A multi-target decision method, system and device for crop rotation ton semi-precision cultivation

Abstract

The application provides a multi-target decision-making method, system and equipment for crop rotation and half-precision cultivation, which comprises the following steps: obtaining annual crop rotation data of all sampling points in a target area; performing unified scale processing on meteorological data and geographical data, and combining the two with management measure data and variety data into a multi-source heterogeneous data set with the same resolution; determining an optimal solution set of key variety parameters by using the multi-source heterogeneous data set and a crop growth model; constructing a gridded data set by using a first deep neural network model; simulating annual yield by using the crop growth model, calculating a simulation value of a joint index based on the annual yield, and combining the simulation value with the gridded data set to construct a training set; constructing a second deep neural network model, training the second deep neural network model by using the training set, and obtaining a decision scheme of the annual crop rotation data corresponding to the target by taking the joint index as the target. The application can achieve the purposes of multi-target collaborative optimization and high-precision spatialization recommendation through annual integrated simulation.

Description

Technical Field

[0001] This invention belongs to the field of agricultural intelligent decision support and crop cultivation optimization technology, specifically involving a multi-objective decision-making method, system and equipment for precision cultivation of 1.5 tons of grain in crop rotation. Background Technology

[0002] As my country's agriculture transforms towards high-quality and sustainable development, model- and data-driven smart agriculture technologies have become a research and application hotspot. In the field of crop cultivation and management optimization, current technologies are mainly developing along two paths, but both exhibit significant shortcomings and challenges in addressing the complex needs of the specific ecological zone of the North China Plain.

[0003] The first approach: simulation optimization based on mechanistic models. Crop growth models, such as APSIM and DSSAT, are based on the mechanisms of crop physiological and ecological processes and can quantitatively simulate crop growth, yield, and water and fertilizer use under different climates, soils, and management practices. The advantage of this approach lies in its clear mechanisms and strong interpretability of results, making it a core tool for a deeper understanding of the interaction between crops and the environment. Existing research has used such models to simulate and optimize management practices at point or small-area scales. However, their inherent computationally intensive nature leads to extremely low simulation efficiency. When extended to vast areas like the North China Plain, and requiring a full-scenario search for global optima across multiple variables such as sowing, density, and water and fertilizer management, the enormous computational cost makes it difficult to achieve high-resolution, real-time, or near-real-time decision support in practical applications.

[0004] The second approach: Data-driven AI prediction models. Machine learning algorithms, especially models like random forests and support vector machines, can learn from historical or remote sensing data and quickly predict key indicators such as crop yield and water requirement. This approach is advantageous due to its high computational efficiency, making it suitable for processing large datasets and performing rapid inferences. Currently, many studies have attempted to use data generated by crop models from the first approach as training sets to construct such AI proxy models, overcoming the computational bottleneck of mechanistic models. However, this approach has fundamental flaws: First, traditional machine learning models, such as random forests, are essentially black-box or gray-box models, lacking a characterization of the crop's internal physiological processes, resulting in weak mechanistic interpretability of their predictions. Their generalization ability and reliability are questionable when encountering new environments or extreme situations not covered by the training data. Second, existing research is mostly limited to predicting single objectives (such as yield) or optimizing single-season crops, failing to systematically address the complex synergy and trade-offs between multiple green objectives such as "year-round high yield" and "water conservation, fertilizer reduction, and emission reduction" in the wheat-maize rotation system of the North China Plain.

[0005] While both approaches have their advantages, current technological development reveals significant gaps and shortcomings when addressing the crucial and specific need for green, half-ton grain production in the North China Plain. On one hand, mechanistic models are accurate but slow, unable to support refined regional recommendations; on the other hand, mainstream AI models (such as random forests) are fast but lack mechanistic understanding and have limited ability to optimize multi-objective collaboration. The two have not yet achieved true deep coupling and complementary advantages. Furthermore, current cultivation technology recommendations are largely based on macro-level experience from ecological zoning or focus only on single aspects, failing to fully integrate high-resolution soil, meteorological, and varietal characteristic data to provide a comprehensive, year-round integrated smart cultivation solution that is high-yielding, efficient, green, and feasible for millions of differentiated fields in the North China Plain. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-objective decision-making method, system, and equipment for precision cultivation of 1.5-ton crops in rotation, which can achieve multi-objective collaborative optimization and high-precision spatial recommendation through year-round integrated simulation.

[0007] This invention provides a multi-objective decision-making method for precision cultivation of 1.5-ton crops per rotation crop, comprising the following steps: Obtain annual crop rotation data from all sampling points within the target area. The annual crop rotation data includes meteorological data, soil data, geographical data, management practice data, and variety data. The meteorological and geographical data are downscaled, and the soil data is upscaled. The two data are then merged with the management measures data and variety data into a multi-source heterogeneous dataset with the same resolution. Using the aforementioned multi-source heterogeneous dataset and crop growth model, the global sensitivity analysis-contribution decomposition method is employed to determine the key variety parameters that play a dominant role in the joint indicators. With the objectives of minimizing the annual yield simulation error and maximizing water and nitrogen use efficiency, the optimal solution set of the key variety parameters is determined. The joint indicators include annual yield, annual nitrogen fertilizer partial productivity, irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system. A first deep neural network model is constructed and a loss function based on a preset penalty term is used to determine the mapping relationship between the optimal solution set and the meteorological data, soil data and geographical data in the multi-source heterogeneous dataset, so as to obtain a rasterized dataset. Using the rasterized dataset as input data, a crop growth model is used to simulate annual yield, and the simulated value of the joint index is calculated based on the annual yield. This simulated value is then merged with the rasterized dataset to construct a training set. A second deep neural network model is constructed and trained using the training set. With the joint index as the objective, a decision scheme for the annual crop rotation data corresponding to the objective is obtained.

