Intelligent calibration and decision system and method for rice fertilization

Abstract

This invention discloses an intelligent calibration and decision-making method for rice fertilization. The method includes: designing an initial experiment using an active learning strategy to acquire a dataset; constructing an instrument-side Bayesian probabilistic surrogate model and a remote sensing-side Gaussian process regression inversion model to achieve an interpretable and probabilistic mapping from raw signals to agronomic parameters; generating a transparent decision function based on multi-objective optimization; and finally, in a closed-loop application, achieving precise sensing through Bayesian fusion and utilizing an online transfer learning mechanism to achieve continuous model evolution. This invention solves the problems of "black box" decision-making, cold start difficulties, lack of physical consistency, and inability to adaptively update existing intelligent agriculture technologies, providing a highly reliable, interpretable, and lifelong learning-capable precision fertilization solution.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to an intelligent calibration and decision-making system and method for rice fertilization. Background Technology

[0002] Precision fertilization is crucial for improving agricultural resource utilization efficiency and achieving sustainable development. Existing technologies suffer from two main problems: first, methods based on fixed mechanistic models or general experience struggle to adapt to field-specific and interannual variations; second, methods based on complex artificial intelligence models such as deep learning, while technologically advanced, exhibit significant "black box" characteristics, including opaque decision-making logic, poor agronomic interpretability, data scarcity and cold-start problems (difficult to deploy in new regions or when data is scarce), and a lack of physical consistency. Furthermore, existing technologies generally treat data collection, instrument calibration, model training, and decision application as separate processes, resulting in unclear sources of error throughout the decision-making chain and hindering the achievement of consistently reliable and precise decisions. Therefore, a novel intelligent fertilization decision-making framework is urgently needed that integrates physical measurement information, quantifies uncertainties throughout the entire process, and possesses continuous adaptive capabilities. Summary of the Invention

[0003] To address the above problems, this invention proposes an intelligent calibration and decision-making system and method for rice fertilization.

[0004] The main contents of this invention include: According to one aspect of the present invention, a method is provided, comprising an initialization phase and an application and evolution phase. The initialization phase acquires a high-quality initial dataset through an active learning strategy. and based on the initial dataset A Bayesian probabilistic surrogate model for the instrument, a Gaussian process regression inversion model for the remote sensing end, and an agronomic decision function are constructed sequentially. In the application and evolution stage, the constructed models are used to conduct collaborative diagnosis and fertilization decisions for new fields, and the models are continuously evolved through transfer learning mechanisms using new online data.

[0005] According to another aspect of the present invention, a system for implementing the above method is provided, comprising: an active experiment design module, an intelligent perception calibration module, a probability inversion and fusion module, a transparent decision optimization module, and a continuous evolutionary learning module.

[0006] Compared with existing technologies, the beneficial effects of the above-mentioned intelligent calibration and decision-making system and method for rice fertilization are as follows: (1) By using interpretable models such as Bayesian linear regression and Gaussian process regression, and embedding the instrument physical measurement model in a probabilistic form, the physical rationality and transparency of the decision are guaranteed.

[0007] (2) Through active learning and Bayesian probabilistic modeling, a highly reliable localized model can be quickly established using a very small amount of carefully designed initial experimental data, which reduces the deployment threshold.

[0008] (3) From instrument measurement and remote sensing inversion to final decision-making, each step outputs probabilistic results and confidence estimates, clarifying the sources of error and supporting robust decision-making with known risks.

[0009] (4) By combining the online transfer learning mechanism with the Elastic Weight Consolidation (EWC) strategy, the system can continuously evolve by utilizing the sparse new data generated during the production process and automatically adapt to environmental changes and variety updates.

[0010] (5) The traditionally fragmented calibration, monitoring, decision-making and updating processes are integrated into an autonomous intelligent closed loop, which greatly improves the collaborative efficiency and practicality of the technology system. Detailed Implementation

[0011] The technical solution protected by this invention will be described in detail below.

[0012] This invention proposes an intelligent calibration and decision-making system and method for rice fertilization. By using interpretable models such as Bayesian linear regression and Gaussian process regression, and embedding the instrument's physical measurement model in a probabilistic form, the physical rationality and transparency of the decision-making are guaranteed.

[0013] Specifically, the present invention provides a method comprising the following steps: S1. Using a hybrid active learning strategy that combines uncertainty sampling and space filling, design and execute the initial field experiment to obtain a structured initial dataset. .

