standardized and automated mapping channel between the diagnosis of complex systems and specific action plans.

[0081] In a preferred embodiment, the method involves generating a fertilization decision scheme that includes nitrogen source type and application rate recommendations, achieved through a trained fertilization scheme recommendation model. The fertilization scheme recommendation model uses the quantitative score of soil nitrogen plant availability, the quantitative score of nitrogen cycle transformation intensity, and the comparison results between the fruit quality prediction interval and the preset quality target as input features. The fertilization scheme recommendation model uses fertilization scheme data that have been validated as preferred in historical agronomic trials and corresponds to the input features as training labels. The output of the model is a recommended nitrogen source type and application rate value that matches the system state represented by the input features.

[0082] First, a dataset is needed to train the fertilization recommendation model. Each sample in this dataset consists of two parts: the first part is the "input features," which are the system evaluation results obtained and output for a specific historical field or experimental plot. Specifically, this includes a quantitative score of soil nitrogen plant availability, a quantitative score of nitrogen cycle transformation intensity, and a comparison of the predicted fruit quality range with the preset target (which can be encoded as numerical differences). The second part is the "training labels," which are the fertilization schemes actually used in the corresponding planting season for that historical field or experimental plot, and subsequently verified by agronomy as performing optimally or effectively under that environmental and system condition. These labels must include the specific nitrogen source type and application rate. Then, a suitable supervised learning model architecture, such as a decision tree ensemble model or a neural network, is selected, and the model is trained using the input features X and the optimal fertilization scheme Y. The learning objective of the model is to fit a complex mapping function from multi-dimensional system state diagnosis to optimal agronomic decisions.

[0083] When generating fertilization plans for new target fields, the first step is to obtain the system evaluation results according to standard procedures and use them to construct the input feature vector. This vector is then input into a pre-trained fertilization plan recommendation model. The model processes and calculates the results using its internally learned mapping rules, directly outputting a structured decision. This result typically includes two parts: the recommended nitrogen source type (which may be presented as a category label or probability distribution) and the suggested application rate (a specific continuous value). This output is the generated fertilization decision plan.

[0084] In some embodiments, field performance data implemented according to the model's recommended scheme can be continuously added to the training dataset as new, labeled samples for regular retraining and updating of the model, so that its decision-making ability can adapt to new varieties, soils, or management practices.

[0085] In this implementation, the decision rules are derived from learning from historical successful case data. They are typically implicit patterns automatically inferred by the model from the data, and their generation process is data-driven and dynamically adjustable. Updating a pre-defined rule knowledge base requires manual modification of the rules, which may struggle to handle newly emerging, undefined state combinations. In contrast, a data-driven recommendation model, when faced with new input feature combinations, can make inferences based on its learned generalization ability, potentially exhibiting stronger adaptability. With the accumulation of new data, the model can self-update through retraining, adjusting its decision boundaries. For the complex, non-linear, high-dimensional mapping relationship between system states and optimal decisions, a knowledge base based on limited manual rules may be insufficient to fully and accurately characterize it. Machine learning models, especially those with a certain expressive power, are theoretically better suited to learning and approximating such complex functional relationships, potentially capturing subtle correlations that are difficult for humans to summarize.

[0086] In a preferred embodiment, the method uses a multilayer perceptron neural network as the fertilization recommendation model. The multilayer perceptron neural network includes an input layer, at least one hidden layer, and an output layer. The output layer is configured to simultaneously output discrete classification results representing different nitrogen source types and continuous values ​​representing recommended application amounts.

[0087] A multilayer perceptron was chosen as the core architecture of the fertilization recommendation model. This network structure comprises at least three basic layers: an input layer with the number of neurons strictly matching the dimension of the input feature vector, used to receive standardized input data consisting of soil nitrogen availability scores, nitrogen cycle intensity scores, and quality variance values; one or more hidden layers, each containing a certain number of neurons and configured with a non-linear activation function, used to extract and combine complex patterns from the input features layer by layer; and an output layer, whose structure requires special design to simultaneously meet two types of output requirements. A portion of the neurons in the output layer output discrete results representing different nitrogen source type selections through a specific activation function; the other portion outputs a continuous value representing the recommended fertilization amount. The connection weights and biases of the entire network are parameters that need to be determined through training.