[0008] Furthermore, the step of downscaling the meteorological and geographical data, upscaling the soil data, and merging them with the management measures data and variety data into a multi-source heterogeneous dataset of the same resolution includes the following steps: A local relationship model between the geographic data and meteorological data is established by geographic weighted regression, and the meteorological data is spatially refined by inverse distance weighted interpolation to complete the initial downscaling process. Systematic bias correction was performed on the meteorological data that had undergone preliminary downscaling, and then the geographic data that had undergone preliminary downscaling were merged to obtain a daily meteorological dataset with a resolution of 250 meters. The resolution of the soil data was increased to 250 meters by co-Kriging spatial interpolation to obtain a daily soil dataset; The daily meteorological dataset, daily soil dataset, management measure data, and variety data are merged into a multi-source heterogeneous dataset with the same resolution.

[0009] Furthermore, the first deep neural network model includes three hidden layers, and the preset penalty term includes developmental time-series logical constraints, annual agronomic linkage constraints, and variety experience value range constraints; the expression of the loss function is: ; Wherein, Loss1 is the data prediction loss, Loss2 is the rule constraint loss in the preset penalty term, and γ is the balancing weight, which is adjusted when the output of the first deep neural network model contradicts the data in the preset penalty term.

[0010] Furthermore, the step of using the rasterized dataset as input data, simulating annual yield using a crop growth model, calculating the simulated value of the joint index based on the annual yield, and merging it with the rasterized dataset to construct a training set includes the following steps: The soil and meteorological data were stratified by stratified random sampling and representative geographic grids were selected. Then, management measure data obtained by Latin hypercube sampling were generated for each geographic grid. Using soil and meteorological data from the geographic raster as input data, and management measures within a preset optimization range as decision variables, a crop growth model is used to simulate annual yield. Based on the annual yield, calculate the annual nitrogen fertilizer partial productivity, annual irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss of the system to obtain a training set consisting of management measures data and simulated values ​​of joint indicators for each geographic grid.

[0011] Furthermore, the second deep neural network model employs a weighted composite loss function, the expression of which is: ; Among them, w j L represents the weighting coefficient of the joint indicator j, n represents the number of joint indicators, and L represents the weighting coefficient of the joint indicator j. j The mean squared error loss of the joint index j; In setting the weighting coefficients, ensuring annual yield is always the primary objective. On this basis, other objectives are optimized in a coordinated manner according to regional conditions: in all scenarios, the annual yield weight is set to the highest level; in water-scarce areas, the weight of irrigation water use efficiency is increased; in environmentally sensitive areas, the weight of apparent nitrogen loss in the system and greenhouse gas emission intensity per unit yield is strengthened; in key nutrient management areas, the weight of annual nitrogen fertilizer partial productivity is increased; in general production areas, a higher annual yield weight is used in conjunction with the weights of other objectives, which are equal in value.

[0012] Furthermore, the key variety parameters for crop rotation are: Wheat: Accumulated temperature during vegetative growth period, photoperiod sensitivity, vernalization sensitivity coefficient; Corn: accumulated temperature from emergence to rapid growth period, accumulated temperature during grain filling period, and photosynthetic stress factors.

[0013] Furthermore, the meteorological data includes daily maximum temperature, daily minimum temperature, daily precipitation, and daily radiation data; The geographic data includes latitude and longitude, altitude, slope, aspect, and distance from water bodies; The soil data includes soil bulk density, pH value, clay content, silt content, sand content, organic matter, and cation conversion rate; The management measures data include year, sowing date, harvest date, planting density, nitrogen application rate, fertilizer application rate, fertilization date, basal and topdressing ratio, N2O emissions, NH3 emissions, and NO3 emissions. - Discharges, irrigation volume, irrigation dates, and annual yield.

[0014] This invention also provides a multi-objective decision-making system for precision cultivation of 1.5-ton crops per rotation crop, comprising: The acquisition module is used to acquire annual crop rotation data from all sampling points within the target area. The annual crop rotation data includes meteorological data, soil data, geographical data, management practice data, and variety data. The unified scale module is used to downscale the meteorological and geographical data, upscale the soil data, and merge the two with the management measures data and variety data into a multi-source heterogeneous dataset with the same resolution. The parameter tuning module is used to determine the key variety parameters that play a dominant role in the joint indicators by using the multi-source heterogeneous dataset and crop growth model and employing the global sensitivity analysis-contribution decomposition method. The goal is to minimize the annual yield simulation error and maximize water and nitrogen use efficiency, and to determine the optimal solution set of the key variety parameters. The joint indicators include annual yield, annual nitrogen fertilizer partial productivity, irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system. The rasterization module is used to construct a first deep neural network model and use a loss function based on a preset penalty term to determine the mapping relationship between the optimal solution set and the meteorological data, soil data and geographical data in the multi-source heterogeneous dataset, so as to obtain a rasterized dataset. The dataset generation module is used to simulate annual yield using a crop growth model with the rasterized dataset as input data, calculate the simulated value of the joint index based on the annual yield, and merge it with the rasterized dataset to construct a training set. The decision module is used to construct a second deep neural network model and train it with the training set, and to obtain a decision scheme for the annual crop rotation data corresponding to the joint index as the objective.

[0015] The present invention also provides a multi-objective decision-making device for precision cultivation of crop rotation with a yield of 1.5 tons, including a processor and a memory, wherein the memory is used to store a computer program and the processor is used to execute the computer program to implement the steps of the above method.

[0016] The beneficial effects of this invention are: 1. The multi-objective synergistic optimization capability of this invention achieves a qualitative improvement: Under the premise of ensuring an annual yield of 1.5 tons of grain (≥1500 kg / mu), the system-recommended "water-saving and emission-reduction" package, compared with the single yield maximization scheme, can increase annual nitrogen fertilizer partial productivity by an average of 10%-15%, irrigation water use efficiency by 15%-25%, and reduce greenhouse gas emission intensity per unit yield by 8%-12% and nitrogen loss risk by 15%-20%. This proves that this invention has successfully solved the problem of the trade-off between high yield and green environmental protection, and achieved a fundamental shift from "trading yield for benefits" to "synergistic gains".

