A machine learning based corn mechanization harvesting decision method

Abstract

The present application relates to the technical field of intelligent agricultural machinery operation decision-making, and particularly relates to a corn mechanized harvesting decision-making method based on machine learning, which comprises the following steps: acquiring observed data of grain moisture content of a target variety, genotype data and meteorological data aligned with the observation time; establishing a grain moisture content dynamic prediction model based on the above data, outputting the grain moisture content probability distribution at a future time, and determining the variety dehydration type from the probability distribution; inputting the future moisture content probability distribution, the dehydration type and the machine operation parameters into a mechanized harvesting operation loss model to calculate the expected value and quantile of the operation loss of ear collection and grain collection, respectively, and determining the harvesting time window and the corresponding harvesting method on the candidate time set by using dynamic programming to output the harvesting decision-making result. The present application can consider the average loss and risk under prediction uncertainty, and reduce the comprehensive loss such as breakage, entrainment, cleaning and drying energy consumption.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural machinery operation decision-making technology, and in particular to a machine learning-based decision-making method for mechanized corn harvesting. Background Technology

[0002] The corn harvesting process directly determines the effectiveness of kernel crushing, inclusion, and cleaning, as well as subsequent drying energy consumption, making it a critical cost and quality control point for large-scale planting. Due to significant differences in the dehydration process of different varieties, frequent fluctuations in field weather, and the uncertainty of moisture content, production often relies on experience or single measurements to set harvesting timing and methods. This can easily lead to problems such as "early harvesting results in high breakage rates, late harvesting increases the risk of ear drop and mold, and moisture content fluctuations cause uncontrolled drying costs." At the same time, different machinery parameters and operating methods have different sensitivities to losses, and there is a lack of a quantitative decision-making path that connects "future moisture content uncertainty—variety dehydration characteristics—operational losses," making it difficult to stably replicate harvesting plans and resulting in insufficient adaptability across plots / years. Summary of the Invention

[0003] This invention provides a machine learning-based decision-making method for mechanized corn harvesting, which addresses the problem of how to quantitatively evaluate and optimize mechanized ear / grain harvesting methods and harvesting time windows to reduce overall operational losses under conditions of uncertain future moisture content evolution and significant differences in variety dehydration.

[0004] In a first aspect, this application provides a machine learning-based decision-making method for mechanized corn harvesting, comprising the following steps: Acquire observational data on kernel moisture content of the target maize variety, genotype data of the target maize variety, and meteorological data time-aligned with the kernel moisture content observation data; A dynamic prediction model for grain moisture content is established based on grain moisture content observation data, genotype data, and meteorological data. The model outputs the probability distribution of grain moisture content at future times and determines the dehydration type of the target maize variety based on the probability distribution of grain moisture content at future times. The mechanized harvesting operation loss model is input based on the probability distribution of grain moisture content at future time and the dehydration type of the target maize variety. The operation loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method are calculated respectively. Based on the operation loss values, the mechanized ear harvesting method or the mechanized grain harvesting method and the harvesting time window are determined, and the mechanized harvesting decision results are output.

[0005] In one possible implementation, after acquiring the grain moisture content observation data of the target maize variety, the genotype data of the target maize variety, and the meteorological data time-aligned with the grain moisture content observation data, the grain moisture content observation data is aligned to the developmental stage, and environmental condition characteristic data corresponding to the grain moisture content observation data is generated from the meteorological data based on the developmental stage alignment results.

[0006] In one possible implementation, a genotype representation vector is generated based on the genotype data of the target maize variety, and the genotype representation vector and environmental condition characteristic data are used together as input to a dynamic prediction model of grain moisture content.

[0007] In one possible implementation, the genotype data of the target maize variety is single nucleotide polymorphism (SNP) marker data, and generating a genotype characterization vector based on the genotype data of the target maize variety includes performing principal component analysis on the SNP marker data to obtain the genotype characterization vector.

[0008] In one possible implementation, the dynamic prediction model for grain moisture content is a state-space model. The state-space model represents grain moisture content as a slow-changing component state and a fast-changing component state, and outputs the probability distribution of grain moisture content at future times based on filtering estimation.

[0009] In one possible implementation, the rapidly changing moisture content component is incrementally updated when a precipitation event or a meteorological process with a continuous increase in relative humidity occurs, as determined by meteorological data, and is attenuatedly updated based on the dry condition characteristics generated from meteorological data when there is no precipitation event and the relative humidity does not increase continuously.

[0010] In one possible implementation, determining the dehydration type of a target maize variety based on the probability distribution of grain moisture content at future times includes calculating the rate of change of moisture content characteristics based on the expected value sequence of the probability distribution of grain moisture content at future times, and calculating the variance characteristics of the probability distribution of moisture content based on the probability distribution of grain moisture content at future times. The rate of change of moisture content characteristics and the variance characteristics of the probability distribution of moisture content are then input into a dehydration type classification model to output the dehydration type.

[0011] In one possible implementation, the inputs to the mechanized harvesting operation loss model include the probability distribution of grain moisture content at future time points, the dehydration type of the target maize variety, the mechanized ear harvesting method or the mechanized grain harvesting method, and the machine operation parameters. The output of the mechanized harvesting operation loss model is the operation loss value, which includes grain breakage loss, entrainment loss, cleaning loss, and drying energy consumption loss.

[0012] In one possible implementation, calculating the operational loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method respectively includes generating a set of moisture content scenarios by sampling the probability distribution of grain moisture content at future time moments, and calculating the expected value and quantile of the operational loss value corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method respectively based on the set of moisture content scenarios.

[0013] In one possible implementation, determining the mechanized ear harvesting method or mechanized grain harvesting method and the harvest time window based on the operation loss value includes determining a set of candidate harvest times, calculating the expected value and quantile of the operation loss value corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method respectively on the set of candidate harvest times, and determining the harvest time window and the mechanized ear harvesting method or mechanized grain harvesting method corresponding to the harvest time window based on dynamic programming.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By jointly modeling grain moisture content observation data, genotype data, and aligned meteorological data, and outputting future moisture content in the form of a probability distribution, we achieved explicit quantification of moisture content uncertainty and controllable risk prediction results. By extracting change rate features and variance features based on the probability distribution of future moisture content and inputting them into a classification model, we achieved interpretable identification of variety dehydration types, reducing the bias in method selection caused by relying solely on experience. By introducing the probability distribution of future moisture content and dehydration type into the operation loss model and conducting expected value and quantile assessments for ear harvest and grain harvest respectively, we achieved method comparison that takes into account both average loss and extreme risk. By using dynamic programming to determine the harvest time window on the candidate harvest time set, we achieved integrated optimization of harvesting methods and time arrangements, reducing overall losses and improving the stability of the operation plan. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the method of the present invention; Figure 2 This is a graph showing the input data and characterization results used to establish a dynamic prediction model for grain moisture content in this embodiment of the invention. Figure 3 This is a diagram showing the decision results of mechanized harvesting methods and harvesting time windows based on the probability distribution of grain moisture content at future times in an embodiment of the present invention. Detailed Implementation

[0016] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0017] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0018] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0019] Mechanized corn harvesting typically includes two operational paths: mechanized ear harvesting and mechanized grain harvesting. The former focuses on harvesting the ears and completing threshing and drying in subsequent stages, suitable for scenarios with high moisture content or large fluctuations in field conditions. The latter directly completes ear picking, threshing, and cleaning in the field, resulting in a shorter and more efficient operational chain, but it is more sensitive to grain moisture content, variety breakage resistance, and machinery parameter matching. In actual production, the choice of harvesting method and timing often needs to simultaneously consider operational efficiency, grain quality, and overall cost, and dynamically changes with variety dehydration characteristics, weather conditions, and machinery status. Based on this, this invention proposes an intelligent decision-making method for mechanized harvesting processes. By predicting and quantifying the uncertainty of future moisture content changes, and combining variety dehydration type and operational loss assessment, it outputs executable harvesting method and harvesting time window decisions.

