A Method and System for Predicting Future Yield of Wheat-Maize Rotation Based on Process Modeling and Machine Learning

Abstract

This invention discloses a method and system for predicting future yields of wheat-maize rotation based on process models and machine learning. It belongs to the interdisciplinary field of agricultural information technology, crop models, and machine learning. The method includes: constructing a spatiotemporally continuous county-level regional dataset; calibrating the parameters of the process model using location data to simulate wheat and maize yields at the site scale; generating a sample set through fertilization scenario experiments and extreme climate combined scenario experiments, dividing the sample set into a training set and a validation set; training and validating a gradient boosting regression tree model with an integrated attention mechanism, constructing the DNDC-GBRT-ATT prediction framework; inputting the county-level regional dataset into the trained gradient boosting regression tree model, and outputting the county-level yield prediction results of the wheat-maize rotation pattern under future climate change. This invention achieves accurate prediction at the county level, providing reliable technical support for regional agricultural production decision-making.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of agricultural information technology, crop modeling and machine learning, and particularly relates to a method and system for predicting future yields of wheat-maize rotation based on process modeling and machine learning. Background Technology

[0002] Current regional crop yield prediction methods mainly fall into four categories: 1. Statistical models, which establish linear empirical relationships based on historical meteorological and yield data; 2. Process models, which simulate the growth process through crop physiological and ecological mechanisms (such as the DNDC model); 3. Remote sensing assimilation methods, which use satellite data to retrieve crop growth indicators; and 4. Machine learning prediction, which uses historical data to train black-box models (such as neural networks and random forests).

[0003] The existing technical methods have the following drawbacks:

[0004] First, there is insufficient coordination between mechanisms and data. Statistical models ignore crop physiological mechanisms and cannot explain the impact of extreme climate on photosynthesis. Machine learning models lack mechanistic interpretability and have weak generalization ability under unknown climate scenarios.

[0005] Secondly, the regional adaptability is poor, the process model parameter calibration is complicated, there is a scale expansion bottleneck, and it is difficult to be efficiently applied to county-level forecasting; the remote sensing data has low temporal resolution, cannot continuously monitor the key period of the growing season, and does not quantify the multi-dimensional interaction effects of soil-climate-management.

[0006] Finally, extreme weather events are not robust, and traditional methods do not design training samples for complex extreme weather scenarios (such as sudden changes from drought to flood), which leads to amplified prediction errors. Statistical models and machine learning models rely on historical data distributions and respond distortedly to extreme events that are outside the training range. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a method and system for predicting future yields of wheat-maize rotation based on process models and machine learning, thereby resolving the issues present in the prior art.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a method for predicting future yields of wheat-maize rotation based on process models and machine learning, comprising:

[0009] Multi-source heterogeneous data of the target area of ​​wheat-maize rotation are obtained, and the multi-source heterogeneous data are integrated through a spatiotemporal alignment algorithm to construct a spatiotemporally continuous county-scale regional dataset.

[0010] The process model was calibrated using site data to simulate wheat and corn yields at the site scale.

[0011] Sample sets were generated at the sites through fertilization scenario experiments and extreme climate combined scenario experiments, and the sample sets were divided into training sets and validation sets.

[0012] Using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon, nitrogen fertilizer application rate, organic fertilizer carbon input, and organic fertilizer nitrogen input at the monitoring points as predictive variables, and the simulated yield of the process model as the target variable, a gradient boosting regression tree model with integrated attention mechanism was trained and validated to construct the DNDC-GBRT-ATT prediction framework.

[0013] By inputting county-scale regional datasets into the DNDC-GBRT-ATT prediction framework, the county-scale yield prediction results of the wheat-maize rotation pattern under future climate change are output.

[0014] Preferably, the multi-source heterogeneous data includes: soil data, agricultural management data, monitoring indicators, satellite remote sensing vegetation index, ground IoT sensor data, historical meteorological data, and future meteorological data;

[0015] The soil data includes soil organic carbon content, soil texture, bulk density, and pH value;

[0016] The agricultural management data includes crop rotation methods, fertilizer types, and irrigation methods;

[0017] The ground-based IoT sensor data includes temperature, humidity, precipitation, wind speed, solar radiation, soil moisture content, and soil temperature.

[0018] Preferably, the spatiotemporal alignment algorithm includes:

[0019] Atmospheric correction and cloud masking were performed on satellite remote sensing data, and Z-score standardization was performed on meteorological station data.

[0020] A spatial relationship map of meteorological elements is constructed based on a graph neural network. The nonlinear relationship between nodes is learned through a graph attention network, and high-resolution meteorological raster data is output.

[0021] A convolutional long short-term memory network is used to fill in missing time series data, and satellite data is fused with the interpolation results.

