A yield-type farmland quality index construction method and device and a computer storage medium

Abstract

The present application provides a yield type cultivated land quality index construction method, comprising: obtaining farmland sample data to form soil index-crop yield paired data set; data preprocessing is carried out on farmland sample data; machine learning yield prediction model is constructed and trained; the interpretability of the yield prediction model is analyzed; the response threshold is identified: the SHAP dependence curve or the cumulative local effect curve is drawn for the yield sensitive soil index; the response type scoring function is constructed: the yield sensitive index is divided according to the response curve form to construct the response scoring function; the yield supporting type cultivated land quality index is calculated: the yield response score of each yield sensitive index is weighted and integrated with SHAP weight to obtain the yield supporting type cultivated land quality index. The present application also provides a yield type cultivated land quality index construction device and a computer storage medium. The present application converts the background quality description in the traditional cultivated land quality evaluation into a functional diagnosis facing crop yield formation.

Description

Technical Field

[0001] This invention belongs to the field of farmland quality evaluation technology, specifically relating to a method, apparatus and computer storage medium for constructing a yield-based farmland quality index. Background Technology

[0002] Farmland quality assessment typically involves collecting soil physical, chemical, and biological indicators, which are then standardized, weighted, and integrated to form a comprehensive soil quality index or farmland quality index. This index characterizes the soil's overall capacity to support crop production, maintain resource cycles, and preserve ecological functions. Such comprehensive indices can be used for regional farmland quality grading, long-term monitoring, and comparison of different management practices.

[0003] Existing traditional farmland quality indices typically involve the following technical process: First, constructing a complete dataset, which includes all soil physical, chemical, and biological indicators in the evaluation; second, forming a minimal dataset through principal component analysis, correlation screening, or expert judgment; third, classifying indicators into positive, negative, or suitability indicators based on their attributes, and using linear or nonlinear functions to convert indicators of different dimensions into evaluation scores within the 0-1 range; finally, obtaining a comprehensive farmland quality index using a weighted quality index, Nemerow quality index, or similar ensemble methods.

[0004] While the aforementioned traditional arable land quality indices are advantageous in characterizing the background quality of arable land, their weighting primarily relies on data variation, principal component contributions, or expert experience. The evaluation objective is typically "comprehensive soil condition," which is not necessarily equivalent to "the crop yield-forming capacity for the current season." In actual production, some plots may have high comprehensive soil quality scores, but their crop yields may not necessarily be high because key yield-limiting factors have not been accurately identified.

[0005] A key reason for these problems is that crop yield's response to soil indicators exhibits significant nonlinearity and context-dependent characteristics. For example, some nutrient indicators may limit yield at low values, entering a plateau region after reaching a threshold; some indicators may have suitable ranges, with excessively low or high values ​​being detrimental to yield; and some indicators may even act as limiting factors at high values. Therefore, simply adopting the traditional scoring method of "the higher the better" or "the lower the better" may lead to a misalignment between the direction of indicator scoring and the direction of yield response.

[0006] Machine learning methods can capture the non-linear relationship between soil indicators and crop yield. However, conventional machine learning prediction models primarily focus on prediction accuracy and struggle to directly explain the positive or negative contribution, threshold positions, and management implications of individual soil indicators to yield prediction. For farmland management, obtaining yield predictions alone is insufficient; the model's interpretation results must be transformed into calculable soil function scores, weakness rankings, and management recommendations. Therefore, while traditional arable land quality indices can comprehensively reflect the background state of arable land, they struggle to reliably explain crop yield and identify yield-limiting factors for the current season. Although machine learning models can improve yield prediction capabilities, their results are difficult to directly translate into arable land quality indices, weakness diagnoses, and management recommendations. Summary of the Invention

[0007] To address the problems in the background technology, this invention introduces a crop yield response calibration mechanism, which transforms machine learning predictions and interpretability analysis results into yield-sensitive indicators, response thresholds, scoring functions, and weighting systems, thereby constructing a yield-supported arable land quality index.

[0008] This invention proposes a method for constructing a farmland quality index based on yield response calibration. The method acquires soil physical, chemical, and biological indicators and corresponding crop yield data from multiple sampling points in a target farmland, forming a soil indicator-crop yield paired dataset. A machine learning prediction model is trained using crop yield as the response variable. SHAP values ​​and permutation importance are used to identify yield-sensitive soil indicators. Based on the SHAP dependency curve, lower threshold, optimal response point, and upper threshold are determined. Yield response scoring functions are established according to positive threshold type, high-value restriction type, and suitable interval type. Finally, yield-oriented weights are calculated using the average absolute SHAP value to obtain a yield-supported farmland quality index.

[0009] The method for constructing a yield-based arable land quality index of the present invention includes: S1, acquiring farmland sample data, including soil physical, chemical and biological indicators and corresponding crop yield data, forming a sample-level soil indicator-crop yield paired dataset; S2, performing data preprocessing on the farmland sample data; S3, constructing and training a machine learning yield prediction model: using crop yield as the response variable and candidate soil indicators as explanatory variables, training multiple machine learning models, and selecting a target yield prediction model based on the training set, test set, and cross-validation results; S4, performing interpretability analysis on the yield prediction model: using SHAP analysis and permutation importance analysis. S5. Analyze the contribution direction and intensity of each soil index to yield prediction, and screen yield-sensitive soil indices by combining the mean absolute SHAP value; S6. Identify response thresholds: Plot SHAP dependency curves or cumulative local effect curves for yield-sensitive soil indices; S7. Construct response-type scoring functions: Divide yield-sensitive indices into positive threshold type, high value restriction type, and suitable interval type according to the shape of the response curve, and construct response scoring functions for each; S8. Calculate the yield-supporting arable land quality index YSSQI: Weight and integrate the yield response scores of each yield-sensitive index with the SHAP weights to obtain the yield-supporting arable land quality index.