[0020] like Figure 1 As shown, a machine learning-based decision-making method for mechanized corn harvesting includes the following steps: Acquire observational data on kernel moisture content of the target maize variety, genotype data of the target maize variety, and meteorological data time-aligned with the kernel moisture content observation data; In one embodiment, a data collection list for the target maize variety is first established. This list includes at least the variety number, test site number, planting season identifier, and developmental stage identifier. Grain moisture content observation data is obtained through field sampling and laboratory testing. Field sampling is stratified according to test site and developmental stage, with each sampling recording the sampling time, sampling location, ear position of the grain source, and sample number. Laboratory testing uses the drying method or a near-infrared moisture meter to measure grain moisture content, and each observation data point is recorded with the sample number, testing method, and test result. Genotype data for the target maize variety is obtained from a breeding material database or through molecular marker detection. Genotype data includes at least the variety number and single nucleotide polymorphism (SNP) marker information, with deletion sites undergoing consistency coding. Meteorological data aligned with the grain moisture content observation data were obtained from automatic weather stations at the test sites or from public meteorological services. The meteorological data included at least temperature, relative humidity, precipitation, wind speed, and solar radiation. The meteorological data were interpolated or summarized according to the sampling time so that each grain moisture content observation data corresponded to a meteorological characteristic record, thereby providing consistent data input for the subsequent dynamic prediction model of grain moisture content.

[0021] After acquiring the grain moisture content observation data of the target maize variety, the genotype data of the target maize variety, and the meteorological data time-aligned with the grain moisture content observation data, the grain moisture content observation data is aligned with the developmental stage, and environmental condition characteristic data corresponding to the grain moisture content observation data are generated from the meteorological data based on the developmental stage alignment results.

[0022] In one embodiment, after acquiring the grain moisture content observation data, genotype data, and meteorological data time-aligned with the grain moisture content observation data for the target maize variety, developmental stage alignment processing is performed on the grain moisture content observation data to eliminate differences in developmental progress caused by different sowing dates and temperature conditions, enabling data from different years and different experimental sites to be compared and modeled on the same developmental scale. Developmental stage alignment prioritizes field phenological records, which must include at least the observation dates of silking and physiological maturity. Each grain moisture content observation data point is associated with the phenological record of the same experimental site. When phenological records are missing, the cumulative effective accumulated temperature is calculated using meteorological data to estimate the silking and physiological maturity dates. The expression for calculating the cumulative effective accumulated temperature is: in, To accumulate effective temperature, For cumulative days, For the first The average daily temperature of the day, The base temperature is used. Subsequently, the sampling date of each grain moisture content observation data is mapped to the developmental interval from silking stage to physiological maturity, and developmental stage labels are generated. These labels can be obtained using developmental interval division rules, which are determined and solidified into a configuration file based on the grain moisture content change trends and phenological records of historical samples, ensuring consistent execution across different experimental sites. Based on the developmental stage alignment results, environmental condition characteristic data corresponding to the grain moisture content observation data are generated from meteorological data. This generation includes determining the meteorological statistical window corresponding to the developmental stage label, and calculating the mean, extreme values, and cumulative values ​​for temperature, relative humidity, precipitation, wind speed, and solar radiation within the meteorological statistical window, while simultaneously recording missing meteorological data markers. When missing meteorological data exists, interpolation using adjacent dates at the same experimental site or data from neighboring meteorological stations on the same day is used to supplement the missing data, and the source marker for the supplementation is retained. Through the above processing, each grain moisture content observation data can obtain consistent developmental stage labels and environmental condition characteristics, thereby providing a directly input, cross-year comparable data representation for the subsequent grain moisture content dynamic prediction model, and reducing the impact of developmental process misalignment on prediction results.

[0023] Genotype representation vectors are generated based on the genotype data of the target maize variety, and these genotype representation vectors, together with environmental condition characteristic data, are used as inputs to a dynamic prediction model for grain moisture content.

[0024] In one embodiment, the process of generating genotype representation vectors based on genotype data of a target maize variety includes data unification, quality control, numerical encoding, and dimensionality reduction mapping. First, the single nucleotide polymorphism (SNP) marker data corresponding to the target maize variety is read, and the variety number is checked for consistency with the variety number in the grain moisture content observation data, eliminating records with inconsistent numbers. Then, quality control processing is performed on the SNP marker data. This quality control process includes eliminating marker sites with detection rates below a preset threshold, eliminating marker sites with excessively low allele frequencies, and imputing missing genotypes. The imputation uses the expected genotype code calculated from the allele frequency of the same marker site in the training samples. After quality control, the genotype of each marker site is numerically encoded according to allele counts, and the encoding results of each marker site are standardized to obtain a variety-level genotype numerical feature matrix. Principal component analysis is performed based on the genotype numerical feature matrix, and a preset number of principal component scores are selected as the genotype representation vector of the target maize variety. The genotype representation vector is then bound to the variety number and stored for subsequent retrieval. Furthermore, the environmental condition characteristic data are aligned according to the sample numbers of the grain moisture content observation data to obtain sample-level environmental condition characteristic sequences. When constructing the input for the dynamic prediction model of grain moisture content, the genotype representation vector of the corresponding variety is read for each sample-level environmental condition characteristic sequence. The genotype representation vector is then fused with the environmental condition characteristic data as dynamic features, using the same genotype representation vector concatenated at each time step of the environmental condition characteristic sequence to form the input feature sequence for model training and inference. Through the above processing, the dynamic prediction model of grain moisture content can learn the environmentally driven moisture content change pattern while using genotype differences to distinguish the dehydration potential of different varieties, thereby improving the predictive consistency and transferability under cross-experimental site and cross-year conditions.

[0025] The genotype data of the target maize variety is single nucleotide polymorphism (SNP) marker data. Genotype characterization vectors are generated based on the genotype data of the target maize variety by performing principal component analysis on the SNP marker data to obtain the genotype characterization vectors.

[0026] In one embodiment, the genotype data of the target maize variety is represented using single nucleotide polymorphism (SNP) marker data. The SNP marker data uses the variety number as an index to record the allele status of each marker locus. To ensure that test results from different batches can be directly used in the same dynamic prediction model for grain moisture content, firstly, site consistency processing is performed on the SNP marker data. This includes standardizing the reference allele direction, removing sites where the allele direction cannot be determined, and verifying the consistency between the variety number and the variety number in the grain moisture content observation data. Subsequently, quality control processing is performed on the SNP marker data. This quality control processing includes removing sites with low detection rates and sites with excessively low allele frequencies, and imputing missing genotypes. The imputation value for missing genotypes is the average allele count of the same marker locus in the training samples, thereby avoiding dimensionality reduction bias caused by missing genotypes.