[0022] Preferably, the fertilization scenario experiment includes nitrogen fertilizer gradient reduction treatment, organic fertilizer replacing chemical fertilizer nitrogen treatment, and straw returning to the field treatment;

[0023] The extreme climate scenario experiments include high temperature and drought treatment, waterlogging stress treatment, and combined extreme treatment.

[0024] Preferably, the combination of drought followed by flooding involves high temperature and drought during the heading stage combined with waterlogging stress during the grain-filling stage;

[0025] The combination of flooding followed by drought, with extreme waterlogging during the seedling stage and mild high temperature and drought during the maturity stage;

[0026] Synchronous compounding, with high temperatures, precipitation fluctuations, and soil anaerobic conditions occurring simultaneously throughout the entire growth period.

[0027] Preferably, the attention mechanism calculates the normalized weights of the predictor variables using a multilayer perceptron, as shown in the formula:

[0028]

[0029] Where, x i Let α be the i-th predictor variable, MLP be a multilayer perceptron, and α be the ith predictor variable. i For normalized weights, exp(MLP(x) i )) represents calculating the exponential attention score of the current element i. This represents the sum of the exponential attention scores for all elements (including i itself), used for normalization.

[0030] Preferably, the hyperparameters of the gradient boosting regression tree model are optimized through grid search cross-validation, including the loss function, maximum number of iterations, minimum number of samples for internal node repartitioning, minimum number of samples for leaf nodes, learning rate, and maximum depth of the decision tree.

[0031] Preferably, the model accuracy is verified by the coefficient of determination, standardized root mean square error, prediction accuracy of extreme events, and error amplification factor of composite events.

[0032] The extreme event prediction accuracy is used to calculate the relative prediction error under high temperature drought or waterlogging stress.

[0033] The combined event error amplification factor is used to evaluate the error increase under combined drought and flood events.

[0034] Preferably, the yield forecast results are accompanied by a 95% confidence interval of the CMIP6 multi-mode set.

[0035] In a second aspect, the present invention also discloses a future crop yield prediction system for implementing the method described in the first aspect, comprising:

[0036] The multi-source data fusion module is used to acquire multi-source heterogeneous data of the target area of ​​wheat-maize rotation, and integrate the multi-source heterogeneous data through a spatiotemporal alignment algorithm to construct a spatiotemporally continuous county-scale regional dataset.

[0037] The model calibration module is used to calibrate the parameters of the process model using point data, simulating wheat and corn yields at the site scale;

[0038] The dataset processing module is used to generate sample sets at the locations through fertilization scenario experiments and extreme climate combined scenario experiments, and divide the sample sets into training sets and validation sets.

[0039] The hybrid modeling module is used to train and validate a gradient boosting regression tree model with an integrated attention mechanism, using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon, nitrogen fertilizer application rate, organic fertilizer carbon input, and organic fertilizer nitrogen input at the location as predictive variables and the simulated yield of the process model as the target variable, and to construct the DNDC-GBRT-ATT prediction framework.

[0040] The yield prediction module is used to input county-scale regional datasets into the DNDC-GBRT-ATT prediction framework and output county-scale yield prediction results for wheat-maize rotation patterns under future climate change.

[0041] Compared with the prior art, the present invention has the following advantages and technical effects:

[0042] This invention provides a method for predicting future yields of wheat-maize rotation based on process models and machine learning, comprising: First, acquiring multi-source heterogeneous data of the target region for wheat-maize rotation, and integrating the multi-source heterogeneous data through a spatiotemporal alignment algorithm to construct a spatiotemporally continuous county-level regional dataset; Second, using site data to calibrate the parameters of the process model to simulate wheat and maize yields at the site scale; Third, generating sample sets at the sites through fertilization scenario experiments and extreme climate combined scenario experiments, and dividing the sample sets into training and validation sets; Further, using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon, nitrogen fertilizer application rate, organic fertilizer carbon input, and organic fertilizer nitrogen input at the sites as prediction variables, and the simulated yield of the process model as the target variable, training and validating a gradient boosting regression tree model with an integrated attention mechanism to construct the DNDC-GBRT-ATT prediction framework; Finally, inputting the county-level regional dataset into the DNDC-GBRT-ATT prediction framework to output the county-level yield prediction results of the wheat-maize rotation pattern under future climate change.

[0043] This invention provides a method and system for predicting future yields in wheat-maize rotation based on process models and machine learning. Through a DNDC-GBRT-ATT hybrid architecture, it combines the physiological process interpretability of the process model (DNDC) with the efficient generalization capability of machine learning (GBRT-ATT). This fully leverages the interpretability of the mechanistic model and the advantages of strong generalization and high computational efficiency of the machine learning model, solving the "black box" problem of pure machine learning and the "poor regional adaptability" of pure mechanistic models. It achieves accurate predictions at the county level, providing reliable technical support for regional agricultural production decisions. Furthermore, it improves the spatial resolution of meteorological data through an innovative spatiotemporal alignment algorithm using an LSTM-GNN hybrid model.