[0010] In one implementation, step S2 includes: identifying outliers, handling missing values, unifying units, standardizing or normalizing the sample-level soil index-crop yield paired data, and diagnosing collinearity among candidate soil indices; for indices with repetitive functions or high correlation, representative indices are retained by combining soil science significance, crop yield formation process and model stability to form a candidate index dataset.

[0011] In one implementation, in step S3, the model performance evaluation index includes the coefficient of determination R. 2 One or more of the following: root mean square error (RMSE), mean absolute error (MAE), residual prediction bias (RPD), and interquartile range performance ratio (RPIQ).

[0012] In one implementation, step S4 includes: S4.1, determining candidate indicators; S4.2, data partitioning: dividing the data into training and testing sets; S4.3, training multiple models; S4.4, performance evaluation: evaluating model performance using the training set, testing set, and cross-validation; S4.5, determining the target model: selecting the model with the best prediction performance; S4.6, SHAP: after determining the target model, using SHAP summary plot, mean absolute SHAP value, and permutation importance to evaluate the contribution of candidate indicators to yield prediction; S4.7, determining core indicators: indicators with high contribution intensity and clear soil science significance are determined as yield-sensitive soil indicators.

[0013] In one implementation, in step S4.3, the trained model has a high R... 2 It has low RPD and RPIQ; low RMSE and MAE; stable cross-validation; and can perform SHAP or permutation importance interpretation.

[0014] In one implementation, step S5 further includes: determining the intersection of the response curve and the zero contribution baseline, as well as the positive contribution peak point, to obtain the lower limit threshold L, the optimal response point O, and the upper limit threshold U.

[0015] In one implementation, step S6 includes: the position where the SHAP value is equal to 0 is used as the zero contribution benchmark point for the indicator to change from negative contribution to positive contribution or from positive contribution to negative contribution; the lower limit threshold L and the upper limit threshold U can be determined by the intersection of the SHAP response curve and the zero contribution benchmark line; the optimal response point O can be determined by the position in the SHAP response curve where the positive contribution reaches its peak or the output contribution is the highest.

[0016] In one implementation, step S7 includes: 1) converting the measured values ​​of each yield-sensitive soil index into a yield response score S in the range of 0-1 according to the yield response scoring function established in step S6. i Y Among them, S i Y The higher the value, the stronger the supporting effect of the current value of the indicator on the formation of the target crop yield; 2) Based on the model interpretability analysis results obtained in step S4, calculate the average absolute SHAP value of each yield-sensitive soil indicator, and normalize it to obtain the yield-oriented weight w. i Y 3) Calculate the yield-supporting arable land quality index YSSQI, YSSQI = , where n is the number of yield-sensitive soil indicators included in the index calculation, and = 1, the value of YSSQI ranges from 0 to 1. The higher the value, the stronger the comprehensive support capacity of the current soil function status of the target farmland for the formation of the target crop yield.

[0017] This invention proposes a yield-based arable land quality index construction device, comprising: a processor and a memory, wherein the memory stores a computer program, and when the processor runs the computer program, it implements the method described above.

[0018] The present invention proposes a computer storage medium, characterized in that the medium stores a computer executable program, which, when the computer executable program is run, implements the method described above.

[0019] The beneficial effects of this invention are as follows: This invention retains traditional indices as background quality evaluation and comparison, and uses crop yield response as the basis for indicator screening, scoring and weighting, thereby constructing a yield-supporting arable land quality index that can identify the shortcomings of yield function in the current season.

[0020] The method of this invention can transform the background quality description in traditional arable land quality assessment into a functional diagnosis oriented towards crop yield formation, and output the comprehensive yield support level, individual limiting factors, shortcoming contribution value and farmland management recommendations for the plot.

[0021] Compared with the traditional arable land quality index, this invention incorporates crop yield response into indicator screening, scoring functions, and weight calculation, which can more directly reflect the supporting role of soil quality in crop yield formation.

[0022] This invention retains the traditional SQI calculation process as a reference, and can simultaneously obtain background quality assessment results and yield function diagnosis results, which is helpful in distinguishing between two different issues: "good overall soil condition" and "strong yield support capacity in the current season".

[0023] This invention identifies the contribution direction and response threshold of the index by interpreting the results using SHAP, avoiding misjudgments caused by fixed scoring directions in traditional indices, and can handle nonlinear response relationships such as threshold type, high value restriction type, and suitable interval type.

[0024] This invention not only outputs a comprehensive index, but also outputs the contribution value and management priority of individual indicators, which can transform the farmland quality evaluation results into plot-level soil improvement, nutrient regulation and cultivation management suggestions.

[0025] This invention can retrain the model and update the threshold library based on the target crop, soil type and regional sample data, and therefore can be extended to the functional diagnosis of arable land quality and yield in different crop systems and regions. Attached Figure Description

[0026] To facilitate understanding of the invention, it will be described in more detail with reference to the specific embodiments shown in the accompanying drawings. These drawings depict only typical embodiments of the invention and should not be considered as limiting the scope of protection of the invention.

[0027] Figure 1 This is a schematic diagram comparing the overall process and technical path of the present invention.

[0028] Figure 2 Flowchart for screening machine learning models and identifying yield-sensitive indicators.

[0029] Figure 3 A schematic diagram illustrating the construction of a yield response scoring function based on the SHAP response threshold.

[0030] Figure 4This is a schematic diagram of the diagnostic and management output for the YSSQI plot.