[0027] After quality control is completed, the genotype of each marker locus is numerically encoded according to allele count, resulting in a genotype numerical feature vector for each variety. To eliminate the influence of dimensional differences between different loci on the principal component analysis results, the genotype numerical feature vector is standardized by locus, and the standardized genotype numerical feature vector is projected onto the principal component space to obtain the genotype representation vector. The core calculation expression for this projection is: in, This is a genotype representation vector. This is the standardized numerical feature vector of the genotype. This is the projection matrix obtained from principal component analysis. The projection matrix is ​​obtained by calculating the covariance matrix of the genotype numerical feature matrix of the training samples and performing eigenvalue decomposition. The eigenvectors corresponding to the principal components whose cumulative explained variance meets preset requirements are selected to form the projection matrix. Genotype representation vectors are bound to variety numbers for storage and, when inputting into the dynamic prediction model for grain moisture content, are concatenated with environmental condition feature data corresponding to the same variety at each developmental stage. This allows the dynamic prediction model for grain moisture content to simultaneously utilize genotype differences and environmental driving information, improving the transferability of grain moisture content changes under different experimental sites and years, and providing a stable variety-level feature representation for subsequent dehydration type identification.

[0028] A dynamic prediction model for grain moisture content is established based on grain moisture content observation data, genotype data, and meteorological data. The model outputs the probability distribution of grain moisture content at future times and determines the dehydration type of the target maize variety based on the probability distribution of grain moisture content at future times. In one embodiment, samples aligned to developmental stages are organized into a time series according to the developmental stage sequence. For each grain moisture content observation record, model input features are constructed. These input features include a genotype representation vector generated from genotype data, environmental condition feature data generated from meteorological data, and the grain moisture content of the previous developmental stage. The grain moisture content of the next developmental stage is used as a supervised label to train a probabilistic regression model. The probabilistic regression model outputs multiple quantile prediction values ​​to represent the probability distribution of grain moisture content at future times. During inference, the multiple quantile prediction values ​​are generated sequentially according to developmental stages, and the median prediction value is used as the input for the grain moisture content of the previous developmental stage in the next developmental stage. Based on the median sequence of the grain moisture content probability distribution, early dehydration rate features and late dehydration rate features are calculated and input into a dehydration type classification model to output the dehydration type of the target maize variety.

[0029] The dynamic prediction model for grain moisture content is a state-space model. The state-space model represents grain moisture content as a slow-changing component state and a fast-changing component state, and outputs the probability distribution of grain moisture content at future times based on filtering estimation.

[0030] In one embodiment, the dynamic prediction model for grain moisture content is implemented using a state-space model to simultaneously characterize the long-term dehydration trend and short-term moisture reabsorption disturbances of the target maize variety within the same framework. The state-space model decomposes the observed grain moisture content at each developmental stage into a slowly varying moisture content component and a rapidly varying moisture content component. Through filtering estimation, it updates both types of states in real time as grain moisture content observation data is received at each stage, thereby outputting the probability distribution of grain moisture content at future time points. To avoid conceptual ambiguity, the slowly varying moisture content component is used to characterize the continuous dehydration baseline dominated by varietal genetic characteristics and developmental processes, while the rapidly varying moisture content component is used to characterize the short-term moisture reabsorption and time-decreasing disturbances caused by precipitation events or meteorological processes of continuously rising relative humidity. The observation relationship of the state-space model is expressed by the following core expression: in, For the first Observed values ​​of grain moisture content during the development stage For the first The slow-change component state of water content during the development stage. For the first The water content changes rapidly during the developmental stage. This represents the observation error.

[0031] In model construction, the sample sequence aligned with developmental stages is used as the input sequence, and genotype representation vectors and environmental condition feature data are used as exogenous inputs to drive the evolution of the slow-varying and fast-varying component states. The prediction and update of the slow-varying component state is achieved using a dehydration update sub-model. The dehydration update sub-model takes genotype representation vectors and environmental condition feature data as inputs, outputs the dehydration amplitude of adjacent developmental stages, and uses the dehydration amplitude to update the slow-varying component state of moisture content. The dehydration update sub-model can be trained using a regression model. The training samples consist of historical grain moisture content observation data and corresponding genotype representation vectors and environmental condition feature data. The training objective is the difference between the grain moisture content observation values ​​of adjacent developmental stages. The prediction update of the rapid change component state is implemented using the backwater update sub-model. The backwater update sub-model identifies precipitation events or meteorological processes with continuously rising relative humidity based on environmental condition characteristic data. When the meteorological process is identified, the rapid change component state of water content is updated incrementally, and when the meteorological process is not identified, the rapid change component state of water content is updated attenuatedly. The magnitude of the incremental update is determined by the cumulative precipitation and the relative humidity level, and the attenuation coefficient of the attenuation update is determined by the wind speed and solar radiation, so that the influence of backwater gradually fades in the subsequent development stage.

[0032] The filtering estimation is performed recursively, specifically comprising two parts: prediction update and observation correction. In the prediction update stage, prior values ​​for the slow-varying and rapid-varying moisture content components are obtained based on the dehydration update sub-model and the rehydration update sub-model, respectively, with uncertainty propagated simultaneously. In the observation correction stage, the observed grain moisture content is compared with the previously predicted grain moisture content. Based on the statistical characteristics of the observation error, the prior values ​​for the two types of states are weighted and corrected, thus obtaining the posterior estimates for the slow-varying and rapid-varying moisture content components. The statistical characteristics of the observation error are obtained from repeated measurements or historical residual statistics and are written into the model configuration file as fixed configuration parameters for the filtering estimation. Through the above filtering estimation, the dehydration baseline can be stably tracked and rehydration disturbances can be separated even under conditions of measurement fluctuations, sudden meteorological changes, and differences between different test sites.

[0033] The probability distribution of grain moisture content at future time moments is obtained by propagating the uncertainty of the posterior state. Specifically, given the environmental condition characteristic data of the future development stage, the slow-changing component state and the fast-changing component state of moisture content are recursively calculated according to the prediction update rule to obtain the predicted mean and predicted variance of grain moisture content at future time moments, and the probability distribution of grain moisture content at future time moments is output accordingly. This probability distribution is used for subsequent dehydration type discrimination and loss model calculation of mechanized harvesting operations, enabling mechanized harvesting decisions to simultaneously consider long-term dehydration type differences and short-term re-moistening risks, thereby reducing misjudgments of harvesting methods caused by relying solely on single-point moisture content measurements.

[0034] The rapid change component of moisture content is incrementally updated when a precipitation event or a meteorological process with a continuous increase in relative humidity occurs as determined by meteorological data, and is attenuatedly updated based on the dry condition characteristics generated by meteorological data when there is no precipitation event and the relative humidity does not increase continuously.

[0035] In one embodiment, the rapidly varying moisture content component state is used to characterize the short-term damping effects caused by precipitation and air humidity, and their dissipation over time. Updates to the rapidly varying moisture content component state are based on meteorological data, identifying wet event markers at each developmental stage, and performing incremental or decay updates accordingly.