[0088] The network was trained using the prepared training dataset. The training process employed the standard backpropagation algorithm and a loss function suitable for multi-task outputs. For the discrete output portion of nitrogen source type selection, a classification loss function was typically used; for the continuous output portion of application amount, a regression loss function was typically used. The optimizer iteratively tuned the network parameters to minimize the overall difference between the model's predicted recommendations and the historically best recommendations labeled in the dataset.

[0089] After training, the network structure and parameters are saved. In application, the system state score of the target field is processed according to the same standards and input into the network. The network performs forward propagation calculations, simultaneously generating recommended nitrogen source types and specific application rates at the output layer, thus completing the decision generation process.

[0090] In some embodiments, a single type of model (such as regression or classification only) can be used to predict nitrogen source and usage separately; or a simpler linear model or a single decision tree can be used as the recommendation model.

[0091] The main difference between the specific embodiments of the present invention described above and these methods is that: Compared to using two separate models to handle classification and regression tasks, this implementation employs a single multilayer perceptron network to simultaneously output results for both classes. This integrated architecture allows the models to share feature representation learning in the hidden layers, potentially enabling information complementarity between the nitrogen source type determination and dosage estimation tasks, thereby improving learning efficiency and prediction consistency.

[0092] Multilayer perceptrons, through their hidden layers and nonlinear activation functions, possess the ability to learn complex nonlinear mappings between inputs and outputs. Compared to simple linear models or shallow decision trees, they are theoretically better able to fit the complex, high-dimensional nonlinear relationships that may exist between the multidimensional scores of system states and the optimal fertilization decision.

[0093] This implementation specifies a special design for the output layer, enabling it to generate heterogeneous outputs within a single forward propagation process. This is more efficient and consistent than calling two separate models. This design specifically applies the concept of multi-task learning to the fertilization decision generation scenario.

[0094] In a preferred embodiment, the preset weight update algorithm used in the process of iteratively optimizing the weight coefficients using validation set performance feedback specifically involves: dynamically adjusting the weight coefficients of each task in the loss function based on the latest prediction error of each task on the validation subset; wherein an error benchmark threshold is set; for tasks whose prediction errors exceed the error benchmark threshold, their loss weights are increased in the next iteration according to the excess ratio, with the increase being proportional to the excess ratio; for tasks whose prediction errors are below the error benchmark threshold, their loss weights remain unchanged.

[0095] During iterative optimization, after obtaining the latest prediction error of each task on the validation subset in each iteration, the following rules are applied for weight calculation and updating: First, a baseline error threshold is set for comparison. This threshold can be the median or average of the prediction errors of all tasks in this round, or it can be a preset fixed value. Second, for each task, its prediction error in this round is compared with this threshold. If the prediction error of the task exceeds the threshold, its corresponding loss function weight coefficient will be increased in the next iteration. The increase is proportional to the proportion by which the prediction error of the task exceeds the threshold; that is, the greater the error exceeds the threshold, the greater the increase in weight. This proportional relationship can be determined by a preset scaling factor. For tasks with prediction errors below or equal to the threshold, their corresponding loss weights are usually kept unchanged in the next iteration to focus optimization attention on tasks with relatively poor performance.

[0096] This rule is invoked as a specific implementation of the preset weight update algorithm. Based on the actual calculated prediction error on the validation set after each iteration, it dynamically and selectively adjusts the loss function weight vector used in the next training round, thereby enabling the model to focus more on improving the performance of tasks with larger current prediction errors in subsequent training.

[0097] In some embodiments, an equal weighting method can be used, that is, all tasks are assigned the same fixed weight in multi-task training, or a fixed non-equal weight can be preset according to the difficulty or importance of the task.

[0098] The main difference between the specific embodiments of the present invention described above and these methods is that: Unlike all fixed-weight methods, the weight adjustment in this implementation is dynamic, based on the latest verification error, after each iteration, forming a closed-loop feedback. Compared to simple proportional scaling, it introduces a clear "error baseline threshold" as the trigger condition and baseline for adjustment, making the adjustment selective rather than indiscriminate.