[0031] Figure 5 This is a schematic diagram of the structure of the device, electronic equipment, and storage medium. Detailed Implementation

[0032] The embodiments of the present invention are described below with reference to the accompanying drawings to enable those skilled in the art to better understand and implement the present invention. However, the listed embodiments are not intended to limit the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments can be combined with each other, wherein the same components are indicated by the same reference numerals.

[0033] Figure 1 This is a schematic diagram comparing the overall process and technical path of the present invention, showing the overall process from sample data acquisition, traditional SQI comparison construction, machine learning modeling, SHAP interpretation, response threshold identification, YSSQI calculation to weakness diagnosis and management recommendation output, and comparing the technical paths of traditional SQI and YSSQI.

[0034] Unlike the traditional SQI (Traditional Soil Quality Index), the YSSQI (Yield Supported Soil Quality Index) does not primarily rely on soil background differences or index variability. Instead, it uses the direction, intensity, and response threshold of soil indicators' contribution to crop yield prediction as the basis for scoring and weighting. The method of this invention includes steps S1-S8.

[0035] S1, Obtain farmland sampling data: Set up multiple spatial sampling points for the target farmland, collect topsoil samples at each sampling point, measure soil physical, chemical and biological indicators, and obtain crop yield at the corresponding location or plot scale to form sample-level soil indicator-yield paired data.

[0036] In one embodiment, a long-term experimental field of cold-region paddy soil was used as the target farmland. The target farmland included deep plowing and straw return treatment and rotary tillage and straw removal treatment. The deep plowing and straw return treatment involved deep plowing to a depth of 25-30 cm and returning all the straw to the field; the rotary tillage and straw removal treatment involved rotary tillage to a depth of 12-15 cm and removing the straw.

[0037] During the rice harvest season, sampling points were randomly arranged in an "S" shape within the two treatment zones, with 50 sampling points set up for each treatment, resulting in a total of 100 sampling points. Soil samples from the 0-20 cm topsoil layer were collected at each sampling point, and a 1 m² quadrat was set up at the corresponding location for actual rice yield measurement. The grain yield was uniformly converted to the yield at standard moisture content.

[0038] The soil quality indicators measured included: bulk density (BD), topsoil thickness (TD), soil organic carbon (SOC), total nitrogen (TN), total phosphorus (TP), total potassium (TK), pH, nitrate nitrogen (NO3-N), ammonium nitrogen (NH4-N), Olsen-P, available potassium (AK), cation exchange capacity (CEC), amorphous iron (Am-Fe), amorphous aluminum (Am-Al), acid phosphatase (ACP), alkaline phosphatase (ALP), phosphodiesterase (PDE), urease (Urease), and sucrase.

[0039] The above indicators can be increased or decreased based on the target crop and regional soil type. For other farmland systems, the crop yield can also be replaced with the yield of wheat, corn, soybeans, vegetables, or other target crops, or with a comprehensive response variable of yield, quality, and economic benefits.

[0040] Table 1. Sampling data fields, indicator types, units, and measurement / acquisition methods Note: The fields in the table are input data fields in one implementation method; in actual applications, they can be added or removed according to the target crop, soil type, and data acquisition conditions.

[0041] In contrast to the method of this invention, this invention also proposes a method for constructing a traditional farmland quality index benchmark. Following the traditional SQI process, a full dataset and a minimum dataset are constructed separately, and linear or nonlinear scores are combined with a weighted quality index or the Nemerow quality index to form multiple traditional farmland quality indices. This stage is used to obtain background quality assessment results and a comparison benchmark.

[0042] (1) Construct the complete dataset TDS. Let the target farmland have m soil quality indicators, then the complete dataset TDS is represented as TDS = {x1, x2, …, x m}. Where m can be the total number of soil physical, chemical and biological indicators.

[0043] (2) Construct the minimum dataset MDS. After standardizing the indicators in the TDS, perform principal component analysis and retain the principal components with eigenvalues ​​greater than 1 and variance contribution rates that meet the preset conditions. Select indicators with higher loadings in each retained principal component. When the correlation between high-loading indicators is strong, retain indicators with high loadings and clear soil science significance. When the correlation is weak, both can be retained to obtain the MDS.

[0044] (3) Linear scoring. Each indicator is converted into a dimensionless score between 0 and 1. For positive indicators, S = (x - x) min ) / (x max - x min For negative indices, use S = (x max- x) / ( x max - x min For the suitability index, S = 1 - |x - x0| / max(x) max - x0, x0- x min The results are limited to the range of 0-1. Here, x is the measured value, and x... min and x max These are the minimum and maximum values ​​of the indicator, respectively, with x0 being the optimal value.

[0045] (4) Nonlinear scoring. For indices with a gradually changing response, a sigmoid function can be used for nonlinear scoring. Positive indices can use S = 1 / [1 + exp(-a(x - c))]; negative indices can use S = 1 / [1 + exp(a(x - c))]; suitability indices can use S = exp[-(x - x0)^2 / (2σ^2)], or other nonlinear functions that can characterize the suitability interval can be used. Where a is the curve steepness parameter, c is the curve center parameter, and σ is the suitability interval width parameter.

[0046] (5) Calculate the weights. Traditional SQI weights can be determined based on principal component communality, variance contribution rate, entropy weight, or expert weighting. In one implementation, w is used. i = C i / ΣC i Calculate the indicator weights, where C i This represents the commonality or overall contribution of the i-th indicator.

[0047] (6) Weighted Quality Index Integration. The weighted quality index WQI is calculated as WQI = Σ(w i × S i ) calculate, where w i S is the weight of the i-th indicator. i Score the i-th indicator.

