Intelligent decision-making system and method for corn fertilization based on mechanism-data dual-drive fusion

Abstract

The invention belongs to the technical field of unmanned aerial vehicle remote sensing and agriculture combination, and discloses an intelligent decision-making system and method for corn fertilization based on mechanism-data dual-drive fusion. The system comprises a mechanism simulation module, a data preprocessing module, a feature engineering module, a modeling module, a visualization module, a decision support module, a dynamic feedback correction module and a report generation and push module. According to the mechanism-data double-drive fusion-based intelligent decision-making system and method for corn fertilization, a large-area corn field block image is obtained in a short time through an unmanned aerial vehicle multispectral system, and the nitrogen diagnosis efficiency is improved; through a lightweight machine learning agent model, corn canopy leaf nitrogen nutrition parameters are accurately predicted, and a basis is provided for accurate fertilization; the system monitors the nitrogen nutrition status of the corn in real time, and provides possibility for dynamically adjusting a fertilization strategy; the multispectral remote sensing technology can perform nitrogen nutrition diagnosis under the condition of not damaging corn plants, and the corn growth environment is protected.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) remote sensing and agricultural technology, and in particular to a smart decision-making system and method for corn fertilization based on mechanism-data dual-drive fusion. Background Technology

[0002] The yield and quality of maize are constrained by nitrogen supply during critical periods, and yield and its components change significantly with increasing nitrogen fertilizer application. Therefore, efficient, accurate, and rational decision-making regarding nitrogen fertilizer application rates, and ensuring sufficient nitrogen supply and demand for maize through different fertilization methods, are crucial in production. However, traditional methods of manual ground measurement or surveys suffer from low efficiency, poor timeliness, and inconsistent standards.

[0003] In the development of modern agriculture, the rise of precision agriculture technology has provided a new path to improve agricultural production efficiency and ensure food security. Among them, unmanned aerial vehicle (UAV) remote sensing technology, as an important component of precision agriculture, has shown great potential in crop growth monitoring. In recent years, this technology has developed rapidly, and its application in the agricultural field has reached an unprecedented breadth and depth, greatly enriching the means of agricultural information acquisition. It has played an important role in crop condition monitoring, disaster assessment, yield surveys, and vegetation phenotyping, especially in the field of nitrogen nutrition diagnosis for maize, where its application has received widespread attention.

[0004] In the maize planting process, drones equipped with multispectral sensors can acquire reflectance spectral information of the maize canopy in different wavelength bands. Based on this information, various vegetation indices can be constructed, such as the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Red Edge Normalized Difference Index (NDRE). These indices can then be used to quantitatively analyze the chlorophyll content, leaf area index (LAI), and biomass of maize, providing data support for nitrogen nutrition diagnosis.

[0005] However, despite the progress made by UAV multispectral remote sensing technology in monitoring nitrogen nutrition in maize, there are still many limitations in its development:

[0006] At the sensor level, the high cost of commercial multispectral cameras undoubtedly hinders the application of the technology by small and medium-sized farmers and limits its widespread adoption.

[0007] Data processing and analysis: On the one hand, the dynamic changes in different years, seasons and maize growth stages, especially the "nitrogen dilution" effect, greatly reduce the accuracy of nitrogen content inference based on spectral information; on the other hand, in the early stages of maize growth, the canopy coverage is low, and background elements such as soil have a significant impact on spectral reflectance in multispectral or hyperspectral images, increasing the uncertainty of nitrogen estimation. Even if specialized vegetation indices and spectral unmixing methods are developed, their practical application effects still need further verification.

[0008] Furthermore, the data acquired by different sensors vary significantly in spatial and temporal resolution, spectral range, and data format, which poses challenges to data integration and analysis, and affects the accuracy and reliability of nitrogen monitoring.

[0009] Therefore, the present invention aims to provide an efficient, accurate, and real-time nitrogen nutrition diagnosis and decision-making system for maize, based on UAV remote sensing technology, to achieve high-throughput, rapid, and non-destructive detection of nitrogen nutrition parameters in maize, providing a basis for precise nitrogen management. Summary of the Invention

[0010] The purpose of this invention is to provide an intelligent decision-making system and method for maize fertilization based on mechanism-data dual-drive fusion. It utilizes a UAV multispectral system to acquire images of large-area maize fields in a short time, improving the efficiency of nitrogen nutrition diagnosis. Through a lightweight machine learning proxy model, it accurately predicts nitrogen nutrition parameters in the maize canopy leaves, providing a basis for precision fertilization. The system monitors the nitrogen nutrition status of maize in real time, promptly identifying problems and enabling adjustments to fertilization strategies. Multispectral remote sensing technology allows for nitrogen nutrition diagnosis without damaging the maize plants, protecting the maize growth environment.

[0011] To achieve the above objectives, this invention provides a maize fertilization intelligent decision-making system based on mechanism-data dual-drive fusion, including a data acquisition module for acquiring real-time data sources and mechanism model input data;

[0012] The mechanism simulation module calls the APSIM model to generate multi-scenario training data;

[0013] The data preprocessing module is used to improve data quality, eliminate noise and system errors, and perform real-time assimilation correction on real-time data sources and APSIM simulation data.

[0014] The feature engineering module constructs a unified feature space;

[0015] The modeling module trains lightweight machine learning agent models.

[0016] The visualization module, based on a lightweight machine learning agent model, constructs inverted remote sensing monitoring images of maize nitrogen nutrition parameters and combines them with APSIM time-series process variables to achieve spatiotemporal visualization of maize nitrogen nutrition parameters.

[0017] The decision support module uses a two-stage optimization mechanism to generate zonal fertilization prescription maps;

[0018] The dynamic feedback correction module dynamically adjusts the initial state parameters of APSIM based on field sensor data and actual yield measurements at harvest, and updates the feature weights of the lightweight machine learning proxy model.