[0036] Specifically, within the meteorological statistics window corresponding to each development stage, precipitation and relative humidity sequences are read, and a wet event flag is generated according to preset judgment rules. When precipitation meets the precipitation event judgment rules, the wet event flag is set to true; when relative humidity shows a continuous increase and the duration meets preset requirements, the wet event flag is set to true; when neither the precipitation event judgment rules nor the continuous increase in relative humidity judgment rules are met, the wet event flag is set to false. The judgment rules are determined by the statistical analysis of the humidification residuals of historical samples and written into the configuration file, thereby ensuring that consistent rules can be applied to different test points.

[0037] When the wet event indicator is true, the wet intensity is calculated based on the cumulative precipitation and relative humidity level within the meteorological statistical window, and the wet intensity is mapped to the increment of the rapid change component of moisture content. When the wet event indicator is false, dry condition characteristics are generated based on the temperature level, wind speed level, solar radiation level, and relative humidity level within the meteorological statistical window, and the dry condition characteristics are mapped to the attenuation intensity of the rapid change component of moisture content, so that the influence of dampness gradually diminishes with the development stage. The update expression for the rapid change component of moisture content is: in, For the first The water content changes rapidly during the developmental stage. As a marker of a wet event, This refers to the cumulative precipitation within the meteorological statistics window. The relative humidity level within the meteorological statistics window. The dry conditions are characterized by temperature level, wind speed level, solar radiation level, and relative humidity level. The weighting coefficient for precipitation in incremental updates. The relative humidity is the weighting factor for incremental updates. This represents the weighting coefficient for the attenuation update based on dry conditions. The dry condition characteristics are generated by normalizing and weighting the temperature, wind speed, and solar radiation levels, and then adding a normalized relative humidity level as a subtraction term to ensure faster attenuation when dry condition characteristics increase. Through this method, the rapid change component of moisture content can promptly reflect moisture re-entry disturbances when humid conditions occur and automatically attenuate under dry conditions, thereby reducing the interference of short-term meteorological fluctuations on dehydration type identification and mechanized harvesting decisions.

[0038] Determining the dehydration type of a target maize variety based on the probability distribution of grain moisture content at future times involves calculating the rate of change of moisture content based on the expected value sequence of the probability distribution of grain moisture content at future times, and calculating the variance of the probability distribution of moisture content based on the probability distribution of grain moisture content at future times. The rate of change of moisture content and the variance of the probability distribution of moisture content are then input into the dehydration type classification model to output the dehydration type.

[0039] In one embodiment, the process of determining the dehydration type of a target maize variety based on the probability distribution of grain moisture content at future times, aiming at the stability and interpretability required for mechanized harvesting decisions, first converts the probability distribution of grain moisture content at future times output by the dynamic prediction model into a computable sequence of expected values ​​and a sequence of dispersion. Then, the rate of change of moisture content is extracted from the expected value sequence, and the rate of change of moisture content and dispersion features are input into the dehydration type classification model to output the dehydration type. Specifically, the probability distribution of grain moisture content at future times is represented as several moisture content values ​​and their corresponding probability weights, and the expected value and variance at each future time are calculated. The calculation expression is as follows:

[0040] in, For the first Expected grain moisture content at future time. For the first Variance of grain moisture content at future time points For the first The first moment in the future The moisture content of each grain is taken as the value. For the first The first moment in the future The probability weight corresponding to each grain moisture content value The number of values ​​is represented discretely. A moisture content change rate feature is calculated based on the expected value sequence of grain moisture content. This feature includes at least the difference in expected values ​​between adjacent future times and the average difference across multiple future times, used to characterize the dehydration progress speed and stage-wise changes. A moisture content probability distribution variance feature is calculated based on the grain moisture content variance sequence. This feature includes at least the mean variance and the variance variation amplitude, used to characterize prediction uncertainty and the risk of moisture reabsorption disturbance. The dehydration type classification model is trained using supervised learning. The training samples consist of the moisture content change rate feature and the moisture content probability distribution variance feature calculated from the probability distribution of grain moisture content at future times corresponding to historical grain moisture content observation data. The training labels are predefined dehydration types, including rapid dehydration, slow dehydration, and late-stage accelerated dehydration, determined by the stage-wise slope pattern of historical dehydration curves. During inference, the moisture content change rate characteristics and moisture content probability distribution variance characteristics of the target maize variety are input into the dehydration type classification model, and the dehydration type of the target maize variety is output. This provides variety-side constraint information for subsequent operation loss assessment and harvest time window determination for mechanized ear harvesting or mechanized grain harvesting.

[0041] The mechanized harvesting operation loss model is input based on the probability distribution of grain moisture content at future time and the dehydration type of the target maize variety. The operation loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method are calculated respectively. Based on the operation loss values, the mechanized ear harvesting method or the mechanized grain harvesting method and the harvesting time window are determined, and the mechanized harvesting decision results are output.

[0042] In one embodiment, the probability distribution of future grain moisture content output from the dynamic prediction model of grain moisture content and the dehydration type of the target maize variety are jointly input into the mechanized harvesting operation loss model. First, a set of candidate harvest times is constructed, and for each candidate harvest time, a set of moisture content scenarios is generated from the probability distribution of future grain moisture content, with each moisture content scenario corresponding to a probability weight. Then, operation loss values ​​are calculated for both mechanized ear harvesting and mechanized grain harvesting methods. The operation loss values ​​consist of grain breakage loss, entrainment loss, cleaning loss, and drying energy consumption loss. Each loss item is jointly determined by the moisture content scenario, machine operation parameters, and the dehydration type of the target maize variety. The dehydration type is used to select a moisture content response curve that matches the variety, reflecting the differences in sensitivity to breakage and drying for different dehydration types. For the same candidate harvest time, the operation loss values ​​for the two harvesting methods on the moisture content scenario set are summed using probability weighting to obtain the operation loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method at that candidate harvest time. Simultaneously, risk indicators related to prediction uncertainty are output. Based on the operational loss value of each candidate harvest time, the harvesting method with the smaller operational loss value and the risk index meeting the preset requirements is determined, and the continuous candidate harvest time interval that meets the requirements is determined as the harvest time window. Finally, the mechanized harvesting decision result is output, which includes at least the recommended harvesting method and the harvest time window.

[0043] The inputs to the mechanized harvesting operation loss model include the probability distribution of grain moisture content at future time points, the dehydration type of the target maize variety, the mechanized ear harvesting method or the mechanized grain harvesting method, and the machine operation parameters. The output of the mechanized harvesting operation loss model is the operation loss value, which includes grain breakage loss, entrainment loss, cleaning loss, and drying energy consumption loss.

[0044] In one embodiment, a mechanized harvesting operation loss model is used to map the probability distribution of grain moisture content at future times, the dehydration type of the target maize variety, the mechanized ear harvesting method or mechanized grain harvesting method, and machine operation parameters into operation loss values, so as to enable comparable quantitative evaluation between different harvesting methods and different candidate harvesting times. The input to the mechanized harvesting operation loss model is organized as a set of scenarios for each candidate harvesting time. The scenario set is discretized from the probability distribution of grain moisture content at future times. Each scenario includes a moisture content value and probability weight, and is bound to the machine operation parameters for the same candidate harvesting time. Machine operation parameters include travel speed, header height, feed rate, threshing drum speed, concave plate gap, cleaning fan speed, and sieve opening. Mechanized ear harvesting methods also include ear picking roller speed and peeling gap; mechanized grain harvesting methods also include conveying parameters related to grain conveying and rewinding. The dehydration type of the target maize variety is used to select a set of moisture content response curves that match the variety, so that the estimated crushing sensitivity and drying load at the same moisture content conforms to variety differences.