[0048] (7) Nemerow Quality Index Integration. The Nemerow Quality Index (NQI) can be calculated as NQI = [(S mean ^2 + S min Calculate using [^2) / 2]^(1 / 2), where S mean S is the mean score of all indicators. min The lowest score is indicated by the indicator.

[0049] (8) By combining dataset types, scoring methods, and index integration methods, eight traditional farmland quality indices were formed: TDS-L-WQI, TDS-L-NQI, TDS-N-WQI, TDS-N-NQI, MDS-L-WQI, MDS-L-NQI, MDS-N-WQI, and MDS-N-NQI. Subsequently, correlation analysis, regression analysis, or path analysis were performed between the above traditional SQIs and crop yield to evaluate the yield explanatory power of the traditional composite indices.

[0050] Taking the implementation of cold-region rice cultivation as an example, the method for constructing a traditional farmland quality index comparison is as follows.

[0051] 1) Construct a complete TDS dataset. All soil physical, chemical, and biological indicators are incorporated into the TDS to characterize the complete arable land background quality.

[0052] 2) Construct a minimal dataset (MDS). After standardizing each indicator in the TDS, perform PCA (principal component analysis) to screen principal components. Screening criteria include: principal components that meet preset conditions (e.g., eigenvalues ​​greater than 1 and variance contribution rates meeting preset conditions) or those that meet high loading criteria (indicators with large absolute loading values ​​or close to the highest loading of the principal component). If the high loading indicators are strongly correlated (which can be calculated using Pearson or Spearman correlation coefficients; the threshold for strong correlation can be preset according to the study area and indicator system), then retain indicators with high loadings and clear soil science significance; if the correlation is weak, then retain both.

[0053] In one implementation, after standardizing the soil quality indicators in the full TDS dataset, principal component analysis or correlation analysis is used to perform dimensionality reduction screening of the indicators. Based on the loadings of each indicator in the principal components, the correlation between indicators, and the representativeness of soil function, indicators that can represent the physical, chemical, and biological functions of the soil are selected to form a minimal dataset MDS. The MDS is used for the construction of a control for traditional arable land quality indices and is not a necessary component for the calculation of yield-supported arable land quality indices.

[0054] 3) Execution of indicator scoring. Indicators in the TDS or MDS are converted into scores within the 0-1 range using both linear and non-linear scoring methods. Linear scoring is categorized into positive, negative, and appropriate scores based on indicator attributes; non-linear scoring uses a sigmoid function or an appropriate interval function to reduce the impact of extreme values ​​on the evaluation results.

[0055] 4) Perform weight calculation and index integration. Traditional weights are determined using PCA communality or variance contribution rate, and index integration is performed using WQI and NQI respectively.

[0056] 5) Eight traditional SQI combinations are formed, namely TDS-L-WQI, TDS-L-NQI, TDS-N-WQI, TDS-N-NQI, MDS-L-WQI, MDS-L-NQI, MDS-N-WQI and MDS-N-NQI.

[0057] 6) Perform Pearson correlation, linear regression, or path model analysis on the eight traditional SQIs and crop yield respectively to evaluate the explanatory power of traditional SQIs on yield. If traditional SQIs can distinguish treatment differences but have a weak relationship with yield, it indicates that traditional SQIs are more suitable for characterizing the background quality of cultivated land and cannot fully identify the functional shortcomings of yield in the current season.

[0058] Table 2. Combination Types and Calculation Methods of Traditional Cultivated Land Quality Indices Note: TDS represents the entire dataset; MDS represents the minimum dataset; L represents linear score; N represents non-linear score; WQI represents weighted quality index; NQI represents Nemerow quality index.

[0059] S2, Data preprocessing of farmland sample data: outlier checks, missing value handling, unit unification, standardization or normalization processing are performed on the paired data, and collinearity between indicators is diagnosed; for indicators with repetitive functions or high correlation, representative indicators are retained in combination with soil science significance, crop yield formation process and model stability.

[0060] The purpose of this step is to further transform the interpretation results of the yield forecasting model into yield-sensitive indicators, response thresholds, scoring functions, and yield-oriented weights, which are used to construct the YSSQI and diagnose yield-limiting factors.

[0061] In one implementation, the sample-level soil index-crop yield paired data undergoes outlier identification, missing value processing, unit unification, standardization, or normalization. Collinearity among candidate soil indices is also diagnosed. For indices with functional duplication or high correlation, representative indices are retained based on soil science significance, crop yield formation process, and model stability to form a candidate index dataset. This step aims to reduce the impact of dimensional differences, outliers, and index redundancy on subsequent yield prediction models, SHAP contribution analysis, and response threshold identification, and provides consistent data input for the construction of yield-supported arable land quality indices.

[0062] S3. Construct and train a machine learning yield prediction model: Using crop yield as the response variable and candidate soil indicators as explanatory variables, train multiple machine learning models and select the target yield prediction model based on the training set, test set, and cross-validation results. Model performance metrics include, but are not limited to, R... 2RMSE, MAE, RPD, and RPIQ.

[0063] The purpose of this step is to further transform the interpretation results of the yield forecasting model into yield-sensitive indicators, response thresholds, scoring functions, and yield-oriented weights, which are used to construct the YSSQI and diagnose yield-limiting factors.

[0064] In one implementation, crop yield is used as the response variable and candidate soil indicators are used as explanatory variables to train at least one machine learning yield prediction model. In another implementation, one or more models selected from Gradient Boosting, ExtraTrees, AdaBoost, XGBoost, Random Forest, Support Vector Regression, Ridge Regression, and StackingEnsemble can be trained, and the target yield prediction model is determined based on the training set, test set, or cross-validation results. Model performance evaluation metrics include the coefficient of determination R0. 2 The target yield prediction model is used not only to predict crop yield, but also to extract the contribution direction, contribution intensity, and response threshold of soil indicators to yield prediction, thereby providing a model basis for the screening of yield-sensitive indicators, the construction of yield response scoring functions, and the calculation of YSSQI.