[0044] This invention integrates data from long-term positioning experimental sites, county-level agricultural statistics, and CMIP6 multi-model climate scenario data through multi-source data fusion technology to construct a high-precision prediction dataset.

[0045] This invention uses a dynamic attention mechanism (ATT) to automatically identify key stress factors, while traditional models use fixed weights.

[0046] This invention significantly enhances the robustness and adaptability of the model under extreme climatic conditions by designing multi-scenario experimental schemes, including nitrogen fertilizer gradient, organic fertilizer substitution, straw return to the field, high temperature and drought, waterlogging stress, and extreme combination. Attached Figure Description

[0047] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0048] Figure 1 This is a schematic diagram of the method for predicting future yield of wheat corn based on process model and machine learning according to an embodiment of the present invention.

[0049] Figure 2 This is a technical roadmap for the DNDC-GBRT-ATT framework according to an embodiment of the present invention. Detailed Implementation

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0052] Example 1

[0053] like Figure 1 As shown, this embodiment provides a method for predicting future yields of wheat-maize rotation based on process models and machine learning, including:

[0054] Step S101: Obtain multi-source heterogeneous data of the target area of ​​wheat-grass rotation, integrate the multi-source heterogeneous data through a spatiotemporal alignment algorithm, and construct a spatiotemporally continuous county-level regional dataset;

[0055] Specifically, the multi-source heterogeneous data includes soil data, agricultural management data, monitoring indicators, satellite remote sensing vegetation indices, ground-based IoT sensor data, historical meteorological data, and future meteorological data. A spatiotemporal alignment algorithm is used to integrate the multi-source heterogeneous data, including satellite remote sensing vegetation indices, ground-based IoT sensor data, meteorological station data, and CMIP6 climate model data, to construct a spatiotemporally continuous county-scale regional dataset.

[0056] Soil data, agricultural management data, ground-based IoT sensor data, and monitoring indicators are from long-term positioning test sites.

[0057] Soil data recorded the SOC content, soil texture, bulk density, pH value, etc. of the topsoil at each test site;

[0058] Agricultural management data recorded the planting and harvesting times of crops, crop rotation methods, fertilizer management and fertilizer types, irrigation methods, and farmland cultivation methods at each experimental site;

[0059] Ground-based IoT sensor data includes field measurements of temperature, humidity, precipitation, wind speed, solar radiation, soil moisture content, and soil temperature; the monitoring indicators are the crop yields of wheat and corn.

[0060] The monitoring indicators are the yields of winter wheat and summer maize.

[0061] The historical meteorological data obtained comes from the local meteorological station's daily maximum temperature, minimum temperature, and rainfall data from 2000 to the present.

[0062] The acquired future meteorological data were obtained from an ensemble of five global climate models in the CMIP6 dataset (https: / / esgf-node.llnl.gov / search / cmip6 / ). These five global climate models are: ACCESS-CM2, CAS-ESM2, CNRM-CM6-1, GFDL-ESM4, and MIROC6. The implementation example selected was a moderate forcing scenario (SSP2-4.5, with radiative intensity stabilizing at 4.5 W / m² in 2100). -2 ) and a high-forcing scenario (SSP5-8.5, with radiation intensity stabilizing at 8.5 W / m² by 2100). -2(Scenario) Future meteorological data from now until 2090.

[0063] The satellite remote sensing data is MODIS land surface temperature (LST) data.

[0064] Both historical and future meteorological data are spatially downscaled using a spatiotemporal alignment algorithm to meet the accuracy requirements of regional agricultural forecasting. This spatiotemporal alignment algorithm employs a missing data imputation method based on LSTM (Laser-Signal-Time Graph) and a null value interpolation method based on Graph Neural Networks (GNNs). The process involves first preprocessing and standardizing the data: atmospheric correction and cloud masking are applied to satellite data (e.g., MODIS LST), and Z-score standardization is used to eliminate dimensional differences in meteorological station data to construct a spatiotemporal grid. Second, spatial interpolation is performed based on GNNs, constructing a spatial relationship graph of meteorological elements (e.g., temperature). Nodes represent stations / grids, and edge weights are determined by distance, elevation difference, and land use type. A Graph Attention Network (GAT) is used to learn the nonlinear relationships between nodes, as shown in the following formula:

[0065]