[0099] This rule clearly distinguishes between tasks that "perform well" (error below the threshold) and those that "need improvement" (error above the threshold), and adopts different strategies for each: maintaining the weight of the former while proportionally increasing the weight of the latter. This strategy aims to systematically direct the model's learning resources to the task that most needs improvement, helping to more effectively balance the learning progress of different tasks and alleviate the problem of some tasks being "overwhelmed" by others.

[0100] By linking the weight increment to the proportion of error exceeding a threshold and using a preset scaling factor, a clear and quantifiable calculation method is provided for the magnitude of weight adjustment. This makes the weight update process more predictable and controllable, avoiding training instability that may result from excessively large or small adjustment magnitudes. Example

[0101] I. Experimental Treatment Four different nitrogen source treatments were set up, with each treatment replicated three times, for a total of 12 experimental plots. The treatments included: urea treatment (150 kg / ha of pure nitrogen), well-rotted sheep manure treatment (150 kg / ha of pure nitrogen), a combination of urea and sheep manure treatment (75 kg / ha of pure nitrogen each), and a control treatment without nitrogen application.

[0102] Sampling period: Simultaneous sampling was conducted during the three main growth stages of wolfberry: the budding and leaf unfolding stage, the new shoot growth stage, and the fruit enlargement stage.

[0103] II. Sample Collection and Measurement (1) Soil sample determination Soil samples were collected during the fruit enlargement period, and the results are as follows: The urea treatment reduced the soil ammonium nitrogen content to 12.5 mg / kg, δ 15 The nitrogen value was +3.2‰; the nitrate nitrogen content was 18.6 mg / kg, and the δ... 15 The N value is +5.8‰. Based on content-weighted calculations, the soil inorganic nitrogen δ... 15 The weighted average of N is 4.7‰.

[0104] Sheep manure treatment reduced soil ammonium nitrogen content to 15.3 mg / kg, δ 15 The nitrogen value was +6.8‰; the nitrate nitrogen content was 22.1 mg / kg, and the δ... 15 The nitrogen value is +8.5‰. Soil inorganic nitrogen δ 15 The weighted average of N is 7.9‰.

[0105] The soil ammonium nitrogen content in the combined application treatment was 14.2 mg / kg, δ 15 The nitrogen value was +5.1‰; the nitrate nitrogen content was 20.4 mg / kg, and the δ... 15 The nitrogen value is +7.2‰. Soil inorganic nitrogen δ 15 The weighted average of N is 6.4‰.

[0106] The control treatment had an ammonium nitrogen content of 8.7 mg / kg, δ 15 The nitrogen value was +4.5‰; the nitrate nitrogen content was 12.3 mg / kg, and the δ... 15 The nitrogen value is +6.1‰. Soil inorganic nitrogen δ 15 The weighted average of N is 5.5‰.

[0107] (2) Soil culture experiment Fresh soil samples were collected and conditioned to 60% field capacity, then incubated at a constant temperature of 25 degrees Celsius for 14 days. The results showed: The net mineralization rate (NMR) of urea treatment was 0.82 kg N / ha / day, and the net nitrification rate (NMR) was 0.51 kg N / ha / day. The NMR of sheep manure treatment was 1.13 kg N / ha / day, and the NMR was 0.68 kg N / ha / day. The combined application treatment had a NMR of 1.03 kg N / ha / day and a NMR of 0.62 kg N / ha / day. The control treatment had a NMR of 0.44 kg N / ha / day and a NMR of 0.28 kg N / ha / day.

[0108] (3) Leaf sample determination Wolfberry leaves during fruit enlargement period δ 15 N values: +5.2‰ for urea treatment, +8.9‰ for sheep manure treatment, +7.5‰ for combined treatment, and +5.8‰ for control treatment.

[0109] (4) Determination of fruit quality indicators Fruits from each plot were harvested at maturity, and quality indicators were measured.

[0110] The soluble solids content of the urea-treated fruit was 21.3%, the total sugar content was 26.8 mg / g, the protein content was 11.2%, and the total amino acid content was 5.8 mg / g.

[0111] The soluble solids content of the fruit treated with sheep manure was 22.5%, the total sugar content was 28.2 mg / g, the protein content was 12.5%, and the total amino acid content was 6.4 mg / g.

[0112] The soluble solids content of the fruit treated with the combined application was 23.1%, the total sugar content was 29.5 mg / g, the protein content was 13.1%, and the total amino acid content was 6.9 mg / g.