[0014] Specifically, in this embodiment, in the first year, the system, within the target ecological zone, based on historical climate and soil data, runs a hybrid active learning algorithm to generate a basal-ear fertilizer ratio experimental scheme containing four nitrogen gradient levels, with a total of 12 plots. After the experiment is executed, the initial dataset is obtained. The initial dataset At least include: raw leaf absorbance data measured using a portable spectrometer during the critical growth period of the experimental plots. Laboratory standard nitrogen content data for corresponding leaf samples Spectral index data extracted from simultaneously acquired UAV multispectral remote sensing images and the yield of each plot measured after harvest. With cost data .

[0015] More specifically, the hybrid active learning strategy of uncertainty sampling and spatial filling mainly includes defining essential test areas through spatial filling and adding high-value areas through uncertainty sampling supplementation. In this embodiment, spatial filling refers to the system dividing the target area into several core feature regions based on data such as soil type, topography, and irrigation conditions. For example, it might divide the area into 3-5 categories, including sandy soil highlands, clay soil plains, and low-lying waterlogged areas. At least 1-2 basic test plots are set up in each category to ensure no key environmental type is missed. Uncertainty sampling supplementation involves adding 1-2 additional test plots in high-uncertainty feature areas, such as the boundary between sandy soil and clay soil, or plains with missing historical data, to ensure "more comprehensive information in key blank areas." This maximizes coverage of key nonlinear regions of crop response. By collecting data from these regions, a structured initial dataset is created. .

[0016] S2, Based on the initial dataset A Bayesian linear regression surrogate model is constructed and trained on the instrument side, enabling it to output agronomic parameters with confidence interval estimates based on the original sensor signals; the agronomic parameters include, but are not limited to, at least one of nitrogen content, chlorophyll content, biomass, leaf area index, and plant water content.

[0017] In this embodiment, a portable CCN nitrogen analyzer was used to collect raw absorbance data of leaf samples from each plot at multiple specific wavelengths. The samples were sent to the laboratory for chemical total nitrogen analysis to obtain the corresponding standard nitrogen content true values. The system uses a Bayesian regression framework to establish the original absorbance vectors measured at multiple characteristic wavelengths. Compared with laboratory standard nitrogen content Mapping relationship between them: ;in, For model parameter vectors, To augment the input vector, For the new absorbance data, the variance is the noise variance. The posterior distribution of the surrogate model parameters is obtained by inputting the trained Bayesian linear regression surrogate model into the instrument, and the posterior mean is downloaded to the instrument to enable it to output. ability.

[0018] S3, Based on the initial dataset Using the output of the Bayesian linear regression surrogate model, a Gaussian process regression inversion model that can directly output the posterior prediction distribution is constructed and trained.

[0019] Specifically, multispectral images from UAVs are acquired, and measurements are obtained using upgraded instruments. As a label, it is related to the extracted spectral index. Gaussian process regression training was performed to obtain the initial remote sensing inversion model. And for the newly input spectral exponent vector Output posterior prediction Gaussian distribution , The mean nitrogen content is obtained from posterior prediction. To characterize The prediction variance of the confidence level.

[0020] S4. Based on the initial dataset Using agronomic and economic output data, interpretable agronomic decision functions are generated through multi-objective constrained optimization. .

[0021] After harvest, yield, quality, and cost were measured in each plot. The initial measurement was based on the panicle differentiation stage. Using output as input and production as output, a response surface is fitted. With the objectives of maximizing overall benefit and minimizing environmental loss, a multi-objective genetic algorithm is run to obtain the Pareto optimal solution set, which is then parameterized as the initial decision function. .

[0022] More specifically, the transparent agronomic decision function is based on the initial dataset. Generated through multi-objective constraint optimization, it can output transparent fertilization recommendations by inputting calibrated crop status and environmental parameters. The specific process is as follows: Step D1: Fit the response surface. In For input, output For the output, support vector machine regression (SVR) is used to fit the output response surface. .

[0023] Step D2: Multi-objective optimization. With the objectives of maximizing net profit and minimizing environmental risk, and the above response surface as the main constraint, run a multi-objective genetic algorithm (NSGA-II, parameter example: population size 100, generation number 200) to solve for the Pareto front.

[0024] Step D3: Function Encapsulation. Parameterize the Pareto optimal solution set into a decision function. Input the critical period nitrogen state N and the environmental economic parameter Env, and output the recommended amount of topdressing fertilizer Rec. Here, the critical period nitrogen state N is the mean of the calibrated distribution obtained through Bayesian fusion, i.e., N = This represents the current nitrogen nutrient level of the field after system calibration. Environmental and economic parameters... This is a parameter vector, which includes at least the current fertilizer market price and the basic nitrogen supply in the soil. These parameters are input by the user through the system interface or automatically obtained by the system through a data interface. The system will then combine the real-time acquired N with... By substituting the values ​​into the decision function, the optimal recommended amount of topdressing fertilizer Rec under the current scenario can be calculated.