[0066] Where σ is a nonlinear activation function, and N(i) is the node v i The set of neighbors (including itself), h i ′ is node v i Updated feature, h j For node v j The original input features, α ij Here, W represents the attention weights, and W is a learnable weight matrix. This outputs high-resolution meteorological raster data. Then, temporal feature enhancement is performed: ConvLSTM is used to capture spatiotemporal dependencies in the time-series data (e.g., daily precipitation), missing date data is filled in, and meteorological physical constraints (e.g., vertical temperature lapse rate) are introduced to ensure the reasonableness of the interpolation. Finally, multi-source data fusion is performed, fusing satellite data (e.g., LST) with the interpolation results through residual concatenation. The calculation method is as follows:

[0067] T final =T GNN +λ·(T LST -T GNN )

[0068] Among them, T final For the final high-resolution meteorological raster data, T GNN T is the high-resolution meteorological raster data output by the GNN. LST The data is the preprocessed MODIS land surface temperature (LST) data, and λ is the satellite data reliability weight (dynamically calculated via MLP).

[0069] The acquired regional data includes county-level data on winter wheat and summer maize planting areas, organic fertilizer carbon and nitrogen content, and chemical fertilizer nitrogen application rates, as well as future meteorological data calculated down to the county level, such as average temperature, rainfall, and atmospheric CO2 concentration. The county-level future meteorological data is obtained by averaging all downscaled raster meteorological data within the administrative unit using ArcGIS's regional statistical function, thus representing the county's meteorological conditions.

[0070] Step S102: Use site data to calibrate the parameters of the process model (DNDC model) to simulate wheat and corn yields at the site scale;

[0071] Specifically, this embodiment uses the DNDC model to simulate crop yield at the station scale. The input parameters include meteorological, soil and management factors, as well as the station's latitude, longitude, altitude, and simulation start / end time.

[0072] Meteorological factors include daily maximum temperature, daily minimum temperature, daily average temperature, daily precipitation, solar radiation, relative humidity, wind speed, and atmospheric CO2 concentration.

[0073] Soil factors include soil texture, soil bulk density, porosity, field water holding capacity, permanent wilting point, saturated hydraulic conductivity, initial soil moisture, initial soil temperature, SOC, total nitrogen, ammonium nitrogen, nitrate nitrogen, pH value, and C / N ratio.

[0074] Management factors include crop type, sowing date, harvest date, planting density, root depth, maximum leaf area index, fertilizer type, fertilizer application rate, fertilization time, fertilization method, irrigation amount, irrigation time, irrigation method, tillage time, tillage depth, tillage method, and straw return amount.

[0075] Step S103: Generate a sample set through fertilization scenario experiments and extreme climate combined scenario experiments, and divide the sample set into a training set and a validation set;

[0076] Specifically, the experimental setups include: nitrogen fertilizer application gradient reduction treatment, organic fertilizer substitution for chemical fertilizer nitrogen treatment, straw return to field treatment, high temperature and drought treatment, waterlogging stress treatment, and combined extreme treatment, generating multi-scenario sample datasets.

[0077] The fertilization scenario experiments that generated the multi-scenario sample dataset included nitrogen fertilizer application gradient reduction treatments: FP (normal treatment), FP_0.9 (nitrogen application reduced by 10%), FP_0.8 (nitrogen application reduced by 20%), FP_0.7 (nitrogen application reduced by 30%), FP_0.6 (nitrogen application reduced by 40%), and FP_0.5 (nitrogen application reduced by 50%).

[0078] Organic fertilizer replacing chemical fertilizer nitrogen treatments: RMN_0.9 (organic fertilizer replacing 10% of chemical fertilizer nitrogen), RMN_0.8 (organic fertilizer replacing 20% ​​of chemical fertilizer nitrogen), RMN_0.7 (organic fertilizer replacing 30% of chemical fertilizer nitrogen), RMN_0.6 (organic fertilizer replacing 40% of chemical fertilizer nitrogen), and RMN_0.5 (organic fertilizer replacing 50% of chemical fertilizer nitrogen);

[0079] Straw return to the field treatment: SFP (100% straw return to the field), SFP_0.8 (80% straw return to the field), SFP_0.6 (60% straw return to the field), SFP_0.4 (40% straw return to the field), SFP_0.2 (20% straw return to the field), and FP treatment with no straw return to the field.