[0065] S4. Interpretability analysis of the machine learning yield prediction model: SHAP analysis and permutation importance analysis are used to quantify the contribution direction and intensity of each soil index to yield prediction, and yield-sensitive soil indexes are screened by combining the mean absolute SHAP value.

[0066] Figure 2 The flowchart shows the process of screening machine learning models and identifying yield-sensitive indicators, demonstrating the construction of candidate indicators, data partitioning, multi-model training, performance comparison, determination of target models, and analysis of SHAP importance and permutation importance.

[0067] Before constructing the YSSQI, candidate yield explanatory indicators are screened from all soil indicators. This invention comprehensively considers multiple linear regression, variance inflation factor, Pearson correlation, Mantel test, and soil function representativeness to reduce the impact of highly collinear and functionally repetitive indicators on model stability.

[0068] S4.1, Identify candidate indicators.

[0069] S4.2, Data Partitioning: Divide the data into training and test sets. Specifically, using a single farmland sampling point or plot as the basic sample unit, divide the sample point-level soil index-crop yield paired data into training and test sets.

[0070] S4.3, train multiple models. Trainable models include: Gradient Boosting, ExtraTrees, AdaBoost, XGBoost, Random Forest, SVR, Ridge, and Stacking Ensemble. Generally, models with high R-values ​​are selected. 2 Low RPD and RPIQ; low RMSE and MAE; stable cross-validation; model that can be interpreted using SHAP or permutation importance.

[0071] S4.4 Performance Evaluation: Model performance is evaluated using the training set, test set, and cross-validation.

[0072] S4.5, Determine the target model: Select the model with the best prediction performance.

[0073] S4.6: SHAP: After determining the target model, use the SHAP summary plot, average absolute SHAP value, and permutation importance to evaluate the contribution of candidate indicators to yield forecasting.

[0074] S4.7 Determine core indicators: Indicators with high contribution intensity and clear soil science significance are identified as yield-sensitive soil indicators and enter the YSSQI construction stage.

[0075] In one implementation, candidate indicators include BD, TD, NO3-N, AK, Olsen-P, pH, Am-Fe, ACP, Urease, and Sucrase, which represent functional dimensions such as soil structure, topsoil space, available nutrients, acid-base environment, mineral buffering, and biotransformation, respectively.

[0076] Using rice yield as the response variable and candidate soil indicators as explanatory variables, models such as GradientBoosting, Extra Trees, AdaBoost, XGBoost, Random Forest, SVR, Ridge, and StackingEnsemble were trained respectively.

[0077] Model performance is evaluated using training set, test set, and cross-validation. Evaluation metrics include R-squared. 2 RMSE, MAE, RPD, and RPIQ. Prioritize the test set R. 2Models with high RMSE and MAE, high RPD and RPIQ, and good interpretable stability were selected as target yield prediction models. After determining the target model, the contribution of candidate indices to yield prediction was evaluated using SHAP summary plots, mean absolute SHAP values, and substitution importance. Indicators with high contribution intensity and clear soil scientific significance were identified as yield-sensitive soil indices. NO3-N, Olsen-P, pH, Urease, and TD were ultimately selected as yield-sensitive soil indices for subsequent response threshold identification and YSSQI construction.

[0078] The method of the present invention further includes step S5, identifying response thresholds: plotting SHAP dependence curves or cumulative local effect curves for yield-sensitive soil indicators, determining the intersection of the response curve with the zero contribution baseline and the positive contribution peak point, and obtaining the lower limit threshold L, the optimal response point O and the upper limit threshold U.

[0079] Specifically, SHAP dependency curves are plotted for yield-sensitive soil indicators. A SHAP value greater than 0 indicates that the indicator makes a positive contribution to yield prediction within the corresponding value range, while a SHAP value less than 0 indicates a negative contribution. When the SHAP curve changes from negative to positive and gradually stabilizes as the indicator value increases, the indicator is classified as a positive threshold type; when the SHAP curve continuously decreases as the indicator value increases, and shows a negative contribution in the high-value range, the indicator is classified as a high-value restricted type; when the SHAP curve only shows a positive contribution in the middle range, while showing a negative contribution in both the low-value and high-value ranges, the indicator is classified as an appropriate range type.

[0080] Taking cold-region paddy soil as an example, pH was identified as a high-value limiting indicator, NO3-N and TD were identified as positive threshold indicators, and Olsen-P and Urease were identified as suitable range indicators.

[0081] Table 3 Performance Comparison of Machine Learning Models Note: R 2 Higher RPD and RPIQ, and lower RMSE and MAE, indicate better model prediction performance. This table shows the test set results for one implementation method.

[0082] The method of this invention further includes step S6, constructing a response-type scoring function: Based on the shape of the response curve, the yield-sensitive index is divided into positive threshold type, high-value restriction type, and suitable interval type, and corresponding response scoring functions are constructed for each. Specifically, the lower limit threshold L, optimal response point O, and upper limit threshold U are determined based on the SHAP response curve of the yield-sensitive index, and the yield-sensitive index is divided into positive threshold type, high-value restriction type, and suitable interval type based on the shape of the response curve, and corresponding yield response scoring functions are constructed for each. The position where the SHAP value equals 0 is used as the zero-contribution benchmark point for the index to change from negative to positive contribution or from positive to negative contribution; the lower limit threshold L and upper limit threshold U can be determined by the intersection of the SHAP response curve and the zero-contribution benchmark line; the optimal response point O can be determined by the position in the SHAP response curve where the positive contribution reaches its peak or the yield contribution is highest.