[0019] The report generation and push module is used to generate and push reports containing key information.

[0020] Preferably, the data acquisition module is deployed in field plots with different nitrogen application rates, and a control experiment is set up; multispectral images are acquired by using a drone equipped with a multispectral camera; ground true value sampling is carried out simultaneously to measure leaf nitrogen nutrition parameters;

[0021] Ground-based data collection included sampling at the corn seedling stage, jointing stage, large trumpet stage, and grain-filling stage.

[0022] Leaf nitrogen nutrient parameters were measured: Five maize plants were randomly selected from each field plot, and the third functional leaf from the top of each plant was taken as the measurement sample; the total nitrogen content was determined by the Kjeldahl method; and the SPAD value of the leaves was measured by a chlorophyll meter.

[0023] Sampling data from plots with different nitrogen application rates were used as input data for the mechanistic model.

[0024] Preferably, the data preprocessing module includes correcting the real-time data and assimilating it with the APSIM simulated data in real time; wherein, the preprocessing includes:

[0025] Image radiometric correction, through whiteboard correction and atmospheric correction, eliminates differences in lighting conditions;

[0026] Geometric correction and stitching, image registration with geographic coordinates;

[0027] Image enhancement and outlier handling improve image clarity and feature contrast while removing invalid data points.

[0028] Preferably, the feature engineering module includes:

[0029] (1) Calculate multiple vegetation indices, including Normalized Difference Vegetation Index (NDVI), Red Edge Normalized Difference Index (NDRE), Enhanced Vegetation Index (EVI), and Soil Adjusted Vegetation Index (SAVI); the calculation formulas are as follows:

[0030]

[0031] Among them, NIR is the 842nm band and Red is the 668nm band;

[0032]

[0033] RedEdge is the 717nm band;

[0034]

[0035] Blue represents the 475nm wavelength band;

[0036]

[0037] Among them, the soil adjustment factor L = 0.5;

[0038] (2) Extract texture features based on gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP);

[0039] (3) Mechanistic feature extraction: Extracting features of the jointing stage and the large trumpet mouth stage from APSIM simulation data;

[0040] (4) Through correlation analysis and random forest (RF) importance ranking, features with correlation |r|>0.6 with nitrogen nutrition parameters were screened.

[0041] (5) Integrate the spectral, texture and mechanism characteristics after screening.

[0042] Preferably, the modeling module includes:

[0043] (1) Dataset partitioning: The APSIM output data and the spectral, texture and mechanism fusion features selected by the feature engineering module are divided into training set and test set in a 7:3 ratio; the training set data is augmented.

[0044] (2) Model training: Train the random forest (RF) fertilizer application prediction model and establish a mapping relationship with the optimal base fertilizer amount; train the support vector machine (SVM) fertility period diagnosis model; train the deep learning CNN spatial feature extraction model.

[0045] (3) Model evaluation and selection: Cross-validation was used, and the root mean square error (RMSE) and the average coefficient of determination (R²) were calculated. 2 Indicators are used to evaluate model stability; the test set R is used. 2 Maximize the optimal model as the criterion;

[0046] (4) Model fusion: Select three trained models as models to be fused, and use a weighted fusion method to fusion the models according to the test set R. 2 Weight fusion;

[0047] (5) Model storage: The fused model is stored as the final model using the open neural network exchange format ONNX.

[0048] Preferably, the visualization module includes:

[0049] (1) Rasterization: The model prediction results are divided into 5m×5m grids; each grid is assigned a predicted value of nitrogen nutrition parameters; based on the prediction results of the above grids, a spatially continuous distribution map is constructed, namely the remote sensing monitoring image of maize nitrogen nutrition parameters.

[0050] (2) Color coding and thematic map generation: The rasterized data is divided into different levels according to the nitrogen nutrient level and displayed by color coding; based on the color coding rules, the rasterized data is generated into TIFF format thematic maps;

[0051] (3) Develop interactive interfaces: Develop web or mobile interfaces that support zoom and query functions and automatically generate PDF reports containing key indicators and recommendations.

[0052] Preferably, the decision support module includes:

[0053] A two-stage optimization mechanism was used to generate a zonal fertilization prescription map, the details of which are as follows:

[0054] Phase 1: When rapidly pre-screening fertilization schemes using the Random Forest (RF) proxy model, the APSIM model is called to generate a simulation database containing N fertilization schemes. Based on initial soil data, variety characteristics, and weather forecasts, fertilization schemes are rapidly pre-screened and preliminary fertilization schemes are output.

[0055] Phase 2: Combining the spatial constraints of deep learning CNNs, the NSGA-II algorithm is used to optimize the balance between economic benefits and environmental risks.

[0056] Among them, when deep learning CNN is used to process UAV multispectral images during the jointing stage, the spatial variation of canopy nitrogen is extracted to generate a field management zoning map as a spatial constraint.

[0057] Support Vector Machine (SVM) is used to diagnose canopy nitrogen status and associated environmental risks in real time.

[0058] The NSGA-II multi-objective algorithm solves the Pareto optimal fertilization scheme by using a genetic algorithm in conjunction with a random forest RF surrogate model, and generates a zonal fertilization prescription map.

[0059] Preferably, the NSGA-II multi-objective algorithm solves the Pareto optimal fertilization scheme through a genetic algorithm in conjunction with a random forest (RF) surrogate model, as detailed below:

[0060] First, RF proxy model pre-training: RF is trained using APSIM simulated data to improve the ability to quickly predict the effects of fertilization programs;

[0061] Secondly, NSGA-II genetic operations: candidate fertilization schemes are generated through encoding, selection, crossover, and mutation, and the RF proxy model is called to evaluate the economic benefits and environmental risks of the schemes;

[0062] Finally, spatial constraint correction: Combining the field partitioning information extracted by CNN, the spatial consistency of the candidate schemes is checked, and the Pareto optimal fertilization scheme that conforms to field heterogeneity is finally output.