[0025] S5. Closed-loop application and evolution stage: The Bayesian linear regression surrogate model, Gaussian process regression inversion model and agronomic decision function constructed in S2 to S4 are applied to new fields for fertilization decisions. The new data generated during the application process are used to continuously optimize and evolve the model and decision function through an online learning mechanism.

[0026] S5.1 For the target field, the remote sensing predicted distribution is obtained using the Gaussian process regression inversion model, and then Bayesian fusion is performed with the ground measured distribution based on the Bayesian probabilistic surrogate model to obtain the calibrated crop status information. For the target field, the remote sensing inversion model and the Bayesian probabilistic surrogate model are used for collaborative diagnosis to obtain calibrated crop status information, specifically including the following steps: For new field block B, acquire its UAV imagery, input it into the Gaussian process regression inversion model (GPR model) to obtain the initial probabilistic distribution map (each pixel contains In field B, representative locations were selected, and measurements were taken using intelligent instruments. ± Assume the remote sensing prediction distribution N( ) and ground-measured distribution N( The independent Bayesian fusion within it is specifically calculated using the following formula to determine the calibrated state distribution. : , in and Let the mean and variance of the remote sensing prediction distribution be denoted as . and These represent the mean and variance of the measured ground distribution. Uncertainty after fusion. Significantly reduced, generating a "high-confidence diagnostic chart".

[0027] S5.2. Input the calibrated crop status information and current environmental economic parameters into the agronomic decision function to obtain fertilization recommendations; that is, the average field fertilization rate... and current environmental parameters input decision function Real-time fertilization recommendations for field B It can also visualize the economic-environmental trade-off curves behind it.

[0028] S5.3. Collect new data generated during the application process and use an online transfer learning strategy based on elastic weight consolidation to incrementally update the Gaussian process regression inversion model and the agronomic decision function. Specifically, this includes: Step 5.3.1: Data Accumulation and Update Triggering. Online Data ( Store the incremental dataset The system sets update trigger conditions (e.g.: (Data volume reaches 50 records, or model prediction error increases by more than 15% for three consecutive quarters).

[0029] Step 5.3.2: Importance Assessment and Constraint Fine-tuning. When an update is triggered, the EWC strategy is adopted: Calculate the parameters of the old GPR model Fisher information matrix diagonal elements This indicates its importance.

[0030] exist Fine-tuning the parameters of the new model When minimizing the loss function: .in The negative log-likelihood of the new data. The importance weight is set (e.g., 100). A penalty term constrains changes in important parameters to prevent catastrophic forgetting.

[0031] Step 5.3.3: Incremental update of the decision function. When Once a certain scale is accumulated, the decision function is triggered synchronously. Incremental learning and updates.

[0032] Step 5.3.4: Evolutionary Application. Subsequent diagnostics and decisions automatically utilize the evolved model, achieving continuous performance improvement.

[0033] Furthermore, this invention proposes an intelligent calibration and decision-making system for rice fertilization, comprising an active experimental design module, an intelligent sensing and calibration module, a probabilistic inversion and fusion module, a transparent decision optimization module, and a continuous evolutionary learning module connected in sequence. These modules jointly execute the aforementioned calibration and decision-making methods. Specifically, the active experimental design module is used to execute the hybrid active learning strategy; the intelligent sensing and calibration module is used to construct and deploy the Bayesian probabilistic surrogate model; the probabilistic inversion and fusion module is used to construct the Gaussian process regression inversion model and perform Bayesian fusion; the transparent decision optimization module is used to generate and execute the agronomic decision function; and the continuous evolutionary learning module is used to implement the online transfer learning based on elastic weight consolidation.

[0034] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent calibration and decision-making in rice fertilization, characterized in that, Includes the following steps: S1. Using a hybrid active learning strategy that combines uncertainty sampling and space filling, design and execute the initial field experiment to obtain a structured initial dataset. ; S2, Based on the initial dataset A Bayesian linear regression surrogate model is constructed and trained on the instrument side, enabling it to output agronomic parameters estimated with confidence intervals based on the original sensor signals; the agronomic parameters include, but are not limited to, at least one of nitrogen content, chlorophyll content, biomass, leaf area index, and plant water content. S3, Based on the initial dataset Using the output of the Bayesian linear regression surrogate model, construct and train a Gaussian process regression inversion model that can directly output the posterior prediction distribution; S4. Based on the initial dataset Using agronomic and economic output data, interpretable agronomic decision functions are generated through multi-objective constrained optimization. ; S5. Closed-loop application and evolution stage: The Bayesian linear regression surrogate model, Gaussian process regression inversion model and agronomic decision function constructed in S2 to S4 are applied to new fields for fertilization decisions. The new data generated during the application process are used to continuously optimize and evolve the model and decision function through an online learning mechanism.