[0083] Figure 3 This diagram illustrates the construction of a yield response scoring function based on SHAP response thresholds, demonstrating the process of identifying L, O, and U thresholds from the SHAP response curve and further constructing positive threshold-type, high-value-restricted, and suitable-interval-type scoring functions.

[0084] In one implementation, the optimal pH response point O is 5.51, the upper limit threshold U is 6.01, and the lower limit threshold L for NO3-N is 53.02 mg / kg. -1 The upper limit response point U is 90.88 mg kg. -1 The lower threshold L of Olsen-P is 23.53 mg / kg. -1 The optimal response point O is 32.49 mg kg. -1 The upper limit threshold U is 41.55 mg kg. -1 The lower threshold L of Urease is 102.84 mg N kg. -1 h -1 The optimal response point O is 113.78 mg N kg. -1 h -1 The upper limit threshold U is 126.44 mg N kg. - 1 h -1 The lower threshold L of TD is 18.07 cm, and the upper response point U is 20.00 cm.

[0085] The thresholds mentioned above are candidate yield response thresholds obtained based on the current samples and management context, and should not be construed as universal agronomic thresholds for all regions and all crops. In other regions or crop systems, the model can be retrained and the threshold library updated according to the same procedure of this invention.

[0086] Table 4. Yield-sensitive indicators, SHAP response types, thresholds, and scoring rules Note: L represents the lower threshold, O represents the optimal response point, and U represents the upper threshold; the threshold is a candidate threshold in the cold-region paddy soil implementation method and should not be understood as a universal threshold for all regions and crops.

[0087] The method of the present invention further includes step S7, calculating the yield-supported arable land quality index: the yield response scores of each yield-sensitive indicator are weighted and integrated with the SHAP weights to obtain the yield-supported arable land quality index. Step S7 is described in detail below.

[0088] 1) Based on the yield response scoring function established in step S6, convert the measured values ​​of each yield-sensitive soil index into yield response scores S in the range of 0-1. i Y Among them, S i Y The higher the value, the stronger the support the current value of the indicator has for the yield of the target crop.

[0089] 2) Based on the model interpretation results obtained in step S4, calculate the average absolute SHAP value of each yield-sensitive soil index, and normalize it to obtain the yield-oriented weight w. i Y In one implementation, , where φ i This represents the SHAP value of the i-th indicator in the target output prediction model.

[0090] 3) Calculate the yield-supporting farmland quality index, YSSQI. YSSQI = , where n is the number of yield-sensitive soil indicators included in the index calculation, and = 1. The value of YSSQI ranges from 0 to 1. The higher the value, the stronger the comprehensive support capacity of the current soil function status of the target farmland for the formation of the target crop yield.

[0091] The core of this invention lies in incorporating crop yield response into the index screening, scoring function construction, and weight calculation processes. Unlike traditional SQI, YSSQI does not primarily rely on soil background differences or index variability, but rather uses the direction, intensity, and response threshold of soil indicators' contribution to crop yield prediction as the basis for scoring and weighting.

[0092] A positive threshold-based scoring function is used to characterize an indicator where a low value limit is met and exceeding the threshold is beneficial to yield. Its calculation method is as follows: when x ≤ L, S i Y= 0; when L<x<U, S i Y = (x - L) / (U - L); when x ≥ U, S i Y = 1.

[0093] The high-value limited scoring function is used to characterize an indicator where a high-value state reduces the yield contribution, and its calculation method is: when x≤ O, S i Y = 1; when O<x<U, S i Y = (U - x) / (U - O); when x ≥ U, S i Y = 0.

[0094] The suitable interval scoring function is used to characterize an indicator where too low or too high is unfavorable to yield, and only being within the suitable interval is favorable to yield, and its calculation method is: when x ≤ L or x ≥ U, S i Y = 0; when L<x<O, S i Y = (x - L) / (O -L); when O ≤ x<U, S i Y = (U - x) / (U - O).

[0095] The yield-oriented weight is obtained by normalizing the mean absolute SHAP value: w i Y = mean(|φi|) / Σmean(|φj|), where φ i is the SHAP value of the i-th yield-sensitive indicator.

[0096] The yield-supporting cultivated land quality index is calculated as YSSQI = Σi =1..n (w i Y × S i Y )calculation, where S i Y is the yield response score of the i-th yield-sensitive indicator, w i Y is the yield-oriented weight of the i-th yield-sensitive indicator, and n is the number of yield-sensitive indicators included in YSSQI.

[0097] The short-board contribution value of a single indicator is calculated as D i = w i Y × (1 - S i Y )calculation. Di The larger the value, the greater the contribution of this indicator to insufficient production support. The system can be configured according to D. i Output the priority of the limiting factors from largest to smallest, and provide corresponding management suggestions.

[0098] Based on the response type and response threshold determined in step S6, calculate the output response score for each output-sensitive indicator. For positive threshold-type indicators, the score is 0 when below the lower threshold and 1 when reaching or exceeding the upper response point; for high-value-restricted indicators, the score is 1 when below the optimal response point and 0 when exceeding the upper threshold; for suitable interval-type indicators, the score is higher only within the suitable interval.

[0099] Figure 4 The diagram shows the output of YSSQI land parcel diagnosis and management, displaying the output of production support level, production-sensitive indicator ranking, threshold status, weight bar chart, and management recommendations.

[0100] The average absolute SHAP value of each output-sensitive indicator is used as the output contribution intensity, and the output-oriented weight is obtained by normalization.