[0113] The control group had a soluble solids content of 18.2%, a total sugar content of 22.4 mg / g, a protein content of 9.3%, and a total amino acid content of 4.5 mg / g.

[0114] III. Soil-Plant δ 15 N-response relationship establishment Soil δ based on fruit enlargement period measurements 15 N-weighted mean and leaf δ 15 N value, in soil δ 15 N-weighted mean is the independent variable, and δ of the blade is the independent variable. 15 N is the dependent variable, and linear regression analysis is performed.

[0115] The regression analysis results are: leaf δ15 The N value equals 0.92 multiplied by the soil δ. 15 Adding 0.85 to the weighted mean of N, the coefficient of determination R² is 0.94, reaching a highly significant level. The regression coefficients a and b are 0.92 and 0.85 respectively.

[0116] IV. Construction of Comprehensive Evaluation Model (1) Scoring Calculation Method Soil nitrogen plant availability is scored on a scale of 0 to 100, based on leaf δ. 15 Calculation of the degree of agreement between measured and predicted N values. Taking urea treatment as an example, leaf δ 15 The predicted value of N is 5.17‰, the actual value is 5.2‰, the difference is 0.03‰, and the calculated score is 99.4 points.

[0117] The nitrogen cycle conversion intensity score also uses a scale of 0 to 100, and is calculated based on the weighted average of net mineralization rate and net nitrification rate relative to the maximum value of all samples, with a weight of 0.6 for mineralization rate and 0.4 for nitrification rate. Taking urea treatment as an example, the calculated score is 73.6.

[0118] (2) Model training A gradient boosting decision tree algorithm was used to construct a comprehensive evaluation model. All 108 data sets were divided into training and validation sets in a 7:3 ratio. The model input features included soil ammonium nitrogen δ0.05. 15 N value, nitrate nitrogen δ 15 The model outputs N value, regression coefficients a and b, net mineralization rate, net nitrification rate, and environmental factor data. The model outputs include soil nitrogen plant availability score, nitrogen cycle transformation intensity score, and fruit quality prediction interval.

[0119] The multi-task loss function uses a linear weighted sum of three prediction error terms. After optimization through cross-validation, the weight coefficients are determined as follows: effectiveness score weight 0.4, conversion intensity score weight 0.3, and quality interval weight 0.3.

[0120] V. Model Validation On the validation set, the coefficient of determination R² for predicting soil nitrogen plant availability score was 0.86, with a mean absolute error of 8.2 points; the coefficient of determination R² for predicting nitrogen cycle transformation intensity score was 0.79, with a mean absolute error of 10.1 points.

[0121] The fruit quality prediction intervals used were 95% confidence intervals. The coverage rates for the soluble solids prediction intervals were 93.5%, total sugar prediction intervals were 91.2%, protein prediction intervals were 92.8%, and amino acid prediction intervals were 90.5%.

[0122] VI. Model Application and Fertilization Decisions Application validation was conducted using adjacent fields that were not included in the modeling. Soil samples were collected during the fruit enlargement stage to determine the ammonium nitrogen δ. 15 The nitrogen value was +4.2‰, and the content was 13.5 mg / kg; nitrate nitrogen δ 15 The nitrogen value was +6.8‰, and the content was 19.8 mg / kg. The soil δ0.05 was calculated using a weighted average. 15 The mean value of N is 5.75‰.

[0123] The net mineralization rate of the field was measured to be 0.95 kg nitrogen / ha / day, and the net nitrification rate was 0.58 kg nitrogen / ha / day. The average temperature over the seven days prior to sampling was 22.5 degrees Celsius, the cumulative rainfall was 35 mm, and the soil volumetric water content was 18.5%.

[0124] Substituting the above input features into the trained comprehensive evaluation model, the output results are as follows: soil nitrogen plant availability score is 82 points, and nitrogen cycle transformation intensity score is 68 points. The predicted range for fruit quality is: soluble solids 20.8% to 23.2%, total sugar 26.5 to 29.1 mg / g, protein 11.2% to 13.4%, and amino acids 5.6 to 6.8 mg / g.