[0080] The extreme climate scenario experiments included high-temperature drought treatments: HTD_2-30 (mild high-temperature drought, daily average temperature +2℃ during the growing season, precipitation reduced by 30%, simulating the IPCC SSP2-4.5 mid-term scenario), HTD_3-40 (moderate high-temperature drought, daily average temperature +3℃, precipitation reduced by 40%, simulating a typical extreme drought year), and HTD_4-50 (severe high-temperature drought, daily average temperature +4℃, precipitation reduced by 50%, simulating the SSP5-8.5 end-of-century scenario); and waterlogging stress treatments: WL_50-S (standard waterlogging, continuous precipitation days increased by 50% during the critical period, soil saturation continued), WL_100-S (extreme waterlogging, precipitation days doubled, simulating a continuous rainstorm event), and WL_50-A (anaerobic waterlogging, based on WL_50-S, with additional...). Soil oxygen deficiency leads to root hypoxia; combined extreme treatments: CEX_HTD-WL (combination of HTD_3-40 and WL_50-S, drought followed by waterlogging, high temperature and drought (+3℃, -40% precipitation) during the heading stage, and waterlogging during the grain-filling stage), CEX_WL-HTD (combination of WL_100-S and HTD_2-30, waterlogging followed by drought, extreme waterlogging during the seedling stage, and mild high temperature and drought during the maturity stage), CEX_SYNC (combination of HTD_4-50 and WL_50-A, synchronous compound treatment, synchronous high temperature (+4℃), precipitation fluctuations (rapid transition from drought to waterlogging) and soil anaerobic conditions throughout the entire growth period).

[0081] The sample set was composed of indicators with high contribution values ​​selected by the feature relative importance evaluation. Finally, the multi-scenario sample dataset includes annual average temperature, annual precipitation, atmospheric CO2 concentration, SOC, nitrogen fertilizer application rate, fertilizer C, fertilizer N (predictors) and the yield simulated by DNDC under the corresponding scenario for all fertilization scenario experiments and extreme climate combination scenario experiments.

[0082] The sample set was randomly divided into a training set and a validation set in a 7:3 ratio.

[0083] Step S104: Using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon (SOC), nitrogen fertilizer application rate, organic fertilizer carbon input (C), and organic fertilizer nitrogen input (N) at the sampling points as predictive variables, and the simulated yield of the process model (DNDC) as the target variable, train and validate the gradient boosting regression tree (GBRT) model with integrated attention mechanism (ATT); combine the site-scale simulated yield of the DNDC model with the trained GBRT-ATT model to construct the DNDC-GBRT-ATT prediction framework;

[0084] Specifically, the training process for the machine learning model (GBRT model) in this embodiment is as follows: 70% of the sample data from the total sample is randomly divided as training data. First, the spatiotemporal weights of each variable are calculated using ATT, and the features are dynamically weighted. The weight calculation method is as follows:

[0085]

[0086] Where, x i Let α be the i-th predictor variable, MLP be a multilayer perceptron, and α be the ith predictor variable. i For normalized weights, exp(MLP(x) i )) represents calculating the exponential attention score of the current element i. This represents the sum of the exponential attention scores of all elements (including i itself), used for normalization. The weighted features are then concatenated with the DNDC simulated output and input into the GBRT model for training. The hyperparameter values ​​are optimized using a grid search cross-validation method, including the loss function, the maximum number of iterations for the weak learner, the minimum number of samples required for internal node re-splits, the minimum number of samples required for leaf nodes, the learning rate, and the maximum depth of the decision tree. This constructs the DNDC-GBRT-ATT prediction framework.

[0087] The hyperparameters of the GBRT model are optimized using a grid search cross-validation method, including the loss function, the maximum number of iterations for the weak learner, the minimum number of samples required for internal node repartitioning, the minimum number of samples required for leaf nodes, the learning rate, and the maximum depth of the decision tree.

[0088] The specific process for validating the prediction results is as follows: 30% of the total sample data is randomly divided as validation data, and the coefficient of determination (R²) is used. 2 The accuracy of the model was verified by evaluating the standardized root mean square error (RMSE) and predictive stability under extreme climate scenarios. Predictive stability under extreme climate scenarios includes the extreme event prediction accuracy (EPA) and the compound event error amplification factor (CEAF). EPA is used to calculate the relative error between the model's predicted values ​​and the DNDC mechanism simulation values ​​under extreme treatments such as high temperature and drought (HTD) and waterlogging stress (WL). Its calculation method is as follows:

[0089]

[0090] Where n is the total number of extreme event samples, Let be the predicted value of the prediction model for the i-th extreme event (such as high temperature and drought (HTD) or waterlogging stress (WL)). The simulated value of the DNDC model for the same event (serving as the baseline true value).

[0091] CEAF is used to evaluate the error increase of a model under combined drought and flood events (CEX), and it is calculated as follows:

[0092]

[0093] Among them, RMSE CEX To predict the root mean square error (RMSE) of the model under a combined drought and flood event (CEX), HTD To predict the root mean square error (RMSE) of the model under high temperature drought (HTD) events, WL The root mean square error of the prediction model under waterlogging stress (WL) events is calculated.