[0101] The YSSQI is obtained by multiplying the yield response score by the yield-oriented weight and then summing the results. The higher the YSSQI, the stronger the support capacity of the current soil function status of the farmland for the formation of the target crop yield.

[0102] The method of the present invention further includes step S8, in the farmland management recommendation output stage, calculating the short-board contribution value of each yield-sensitive indicator, determining the priority of limiting factors according to the short-board contribution value from high to low, and outputting the control direction and management suggestions according to the indicator response type.

[0103] To achieve site-level diagnosis, this invention further calculates the contribution value D of the weakest link in each individual indicator. i = w i Y × (1 -S i Y When a certain indicator scores low and has a high weight, its D... i A relatively large value indicates that this indicator should be a priority for management.

[0104] In one implementation, YSSQI can be divided into four levels: YSSQI ≥ 0.80 is the high-yield support level, 0.60 ≤ YSSQI < 0.80 is the relatively high-yield support level, 0.40 ≤ YSSQI < 0.60 is the medium-yield support level, and YSSQI < 0.40 is the low-yield support level. These level thresholds can be adjusted based on regional yield distribution or management objectives.

[0105] To demonstrate the technical effectiveness of this invention compared to traditional SQI, this embodiment performs correlation analysis, linear regression, and path model analysis on traditional SQI and YSSQI with crop yield, respectively.

[0106] In one implementation, eight conventional SQIs can distinguish the overall background state of cultivated land under different tillage management treatments, but their explanatory power with rice yield is low, indicating that conventional SQIs mainly reflect the overall soil condition and are insufficient to accurately characterize the functional shortcomings of yield in the current season.

[0107] YSSQI is built on yield-sensitive indicators, SHAP response thresholds, and yield-oriented weights, which can significantly improve the explanatory power of crop yield and more effectively connect tillage management, key soil functions, and crop yield formation processes.

[0108] Therefore, this invention does not simply replace the traditional SQI, but adds a yield function diagnostic layer to the traditional SQI background quality assessment, so that the farmland quality assessment results can simultaneously serve background quality classification and seasonal yield increase management.

[0109] Table 5 Comparison of the output explanatory power of traditional SQI and YSSQI Note: Traditional SQI is mainly used for background quality assessment; YSSQI uses yield response as the calibration target and is more suitable for identifying and managing yield limiting factors in the current season.

[0110] The recommended rules for the diagnosis and management of yield-limiting factors are as follows.

[0111] Based on the response type and bottleneck contribution value of each output-sensitive indicator, the system outputs management recommendation rules. For positive threshold-type indicators, when the measured value is below the lower threshold, the system suggests taking supplementary, enhancement, or improvement measures; when the measured value reaches the response platform, the system suggests maintaining the current level to avoid excessive investment.

[0112] For high-value restricted indicators, when the measured value exceeds the upper limit threshold, the system suggests taking measures to reduce, buffer, or neutralize; when the measured value is in an appropriate or low-restriction range, the system suggests maintaining it.

[0113] For appropriate range indicators, when the measured value is below the lower threshold, the system suggests supplementing or enhancing it; when the measured value is above the upper threshold, the system suggests controlling the amount or reducing the risk; when the measured value is within the appropriate range, the system suggests maintaining it.

[0114] In the implementation of cold-region paddy soil, if the pH is in the high-limit range, it is recommended to regulate the soil acid-base environment; if NO3-N is below the response threshold, it is recommended to optimize nitrogen supply; if TD is below the response threshold, it is recommended to improve the topsoil thickness or deep loosening and plowing; if Olsen-P or Urease deviates from the suitable range, it is recommended to optimize phosphorus supply or nitrogen conversion environment respectively.

[0115] The system can prioritize the management of the 1-3 indicators with the highest contribution value to the weakest link and form a management list at the plot level.

[0116] Table 6. Types of Indicator Shortcomings and Recommended Management Rules Note: L is the lower threshold, O is the optimal response point, U is the upper threshold; D i The contribution value for the weakest link in a single indicator. The management recommendations are sample rules and can be adjusted according to regional production systems and management objectives.

[0117] The present invention also provides an apparatus comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the method described in any of the foregoing embodiments. This apparatus is capable of constructing a yield limiting factor based on a cropland quality index calibrated with yield response.

[0118] Figure 5 This diagram illustrates the structure of the device, electronic equipment, and storage medium, showing the connections between the data acquisition module, traditional index comparison module, model training module, model interpretation module, threshold identification module, index calculation module, bottleneck diagnosis module, and result output module, as well as the implementation methods of the electronic equipment and storage medium.

[0119] The device of the present invention includes a data acquisition module, a data preprocessing module, a traditional index comparison module, a model training module, a model interpretation module, a threshold identification module, an index calculation module, a bottleneck diagnosis module, and a result output module.

[0120] The data acquisition module is used to acquire soil index data and crop yield data from target farmland sample points; the data preprocessing module is used for outlier identification, missing value handling, standardization, and collinearity diagnosis; the traditional index comparison module is used to construct the traditional SQI and compare it with the YSSQI; the model training module is used to build a machine learning yield prediction model; the model interpretation module is used to perform SHAP analysis and permutation importance analysis; the threshold identification module is used to determine L, O, and U; the index calculation module is used to calculate the YSSQI; the bottleneck diagnosis module is used to calculate the bottleneck contribution value of individual indicators; and the result output module is used to output the yield support level, the ranking of limiting factors, and the management recommendation results.

[0121] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the method described in any of the foregoing embodiments.

[0122] This invention is not limited to cold-region paddy soil, but can also be applied to dryland crops, greenhouse vegetables, corn in black soil areas, paddy soil or other farmland systems.