[0017] 2. The decision-making accuracy, reliability, and executability of this invention are comprehensively improved: Due to the use of deep neural network model agents and multi-objective Pareto optimization, the field validation consistency of the recommended solutions is improved by an average of about 20% compared to the basic patented solutions. The generated "digital agronomic prescription" includes specific water and nitrogen management dates and dosages, which increases farmers' understanding and willingness to adopt the solutions by more than 30%, and achieves 100% compatibility with the instructions of mainstream intelligent agricultural machinery operation systems. Decision-making has been upgraded from "suggesting what to do" to "informing when and how much to do," significantly reducing the threshold and uncertainty of technology implementation.

[0018] 3. The system computation and optimization efficiency in this invention meets the requirements of real-time decision-making: the trained multi-task deep learning agent model has a single scenario evaluation time of less than 5 milliseconds. Based on this, a Pareto optimization search of the entire North China Plain with a 250-meter resolution and full-dimensional management space can be completed within 1 hour on a standard high-performance computing cluster, possessing practical efficiency to support the formulation of annual planting plans. This efficiency improvement makes it possible to perform comprehensive and refined optimization of complex agronomic systems involving water and nitrogen management, meeting the real-time planning needs of regional agricultural management departments.

[0019] 4. This invention promotes green and high-quality development of regional agriculture: Simulation results show that if the recommended scheme of this system is fully implemented in the North China Plain, it can achieve significant environmental benefits such as reducing the total amount of chemical fertilizers used in the region by about 5-8% and reducing agricultural irrigation water consumption by about 6-10%, while maintaining the current total grain output. This invention is not only a technical tool, but also provides a quantitative technical path and decision-making basis for the green transformation of regional agricultural systems, with far-reaching social and ecological benefits. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method in this invention; Figure 2 This is a schematic diagram of the annual yield prediction results of the crop rotation ton-and-a-half grain system in this invention (where A is a schematic diagram of the wheat yield potential prediction results and B is a schematic diagram of the corn yield potential prediction results). Figure 3 This is a schematic diagram of the recommended sowing dates for the crop rotation 1.5-ton-per-grain system in this invention (where A is a schematic diagram of the recommended sowing dates for wheat, and B is a schematic diagram of the recommended sowing dates for corn). Figure 4 This is a schematic diagram of the recommended planting density results for the crop rotation 1.5-ton-per-grain system in this invention (where A is a schematic diagram of the recommended planting density results for wheat, and B is a schematic diagram of the recommended planting density results for maize). Figure 5 This is a schematic diagram of the recommended nitrogen application rate for the crop rotation system of 1.5 tons of grain in this invention (where A is a schematic diagram of the recommended nitrogen application rate for wheat and B is a schematic diagram of the recommended nitrogen application rate for maize). Figure 6 This is a schematic diagram of the recommended irrigation amount for the crop rotation ton-and-a-half grain system in this invention (where A is a schematic diagram of the recommended irrigation amount for wheat and B is a schematic diagram of the recommended irrigation amount for maize). Figure 7 This is a schematic diagram showing the recommended number of nitrogen application times for the crop rotation system of 1.5 tons per ton in this invention (where A is a schematic diagram showing the recommended number of nitrogen application times for wheat, and B is a schematic diagram showing the recommended number of nitrogen application times for maize). Figure 8 This is a schematic diagram illustrating the recommended irrigation frequency results for the crop rotation 1.5-ton-per-grain system in this invention (where A represents the recommended irrigation frequency results for wheat, and B represents the recommended irrigation frequency results for maize). Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0022] like Figures 1 to 8 As shown, a multi-objective decision-making method for precision cultivation of 1.5-ton grain crops in crop rotation includes the following steps: S1. Obtain annual crop rotation data from all sampling points within the target area. This data includes meteorological, soil, geographical, management practice, and crop variety data. Specifically, the meteorological data includes daily maximum temperature, daily minimum temperature, daily precipitation, and daily radiation data, derived from gridded data from the China Meteorological Administration's Land Surface Data Assimilation System (CLDAS-V2.0), covering the North China Plain with a daily meteorological driving field at a spatial resolution of 250 meters.

[0023] The geographic data includes latitude and longitude, altitude, slope, aspect, and distance from water bodies, derived from a digital elevation model.

[0024] Soil data include soil bulk density, pH value, clay content, silt content, sand content, organic matter, and cation conversion rate, derived from raster data from the "Sun Yat-sen University China 1km Soil Attribute Dataset" and point data collected from a special soil survey in North China.

[0025] Management data includes year, sowing date, harvest date, planting density, nitrogen application rate, fertilizer application rate, fertilization date, basal and topdressing ratio, N2O emissions, NH3 emissions, and NO3 emissions. - Data on emissions, irrigation volume, irrigation dates, and annual yield were obtained from academic databases such as CNKI and Web of Science. The system retrieved and extracted field trial literature data from the main wheat-maize producing areas of the North China Plain over the past 20 years.

[0026] S2. Downscale the meteorological and geographical data, upscale the soil data, and merge them with the management practice data and variety data into a multi-source heterogeneous dataset of the same resolution. Specifically, this step includes the following steps: S201. Establish a local relationship model between geographic data and meteorological data through geographic weighted regression, and refine the spatial scale of meteorological data by using inverse distance weighted interpolation to complete the initial downscaling process.

[0027] S202. Perform systematic bias correction on the meteorological data that has undergone preliminary downscaling, and then merge the geographic data that has undergone preliminary downscaling to obtain a daily meteorological dataset with a resolution of 250 meters.