[0094] Step S105: Input the county-scale regional dataset into the trained gradient boosting regression tree model and output the county-scale yield prediction results of the wheat-maize rotation pattern under future climate change.

[0095] Specifically, the regional yield prediction process is as follows: Based on the constructed DNDC-GBRT-ATT framework, the annual average temperature, annual precipitation, atmospheric CO2 concentration, SOC, nitrogen fertilizer application rate, fertilizer C, and fertilizer N at the regional scale are used as prediction variables. After being downscaled by the spatiotemporal alignment algorithm, they are input into the trained GBRT-ATT model for prediction. The yield prediction results of the wheat-maize rotation pattern under future climate change are the yield at the county level, with a 95% confidence interval of the CMIP6 multi-model ensemble attached.

[0096] DNDC-GBRT-ATT framework technology roadmap, such as Figure 2 As shown, the advantages of the DNDC-GBRT-ATT framework lie in its innovative methodological architecture (hybrid modeling) and technological implementation (ATT layer, spatiotemporal alignment algorithm), which systematically solves the three major technical bottlenecks in agricultural yield forecasting: "difficulty in coordinating mechanisms and data," "low accuracy in scale scaling," and "poor response to extreme climates." Using the DNDC-GBRT-ATT framework for regional yield forecasting offers significant advantages in improving forecast accuracy, enhancing model generalization ability, increasing computational efficiency, and providing better interpretability and understandability.

[0097] Example 2

[0098] Based on the same inventive concept, this embodiment also provides a future crop yield prediction system that implements the method described in Embodiment 1, including:

[0099] The multi-source data fusion module is used to acquire multi-source heterogeneous data of the target area of ​​wheat-maize rotation, and integrate the multi-source heterogeneous data through a spatiotemporal alignment algorithm to construct a spatiotemporally continuous county-scale regional dataset.

[0100] The model calibration module is used to calibrate the parameters of the process model using point data, simulating wheat and corn yields at the site scale;

[0101] The dataset processing module is used to generate sample sets at the locations through fertilization scenario experiments and extreme climate combined scenario experiments, and divide the sample sets into training sets and validation sets.

[0102] The hybrid modeling module is used to train and validate a gradient boosting regression tree model with an integrated attention mechanism, using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon, nitrogen fertilizer application rate, organic fertilizer carbon input, and organic fertilizer nitrogen input at the location as predictive variables and the simulated yield of the process model as the target variable, and to construct the DNDC-GBRT-ATT prediction framework.

[0103] The yield prediction module is used to input county-scale regional datasets into the DNDC-GBRT-ATT prediction framework and output county-scale yield prediction results for wheat-maize rotation patterns under future climate change.

[0104] Furthermore, the multi-source data fusion module includes:

[0105] The data acquisition unit is used to acquire soil, meteorological, remote sensing, sensor-measured, monitoring indicators, and agricultural management data, and communicates with the CMIP6 database. Soil data records the topsoil SOC content, soil texture, bulk density, and pH value at each test site; agricultural management data records the planting and harvesting times of crops, crop rotation methods, fertilizer management and fertilizer types, irrigation methods, and tillage methods at each test site; ground-based IoT sensor data includes field measurements of temperature, humidity, precipitation, wind speed, solar radiation, soil moisture content, and soil temperature; monitoring indicators include wheat and corn yields. Crop yields are measured from wheat and corn samples collected at different test sites each growing season; satellite remote sensing data includes MODIS land surface temperature (LST) data; ground meteorological station data includes daily maximum and minimum temperatures and rainfall data collected from national-level ground meteorological stations; and future climate scenario data (SSP2-4.5, SSP5-8.5) are obtained by accessing the CMIP6 database.

[0106] Spatial alignment cells are created using a hybrid algorithm based on LSTM-GNN to unify multi-source data into a higher-resolution grid. Then, ArcGIS spatial analysis tools are used to extract the rasterized region data into administrative units.

[0107] Both historical and future meteorological data are spatially downscaled using a spatiotemporal alignment algorithm to meet the accuracy requirements of regional agricultural forecasting. This spatiotemporal alignment algorithm employs a missing data imputation method based on LSTM (Laser-Signal-Time Graph) and a null value interpolation method based on Graph Neural Networks (GNNs). The process involves first preprocessing and standardizing the data: atmospheric correction and cloud masking are applied to satellite data (e.g., MODIS LST), and Z-score standardization is used to eliminate dimensional differences in meteorological station data to construct a spatiotemporal grid. Second, spatial interpolation is performed based on GNNs, constructing a spatial relationship graph of meteorological elements (e.g., temperature). Nodes represent stations / grids, and edge weights are determined by distance, elevation difference, and land use type. A Graph Attention Network (GAT) is used to learn the nonlinear relationships between nodes, as shown in the following formula:

[0108]

[0109] Where σ is a nonlinear activation function, and N(i) is the node v i The set of neighbors (including itself), h i ′ is node v i Updated feature, h j For node v j The original input features, α ij The attention weights are defined by W, which is a learnable weight matrix, resulting in high-resolution meteorological raster data. Then, temporal feature enhancement is performed: ConvLSTM is used to capture spatiotemporal dependencies in the time-series data (e.g., daily precipitation), missing date data is filled in, and meteorological physical constraints (e.g., vertical temperature lapse rate) are introduced to ensure the reasonableness of the interpolation. Finally, multi-source data fusion is performed, fusing satellite data (e.g., LST) with the interpolation results through residual concatenation. The calculation method is as follows:

[0110] T final =T GNN +λ·(T LST -T GNN )

[0111] Among them, T final For the final high-resolution meteorological raster data, T GNN T is the high-resolution meteorological raster data output by the GNN. LST The data is the preprocessed MODIS land surface temperature (LST) data, and λ is the satellite data reliability weight (dynamically calculated via MLP).

[0112] The acquired regional data includes county-level data on winter wheat and summer maize planting areas, organic fertilizer carbon and nitrogen content, and chemical fertilizer nitrogen application rates, as well as future meteorological data calculated down to the county level, such as average temperature, rainfall, and atmospheric CO2 concentration. The county-level future meteorological data is obtained by averaging all downscaled raster meteorological data within the administrative unit using ArcGIS's regional statistical function, thus representing the county's meteorological conditions.

[0113] Furthermore, the hybrid modeling module includes:

[0114] The DNDC mechanism simulation unit takes a complete set of parameters at the station scale as input and outputs calibrated station yield simulation results for GBRT-ATT training.

[0115] DNDC (Digital Density Conversion) mechanism simulation, by inputting parameters such as meteorological, soil, crop, and management practices, uses the DNDC model to simulate crop processes such as photosynthesis, nutrient cycling, and water transport, and outputs crop growth indicators such as yield. Combined with measured yield data, parameter sensitivity analysis is used to optimize model parameters and improve simulation accuracy.

[0116] The DNDC model was used to simulate crop yield at the site scale. Input parameters included meteorological, soil, and management factors. Meteorological factors included daily maximum temperature, daily minimum temperature, daily average temperature, daily precipitation, solar radiation, relative humidity, wind speed, and atmospheric CO2 concentration. Soil factors included soil texture, soil bulk density, porosity, field capacity, permanent wilting point, saturated hydraulic conductivity, initial soil moisture, initial soil temperature, SOC, total nitrogen, ammonium nitrogen, nitrate nitrogen, pH, and C / N ratio. Management factors included crop type, sowing date, harvest date, planting density, root depth, maximum leaf area index, fertilizer type, fertilizer application rate, fertilization time, fertilization method, irrigation amount, irrigation time, irrigation method, tillage time, tillage depth, tillage method, and straw return amount. The model also included the site's latitude, longitude, altitude, and simulation start / end time.

[0117] The GBRT-ATT prediction unit integrates Scikit-learn and PyTorch, and includes ATT and GBRT. It takes regional-scale core variables as input and outputs regional yield forecasts and uncertainty quantification.

[0118] Input feature variables include DNDC simulation results, agricultural management data, soil data, and meteorological factors. Key variables are selected through feature importance analysis. Training is performed using training sample data, and 10-fold cross-validation is employed to prevent overfitting. Crop yields under future climate scenarios are generated using regional data.

[0119] This implementation also includes a visualization unit, which uses Python's Matplotlib to generate spatial distribution maps of yield, climate scenario comparison curves, and risk warning maps. The spatial distribution maps of yield display the predicted yields of each county using different color gradients, the climate scenario comparison curves are used to show the interannual variations of yields under future climate scenarios, and the risk warning maps display the degree of risk of reduced grain production in different regions using different color gradients.

[0120] This implementation also includes an interactive unit that allows users to filter data by crop type and time range, adjust management parameters, obtain prediction results in real time, and export CSV / Shapefile format for further analysis.

[0121] Through interactive units, users can dynamically adjust management measures (such as fertilizer application) and obtain forecast results instantly; existing systems only support static output. The entire system features high forecast accuracy and fast processing speed, and can be widely applied in fields such as agricultural yield forecasting and climate change impact assessment.