[0045] For each scenario, the mechanized harvesting operation loss model calculates grain breakage loss, entrainment loss, cleaning loss, and drying energy consumption loss. Grain breakage loss is determined by the moisture content, dehydration type, and threshing equipment parameters. A breakage rate curve calibrated through bench and field comparative tests is used to map the combination of threshing drum speed and concave plate gap to an impact intensity level. The breakage rate is then read at the moisture content level and converted to the loss amount by the yield per unit area or feed rate. Entrainment loss is determined by the moisture content, dehydration type, and separation equipment parameters. An entrainment rate curve is used to map the feed rate and moisture content to a separation load, and the entrainment rate is corrected by the travel speed and crop flow rate. Cleaning loss is determined by the moisture content and cleaning equipment parameters. A cleaning loss curve is used to map the cleaning fan speed and sieve opening to a cleaning airflow level, and the cleaning loss rate is read at the moisture content level. The energy loss during drying is determined by the moisture content value and the dehydration type. The difference between the moisture content value and the target storage moisture content is mapped to the energy consumption per unit mass using the drying energy consumption curve. The energy loss is obtained by combining the expected yield at the candidate harvest time. The dehydration type is used to select the energy consumption correction curve when the risk of rehydration is higher or the dehydration is slower in the later stage.

[0046] For the same candidate harvest time and the same harvest method, the operational loss value is obtained by the probability-weighted result of each scenario, and its core calculation expression is: in, This is the value of the work loss. For the number of scenarios, For the first The probability weights of each scenario For the first Grain breakage loss in this scenario For the first The accompanying losses in this scenario For the first Cleaning losses in this scenario For the first The drying energy loss in this scenario This is the weighting coefficient for grain breakage loss. For the entrainment loss weighting coefficient, For the cleaning loss weighting coefficient, This refers to the weighting coefficient for drying energy consumption loss. The weighting coefficient is determined by the operational scenario constraints and fixed as a configuration item, ensuring consistent assessment of loss composition across different regions or operational objectives. Through this method, the mechanized harvesting operation loss model, while maintaining clear inputs and outputs, incorporates uncertain future moisture content into the loss assessment through scenario weighting. This allows subsequent harvesting method selection and harvesting time window determination to simultaneously consider the risk differences arising from average loss and moisture content fluctuations.

[0047] The calculation of the operational loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method includes sampling the probability distribution of grain moisture content at future time points to generate a set of moisture content scenarios, and calculating the expected value and quantile of the operational loss value corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method based on the set of moisture content scenarios.

[0048] In one embodiment, when calculating the operational loss values ​​corresponding to mechanized ear harvesting and mechanized grain harvesting, for each candidate harvest time, the probability distribution of future grain moisture content is first converted into a set of moisture content scenarios. The probability distribution of future grain moisture content is output by a dynamic prediction model of grain moisture content, and the probability distribution is represented by several moisture content values ​​and their probability weights. The generation of the moisture content scenario set adopts a stratified sampling rule, dividing the cumulative probability interval into several sub-intervals, and selecting the corresponding moisture content value in each sub-interval as a moisture content scenario, while using the probability quality of the sub-interval as the probability weight of the moisture content scenario; when the probability distribution is represented by discrete values, the cumulative probabilities are accumulated from small to large and systematic sampling is performed to obtain the set of moisture content scenarios, thereby ensuring that low-probability, high-risk moisture content scenarios are included in the evaluation.

[0049] Subsequently, based on the moisture content scenario set, the scenario operation loss values ​​for mechanized ear harvesting and mechanized grain harvesting were calculated separately. For each moisture content scenario, the moisture content value, the dehydration type of the target maize variety, the corresponding harvesting method, and the machine operation parameters were input into the mechanized harvesting operation loss model. The operation loss value under that moisture content scenario was output, forming a scenario operation loss value sequence for mechanized ear harvesting and a scenario operation loss value sequence for mechanized grain harvesting. Furthermore, the expected value and quantiles of the operation loss values ​​were calculated for both harvesting methods. The expression for calculating the expected value is: in, The expected value of the task loss. For the number of scenarios with moisture content, For the first The probability weights for each moisture content scenario For the first The operational loss values ​​corresponding to each moisture content scenario are calculated. Quantiles are determined using a probability-weighted cumulative method. The operational loss values ​​for each moisture content scenario are sorted from smallest to largest, and the probability weights are simultaneously rearranged. The cumulative probability weight is calculated, and the operational loss value whose cumulative probability weight is first not less than a preset quantile level is taken as the quantile. By simultaneously outputting the expected value and quantiles, the comparison of harvesting methods not only reflects the average loss level but also the upper bound of risk caused by moisture content uncertainty, providing a basis for the joint determination of subsequent harvesting time windows and harvesting methods.

[0050] Determining the mechanized ear harvesting method or mechanized grain harvesting method and the harvest time window based on the operation loss value includes determining the candidate harvest time set, calculating the expected value and quantile of the operation loss value corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method on the candidate harvest time set, and determining the harvest time window and the mechanized ear harvesting method or mechanized grain harvesting method corresponding to the harvest time window based on dynamic programming.

[0051] In one embodiment, when determining the mechanized ear harvesting method or mechanized grain harvesting method and the harvest time window based on the operation loss value, a candidate harvest time set is first determined. The candidate harvest time set is generated from future times corresponding to the probability distribution of grain moisture content at future times, and these future times are mapped to a calendar date sequence. Simultaneously, an operation accessibility flag is generated based on meteorological data. This flag is used to eliminate dates where rainfall or surface humidity is consistently high, preventing machinery from entering the field, thus ensuring that the candidate harvest time set meets the feasible operation conditions. Subsequently, the expected value and quantile of the operation loss value for both the mechanized ear harvesting method and the mechanized grain harvesting method are calculated on the candidate harvest time set. The calculation of the expected value and quantile is obtained using the probability weighting method of the aforementioned moisture content scenario set, and the expected value and quantile of the two harvesting methods at each candidate harvest time are written into the same evaluation sequence.

[0052] To jointly determine the harvest time window and harvesting method, the expected value and quantile of each candidate harvest time are integrated into a single-time comprehensive loss index. This comprehensive loss index simultaneously reflects average loss and uncertainty risk, and its core calculation expression is as follows: in, For the first The comprehensive loss index for each candidate harvest time. For the first The expected value of the job loss at each candidate harvest time. For the first quantiles of job loss values ​​at each candidate harvest time The expected value weighting coefficient, These are quantile weighting coefficients. These weighting coefficients are determined by operational objectives and written into the configuration file, enabling different regions or different operating entities to adopt a consistent risk appetite in decision-making.