[0123] This invention is not limited to the Gradient Boosting model; it may also employ random forest, extreme gradient boosting, support vector regression, neural networks, or other models capable of establishing the relationship between soil indicators and crop yield.

[0124] This invention is not limited to the SHAP method; alternatively, permutation importance, ALE curves, PDP curves, or other interpretable machine learning methods can be used to identify the contribution direction of indicators and response thresholds.

[0125] This invention is not limited to the listed soil indicators, and remote sensing indicators, topographic indicators, moisture indicators, soil electrical conductivity, near-ground sensor indicators or agricultural machinery operation parameters can be added according to the target crop and management objectives.

[0126] The embodiments described above are merely preferred embodiments of the present invention. The terms "in one embodiment," "in another embodiment," "in yet another embodiment," or "in still another embodiment" used in this specification all refer to one or more of the same or different embodiments according to this disclosure. Ordinary variations and substitutions made by those skilled in the art within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a yield-based arable land quality index, comprising: S1. Obtain farmland sample data, including soil physical, chemical and biological indicators and corresponding crop yield data, to form a sample-level soil indicator-crop yield paired dataset. S2, data preprocessing of farmland sample data; S3, Construct and train machine learning yield prediction models: Using crop yield as the response variable and candidate soil indicators as explanatory variables, train multiple machine learning models and select the target yield prediction model based on the training set, test set and cross-validation results. S4. Interpretability analysis of the yield prediction model: SHAP analysis and substitution importance analysis are used to quantify the contribution direction and intensity of each soil index to yield prediction, and yield-sensitive soil indexes are screened by combining the mean absolute SHAP value. S5, Identify response thresholds: Plot SHAP dependency curves or cumulative local effect curves for yield-sensitive soil indicators; S6, Constructing Response Scoring Functions: Based on the shape of the response curve, the yield-sensitive indicators are divided into positive threshold type, high-value restriction type, and suitable range type, and response scoring functions are constructed for each type. S7, Calculate the yield-supported arable land quality index YSSQI: The yield response scores of each yield-sensitive indicator are weighted and integrated with the SHAP weights to obtain the yield-supported arable land quality index.

2. The method for constructing a yield-based arable land quality index according to claim 1, characterized in that, Step S2 includes: The sample-level soil index-crop yield paired data are subjected to outlier identification, missing value processing, unit unification, standardization or normalization processing, and collinearity diagnosis is performed between candidate soil indices. For indicators with redundant or highly correlated functions, representative indicators are retained by combining soil science significance, crop yield formation process and model stability to form a candidate indicator dataset.

3. The method for constructing a yield-based arable land quality index according to claim 2, characterized in that, In step S3, the model performance evaluation index includes the coefficient of determination R. 2 One or more of the following: root mean square error (RMSE), mean absolute error (MAE), residual prediction bias (RPD), and interquartile range performance ratio (RPIQ).

4. The method for constructing a yield-based arable land quality index according to claim 3, characterized in that, Step S4 includes: S4.1, Identify candidate indicators; S4.2, Data partitioning: Divide the data into training set and test set; S4.3, train multiple models; S4.4 Performance Evaluation: Model performance is evaluated using the training set, test set, and cross-validation. S4.5, Determine the target model: Select the model with the best prediction performance; S4.6: SHAP: After determining the target model, use the SHAP summary plot, mean absolute SHAP value, and permutation importance to evaluate the contribution of candidate indicators to yield forecasting; S4.7 Determine core indicators: Indicators with high contribution intensity and clear soil science significance are determined as yield-sensitive soil indicators.

5. The method for constructing a yield-based arable land quality index according to claim 4, characterized in that, In step S4.3, the trained model has a high R-value. 2 Low RPD and RPIQ; low RMSE and MAE; stable cross-validation; model that can be interpreted using SHAP or permutation importance.

6. The method for constructing a yield-based arable land quality index according to claim 5, characterized in that, Step S5 also includes: The intersection of the response curve with the zero contribution baseline and the peak point of the positive contribution are determined to obtain the lower threshold L, the optimal response point O, and the upper threshold U.

7. The method for constructing a yield-based arable land quality index according to claim 6, characterized in that, Step S6 includes: The position where the SHAP value equals 0 serves as the zero contribution baseline point for the indicator to change from negative to positive or from positive to negative contribution; the lower threshold L and the upper threshold U can be determined by the intersection of the SHAP response curve and the zero contribution baseline; the optimal response point O can be determined by the position in the SHAP response curve where the positive contribution reaches its peak or the output contribution is the highest.

8. The method for constructing a yield-based arable land quality index according to claim 7, characterized in that, Step S7 includes: 1) Based on the yield response scoring function established in step S6, convert the measured values ​​of each yield-sensitive soil index into yield response scores S in the range of 0-1. i Y Among them, S i Y The higher the value, the stronger the supporting effect of the current value of the indicator on the yield of the target crop. 2) Based on the model interpretability analysis results obtained in step S4, calculate the average absolute SHAP value of each yield-sensitive soil index, and normalize it to obtain the yield-oriented weight w. i Y ; 3) Calculate the yield-supporting arable land quality index YSSQI, YSSQI = , where n is the number of yield-sensitive soil indicators included in the index calculation, and = 1, the value of YSSQI ranges from 0 to 1. The higher the value, the stronger the comprehensive support capacity of the current soil function status of the target farmland for the formation of the target crop yield.

9. A device for constructing a yield-based arable land quality index, characterized in that, include: A processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1-8.

10. A computer storage medium, characterized in that, The medium stores a computer-executable program that, when run, implements the method as described in any one of claims 1-8.

Images