[0122] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting future yield of wheat-maize rotation based on process models and machine learning, characterized in that, Includes the following steps: Multi-source heterogeneous data of the target area of ​​wheat-maize rotation are obtained, and the multi-source heterogeneous data are integrated through a spatiotemporal alignment algorithm to construct a spatiotemporally continuous county-scale regional dataset. The process model was calibrated using site data to simulate wheat and corn yields at the site scale. Sample sets were generated at the sites through fertilization scenario experiments and extreme climate combined scenario experiments, and the sample sets were divided into training sets and validation sets. Using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon, nitrogen fertilizer application rate, organic fertilizer carbon input, and organic fertilizer nitrogen input at the monitoring points as predictive variables, and the simulated yield of the process model as the target variable, a gradient boosting regression tree model with integrated attention mechanism was trained and validated to construct the DNDC-GBRT-ATT prediction framework. The attention mechanism calculates the normalized weights of the predictor variables using a multilayer perceptron, as shown in the formula: Where, x i Let be the i-th predictor variable, and MLP be a multilayer perceptron. To normalize the weights, This indicates the calculation of the exponential attention score for the current element i. This represents the sum of the exponential attention scores for all elements, used for normalization; The hyperparameters of the gradient boosting regression tree model are optimized through grid search cross-validation, including the loss function, maximum number of iterations, minimum number of samples for internal node repartition, minimum number of samples for leaf nodes, learning rate, and maximum depth of the decision tree. The model accuracy was verified using the coefficient of determination, standardized root mean square error, prediction accuracy of extreme events, and error amplification factor of composite events. The extreme event prediction accuracy is used to calculate the relative prediction error under high temperature drought or waterlogging stress. The combined event error amplification factor is used to evaluate the error increase under combined drought and flood events; The county-scale regional dataset is input into the DNDC-GBRT-ATT prediction framework, which outputs the county-scale yield prediction results of the wheat-maize rotation pattern under future climate change; the yield prediction results are accompanied by the 95% confidence interval of the CMIP6 multi-model ensemble.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data includes: soil data, agricultural management data, monitoring indicators, satellite remote sensing vegetation index, ground IoT sensor data, historical meteorological data, and future meteorological data; The soil data includes soil organic carbon content, soil texture, bulk density, and pH value; The agricultural management data includes crop rotation methods, fertilizer types, and irrigation methods; The ground-based IoT sensor data includes temperature, humidity, precipitation, wind speed, solar radiation, soil moisture content, and soil temperature.

3. The method according to claim 1, characterized in that, The spatiotemporal alignment algorithm includes: Atmospheric correction and cloud masking were performed on satellite remote sensing data, and Z-score standardization was performed on meteorological station data. A spatial relationship map of meteorological elements is constructed based on a graph neural network. The nonlinear relationship between nodes is learned through a graph attention network, and high-resolution meteorological raster data is output. A convolutional long short-term memory network is used to fill in missing time series data, and satellite data is fused with the interpolation results.

4. The method according to claim 1, characterized in that, The fertilization scenario experiments include nitrogen fertilizer gradient reduction treatment, organic fertilizer replacing chemical fertilizer nitrogen treatment, and straw returning to the field treatment; The extreme climate scenario experiments include high temperature and drought treatment, waterlogging stress treatment, and combined extreme treatment.

5. The method according to claim 4, characterized in that, The composite extreme treatment includes: The combination of drought followed by flooding, with high temperature and drought during the heading stage superimposed on waterlogging stress during the grain-filling stage; The combination of flooding followed by drought, with extreme waterlogging during the seedling stage and mild high temperature and drought during the maturity stage; Synchronous compounding, with high temperatures, precipitation fluctuations, and soil anaerobic conditions occurring simultaneously throughout the entire growth period.

6. A future crop yield prediction system implementing the method of any one of claims 1-5, characterized in that, include: The multi-source data fusion module is used to acquire multi-source heterogeneous data of the target area of ​​wheat-maize rotation, and integrate the multi-source heterogeneous data through a spatiotemporal alignment algorithm to construct a spatiotemporally continuous county-scale regional dataset. The model calibration module is used to calibrate the parameters of the process model using point data, simulating wheat and corn yields at the site scale; The dataset processing module is used to generate sample sets at the locations through fertilization scenario experiments and extreme climate combined scenario experiments, and divide the sample sets into training sets and validation sets. The hybrid modeling module is used to train and validate a gradient boosting regression tree model with an integrated attention mechanism, using the annual average temperature, annual precipitation, atmospheric CO2 concentration, soil organic carbon, nitrogen fertilizer application rate, organic fertilizer carbon input, and organic fertilizer nitrogen input at the location as predictive variables and the simulated yield of the process model as the target variable, and to construct the DNDC-GBRT-ATT prediction framework. The yield prediction module is used to input county-scale regional datasets into the DNDC-GBRT-ATT prediction framework and output county-scale yield prediction results for wheat-maize rotation patterns under future climate change.

Images