[0028] S203. The resolution of the soil data is increased to 250 meters by co-kriging spatial interpolation to obtain the daily soil dataset.

[0029] S204. Merge the daily meteorological dataset, daily soil dataset, management measures data, and variety data into a multi-source heterogeneous dataset with a resolution of 250 meters.

[0030] S3. A wheat-maize year-round rotation simulation scenario is constructed using multi-source heterogeneous datasets and a crop growth model (APSIM model). A global sensitivity analysis-contribution decomposition method is employed to determine the key variety parameters that dominate the joint indicators (this method is existing technology and will not be elaborated upon). With the objectives of minimizing the annual yield simulation error and maximizing water and nitrogen use efficiency, the optimal solution set for the key variety parameters is determined, namely: wheat: accumulated temperature during the vegetative growth period, photoperiod sensitivity, and vernalization sensitivity coefficient. This step abandons the traditional extensive approach of configuring uniform variety parameters for a single crop or a large region, and innovatively proposes a "year-round collaborative-spatial customization" variety parameter optimization method. The core of this method is to treat the variety parameters of wheat and maize as an interconnected holistic system, and for each 250-meter spatial grid cell in the North China Plain, collaboratively optimize and generate a set of variety parameter combinations that localizes the year-round system (wheat-maize rotation).

[0031] Maize: accumulated temperature from emergence to rapid growth, accumulated temperature during grain filling, and photosynthetic stress factors. These parameters collectively regulate resource competition and synergistic allocation in dual-crop crops within a limited growing season. Joint indicators include annual yield, annual nitrogen fertilizer partial productivity, irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system.

[0032] Using a database established by S2 containing multi-year, multi-location wheat-maize rotation field trial data from North China, the parameters of the aforementioned key varieties were synergistically calibrated. With the goal of minimizing annual yield simulation error while maximizing the simulation accuracy of water and nitrogen use efficiency, the optimal solution set for the key variety parameters was determined. This process ensured that the optimized parameter combination could simultaneously reproduce the synergistic effect of high yield and high efficiency at the field measurement level.

[0033] S4. Construct a first deep neural network model and use a loss function based on a preset penalty term to determine the mapping relationship between the optimal solution set and the meteorological, soil, and geographical data in the multi-source heterogeneous dataset, thus obtaining a rasterized dataset. To achieve the transition from discrete site calibration to continuous spatial extrapolation, this step constructs a first deep neural network model capable of integrating data patterns and domain knowledge to learn the highly complex and non-stationary mapping relationship between the high-resolution multi-source heterogeneous dataset and the optimal solution set, thereby obtaining a rasterized dataset.

[0034] The model architecture and data flow of the first deep neural network model are as follows: Input layer: Receives multidimensional environmental feature vectors for each 250-meter grid cell, including preprocessed meteorological data, soil data, and geographic data.

[0035] Hidden Layers and Learning Process: A deep neural network model with three hidden layers is used: the first hidden layer (number of neurons H1=128), the second hidden layer (number of neurons H2=64), and the third hidden layer (number of neurons H3=64). Through training, the model automatically extracts high-level abstract features from massive environmental features and learns the complex nonlinear relationship between these features and the target variety parameters.

[0036] Output layer: Directly outputs the collaboratively optimized key variety parameter values ​​for wheat and maize corresponding to this raster cell.

[0037] Training process: The Adam optimizer was used with an initial learning rate η = 0.001, parameters β1 = 0.9, and β2 = 0.999. The batch size was set to 128. The number of training epochs was set to 500, and an early stopping strategy was adopted (training stopped when the validation set loss did not decrease for 10 consecutive epochs). Parameter initialization: The weight matrix was initialized using He normality, and the bias was initialized to 0. Regularization: Dropout (dropout rate 0.2) and L2 weight decay (coefficient 1e-4) were used in the hidden layers to prevent overfitting.

[0038] To ensure that the parameter spatial distribution predicted by the model not only fits the data but also conforms to crop physiological patterns and actual agronomic practices in North China, this invention introduces an agronomic and physiological logical constraint mechanism during the training process of the neural network. Specifically, based on the standard mean squared error loss function, a preset penalty term is designed and added to form a composite loss function: .

[0039] Where Loss1 is the data prediction loss, Loss2 is the rule constraint loss in the preset penalty term, and γ is the balancing weight, which is adjusted when the output of the first deep neural network model contradicts the data in the preset penalty term. The preset penalty term is as follows: Developmental timing logic constraint: penalize parameter combinations that violate the sequential physiological order of accumulated temperature required for each stage of crop development (e.g., tt_emerg_to_endjuv).

[0040] Year-round agronomic linkage constraints: Based on the predicted wheat variety development parameters, the maturity period is simulated, and the predicted results that do not meet the necessary agricultural operation intervals between the wheat and the typical maize sowing period in North China are penalized to ensure the feasibility of year-round agronomy.

[0041] Variety empirical value range constraint: Limit the predicted values ​​of parameters to a reasonable physiological range based on extensive observations and literature reports of the main cultivated varieties in North China, so as to avoid outliers that are detached from actual biological significance.

[0042] By embedding the knowledge derived from crop physiology and regional production experience into the model as soft constraints for training, this method guides the deep learning model to conduct a more reasonable search in the data space. The final predicted parameter space distribution map has both high accuracy driven by data and high reliability guided by mechanism.

[0043] Through the above steps, a spatial database of wheat-maize year-round system-specific variety parameters was generated, covering the entire North China Plain with a spatial resolution of 250 meters. This database is not a simple compilation of parameters, but a core database that deeply integrates regional environmental heterogeneity, year-round agronomic systematicity, and multi-objective optimization orientation. Using geographic rasters as the basic storage unit, each effective agricultural raster contains a set of corresponding, co-optimized wheat and maize variety parameters, totaling millions of highly localized parameter records, forming an unprecedentedly fine-scale digital map of variety adaptability.