[0053] Based on this, dynamic programming is used to search for continuous time windows on the candidate harvest time set to minimize the sum of comprehensive loss indicators within the time window, and simultaneously determine the harvest method corresponding to the time window. Dynamic programming uses the index of the candidate harvest time sequence as the time dimension and the time window length as the length dimension, defining the cumulative loss recursive relationship as follows: in, For the first The candidate harvest times are taken as the end of the time window, and the time window length is [value missing]. The cumulative loss value, For the first The comprehensive loss index for each candidate harvest time. The range of the time window length is limited by a preset minimum operating day and a preset maximum operating day. The preset minimum operating day ensures that the target workload can be completed within the time window, while the preset maximum operating day prevents the harvest timing from shifting due to an excessively wide time window. A set of cumulative loss value sequences is generated for both mechanized ear harvesting and mechanized grain harvesting methods. The combination with the smallest cumulative loss value among all cumulative loss values ​​that satisfy the length constraint is selected to obtain the harvest time window and the corresponding mechanized ear harvesting or mechanized grain harvesting method. The final output of the mechanized harvesting decision result includes at least the harvesting method, the start date of the harvest time window, and the end date of the harvest time window, and retains the corresponding expected value and quantile of the operation loss value to facilitate consistent execution of operation organization and risk assessment.

[0054] In one specific embodiment of the present invention, the experimental plot is located at experimental point A, and the target maize variety is V001. The silking date is used as the developmental stage anchor point, recorded as August 10, 2024. Sampling is conducted in stages after silking, and the grain moisture content is determined using the drying method. For each sampling, the sample number, sampling date, developmental stage identifier, and moisture content measurement result are recorded. A time series suitable for modeling is formed from August 20, 2024 to September 24, 2024, as recorded below: A24-01, 2024-08-20, 10 days after yarn production, 38.6 A24-02, 2024-08-27, 17 days after yarn production, 36.1 A24-03, 2024-09-03, 24 days after silk production, 33.4 A24-04, 2024-09-10, 31 days after yarn production, 31.2 A24-05, 2024-09-17, 38 days after yarn production, 29.7 A24-06, 2024-09-24, 45 days after yarn production, 28.9 Meteorological data time-aligned with grain moisture content observation data comes from the plot's automatic weather station. The original records are hourly data on temperature, relative humidity, precipitation, wind speed, and solar radiation. To ensure that the meteorological information shares the same temporal semantics as the grain moisture content observations, this embodiment uses the seven days preceding each sampling date as a statistical window. Meteorological records within the window are summarized to generate environmental condition characteristic data. The summarization method includes the window's average temperature, average relative humidity, cumulative precipitation, average wind speed, and cumulative solar radiation. The resulting environmental condition characteristic data corresponding to each grain moisture content observation is as follows: A24-01,24.8,71.0,18.2,2.1,128.0 A24-02,23.5,69.4, 6.0,2.3,141.5 A24-03,22.1,67.2, 0.8,2.6,156.2 A24-04,20.7,74.8,22.4,1.8,118.6 A24-05,19.9,65.5, 1.6,2.9,162.4 A24-06,18.7,62.1, 0.0,3.1,170.8 The genotype data of the target maize variety V001 is single nucleotide polymorphism (SNP) marker data. After the genotype data is input, site consistency and sample consistency checks are performed first, followed by deletion imputation for missing sites and allele counting and encoding, with each site taking a value of 0, 1, or 2. Subsequently, principal component analysis is performed on the SNP marker data of all breeding materials, and the first few principal components are retained as genotype representation vectors. In this embodiment, the first 8 dimensions of the genotype representation vector for V001 are retained for subsequent dynamic prediction model input, and are recorded as follows: V001 genotype representation vector, 1.42, -0.36, 0.11, 0.84, -1.05, 0.27, 0.03, -0.58 like Figure 2 As shown, the data includes: the grain moisture content observation sequence (panel A), the results of the changes in environmental characteristic statistics generated from meteorological data with the observation time (panel B), and the weight distribution of genotype characterization vector components obtained by principal component analysis based on single nucleotide polymorphism marker data (panel C). Figure 2 Panel A shows the observed grain moisture content of the target maize variety on different sampling dates and their time alignment. The horizontal axis represents the sampling date, and the vertical axis represents the percentage of grain moisture content. The corresponding values ​​are as follows: 38.6% on August 20, 2024; 36.1% on August 27, 2024; 33.4% on September 3, 2024; 31.2% on September 10, 2024; 29.7% on September 17, 2024; and 28.9% on September 24, 2024. The parentheses below each date are used to identify the meteorological statistics window corresponding to that sampling, and the alignment mark above the parentheses is used to indicate the alignment position of the statistics window with the sampling date. Figure 2Panel B presents environmental condition data corresponding to each sampling date. The five bars, arranged in order from A to E, correspond to the window's average temperature, average relative humidity, cumulative precipitation, average wind speed, and cumulative solar radiation, respectively. The values ​​for A to E for each date are as follows: 2024-08-20: 24.8, 71.0, 18.2, 2.1, 128.0; 2024-08-27: 23.5, 69.4, 6... 0, 2.3, 141.5; 2024-09-03 corresponds to 22.1, 67.2, 0.8, 2.6, 156.2; 2024-09-10 corresponds to 20.7, 74.8, 22.4, 1.8, 118.6; 2024-09-17 corresponds to 19.8, 65.5, 1.6, 2.8, 162.4; 2024-09-24 corresponds to 18.7, 62.1, 0.0, 3.1, 170.8. Figure 2 Panel C presents the principal component components of the genotype representation vector of the target maize variety. The values ​​for PC1 to PC8 are 1.42, -0.36, 0.11, 0.84, -1.05, 0.27, 0.03, and -0.58, respectively. Figure 2 Panel A to Figure 2 Panel C collectively illustrates the correspondence between grain moisture content observation data, time-aligned environmental condition characteristic data, and genotype characterization vectors in the same embodiment, providing a consistent input basis for subsequent dynamic prediction of grain moisture content and harvest decisions.

[0055] After the above data preparation is completed, a dynamic prediction model for grain moisture content is established based on grain moisture content observation data, genotype data, and meteorological data. To balance the inherent dehydration trend of the variety with short-term rehydration disturbances, this embodiment decomposes grain moisture content into slow-changing and fast-changing components, and uses a filtering estimation method to achieve updates and forward predictions based on observations. The core representation relationship is as follows: in, This represents the grain moisture content at the corresponding moment in the developmental stage. This is the component of slowly changing moisture content. This represents the rapidly changing component of moisture content. The slowly changing component is used to characterize the continuous dehydration that occurs as development progresses, while the rapidly changing component is used to characterize the short-term rehydration caused by precipitation or sustained high humidity.

[0056] The slow-varying component is updated using developmental stage increments, environmental condition characteristic data, and genotype representation vectors as inputs. A trainable regression model is employed to provide the dehydration propulsion, allowing different varieties to exhibit different dehydration types under the same meteorological conditions. The fast-varying component is updated based on meteorological processes: when a precipitation event confirmed by meteorological data occurs, or when relative humidity shows a continuous upward trend, the fast-varying component is updated incrementally; when no precipitation event occurs and relative humidity does not show a continuous upward trend, the fast-varying component is updated at a decay rate based on aridity characteristics. In the engineering implementation, precipitation events are determined by the window's cumulative precipitation and hourly precipitation peak; a continuous upward trend in relative humidity is determined by the monotonicity and duration of hourly relative humidity; and aridity characteristics are jointly generated by the window's average relative humidity, window's average wind speed, and window's cumulative solar radiation.