[0063] Preferably, the report generation and push module includes:

[0064] (1) Generation of basic field information: Output field area, coordinates, and soil type;

[0065] (2) Nitrogen nutrition status analysis: Calculate the mean and spatial variation coefficient of nitrogen nutrition parameters to form the status analysis results;

[0066] (3) Fertilizer application recommendation table: Generate a recommendation table for nitrogen application amount, fertilization time and fertilization method for each zone;

[0067] (4) Prediction performance evaluation: Presenting the model error range;

[0068] (5) Push mechanism: The report is automatically generated within 24 hours after sampling and pushed to the farmer's account through the APP. It also supports sending the report link from the client.

[0069] The intelligent decision-making method for maize fertilization based on mechanism-data dual-drive fusion includes the following steps:

[0070] S1. Acquire multispectral images of cornfields, ground truth sampling, and mechanistic model input data; use APSIM to generate a simulation database containing N fertilization schemes;

[0071] S2. Preprocess the acquired data and perform real-time assimilation and correction with the APSIM simulation database;

[0072] S3. Train the random forest (RF) model and establish the mapping relationship between initial soil nitrogen content, weather forecast, variety characteristics and optimal base fertilizer amount;

[0073] S4. Analyze drone images during the jointing stage using CNN to divide the field into management zones;

[0074] S5. Integrate the nitrogen deficit index and zoning map from SVM diagnosis to generate an optimized topdressing scheme;

[0075] S6. Update the surrogate model weights based on the actual yield feedback at harvest time.

[0076] Therefore, the intelligent decision-making system and method for maize fertilization based on the mechanism-data dual-drive fusion described above have the following beneficial effects:

[0077] (1) High efficiency: The present invention uses a UAV multispectral system, which can acquire images of large areas of cornfields in a short time, significantly improving the efficiency of nitrogen nutrition diagnosis.

[0078] (2) Accuracy: This invention can accurately predict the nitrogen nutrition parameters of maize canopy leaves through a lightweight machine learning proxy model, providing real-time support for precision fertilization: The system can monitor the nitrogen nutrition status of maize in real time, promptly identify problems of insufficient or excessive nutrition, and provide the possibility for timely adjustment of fertilization strategies.

[0079] (3) Non-destructive: The present invention uses multispectral remote sensing technology, which can perform nitrogen nutrition diagnosis without damaging the corn plant, thus protecting the corn growth environment.

[0080] (4) Adaptability: This invention is applicable to corn planting areas of all sizes, especially areas that require efficient and precise nitrogen fertilizer management.

[0081] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of the intelligent decision-making system for corn fertilization based on mechanism-data dual-drive fusion of the present invention;

[0083] Figure 2 This is a schematic diagram of the modeling module structure in an embodiment of the present invention;

[0084] Figure 3 This is a schematic diagram of the decision support module structure in an embodiment of the present invention;

[0085] Figure 4 This is a flowchart of the NSGA-II algorithm in the decision support module of this invention.

[0086] Figure 5 This is a flowchart of the intelligent decision-making method for corn fertilization based on the fusion of mechanism and data in this invention. Detailed Implementation

[0087] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0088] like Figure 1 As shown, the intelligent decision-making system for corn fertilization based on mechanism-data dual-drive fusion includes: a data acquisition module, used to acquire real-time data sources and mechanism model input data;

[0089] The mechanism simulation module calls the APSIM model to generate multi-scenario training data;

[0090] The data preprocessing module is used to improve data quality, eliminate noise and system errors, and perform real-time assimilation correction on real-time data sources and APSIM simulation data.

[0091] The feature engineering module constructs a unified feature space;

[0092] The modeling module trains lightweight machine learning agent models.

[0093] The visualization module, based on a lightweight machine learning agent model, constructs inverted remote sensing monitoring images of maize nitrogen nutrition parameters and combines them with APSIM time-series process variables to achieve spatiotemporal visualization of maize nitrogen nutrition parameters.

[0094] The decision support module uses a two-stage optimization mechanism to generate zonal fertilization prescription maps;

[0095] The dynamic feedback correction module dynamically adjusts the initial state parameters of APSIM based on field sensor data and actual yield measurements at harvest, and updates the feature weights of the lightweight machine learning proxy model.

[0096] The report generation and push module is used to generate and push reports containing key information.

[0097] Example

[0098] This invention is a corn fertilization intelligent decision-making system based on mechanism-data dual-drive fusion, specifically including the following:

[0099] 1. Data acquisition module, used to acquire real-time data sources and input data for the mechanism model.

[0100] 1. Field plot setup, details are as follows:

[0101] 1.1 Plot Division Rules: The maize field to be tested was divided into standard plots of 20m × 20m, with 5 nitrogen application gradients set up, with gradient values ​​of 0 kg / ha, 60 kg / ha, 120 kg / ha, 180 kg / ha, and 240 kg / ha respectively; each gradient was replicated 3 times, resulting in a total of 15 plots. A 1m wide isolation strip was set between the plots to prevent nutrient migration.

[0102] 1.2 Basic Parameter Recording: Before planting, a mixed soil sample of the 0-20cm soil layer of each plot was collected using the five-point sampling method; the basic physicochemical properties of the soil sample were determined using a soil nutrient rapid tester, such as pH value: 5.5-8.5, organic matter: 15-30g / kg, available nitrogen: 50-150mg / kg; the soil nutrient rapid tester was TPY-6PC.