[0044] The parameters in this database are the result of collaborative optimization of the wheat-maize rotation system as a whole, ensuring efficient matching of developmental characteristics between successive crop varieties and supporting year-round integrated optimization from the source.

[0045] The 250-meter high resolution can significantly characterize the subtle changes in variety adaptability caused by micro-topography, soil texture patches, and local climate differences within the North China Plain, providing underlying parameter support for achieving true "grid-level" precision management.

[0046] The parameter optimization process implicitly considers the multiple objectives of "green half-ton grain" (high yield, high efficiency, and low emissions), which enables the simulation and optimization based on this database to naturally guide the cultivation scheme that conforms to the regional green development strategy, forming a consistent transmission of objectives from variety adaptability to cultivation strategy.

[0047] S5. Using the rasterized dataset as input, simulate annual yield using the APSIM crop growth model, and calculate the simulated values ​​of joint indicators based on the annual yield. Merge these simulated values ​​with the rasterized dataset to construct a training set. This step aims to construct a high-quality, multi-objective dataset for training subsequent deep learning proxy models through mechanistic model simulation. The construction of the integrated annual simulation system is based on the previously generated 250-meter resolution environmental driving field (meteorology, soil) for the North China region and the "wheat-maize co-optimization variety parameter database." A continuous annual simulation scenario of wheat-maize is constructed within the APSIM model platform. The core innovation of this scenario lies in the system coupling of the simulation process: using soil data at the end of the previous wheat simulation as the precise initial condition for the subsequent maize simulation, thus truly reflecting the continuous impact of crop rotation on water and fertilizer resource utilization in the double-cropping system of the North China Plain, providing a mechanistic basis for the optimal allocation of resources at the annual scale. The core data generation in this step focuses on the "North China wheat-maize annual rotation system" and directly serves the multi-objective co-optimization of "green 1.5 tons of grain." Specifically, this step also includes the following steps: S501. Soil and meteorological data are stratified using stratified random sampling and representative geographic grids are selected. Then, management measure data obtained by Latin hypercube sampling are generated for each geographic grid.

[0048] S502. Using soil and meteorological data from geographic lattices as input data, and management measures within a preset optimization range (including sowing date, planting density, fertilizer application rate, basal and topdressing ratio and timing, irrigation amount and critical irrigation period for wheat and corn) as decision variables, the APSIM crop growth model is used to simulate annual yield.

[0049] Based on mainstream high-yield and high-efficiency agronomic practices in the North China Plain, and combined with literature analysis and professional knowledge, optimized ranges were set for each management variable to suit the actual production conditions of the region: Planting density: 6-10 plants / m² for summer maize, and 180-300 plants / m² for winter wheat. Sowing window: October 1st to 30th for winter wheat, and the sowing period for summer maize after the previous wheat harvest is immediately to within 7 days (approximately May 25th to June 25th). Nitrogen fertilizer management: The total nitrogen application rate was set within a range (50-300 kg / ha per crop), and different base-topdressing ratios (e.g., 5:5, 3:7) and key periods for topdressing (wheat jointing stage, maize tasseling stage). Irrigation system: The total irrigation volume was set within a range (50-300 mm per crop), and the combined effects of different irrigation frequencies (e.g., 1-4 times) and key water supply periods (e.g., winter irrigation for wheat, jointing irrigation) were simulated.

[0050] S503. Based on annual yield, calculate annual nitrogen fertilizer partial productivity, annual irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system to obtain a training set consisting of management measure data and simulated values ​​of joint indicators for each geographic grid. The training set is divided into training, validation, and test sets in a 7:2:1 ratio to ensure the representativeness of the spatiotemporal distribution. The formula for calculating annual nitrogen fertilizer partial productivity is as follows: Among them, PFP i For the annual nitrogen fertilizer partial productivity in year i, Y i For the annual production in year i, F i This represents the amount of fertilizer applied in year i.

[0051] The formula for calculating annual irrigation water use efficiency is: Among them, IWUE ,i Y represents the annual irrigation water use efficiency in year i. i For the annual production in year i, W i Let be the irrigation amount in year i.

[0052] The formula for calculating greenhouse gas emission intensity per unit output is: Among them, GHG i Let G be the greenhouse gas emission intensity per unit output in year i. N2O,i Let G be the N2O emissions in year i. NH3,i Let G be the NH3 emissions in year i. NO3-,i NO3 in year i - Emissions.

[0053] The formula for calculating the apparent nitrogen loss of the system is: ; among which, Nr loss,iFor the apparent nitrogen loss of the system in year i, FN i Let EF be the nitrogen application rate in year i. NH3,i Let EF be the NH3 loss coefficient in year i. N2O,i Let EF be the N2O loss coefficient for year i. NO3-,i NO3 in year i - Loss coefficient.

[0054] Thus, each record in the generated structured dataset constitutes a complete mapping relationship: environmental characteristics, combination of annual management measures → five target variable values ​​for green grain production. This dataset is the fundamental data guarantee for achieving multi-objective collaborative intelligent decision-making.