2. The intelligent calibration and decision-making method for rice fertilization according to claim 1, characterized in that, The initial dataset At least include: raw leaf absorbance data measured using a portable spectrometer during the critical growth period of the experimental plots. Laboratory standard nitrogen content data for corresponding leaf samples Spectral index data extracted from simultaneously acquired UAV multispectral remote sensing images and the yield of each plot measured after harvest. With cost data .

3. The intelligent calibration and decision-making method for rice fertilization according to claim 1, characterized in that, In step S2, the Bayesian probabilistic surrogate model is a Bayesian linear regression model, which establishes the original absorbance vector measured by the instrument at multiple characteristic wavelengths. Compared with laboratory standard nitrogen content Mapping relationship between them: ;in, For model parameter vectors, To augment the input vector, For noise variance; After deployment, the Bayesian probabilistic proxy model will adapt to new inputs. Output point estimates of agronomic parameters and its confidence interval , The half-width of the confidence interval at the preset confidence level.

4. The intelligent calibration and decision-making method for rice fertilization according to claim 3, characterized in that, In step S3, the specific steps for constructing the remote sensing inversion model include: The initial dataset is processed using the Bayesian probabilistic surrogate model. Based on the aforementioned remote sensing spectral features As input, point estimation To construct training labels, and for new input spectral index vectors Output posterior prediction Gaussian distribution , The mean nitrogen content is obtained from the posterior prediction. To characterize The prediction variance of the confidence level.

5. The intelligent calibration and decision-making method for rice fertilization according to claim 1, characterized in that, The S5 steps specifically include: S5.1 For the target field, the remote sensing predicted distribution is obtained using the Gaussian process regression inversion model, and then Bayesian fusion is performed with the ground measured distribution based on the Bayesian probabilistic surrogate model to obtain the calibrated crop status information. S5.2 Input the calibrated crop status information and current environmental economic parameters into the agronomic decision function to obtain fertilization recommendations; S5.

3. Collect new data generated during the application process, and use an online transfer learning strategy based on elastic weight consolidation to incrementally update the Gaussian process regression inversion model and the agronomic decision function.

6. The intelligent calibration and decision-making method for rice fertilization according to claim 5, characterized in that, The Bayesian fusion in S5.1 is specifically calculated using the following formula to determine the calibrated state distribution. : ， in and Let the mean and variance of the remote sensing prediction distribution be denoted as . and These represent the mean and variance of the measured distribution on the ground.

7. The intelligent calibration and decision-making method for rice fertilization according to claim 5, characterized in that, The online transfer learning strategy based on elastic weight consolidation in S5.3 has the following loss function for updating the Gaussian process regression model: ， in old model parameters The diagonal elements of the Fisher information matrix.

8. A smart calibration and decision-making system for rice fertilization, characterized in that, The system is used to execute the intelligent calibration and decision-making method for rice fertilization as described in any one of claims 1 to 7. The system comprises the following components connected in sequence: An active experiment design module is used to execute the hybrid active learning strategy; The intelligent perception calibration module is used to construct and deploy the Bayesian probabilistic proxy model; The probability inversion and fusion module is used to construct the Gaussian process regression inversion model and perform the Bayesian fusion. The transparent decision optimization module is used to generate and execute the agronomic decision function. The continuous evolutionary learning module is used to implement the online transfer learning based on elastic weight consolidation.

9. The intelligent calibration and decision-making system for rice fertilization according to claim 8, characterized in that, The Bayesian probabilistic surrogate model constructed by the intelligent perception calibration module is a Bayesian linear regression model, and the posterior distribution of its model parameters is based on the initial dataset. The learned information is used to calculate the confidence interval of the output.

10. The intelligent calibration and decision-making system for rice fertilization according to claim 8, characterized in that, The remote sensing inversion model constructed by the probability inversion and fusion module is a Gaussian process regression model, which takes spectral index features as input and agronomic parameter estimates with confidence intervals from the intelligent sensing calibration module as training labels.

Images