[0103] 2. Data collection, as shown below:

[0104] 2.1. UAV multispectral image acquisition.

[0105] First, the drone configuration: a DJI Matrice 300 RTK equipped with a MicaSense RedEdge-MX multispectral camera, covering 5 bands, specifically wavelengths of 475nm, 560nm, 668nm, 717nm, and 842nm, with a resolution of 1280×960; the RTK module is installed to achieve centimeter-level positioning (horizontal accuracy ±1cm+1ppm, vertical accuracy ±2cm+1ppm).

[0106] Secondly, the flight parameter settings are as follows:

[0107] Flight altitude: 100m (ground resolution 10cm / pixel).

[0108] Flight speed: 5 m / s.

[0109] Forward overlap: 80%, Lateral overlap: 70%.

[0110] Flight time: Choose a clear, cloudless day between 9:00 and 15:00, with a solar altitude angle ≥45°.

[0111] 2.2 Data storage: The image format is TIFF, with GPS coordinates (WGS84 coordinate system) and shooting timestamp. The data volume of a single flight is about 8GB, which is transmitted back to the server in real time via a 5G module.

[0112] 2.3 Ground truth data collection, details of which are as follows:

[0113] Sampling frequency: once each during the corn seedling stage (6-leaf stage), jointing stage, large trumpet stage, and grain-filling stage.

[0114] Leaf nitrogen nutrient parameters were measured: Five maize plants were randomly selected from each plot, and the third functional leaf from the top of each plant was taken; the total nitrogen content was determined by the Kjeldahl method (FOSS Kjeltec 8400), with a detection accuracy of ±0.01%; the SPAD-502Plus chlorophyll meter was used to measure the chlorophyll content at three points on each leaf: leaf tip, leaf middle, and leaf base, and the average value was taken as the SPAD value of that leaf, with an accuracy of ±1.

[0115] Data matching: The coordinates of the sampling points were recorded using a handheld GPS device (accuracy ±30cm), and matched one-to-one with the pixels in the drone image.

[0116] Second, the data preprocessing module corrects the real-time data and assimilates it with the APSIM simulated data in real time.

[0117] 1. Image radiometric correction, details of which are as follows:

[0118] First, whiteboard calibration: A diffuse whiteboard with a reflectivity of 99% was photographed before flight to provide baseline data for subsequent atmospheric calibration.

[0119] Secondly, atmospheric correction: atmospheric correction was performed using the FLAASH module of ENVI 5.6 software. The input atmospheric model was a mid-latitude summer atmospheric model, and the aerosol type was rural. After correction, the reflectivity error of each band was ≤3%.

[0120] 2. Geometric correction and splicing, details of which are as follows:

[0121] Control point selection: Mark prominent features in the image, such as field ridges and utility poles, and match them with high-resolution Google Earth imagery. Select ≥8 control points for each image, with a correction error ≤1 pixel.

[0122] Mosaic: Using Agisoft Metashape, images are stitched together using the bundle adjustment method to generate a seamless DOM (digital orthophoto map). The overall deviation of the stitched image is ≤2 pixels.

[0123] 3. Image enhancement and outlier handling, details of which are as follows:

[0124] Enhancement processing: Gaussian filtering with a standard deviation of 1.0 is used to remove noise; histogram equalization is used to improve contrast; vegetation features are preserved.

[0125] Outlier removal: Z-score test was performed on the reflectance data, and values ​​|Z|>3 were considered outliers; inverse distance weighted interpolation (IDW) was used to complete the data and ensure data integrity >99%.

[0126] 4. Perform real-time assimilation and correction between the preprocessed real-time data and the APSIM simulation data.

[0127] III. Feature Engineering Module, used to extract features that are highly correlated with nitrogen nutrition.

[0128] 1. Feature extraction.

[0129] Based on Python's GDAL library, this tool calculates various vegetation indices, including the Normalized Difference Vegetation Index (NDVI), the Red Edge Normalized Difference Index (NDRE), the Enhanced Vegetation Index (EVI), and the Soil-Adjusted Vegetation Index (SAVI), as detailed below:

[0130]

[0131] Among them, NIR is the 842nm band and Red is the 668nm band;

[0132]

[0133] RedEdge is the 717nm band;

[0134]

[0135] Blue represents the 475nm wavelength band;

[0136]

[0137] Among them, the soil adjustment factor L = 0.5.

[0138] 2. Extract texture features, the details of which are as follows:

[0139] First, contrast, correlation, energy, and entropy are calculated based on the gray-level co-occurrence matrix (GLCM). When calculating contrast, a distance of 1 and angles of 0°, 45°, 90°, and 135° are used, and the average value of the calculation results for each angle is taken. The calculation window for correlation, energy, and entropy is 5×5 pixels.

[0140] Secondly, the texture histogram is extracted using Local Binary Pattern (LBP) with parameters set to a radius of 3 pixels and a neighborhood of 8 points.

[0141] 3. Mechanistic feature extraction: Extract features of the jointing stage and the large trumpet mouth stage from APSIM simulation data.

[0142] 4. Feature filtering, details are as follows:

[0143] First, correlation analysis: calculate the Pearson correlation coefficient between each feature and leaf nitrogen content, and screen features with |r|>0.6, such as NDRE, EVI, and contrast.

[0144] Secondly, Random Forest (RF) importance ranking: Train the initial Random Forest (RF) model and remove features with an importance ratio of <10% to reduce redundant variables.

[0145] Finally, collinearity test: The variance inflation factor (VIF) test is used to remove collinearity features with VIF>10. When NDVI and EVI are collinear, EVI is retained.