[0125] The preset quality targets are: soluble solids not less than 20%, total sugar not less than 25 mg / g, protein not less than 10%, and amino acids not less than 5 mg / g. The comparison results show that the lower limit of the predicted range for all quality indicators is higher than the target values, indicating that no additional fertilization is necessary under the current soil nitrogen conditions.

[0126] According to the decision mapping knowledge base rules, when the soil nitrogen plant availability score is above 70, the soil nitrogen supply capacity is sufficient, and fertilization is unnecessary. Therefore, the generated fertilization decision plan is: do not apply nitrogen fertilizer for the time being, and it is recommended to consider supplementing with a small amount of foliar fertilizer in the later stage of fruit enlargement, depending on the actual situation.

[0127] Fruit quality was measured after actual harvest: soluble solids 22.1%, total sugar 27.8 mg / g, protein 12.3%, and amino acids 6.2 mg / g, all within the predicted range, verifying the effectiveness of the model.

[0128] VII. Benefit Analysis Compared to traditional soil testing and fertilizer recommendation, precision fertilization can improve nitrogen fertilizer utilization, increase the rate of high-quality fruit by 13%, and increase yield by 5.6%. This method is suitable for large-scale goji berry planting bases, especially green and organic goji berry production bases with high requirements for fruit quality. For small-scale farmers, a simplified regional model can be used to reduce implementation costs by sharing testing resources.

[0129] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, the present invention is not limited to the specific details without departing from the general concept defined by the claims and their equivalents.

Claims

1. A method based on δ 15 A method for evaluating the nitrogen cycle system of Lycium barbarum soil, characterized by N, is described. include: Field trials with different nitrogen source treatments were conducted in the wolfberry planting area. Each treatment corresponded to a fertilization program that included a specific nitrogen source type and application rate. During the main growth stages of wolfberry, the ammonium nitrogen and nitrate nitrogen content and their δ-values ​​in the soil under each treatment were measured simultaneously. 15 N value, changes in inorganic nitrogen content in soil before and after incubation to reflect nitrogen transformation dynamics, and δ0.05 of functional leaves of wolfberry plants. 15 The N value was obtained, and the quality indicators of wolfberry fruit in the corresponding experimental plots were acquired. The fruit quality indicators included soluble solids content, total sugar content, protein content, and amino acid content. Based on the measured soil inorganic nitrogen δ 15 N value and δ of wolfberry leaves 15 N value, established by regression analysis for soil inorganic nitrogen δ 15 N-mean and δ of wolfberry leaves 15 Quantitative functional relationship between N values; Based on the changes in inorganic nitrogen content before and after soil incubation, the apparent net mineralization rate and apparent net nitrification rate of soil nitrogen were calculated. The regression coefficients in the quantitative functional relationship, the values ​​of the apparent net mineralization rate and net nitrification rate, and the fruit quality indicators are used to construct a δ-value for soil ammonium nitrogen and nitrate nitrogen. 15 The N-value is the input, and the comprehensive evaluation model outputs a quantitative score of soil nitrogen plant availability, a quantitative score of nitrogen cycle transformation intensity, and a fruit quality prediction range. The fruit quality prediction range includes the prediction range of soluble solids content, total sugar content, protein content, and amino acid content. The target wolfberry field was evaluated based on the comprehensive evaluation model. The evaluation included: detecting the current δ¹⁸O values ​​of ammonium nitrogen and nitrate nitrogen in the soil of the field. 15 The N value is input into the model to obtain the soil nitrogen plant availability quantitative score, nitrogen cycle transformation intensity quantitative score, and fruit quality prediction range for the field. Based on the comparison results between the fruit quality prediction range and the preset quality target, as well as the values ​​of the soil nitrogen plant availability quantitative score and nitrogen cycle transformation intensity quantitative score, a fertilization decision scheme including nitrogen source type and application rate recommendations is generated. The preset quality target includes the baseline values ​​of soluble solids content, total sugar content, protein content, and amino acid content.

2. The method according to claim 1, characterized in that, During the same period of the field trial and the corresponding period for evaluating the target field, at least one of the following environmental factors was monitored and acquired simultaneously: Meteorological data, including average temperature and rainfall; Soil physical data, including soil volumetric water content; When constructing the comprehensive evaluation model, the acquired environmental factor data are compared with the δ values ​​of soil ammonium nitrogen and nitrate nitrogen. 15 The N values ​​are combined to form the input feature set of the model.