[0057] After receiving observations from A24-01 to A24-06, the model outputs a probability distribution of grain moisture content for future dates. In this example, September 18, 2024 to September 30, 2024 is used as the candidate prediction interval, and a discretized probability distribution is output with a 3-day step size, as recorded below: 2024-09-18:{31.0:0.10,30.0:0.25,29.0:0.35,28.0:0.20,27.0:0.10} 2024-09-21:{30.0:0.10,29.0:0.25,28.0:0.35,27.0:0.20,26.0:0.10} 2024-09-24:{29.5:0.10,28.5:0.25,27.5:0.35,26.5:0.20,25.5:0.10} 2024-09-27:{29.0:0.10,28.0:0.25,27.0:0.35,26.0:0.20,25.0:0.10} 2024-09-30:{28.5:0.10,27.5:0.25,26.5:0.35,25.5:0.20,24.5:0.10} To facilitate subsequent dehydration type identification and operational loss assessment, the expected value and variance of the probability distribution for each future date are calculated using the following formula:

[0058] in, This represents the expected grain moisture content for a future date. The variance of grain moisture content for future dates. For discrete moisture content values, For the corresponding probability weights, This represents the number of discrete values. Taking September 24, 2024 as an example, the expected value is calculated as follows: The expected value serves as the basic sequence input for the dehydration type discrimination process, which is characterized by the rate of change of moisture content. Meanwhile, the variance is used to characterize the risk of short-term rewetting and the level of prediction uncertainty.

[0059] When determining the dehydration type of a target maize variety based on the probability distribution of grain moisture content at future times, this embodiment extracts two types of features. The first type is the moisture content change rate feature, which calculates the ratio of the difference between adjacent dates to the difference between preceding and following periods based on the expected value sequence of future dates, used to distinguish the stage characteristics of dehydration progress. The second type is the variance feature of the moisture content probability distribution, which calculates the mean and peak value based on the variance sequence, used to distinguish the strength of moisture reabsorption disturbance under the same dehydration trend. These features are input into the dehydration type classification model, and the output V001 corresponds to the late-stage accelerated dehydration type. During the training phase, the classification model uses multi-variety, multi-year, multi-location data, inputs isomorphic features, and uses manually labeled or clustered results as category labels. During engineering deployment, only forward inference is performed.

[0060] After obtaining the probability distribution of grain moisture content and dehydration type at future time moments, the probability distribution of grain moisture content at future time moments, the dehydration type of the target maize variety, the mechanized ear harvesting method or mechanized grain harvesting method, and the machine operation parameters are input into the mechanized harvesting operation loss model. The machine operation parameters are obtained from the machine harvesting operation records, including the operation setpoints for parameters such as travel speed, drum speed, concave plate gap, fan speed, and screen opening. The operation loss model outputs the operation loss value, which consists of grain breakage loss, entrainment loss, cleaning loss, and drying energy consumption loss, and is synthesized according to a unified dimension. To ensure feasibility, this embodiment adopts a combination of calibration curves and regularized calculations: for grain breakage and cleaning losses, the results are obtained by looking up tables or interpolating based on the moisture content range and key operating parameters; for entrainment losses, the results are calculated based on the matching degree between moisture content, fan speed, and sieve opening; for drying energy consumption losses, the energy consumption is calculated based on the difference between the expected moisture content and the target moisture content at the warehouse, and when the dehydration type is accelerated dehydration in the later stage, stability constraints are added to the predicted energy consumption items in the later stage to avoid ignoring the energy consumption fluctuations caused by the high moisture content scenario at the tail end simply because the expected value decreases.

[0061] When calculating the operational losses for mechanized ear harvesting and mechanized grain harvesting, a set of moisture content scenarios is generated by sampling from the probability distribution of grain moisture content at future time points. The sampling method involves repeatedly sampling moisture content values ​​according to discrete probability weights to form the moisture content scenario set. Then, each scenario is substituted into the operational loss model to calculate the scenario's operational loss value, and the expected value is obtained by weighting the values ​​according to probability weights. Simultaneously, quantiles are calculated for risk control. The expected value calculation follows the following principles: in, The expected value of the task loss. This represents the scenario-based task loss value. The probability weights are for different scenarios. Calculations were performed for candidate dates September 21, 24, and 27, 2024. The expected operational losses for grain harvesting and ear harvesting methods showed an alternating pattern. On September 21, the breakage loss term for grain harvesting increased significantly under the high moisture content scenario, with both the expected value and high quantiles being relatively high. On September 24, the expected value for grain harvesting decreased and entered a controllable range. On September 27, the expected value for grain harvesting decreased further and the quantiles converged. Ear harvesting, however, was affected by drying energy consumption losses during this stage, resulting in a smaller decrease in the overall value.

[0062] When determining the mechanized ear harvesting or mechanized grain harvesting method and the harvest time window based on the operational loss value, the candidate harvest time set is first determined as a discrete date set from September 18, 2024 to September 30, 2024. Then, the expected value and quantile of the operational loss value for each candidate date are calculated for both harvesting methods, forming two candidate sequences. Subsequently, the harvest time window and the corresponding harvesting method are determined based on dynamic programming to minimize the comprehensive cost within the time window. The comprehensive cost consists of the expected value and quantile of the operational loss value, and is superimposed with time window continuity constraints and mode switching penalties to ensure that the output results conform to the actual mechanized harvesting organization method. The calculation results show that the harvest time window is from September 24, 2024 to September 27, 2024, and the corresponding harvesting method is mechanized grain harvesting.

[0063] like Figure 3 As shown, this includes: the moisture content curve during the observation stage and the probability distribution of moisture content during the prediction stage (panel A); the expected value and quantile range of the operation loss of mechanized ear harvesting and mechanized grain harvesting methods as a function of time, with the harvest time window marked (panel B); and a comparison of the composition of total operation loss, breakage loss, entrainment loss, cleaning loss and drying energy consumption loss for the two methods (panel C). Figure 3 Panel A, with date on the horizontal axis and grain moisture content percentage on the vertical axis, shows the changes in moisture content of the target variety during the observation period: 38.6% on August 20, 2024; 36.1% on August 27, 2024; 33.4% on September 3, 2024; 31.2% on September 10, 2024; 29.7% on September 17, 2024; and 28.9% on September 24, 2024. After the “Obs | Pred” boundary, violin plots are used to represent the probability distribution of future grain moisture content on October 1, 2024; October 8, 2024; and October 15, 2024, respectively. The center point and vertical line segments are used to indicate the central position and dispersion of the distribution. Figure 3Panel B uses unit loss as the vertical axis and date as the horizontal axis to compare the unit loss curves of mechanized ear harvesting and mechanized grain harvesting over time. It shows the expected loss curves for the two methods and the corresponding 10th and 90th percentile ranges (labeled as q10 / q90 in the legend). The “Time window” marked in parentheses in the figure is the specific harvest time window, and “Window end” is marked with a vertical line at 2024-09-24. Figure 3 Panel C presents the loss composition and total amount for both scenarios using a stacked bar chart: the total loss for ear harvesting is 85 (loss units), with sub-items valued at 43, 13, 11, and 18 respectively; the total loss for grain harvesting is 72 (loss units), with sub-items valued at 21, 20, 13, and 18 respectively. The legend on the right explains the meaning of each item, with 1–4 corresponding to Breakage, Entrainment, Cleaning, and Drying, respectively. Within this time window, the expected moisture content has decreased to a range suitable for grain harvesting, while the probability of a high moisture content scenario at the tail end has decreased. The combined risk of breakage loss and drying energy consumption loss is controllable, meeting the decision-making requirement of minimizing losses.