[0146] 4. Feature fusion, details of which are as follows:

[0147] First, data standardization: The Min-Max normalization method is used to scale the feature values ​​to the [0,1] interval, as shown below:

[0148] x′=max(x)-min(x)x-min(x);

[0149] Where x′ is the normalized value of the original data x; max(x) is the maximum value of the original dataset; min(x) is the minimum value of the original dataset; and x is the value of a single sample in the original data.

[0150] Secondly, Principal Component Analysis (PCA): This method reduces the dimensionality of the standardized features and retains the principal components with a cumulative variance contribution rate of ≥95%. The number of principal components is usually the first 3-5 to reduce the complexity of the model.

[0151] IV. Modeling module: Training lightweight machine learning agent models, such as... Figure 2 As shown.

[0152] 1. Dataset partitioning, details are as follows:

[0153] First, the APSIM output data and the spectral, texture, and mechanism fusion features selected by the feature engineering module are divided into training and testing sets in a 7:3 ratio. Stratified sampling is used to ensure that the sample ratio of each nitrogen application gradient is consistent.

[0154] Secondly, the training set data is augmented: the training set is randomly rotated (±15°), scaled (0.8-1.2 times), and noise is added (Gaussian noise σ=0.01) to expand the sample size to twice the original size.

[0155] 2. Model Training: Train the Random Forest (RF) fertilizer application prediction model and establish its mapping relationship with the optimal base fertilizer amount; train the Support Vector Machine (SVM) fertilization period diagnosis model; train the Deep Learning CNN spatial feature extraction model.

[0156] 3. Model Evaluation and Selection: Five-fold cross-validation is used, and the root mean square error (RMSE) and average coefficient of determination (R²) are calculated. The standard deviation is also calculated (standard deviation < 0.05 to ensure model stability). The best model is selected based on maximizing the R² of the test set. Random forest (RF) is usually preferred (balancing accuracy and computational efficiency). When the amount of data is ≥ 1000 samples, deep learning CNN is selected, as shown in Table 1.

[0157] Table 1. Parameters and Evaluation Indicators for Different Regression Algorithms

[0158]

[0159] 4. Model fusion, details of which are as follows:

[0160] First, the three best-performing models were selected as the models to be fused, including deep learning CNN, random forest RF, and support vector machine SVM.

[0161] Then, a weighted fusion method is used to combine the models to be fused according to the test set R. 2 Weight fusion is performed with a weight ratio of 0.5:0.3:0.2. The resulting R value is... 2 Increase by ≥2%.

[0162] 5. Model storage: The fused model is stored as the final model using the open neural network exchange format ONNX. The final model size is controlled within 50MB to ensure fast loading on mobile devices.

[0163] V. Visualization Module: Based on a lightweight machine learning proxy model, remote sensing monitoring images of maize nitrogen nutrient parameters are constructed. Combined with APSIM time-series process variables, information visualization of maize nitrogen nutrient parameters is achieved.

[0164] First, the model prediction results are divided into 5m×5m grids; each grid is assigned a predicted value for nitrogen nutrition parameters; based on the prediction results of the above grids, a spatially continuous distribution map is constructed, which is the inversion remote sensing monitoring image of maize nitrogen nutrition parameters.

[0165] Secondly, color coding and thematic map generation: the rasterized data is divided into different levels according to nitrogen nutrient levels and displayed by color coding; based on the color coding rules, the rasterized data is generated into TIFF format thematic maps; the color coding rules are shown in Table 2.

[0166] Table 2. Color Coding Rules for Nitrogen Content

[0167] Nitrogen content Encoded Colors State Conclusion <2.0％ red Severe nitrogen deficiency 2.0％-3.0％ yellow Mild nitrogen deficiency 3.0％-4.0％ green normal >4.0％ blue Nitrogen surplus

[0168] Finally, an interactive interface was developed to support functions such as zooming and querying. Diagnostic reports are automatically generated, producing PDF reports containing key indicators and recommendations, as detailed below:

[0169] Web-based: A map component developed using Leaflet supports 1-20 levels of zoom, allows users to select and query the average nitrogen content of a region, and overlays nitrogen data from different growth stages for historical comparison.

[0170] Mobile devices: Integrate offline maps into the Flutter app, supporting the display of nitrogen content details by clicking on any location, with a response time of less than 1 second.

[0171] VI. Decision Support Module: Employs a two-stage optimization mechanism to generate zonal fertilization prescription maps, such as... Figure 3 As shown.

[0172] Phase 1: Rapidly pre-screen fertilization schemes using a random forest (RF) proxy model.

[0173] First, the APSIM model was called to generate more than 1,000 sets of simulation data, covering different soil nitrogen content, nitrogen application rate, and meteorological conditions.

[0174] Key output indicators: Dependent variables (related to optimization objectives) include maize yield (kg / ha) and nitrogen leaching amount (kg / ha), which characterizes environmental risk.

[0175] The independent variables (fertilization program parameters) include: basal fertilizer amount (N1), topdressing amount at the jointing stage (N2), topdressing amount at the large bell stage (N3), and fertilization time (T1-T3).

[0176] Secondly, the training and optimization of the RF proxy model.

[0177] Model structure: Input layer: 4 parameters including N1, N2, N3, and average temperature; Output layer: 2 target values ​​including yield and nitrogen leaching.

[0178] Training optimization: Grid search was used for parameter tuning (n_estimators=200, max_depth=15) to ensure that the prediction error RMSE of RF to APSIM is ≤5%. By replacing the calculation of APSIM, the evaluation time of a single scheme was reduced from 2 hours to 0.1 seconds.

[0179] Finally, based on initial soil data, variety characteristics, and weather forecasts, a preliminary fertilization plan is generated.