[0055] S6. Construct a second deep neural network model and train it using the training set. Using the joint index as the objective, obtain a decision scheme for the annual crop rotation data corresponding to this objective. This step aims to construct an intelligent agent model capable of accurately and quickly predicting the multi-objective response of the North China wheat-maize annual crop rotation system, replacing the computationally expensive APSIM mechanism model for subsequent optimization. Compared to the single-output random forest model, the core innovation of this step lies in constructing a "multi-task deep neural network model," realizing a paradigm shift from predicting a single yield to simultaneously predicting five key indicators for green 1.5-ton grain production, and providing a core engine for multi-objective collaborative optimization. To achieve multi-objective collaborative optimization in the wheat-maize annual crop rotation system, this invention designs a dedicated multi-task learning architecture, aiming to accurately fit and balance the complex and nonlinear response relationships among the five key objectives. These objectives include: annual yield, annual partial productivity of nitrogen fertilizer (PFPN), irrigation water use efficiency (IWUE), greenhouse gas emission intensity per unit yield (GHG), and system apparent nitrogen loss (Nr loss). Together, these elements constitute a comprehensive evaluation system for "green 1.5 tons of grain," encompassing three dimensions: high yield, resource efficiency, and environmental friendliness. Furthermore, trade-offs and conflicts often exist between these objectives (e.g., high yield may increase environmental impact). To effectively capture their inherent relationships and achieve overall optimization, this invention constructs a second deep neural network model, as follows: The model architecture employs a shared-branch multi-task learning framework, specifically designed for multi-objective coupled optimization. The input layer is designed as a high-dimensional vector, fully encompassing six categories of management decision variables across both seasons: planting density (separately set for wheat and corn), sowing date (independently adjustable for both seasons), irrigation amount (total irrigation and seasonal allocation), irrigation frequency (1-4 times, timing adjustable), nitrogen application rate (total nitrogen and base-topdressing ratio), and nitrogen application frequency (1-4 times, topdressing timing adjustable). It also integrates initial meteorological conditions and soil property parameters. The bottom-level shared feature encoder consists of multiple fully connected layers (containing three hidden layers with 256, 128, and 64 neurons respectively), responsible for extracting general high-level features related to all five objectives from this high-dimensional input. Its core function is to implicitly learn the intrinsic relationships and trade-offs between multiple objectives and capture complex interaction effects between multi-dimensional decision variables (such as the synergy between density and nitrogen application, and the coupling relationship between water and fertilizer). The upper layers consist of five independent regression task branches, each corresponding to an optimization objective, responsible for learning fine-grained mapping relationships for specific objectives based on shared system features. This "shared-independent" structure allows the model to simultaneously leverage the correlation between multiple objectives to improve prediction consistency while retaining the ability to model the specificity of each objective, thus providing an accurate and coordinated high-dimensional objective response surface for subsequent multi-objective optimization algorithms.

[0056] Regarding the training strategy and loss function, to coordinate multi-objective learning and explicitly define "achieving a ton and a half of grain" as the primary output objective, this invention employs a weighted composite loss function for model training. The expression of the weighted composite loss function is as follows: ; Among them, w j L represents the weighting coefficient of the joint indicator j, n represents the number of joint indicators, and L represents the weighting coefficient of the joint indicator j. jThis represents the mean squared error loss of the joint index j. In weighting, the core focus is always on ensuring the annual yield target (1.5 tons of grain). Based on this, other objectives are optimized collaboratively according to regional conditions: In all scenarios, the annual yield weight w1 is set at a relatively high level (e.g., 1.5~2.0) to ensure that the model always uses "achieving 1.5 tons of grain" as the optimization basis; in water-scarce areas (e.g., annual rainfall below 500 mm), while ensuring the annual yield, the weight w3 of irrigation water use efficiency (IWUE) can be appropriately increased (e.g., 1.2~1.5) to promote water conservation; in environmentally sensitive areas (e.g., near water sources), while prioritizing the annual yield target... Based on this, the weights w5 and w4 of apparent nitrogen loss (Nr loss) and greenhouse gas emission intensity per unit yield (GHG) can be appropriately increased (e.g., 1.3-1.6) to coordinate annual yield and environmental sustainability. In key nutrient management areas, the weight w2 of annual nitrogen fertilizer partial productivity (PFPN) can be increased simultaneously (e.g., 1.2-1.5) to improve nitrogen fertilizer utilization efficiency while stabilizing annual yield. In general production areas, a balanced weighting model dominated by annual yield can be adopted, that is, using a higher w1 combined with other target weights of about 1.0 to achieve synergistic promotion of half a ton of grain and green production.

[0057] The training data comes from the training set in S5. After hyperparameter tuning, the model's average prediction R² for five targets on the test set exceeds 0.89. While maintaining the accuracy of the mechanistic model, it improves the single-point evaluation speed by 4 orders of magnitude, laying the foundation for efficiently exploring the high-dimensional multi-objective optimization space.

[0058] The trained multi-task agent model is serialized and encapsulated into a "multi-dimensional, multi-objective response surface simulator," serving as the core evaluation engine for the multi-objective optimization algorithm. This simulator receives a vector containing all decision variables (planting density, sowing date, irrigation amount and frequency, nitrogen application amount and frequency, etc.) and synchronously returns its predicted values ​​for all five objectives within approximately 1-3 milliseconds, completely replacing the previously time-consuming APSIM simulation. This makes it possible to perform large-scale, multi-objective optimization searches on high-dimensional management spaces that include dual-season water and nitrogen management and sowing date and quantity control. The optimization algorithm utilizes this engine to efficiently explore the complex mapping relationships between management measures and multiple objectives within the decision space composed of the aforementioned multi-dimensional variables, automatically finding the Pareto optimal solution set that achieves the best balance among the objectives, providing decision-makers with multiple options ranging from "high yield and high efficiency" to "green and low-carbon." This design marks a leap from single-objective or piecewise optimization to full-cycle, multi-dimensional, multi-objective collaborative intelligent optimization, representing a core technological breakthrough in achieving multi-objective balance decision-making for green grain production.