[0064] To illustrate the effectiveness, this embodiment uses a single-threshold empirical method as a control. It determines the harvest date solely based on the expected moisture content of future dates being lower than a fixed threshold, without introducing quantile constraints or time window optimization. The control method, under this data, tends to postpone the harvest to around September 30th. Although the expected value is lower, this decision is insensitive to high moisture content scenarios at the tail end of the harvest and ignores the organizational constraints of continuous operation time windows. In practical applications, if rainfall or reduced field accessibility occurs, the operable days can be passively shortened, increasing carryover losses and cleaning losses, and increasing fluctuations in temporary drying energy consumption. This invention incorporates moisture content uncertainty and operational loss risks into the optimization process simultaneously, linking the selection of harvesting methods and harvest time windows in the solution process. The output results are more robust at similar expected moisture content levels.

[0065] To facilitate the reproduction of key calculation processes by those skilled in the art, the following code can be used to calculate the expected value, variance, expected and quantiles of job loss, and time window selection. In actual engineering deployment, the discrete probability distribution can be replaced with the model output: import numpy as np dist = { "2024-09-21": {30.0:0.10,29.0:0.25,28.0:0.35,27.0:0.20,26.0:0.10}, "2024-09-24": {29.5:0.10,28.5:0.25,27.5:0.35,26.5:0.20,25.5:0.10}, "2024-09-27": {29.0:0.10,28.0:0.25,27.0:0.35,26.0:0.20,25.0:0.10}, } def mean_var(d): m = np.array(list(d.keys())) p = np.array(list(d.values())) E = np.sum(p*m) V = np.sum(p*(m-E)**2) return E, V def loss_grain(m): # Calibration function example for grain harvesting method crush = max(0.0, (m-28.0)*0.9) clean = max(0.0, (26.0-m)*0.4) carry = max(0.0, (25.5-m)*0.2) dry = max(0.0, (m-24.0)*0.3) return crush+clean+carry+dry def weighted_mean_q90(d, loss_fn): m = np.array(list(d.keys())) p = np.array(list(d.values())) J = np.array([loss_fn(x) for x in m]) EJ = np.sum(p*J) idx = np.argsort(J) J2, p2 = J[idx], p[idx] cdf = np.cumsum(p2) q90 = J2[np.searchsorted(cdf, 0.90)] return EJ, q90 for day in dist: E,V = mean_var(dist[day]) EJ,Q90 = weighted_mean_q90(dist[day], loss_grain) print(day, "E=",E, "V=",V, "EJ=",EJ, "Q90=",Q90) The above embodiments provide a complete execution chain, from data acquisition and developmental stage alignment, environmental condition feature generation, genotype characterization vector generation, to dynamic prediction of output probability distribution, dehydration type identification, operational loss assessment, and time window optimization. In practical implementation, establishing a correspondence between sample numbers and moisture content measurement records, weather station logs, genotype testing reports, machinery operation parameter records, and machine harvesting loss sampling records creates a traceable input-output chain, meeting the reproducibility requirements for engineering verification and industrial applications.

[0066] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0067] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A machine learning-based decision-making method for mechanized corn harvesting, characterized in that, Includes the following steps: Acquire observational data on kernel moisture content of the target maize variety, genotype data of the target maize variety, and meteorological data time-aligned with the kernel moisture content observation data; A dynamic prediction model for grain moisture content was established based on grain moisture content observation data, genotype data, and meteorological data. The dehydration type of the target maize variety was determined based on the probability distribution of grain moisture content at future times. Among them, the dynamic prediction model of grain moisture content is a state-space model. The state-space model represents the grain moisture content as a slow-changing component state and a fast-changing component state, and outputs the probability distribution of grain moisture content at future time based on filtering estimation. Among them, the rapid change component of moisture content is incrementally updated when a precipitation event or a meteorological process with a continuous increase in relative humidity occurs as determined by meteorological data, and is attenuatedly updated based on the dry condition characteristics generated by meteorological data when there is no precipitation event and the relative humidity does not increase continuously. Based on the probability distribution of grain moisture content at future time points and the dehydration type of the target maize variety, the mechanized harvesting operation loss model is input, and the operation loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method are calculated respectively. Based on the operation loss values, the mechanized ear harvesting method or the mechanized grain harvesting method and the harvesting time window are determined, and the mechanized harvesting decision results are output. The calculation of the operation loss values ​​corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method includes generating a set of moisture content scenarios by sampling the probability distribution of grain moisture content at future time, and calculating the expected value and quantile of the operation loss value corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method based on the set of moisture content scenarios. The process of determining the mechanized ear harvesting method or mechanized grain harvesting method and the harvest time window based on the operation loss value includes determining a set of candidate harvest times, calculating the expected value and quantile of the operation loss value corresponding to the mechanized ear harvesting method and the mechanized grain harvesting method on the set of candidate harvest times, and determining the harvest time window and the mechanized ear harvesting method or mechanized grain harvesting method corresponding to the harvest time window based on dynamic programming.

2. The method according to claim 1, characterized in that, After acquiring the grain moisture content observation data of the target maize variety, the genotype data of the target maize variety, and the meteorological data time-aligned with the grain moisture content observation data, the grain moisture content observation data is aligned with the developmental stage, and environmental condition characteristic data corresponding to the grain moisture content observation data are generated from the meteorological data based on the developmental stage alignment results.

3. The method according to claim 2, characterized in that, Genotype representation vectors are generated based on the genotype data of the target maize variety, and these genotype representation vectors, together with environmental condition characteristic data, are used as inputs to a dynamic prediction model for grain moisture content.

4. The method according to claim 3, characterized in that, The genotype data of the target maize variety is single nucleotide polymorphism (SNP) marker data. Genotype characterization vectors are generated based on the genotype data of the target maize variety by performing principal component analysis on the SNP marker data to obtain the genotype characterization vectors.

5. The method according to claim 1, characterized in that, Determining the dehydration type of a target maize variety based on the probability distribution of grain moisture content at future times involves calculating the rate of change of moisture content based on the expected value sequence of the probability distribution of grain moisture content at future times, and calculating the variance of the probability distribution of moisture content based on the probability distribution of grain moisture content at future times. The rate of change of moisture content and the variance of the probability distribution of moisture content are then input into the dehydration type classification model to output the dehydration type.

6. The method according to claim 1, characterized in that, The inputs to the mechanized harvesting operation loss model include the probability distribution of grain moisture content at future time points, the dehydration type of the target maize variety, the mechanized ear harvesting method or the mechanized grain harvesting method, and the machine operation parameters. The output of the mechanized harvesting operation loss model is the operation loss value, which includes grain breakage loss, entrainment loss, cleaning loss, and drying energy consumption loss.

Images