[0180] The second stage involves combining deep learning CNN spatial constraints to overcome the limitations of uniform fertilization and make the scheme conform to field heterogeneity; and using the NSGA-II algorithm to optimize the balance between economic benefits and environmental risks.

[0181] First, when deep learning CNN is used to process UAV multispectral images during the jointing stage, the spatial variation of canopy nitrogen is extracted to generate a field management zoning map as a spatial constraint; then, support vector machine (SVM) is used to diagnose the canopy nitrogen status and associated environmental risks in real time.

[0182] Secondly, the NSGA-II algorithm is used to optimize economic benefits and environmental risks. Through a genetic algorithm framework, the NSGA-II algorithm uses the RF surrogate model as an "efficient substitute for the mechanistic model," achieving closed-loop optimization of "fast search - accurate evaluation - spatial constraints," such as... Figure 4 As shown.

[0183] (1) A fertilization scheme using real number encoding, where each chromosome corresponds to a partition, such as [N1_partition 1, N2_partition 1, N3_partition, N1_partition 2, N2_partition 2, N3_partition 2, ...];

[0184] The partitions are generated by CNN processing of drone images, and there are 3-5 partitions in total, reflecting the spatial variation of nitrogen.

[0185] Coding constraints: Fertilizer application range [N_min, N_max], such as during the jointing stage, calculated based on soil carrying capacity, N2∈[30,120]kg / ha).

[0186] (2) Initial population generation: 100 initial schemes are randomly generated. Through spatial smoothness filtering, the difference in fertilization amount between adjacent zones is calculated. Schemes with a difference of ≤20kg / ha are retained to avoid abrupt changes in fertilization at the zone boundary, which conforms to the field nutrient diffusion law.

[0187] (3) RF proxy model invocation: For each candidate scheme (chromosome), input the fertilization parameters and meteorological data for the corresponding region. RF outputs the yield and nitrogen leaching amount of the corresponding region within 0.1 seconds, and summarizes them to obtain the first global objective function f1 (representing economic benefits, corresponding to the corn yield under the candidate scheme, unit: kg / ha) and the second objective function f2 (representing environmental risk, corresponding to the nitrogen leaching amount under the candidate scheme, unit: kg / ha). Compared with the APSIM model, which requires 2 hours per scheme for computation, the optimization efficiency is improved by 7200 times.

[0188] (4) Genetic operations: Combining biological constraints of selection-crossover-mutation, non-dominated sorting and crowding calculation: Sort the population according to the NSGA-II standard procedure and retain non-dominated solutions; when calculating crowding, penalize schemes with fertilizer application at the jointing stage > N3 (large trumpet stage) (based on the nitrogen requirement law of maize: nitrogen requirement is highest at the large trumpet stage) to avoid biological inconsistencies.

[0189] Selection operator: Tournament selection (tournament size=3) is adopted, giving priority to schemes with high f1 and low f2, while ensuring that the fertilizer application in each zone meets the nitrogen threshold in Table 3.

[0190] Table 3 Nitrogen Nutrition Thresholds During Maize Growth Period

[0191] reproductive period Suitable nitrogen content Seedling stage 3.0％-3.5％ Propagation period 3.5％-4.0％ Big trumpet mouth period 4.0％-4.5％ Grouting period: 3.5％-4.0％

[0192] Crossover operator: Simulated binary crossover (SBX) is adopted with a crossover probability of 0.8. Gene segments at the partition boundary are forced to crossover. After crossover, the difference in fertilization amount between adjacent partitions is ≤15kg / ha, thus maintaining spatial continuity.

[0193] Mutation operator: polynomial mutation, mutation probability 0.1, the mutation amplitude is related to the maize growth stage (fertilization amplitude at the large trumpet stage is ±20%, which is higher than ±10% at the seedling stage), which is consistent with the nitrogen sensitivity of critical periods.

[0194] By incorporating the nitrogen requirement patterns of maize into the sorting and mutation steps, such as the peak nitrogen requirement during the large trumpet stage, biologically invalid solutions are eliminated, thereby improving the practical operability of the scheme.

[0195] (5) Convergence judgment and optimal solution output.

[0196] Convergence criteria: ≥100 iterations; Pareto front crowding rate <5% over 20 consecutive iterations to ensure solution stability; standard deviation of fertilizer application rate in key regions <5 kg / ha to avoid excessive fluctuations in the solution.

[0197] Solution selection: From the Pareto optimal set, based on user preferences (e.g., prioritizing economic benefits or environmental friendliness), recommend three alternative solutions and label them:

[0198] Spatial adaptability: The matching degree between each scheme and the CNN partition (≥85% is excellent);

[0199] Feasibility: Based on the meteorological data for the next 7 days (during periods without precipitation), mark the fertilization window period.

[0200] The fertilization plan has been refined.

[0201] First, calculate the fertilizer application rate: Based on the nutrient deficit and fertilizer utilization rate, calculate the optimal fertilizer application rate as follows:

[0202]

[0203] (Note: The fertilizer requirement calculation for the target yield involves an estimated leaf biomass of 15,000 kg / ha and a nitrogen use efficiency of 0.4).

[0204] Secondly, the recommended fertilization time is as follows: Based on the corn growth period and meteorological data, the best fertilization time is recommended: Use the China Weather Network API to obtain meteorological data, select a period with no precipitation in the next 3 days and wind speed <3m / s, and prioritize fertilization between 9-11 am.

[0205] Finally, fertilization methods are recommended: fertilization methods such as broadcasting and drip irrigation are recommended based on the terrain and equipment conditions, as shown in Table 4.

[0206] Table 4 Recommended Nitrogen Fertilization Scheme for Maize

[0207]

[0208]

[0209] VII. Dynamic Feedback Correction Module: Based on field sensor data and actual yield measurements at harvest, the module dynamically adjusts the initial state parameters of APSIM and updates the feature weights of the lightweight machine learning proxy model.