[0059] Example 2: Based on the same inventive concept, the present invention also provides a multi-objective decision-making system for precision cultivation of 1.5-ton-per-crop rotation, comprising: The acquisition module is used to acquire annual crop rotation data from all sampling points within the target area. The annual crop rotation data includes meteorological data, soil data, geographical data, management practice data, and variety data. The unified scale module is used to downscale meteorological and geographical data and upscale soil data, merging them with management measures data and variety data into a multi-source heterogeneous dataset with the same resolution. The parameter tuning module is used to determine the key variety parameters that play a dominant role in the joint indicators by using multi-source heterogeneous datasets and crop growth models and employing global sensitivity analysis-contribution decomposition method. The goal is to minimize the annual yield simulation error and maximize water and nitrogen use efficiency, and to determine the optimal solution set of key variety parameters. The joint indicators include annual yield, annual nitrogen fertilizer partial productivity, irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system. The rasterization module is used to construct the mapping relationship between the optimal solution set and meteorological data, soil data and geographical data in the multi-source heterogeneous dataset using the first deep neural network model and a loss function based on a preset penalty term, so as to obtain the rasterized dataset. The dataset generation module is used to simulate annual yield using a crop growth model with a rasterized dataset as input data, calculate the simulated value of the joint index based on the annual yield, and merge it with the rasterized dataset to construct a training set. The decision module is used to construct a second deep neural network model and train it with a training set to obtain a decision scheme for the annual crop rotation data corresponding to the joint index.

[0060] Example 3: Based on the same inventive concept, the present invention also provides a multi-objective decision-making device for precision cultivation of 1.5-ton-per-crop rotation, including a processor and a memory. The memory is used to store computer programs, and the processor is used to execute the computer programs to implement the steps of the method in Example 1.

[0061] The innovation of this invention lies in: 1. Variety-environment synergistic optimization and spatial customization for year-round systems.

[0062] This invention pioneers a spatialized customization method for variety parameters in a year-round wheat-maize system. Unlike traditional or existing methods that allocate uniform parameters for single crops or large regions, this method treats the variety parameters of wheat and maize as a whole driven by a shared environment and constrained by agronomic timelines. Through multi-objective optimization, a key "coupled parameter set" is determined, and a deep learning model incorporating agronomic logic constraints is used to generate a dedicated parameter spatial distribution with a resolution of 250 meters. Its breakthrough lies in achieving, for the first time, not only precise matching of variety characteristics with the local environment but also adaptation to the year-round cropping system, ensuring the agronomic feasibility of the optimization scheme from the outset.

[0063] 2. Deep integration of mechanistic models and multi-task deep learning models.

[0064] To address the specific needs of multi-objective collaborative optimization for green grain production, this invention proposes a novel "mechanism-multi-task AI" fusion architecture. This architecture uses annual system simulation data containing multi-objective responses generated by the APSIM mechanistic model as its training foundation, constructing a multi-task deep neural network as a surrogate model. Compared to the single-output random forest used in the basic patent, this model can simultaneously and accurately predict five key indicators: annual yield, resource efficiency, and environmental effects. While maintaining the credibility of the mechanism, it improves the evaluation efficiency of high-dimensional management scenarios by four orders of magnitude, providing a core intelligent engine for resolving multi-objective conflicts and conducting efficient Pareto frontier searches.

[0065] 3. Pareto optimization of comprehensive management measures and package-style decision generation.

[0066] This invention represents a paradigm shift from "single-objective, few-variable" optimization to "multi-objective, multi-dimensional, package-style" decision-making. The optimization variables have expanded from traditional sowing date and density to a complete set of agronomic measures, including the timing, amount, and frequency of nitrogen application and irrigation, forming a high-dimensional decision space. By integrating the aforementioned multi-task proxy model with a multi-objective evolutionary algorithm, the Pareto optimal solution set is solved for each geographical unit, and further, 2-3 sets of "green half-ton grain decision packages" with clear orientations (such as high-yield and high-efficiency, water-saving and emission-reduction types) are extracted. This method, for the first time, provides users with the ability to quantitatively weigh and autonomously choose among different green objectives at a regional scale.

[0067] 4. Generation of executable "digital agronomic prescriptions" from data to the field.

[0068] This invention bridges the "last mile" in implementing intelligent decision-making, delivering results that go beyond simple potential distribution maps. Based on the aforementioned optimization results, the system utilizes spatial information technology to generate a 250-meter resolution spatial explicit raster layer for each refined agronomic variable in each recommended "package" (e.g., nitrogen application rate during wheat jointing stage, irrigation date during corn tasseling stage). This forms a "digital agronomic prescription map" that can directly drive precision equipment such as variable-rate fertilizer applicators and intelligent irrigation systems. Through a WebGIS platform, users can query and download complete, quantitative, and executable technical solutions for their fields, achieving end-to-end digital control from macro-strategic analysis to micro-field operations.