3. The method according to claim 2, characterized in that, The comprehensive evaluation model is constructed using the gradient boosting decision tree algorithm; The input features of the gradient boosting decision tree model include: the soil ammonium nitrogen δ 15 N value, nitrate nitrogen δ 15 The N value, the regression coefficient in the quantitative function relationship, the numerical values ​​of the apparent net mineralization rate and net nitrification rate, and the key environmental factor data; The model is configured to simultaneously predict the output soil nitrogen plant availability quantification score, nitrogen cycle transformation intensity quantification score, and fruit quality prediction range based on the input features.

4. The method according to claim 3, characterized in that, The gradient boosting decision tree model is trained by minimizing a multi-task loss function; The multi-task loss function is a linear weighted sum of the prediction error terms corresponding to the soil nitrogen plant availability quantification score, the nitrogen cycle transformation intensity quantification score, and the fruit quality prediction interval, respectively. The weight coefficients in the linear weighted sum are determined by iteratively optimizing the weight coefficients using validation set performance feedback during model training.

5. The method according to claim 4, characterized in that, The determination of the weight coefficients through iterative optimization using validation set performance feedback is achieved by executing a process comprising the following steps: Assign a set of preset initial values ​​to the weight coefficients, and divide the field test dataset into a non-overlapping training subset and a validation subset; Multiple optimization loops are performed. In each loop, the gradient boosting decision tree model is trained on the training subset using the current weight coefficients. The trained model is used to make predictions on the validation subset and the prediction errors of the soil nitrogen plant availability quantification score, nitrogen cycle transformation intensity quantification score, and fruit quality prediction interval are calculated respectively. Then, a set of new weight coefficient candidate values ​​is calculated according to the prediction error and the preset weight update algorithm. Substitute the new candidate values ​​of the weight coefficients into the model and re-evaluate the overall prediction performance on the validation subset. If the overall performance is better than the performance using the current weight coefficients, then update the weight coefficients to the candidate values. Repeat the optimization loop and evaluation update process until the preset optimization termination condition is reached, and use the final obtained weight coefficient as the determined value in the linear weighted sum.

6. The method according to any one of claims 1 to 5, characterized in that, The generation of a fertilization decision scheme that includes suggestions on nitrogen source type and application rate is accomplished by querying a preset decision mapping knowledge base. The decision mapping knowledge base uses the quantitative score of soil nitrogen plant availability, the quantitative score of nitrogen cycle transformation intensity, and the comparison results of the fruit quality prediction interval with the preset quality target as joint input conditions. The knowledge base pre-stores the mapping relationship between multiple sets of input condition combinations and corresponding output decisions, where each set of output decisions includes a specific recommended nitrogen source type and application amount.

7. The method according to claim 6, characterized in that, The generation of a fertilization decision scheme that includes suggestions on nitrogen source type and application amount is achieved through a trained fertilization scheme recommendation model; The fertilization program recommendation model uses the quantitative score of soil nitrogen availability to plants, the quantitative score of nitrogen cycle transformation intensity, and the comparison results between the fruit quality prediction interval and the preset quality target as input features. The fertilization scheme recommendation model uses fertilization scheme data that have been verified as preferred in historical agronomic trials and correspond to the input features as training labels; The model outputs a recommended nitrogen source type and application rate that matches the system state represented by the input features.

8. The method according to claim 7, characterized in that, The recommended fertilization scheme model is a multilayer perceptron neural network. The multilayer perceptron neural network includes an input layer, at least one hidden layer, and an output layer; wherein the output layer is configured to simultaneously output discrete classification results representing different nitrogen source types and continuous numerical values ​​representing recommended application amounts.

9. The method according to claim 5, characterized in that, The preset weight update algorithm used in the process of iteratively optimizing the weight coefficients using validation set performance feedback is as follows: Based on the latest prediction error of each task on the validation subset, dynamically adjust its weight coefficient in the loss function; Specifically, an error benchmark threshold is set; for tasks whose prediction error exceeds the error benchmark threshold, their loss weight is increased in the next iteration according to the excess ratio, and the increase is proportional to the excess ratio; for tasks whose prediction error is below the error benchmark threshold, their loss weight remains unchanged.