[0210] 8. Based on the above, generate a report and push it to farmers. The specific content is as follows:

[0211] (1) Basic information of the field: Output basic parameters such as area, coordinates, and soil type.

[0212] (2) Analysis of the current status of nitrogen nutrition: Calculate the mean value of nitrogen nutrition parameters and the coefficient of spatial variation to form the results of the current status analysis.

[0213] (3) Fertilizer application suggestion table: Generate a suggestion table for nitrogen application amount, fertilization time and fertilization method for each zone.

[0214] (4) Prediction effect evaluation: Present the model error range.

[0215] (5) Push mechanism: The report is automatically generated within 24 hours after sampling and pushed to the farmer's account through the APP. The report link can also be sent via WeChat or SMS.

[0216] In this embodiment, the system deployment and maintenance are implemented as follows:

[0217] First, server configuration: Hardware: Alibaba Cloud CS server, with specific parameters of 8-core CPU, 16GB memory, and 1TB SSD; Software deployment: Docker containerized application web services, database, and model inference engine.

[0218] Secondly, the drone ground station is equipped with an industrial-grade tablet computer running Android 11, a 4G module, and data acquisition and control software that supports offline caching of data from 1,000 acres of farmland.

[0219] Finally, sensor calibration: the multispectral camera is calibrated with a standard whiteboard every 3 months, and the GPS module is synchronized with base station data once a month to ensure positioning accuracy.

[0220] A smart decision-making method for maize fertilization based on mechanism-data dual-drive fusion, such as Figure 5 As shown, it includes the following steps:

[0221] S1. Acquire multispectral images of cornfields, ground truth sampling, and mechanistic model input data; use APSIM to generate a simulation database containing N fertilization schemes.

[0222] S2. Preprocess the acquired data and perform real-time assimilation and correction with the APSIM simulation database.

[0223] S3. Train the random forest (RF) model to establish the mapping relationship between initial soil nitrogen content, weather forecast, variety characteristics and optimal base fertilizer amount.

[0224] S4. Analyze drone images during the jointing stage using CNN to divide the field into management zones.

[0225] S5. Integrate the nitrogen deficit index and zoning map from SVM diagnosis to generate an optimized topdressing scheme.

[0226] S6. Update the surrogate model weights based on the actual yield feedback at harvest time.

[0227] Therefore, the intelligent decision-making system and method for maize fertilization based on the mechanism-data dual-drive fusion described above have the following beneficial effects:

[0228] (1) High efficiency: The present invention uses a UAV multispectral system, which can acquire images of large areas of cornfields in a short time, significantly improving the efficiency of nitrogen nutrition diagnosis.

[0229] (2) Accuracy: This invention can accurately predict the nitrogen nutrition parameters of maize canopy leaves through a lightweight machine learning proxy model, providing real-time support for precision fertilization: The system can monitor the nitrogen nutrition status of maize in real time, promptly identify problems of insufficient or excessive nutrition, and provide the possibility for timely adjustment of fertilization strategies.

[0230] (3) Non-destructive: The present invention uses multispectral remote sensing technology, which can perform nitrogen nutrition diagnosis without damaging the corn plant, thus protecting the corn growth environment.

[0231] (4) Adaptability: This invention is applicable to corn planting areas of all sizes, especially areas that require efficient and precise nitrogen fertilizer management.

[0232] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A corn fertilization intelligent decision-making system based on mechanism-data dual-drive fusion, characterized in that, Includes a data acquisition module, used to acquire real-time data sources and input data for the mechanism model; The mechanism simulation module calls the APSIM model to generate multi-scenario training data; The data preprocessing module is used to improve data quality, eliminate noise and system errors, and perform real-time assimilation correction on real-time data sources and APSIM simulation data. The feature engineering module constructs a unified feature space; The modeling module trains lightweight machine learning agent models. The visualization module, based on a lightweight machine learning agent model, constructs inverted remote sensing monitoring images of maize nitrogen nutrition parameters and combines them with APSIM time-series process variables to achieve spatiotemporal visualization of maize nitrogen nutrition parameters. The decision support module uses a two-stage optimization mechanism to generate zonal fertilization prescription maps; The dynamic feedback correction module dynamically adjusts the initial state parameters of APSIM based on field sensor data and actual yield measurements at harvest, and updates the feature weights of the lightweight machine learning proxy model. The report generation and push module is used to generate and push reports containing key information.

2. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The data acquisition module is deployed in field plots with different nitrogen application rates to set up control experiments; multispectral images are acquired using a drone equipped with a multispectral camera; ground-based true value sampling is performed simultaneously to measure leaf nitrogen nutrition parameters; Ground-based data collection included sampling at the corn seedling stage, jointing stage, large trumpet stage, and grain-filling stage. Leaf nitrogen nutrition parameter determination: Five maize plants were randomly selected from each field plot, and the third functional leaf from the top of each plant was taken as the test sample. Total nitrogen content was determined using the Kjeldahl method; SPAD values ​​of leaves were determined using a chlorophyll meter. Sampling data from plots with different nitrogen application rates were used as input data for the mechanistic model.

3. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The data preprocessing module includes correcting real-time data and assimilating it with APSIM simulated data in real time; the preprocessing includes: Image radiometric correction, through whiteboard correction and atmospheric correction, eliminates differences in lighting conditions; Geometric correction and stitching, image registration with geographic coordinates; Image enhancement and outlier handling improve image clarity and feature contrast while removing invalid data points.

4. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The feature engineering module includes: (1) Calculate multiple vegetation indices, including Normalized Difference Vegetation Index (NDVI), Red Edge Normalized Difference Index (NDRE), Enhanced Vegetation Index (EVI), and Soil Adjusted Vegetation Index (SAVI). The calculation formulas are as follows: Among them, NIR is the 842nm band and Red is the 668nm band; RedEdge is the 717nm band; Blue represents the 475nm wavelength band; Among them, the soil adjustment factor L = 0.5; (2) Extract texture features based on gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP); (3) Mechanistic feature extraction: Extracting features of the jointing stage and the large trumpet mouth stage from APSIM simulation data; (4) Through correlation analysis and random forest (RF) importance ranking, features with correlation |r|>0.6 with nitrogen nutrition parameters were screened. (5) Integrate the spectral, texture and mechanism characteristics after screening.

5. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 4, characterized in that, The modeling module includes: (1) Dataset partitioning: The APSIM output data and the spectral, texture and mechanism fusion features selected by the feature engineering module are divided into training set and test set in a 7:3 ratio; the training set data is augmented. (2) Model training: Train the random forest (RF) fertilizer application prediction model and establish a mapping relationship with the optimal base fertilizer amount; train the support vector machine (SVM) fertility period diagnosis model; train the deep learning CNN spatial feature extraction model. (3) Model evaluation and selection: Cross-validation was used, and the root mean square error (RMSE) and the average coefficient of determination (R²) were calculated. 2 Indicators are used to evaluate model stability; the test set R is used. 2 Maximize the optimal model as the criterion; (4) Model fusion: Select the trained model as the model to be fused, and use the weighted fusion method to fuse the models to be fused according to the test set R2 weight; (5) Model storage: The fused model is stored as the final model using the open neural network exchange format ONNX.

6. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The visualization module includes: (1) Rasterization: The model prediction results are divided into 5m×5m grids; each grid is assigned a predicted value of nitrogen nutrition parameters; based on the prediction results of the above grids, a spatially continuous distribution map is constructed, namely the remote sensing monitoring image of maize nitrogen nutrition parameters. (2) Color coding and thematic map generation: The rasterized data is divided into different levels according to the nitrogen nutrient level and displayed by color coding; based on the color coding rules, the rasterized data is generated into TIFF format thematic maps; (3) Develop interactive interfaces: Develop web or mobile interfaces that support zoom and query functions and automatically generate PDF reports containing key indicators and recommendations.

7. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The decision support module includes: A two-stage optimization mechanism was used to generate a zonal fertilization prescription map, the details of which are as follows: Phase 1: When pre-screening fertilization schemes using the Random Forest (RF) proxy model, the APSIM model is called to generate a simulation database containing N fertilization schemes. Based on initial soil data, variety characteristics, and weather forecasts, fertilization schemes are pre-screened and preliminary fertilization schemes are output. Phase 2: Combining the spatial constraints of deep learning CNNs, the NSGA-II algorithm is used to optimize the balance between economic benefits and environmental risks. Among them, when deep learning CNN is used to process UAV multispectral images during the jointing stage, the spatial variation of canopy nitrogen is extracted to generate a field management zoning map as a spatial constraint. The support vector machine (SVM) model is used to diagnose canopy nitrogen status and associated environmental risks in real time. The NSGA-II multi-objective algorithm solves the Pareto optimal fertilization scheme by using a genetic algorithm in conjunction with a random forest RF surrogate model, and generates a zonal fertilization prescription map.

8. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The NSGA-II multi-objective algorithm solves for the Pareto optimal fertilization scheme using a genetic algorithm in conjunction with a random forest (RF) surrogate model. The specific process is as follows: First, the random forest RF surrogate model is pre-trained: the random forest RF surrogate model is trained with APSIM simulated data to enable it to quickly predict the effects of fertilization programs. Secondly, NSGA-II genetic operations: candidate fertilization schemes are generated through encoding, selection, crossover, and mutation, and the economic benefits and environmental risks of the schemes are evaluated by calling the random forest RF surrogate model; Finally, spatial constraint correction: Combining the field partitioning information extracted by deep learning CNN, the spatial consistency of the candidate schemes is checked, and the Pareto optimal fertilization scheme that conforms to field heterogeneity is finally output.

9. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion as described in claim 1, characterized in that, The report generation and push module includes: (1) Generation of basic field information: Output field area, coordinates, and soil type; (2) Nitrogen nutrition status analysis: Calculate the mean and spatial variation coefficient of nitrogen nutrition parameters to form the status analysis results; (3) Fertilizer application recommendation table: Generate a recommendation table for nitrogen application amount, fertilization time and fertilization method for each zone; (4) Prediction performance evaluation: Presenting the model error range; (5) Push mechanism: The report is automatically generated within 24 hours after sampling and pushed to the farmer's account through the APP. It also supports sending the report link from the client.

10. The intelligent decision-making system for maize fertilization based on mechanism-data dual-drive fusion according to claims 1-9, applied to the intelligent decision-making method for maize fertilization based on mechanism-data dual-drive fusion, characterized in that, Includes the following steps: S1. Acquire multispectral images of cornfields, ground truth sampling, and mechanistic model input data; use APSIM to generate a simulation database containing N fertilization schemes; S2. Preprocess the acquired data and perform real-time assimilation and correction with the APSIM simulation database; S3. Train the random forest (RF) model and establish the mapping relationship between initial soil nitrogen content, weather forecast, variety characteristics and optimal base fertilizer amount; S4. Use deep learning CNN to analyze drone images during the jointing stage and divide the field into management zones; S5. Integrate the nitrogen deficit index and partition map diagnosed by support vector machine (SVM) to generate an optimized topdressing scheme. S6. Update the surrogate model weights based on the actual yield feedback at harvest time.

Images