[0069] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-objective decision-making method for precision cultivation of 1.5-ton grain crops in crop rotation, characterized in that, Includes the following steps: Obtain annual crop rotation data from all sampling points within the target area. The annual crop rotation data includes meteorological data, soil data, geographical data, management practice data, and variety data. The meteorological and geographical data are downscaled, and the soil data is upscaled. The two data are then merged with the management measures data and variety data into a multi-source heterogeneous dataset with the same resolution. Using the aforementioned multi-source heterogeneous dataset and crop growth model, the global sensitivity analysis-contribution decomposition method is employed to determine the key variety parameters that play a dominant role in the joint indicators. With the objectives of minimizing the annual yield simulation error and maximizing water and nitrogen use efficiency, the optimal solution set of the key variety parameters is determined. The joint indicators include annual yield, annual nitrogen fertilizer partial productivity, irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system. A first deep neural network model is constructed, and a loss function based on a preset penalty term is used to determine the mapping relationship between the optimal solution set and the meteorological data, soil data, and geographical data in the multi-source heterogeneous dataset, thus obtaining a rasterized dataset. The first deep neural network model includes three hidden layers, and the preset penalty term includes developmental time-series logical constraints, annual agronomic linkage constraints, and variety empirical value range constraints. The expression of the loss function is: ； Wherein, Loss1 is the data prediction loss, Loss2 is the rule constraint loss in the preset penalty term, and γ is the balancing weight, which is adjusted when the output of the first deep neural network model contradicts the data in the preset penalty term; Using the rasterized dataset as input data, a crop growth model is used to simulate annual yield, and the simulated value of the joint index is calculated based on the annual yield. This simulated value is then merged with the rasterized dataset to construct a training set. This step includes: using stratified random sampling to stratify the soil data and meteorological data and selecting representative geographic rasters, and then generating management measure data obtained by Latin hypercube sampling for each geographic raster. Using soil and meteorological data from the geographic raster as input data, and management measures within a preset optimization range as decision variables, a crop growth model is used to simulate annual yield. Based on the annual yield, calculate the annual nitrogen fertilizer partial productivity, annual irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss of the system to obtain a training set consisting of management measures data and simulated values ​​of joint indicators for each geographic grid. A second deep neural network model is constructed and trained using the training set. With the joint index as the objective, a decision scheme for the annual crop rotation data corresponding to this objective is obtained. The second deep neural network model employs a weighted composite loss function, the expression of which is: ； Where wj represents the weight coefficient of joint indicator j, n represents the number of joint indicators, and Lj is the mean square error loss of joint indicator j. In setting the weighting coefficients, ensuring annual yield is always the primary objective. On this basis, other objectives are optimized in a coordinated manner according to regional conditions: in all scenarios, the annual yield weight is set to the highest level; in water-scarce areas, the weight of irrigation water use efficiency is increased; in environmentally sensitive areas, the weight of apparent nitrogen loss in the system and greenhouse gas emission intensity per unit yield is strengthened; in key nutrient management areas, the weight of annual nitrogen fertilizer partial productivity is increased; in general production areas, a higher annual yield weight is used in conjunction with the weights of other objectives, which are equal in value.

2. The multi-objective decision-making method for precision cultivation of 1.5-ton grain crops in crop rotation according to claim 1, characterized in that, The steps of downscaling the meteorological and geographical data, upscaling the soil data, and merging them with the management measures data and variety data into a multi-source heterogeneous dataset of the same resolution include the following: A local relationship model between the geographic data and meteorological data is established by geographic weighted regression, and the meteorological data is spatially refined by inverse distance weighted interpolation to complete the initial downscaling process. Systematic bias correction was performed on the meteorological data that had undergone preliminary downscaling, and then the geographic data that had undergone preliminary downscaling were merged to obtain a daily meteorological dataset with a resolution of 250 meters. The resolution of the soil data was increased to 250 meters by co-Kriging spatial interpolation to obtain a daily soil dataset; The daily meteorological dataset, daily soil dataset, management measure data, and variety data are merged into a multi-source heterogeneous dataset with the same resolution.

3. The multi-objective decision-making method for precision cultivation of 1.5-ton grain crops in crop rotation according to claim 1, characterized in that, The key variety parameters for crop rotation are: Wheat: Accumulated temperature during vegetative growth period, photoperiod sensitivity, vernalization sensitivity coefficient; Corn: accumulated temperature from emergence to rapid growth period, accumulated temperature during grain filling period, and photosynthetic stress factors.

4. The multi-objective decision-making method for precision cultivation of 1.5-ton grain crops in crop rotation according to claim 1, characterized in that, The meteorological data includes daily maximum temperature, daily minimum temperature, daily precipitation, and daily radiation data; The geographic data includes latitude and longitude, altitude, slope, aspect, and distance from water bodies; The soil data includes soil bulk density, pH value, clay content, silt content, sand content, organic matter, and cation conversion rate; The management data includes year, sowing date, harvest date, planting density, nitrogen application rate, fertilizer application rate, fertilizer application date, basal and topdressing ratio, N2O emissions, NH3 emissions, NO3- emissions, irrigation amount, irrigation date, and annual yield.

5. A multi-objective decision-making system for precision cultivation of 1.5-ton-per-crop rotation, used in the multi-objective decision-making method for precision cultivation of 1.5-ton-per-crop rotation as described in claim 1, characterized in that, include: The acquisition module is used to acquire annual crop rotation data from all sampling points within the target area. The annual crop rotation data includes meteorological data, soil data, geographical data, management practice data, and variety data. The unified scale module is used to downscale the meteorological and geographical data, upscale the soil data, and merge the two with the management measures data and variety data into a multi-source heterogeneous dataset with the same resolution. The parameter tuning module is used to determine the key variety parameters that play a dominant role in the joint indicators by using the multi-source heterogeneous dataset and crop growth model and employing the global sensitivity analysis-contribution decomposition method. The goal is to minimize the annual yield simulation error and maximize water and nitrogen use efficiency, and to determine the optimal solution set of the key variety parameters. The joint indicators include annual yield, annual nitrogen fertilizer partial productivity, irrigation water use efficiency, greenhouse gas emission intensity per unit yield, and apparent nitrogen loss in the system. The rasterization module is used to construct a first deep neural network model and use a loss function based on a preset penalty term to determine the mapping relationship between the optimal solution set and the meteorological data, soil data and geographical data in the multi-source heterogeneous dataset, so as to obtain a rasterized dataset. The dataset generation module is used to simulate annual yield using a crop growth model with the rasterized dataset as input data, calculate the simulated value of the joint index based on the annual yield, and merge it with the rasterized dataset to construct a training set. The decision module is used to construct a second deep neural network model and train it with the training set, and to obtain a decision scheme for the annual crop rotation data corresponding to the joint index as the objective.

6. A multi-objective decision-making device for precision cultivation of 1.5-ton grain crops in crop rotation, characterized in that, It includes a processor and a memory, the memory being used to store a computer program, and the processor being used to execute the computer program to implement the steps of the method of claim 1.

Images