Farmland sand table deduction method and system based on 3DGS

Abstract

The application discloses a farmland sand table deduction method and system based on 3DGS, belongs to the technical field of virtual modeling, and comprises the following steps: acquiring multi-view images of a target farmland, combining 3DGS technology to construct a three-dimensional virtual farmland; acquiring attribute data sets of the target farmland, and constructing a growth coupling model; optimizing initial parameters of the growth coupling model through measured phenotype parameter values in the three-dimensional virtual farmland; using regression analysis to establish a response function of different management measure parameters on crop growth parameters to embed the growth coupling model; according to phenotype parameter prediction values generated by the growth coupling model, performing incremental updating on the three-dimensional virtual farmland; using an NSGA-II multi-objective optimization algorithm to generate a plurality of groups of candidate management measures; simulating future growth processes of crops of the candidate management measures; and using a constructed multi-dimensional evaluation system to screen optimal management measures, so that a closed loop from static three-dimensional modeling to dynamic intelligent decision-making is realized.

Description

Technical Field

[0001] This invention relates to a method and system for farmland sand table simulation based on 3DGS, belonging to the field of virtual modeling technology. Background Technology

[0002] With the rapid advancement of digital twin and smart city construction, 3D geographic environment simulation technology is increasingly widely used in disaster simulation, urban planning, traffic management and other fields. Through the deep integration of computer vision and computer graphics, high-fidelity digital reconstruction of the real geographic environment can be achieved.

[0003] A Chinese patent with publication number CN121280655B discloses a method for real-time intelligent inference of three-dimensional geographic environment based on a multimodal large model. The method includes: generating sparse point clouds by solving camera pose using a motion recovery structure algorithm and extracting motion features from dynamic foreground objects in surveillance videos; subsequently, initializing a three-dimensional Gaussian distribution set based on the sparse point clouds, extracting semantic features through a visual encoder and mapping them to the corresponding Gaussian distributions; then, constructing a topological graph structure of the Gaussian distributions, using motion features as initial stimuli, iteratively updating the coordinate offsets and appearance changes of each distribution through message passing, dynamically updating the Gaussian distribution attributes, and synthesizing a continuous sequence of inferred images through differentiable rasterization rendering.

[0004] Although existing technologies have achieved highly realistic real-time simulations of the dynamic evolution of three-dimensional geographic environments, there are still shortcomings. Specifically, especially in farmland scenarios, crop leaves exhibit high homogeneity and severe mutual occlusion. Uneven lighting conditions and limited shooting angles can significantly reduce the performance of typical 3D reconstruction methods, leading to incomplete or inaccurate scene modeling. At the same time, morphological changes during the crop growth cycle involve non-rigid, long-term dynamic processes, posing challenges to dynamic scene modeling. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a farmland sand table simulation method and system based on 3DGS. The farmland three-dimensional model constructed based on 3DGS technology integrates multi-source attribute data, achieves dynamic updates through a growth model, and combines multi-strategy parallel simulation and evaluation to solve the problems of low farmland reconstruction accuracy, difficulty in dynamic growth simulation, and insufficient decision support.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The farmland sand table simulation method based on 3DGS includes:

[0008] Acquire multi-view images of the target farmland and construct a three-dimensional virtual farmland using 3DGS technology;

[0009] Obtain the attribute dataset of the target farmland, and construct the mapping relationship between each Gaussian in the three-dimensional virtual farmland and the attribute dataset;

[0010] A growth coupling model was constructed to predict the growth status of crops under different environmental conditions and management measures.

[0011] The measured phenotypic parameter values ​​of crops in the three-dimensional virtual farmland are extracted, and the initial parameters of the growth coupling model are optimized by gradient descent.

[0012] Obtain field experiment data and historical production data of the target farmland, use regression analysis to establish response functions of different management measures parameters to crop growth parameters, and embed the response functions into the growth coupling model;

[0013] Based on the predicted phenotypic parameters generated by the growth coupling model, the three-dimensional virtual farmland is incrementally updated.

[0014] Configure the simulation data source, combine it with the NSGA-II multi-objective optimization algorithm to generate several sets of candidate management measures. For each candidate management measure, combine meteorological forecast data with the growth coupling model to simulate the future growth process of the current crop. Use the constructed multi-dimensional evaluation system to evaluate all candidate management measures and select the optimal management measure.

[0015] Specifically, the steps for constructing the growth coupling model include:

[0016] Soil parameters and meteorological parameters in the attribute dataset are used as environmental parameters. At the same time, management measures parameters and crop status parameters are set as input parameters of the growth coupling model, and phenotypic parameters are set as output parameters of the growth coupling model.

[0017] It adopts a two-layer cascaded architecture, including a basic mechanism layer and a correction layer;

[0018] The basic mechanism layer obtains crop growth patterns through the WOFOST and Logistic growth equations, and obtains basic predicted values ​​of phenotypic parameters through effective accumulated temperature calculation, dry matter production and distribution, and environmental stress correction.

[0019] The correction layer is used to extract spatiotemporal features. The correction coefficient is calculated through a fully connected layer and multiplied by the base prediction value to obtain the phenotypic parameter prediction value.

[0020] Specifically, the basic mechanism layer is configured with a temperature accumulation layer, a distribution layer, and an environmental correction layer;

[0021] The accumulated temperature layer is used to calculate the daily growth rate using the three cardinal temperature points and the daily average temperature, and to calculate the daily effective accumulated temperature by combining the lowest growth temperature among the three cardinal temperature points. The cumulative effective accumulated temperature is obtained by accumulating these values.

[0022] The allocation layer is used to calculate the total daily photosynthetic production and respiration consumption using WOFOST, thereby obtaining the daily net dry matter accumulation. Combined with the Logistic growth equation, the current growth stage is determined based on the cumulative effective accumulated temperature, and the net dry matter is allocated to generate the initial predicted values ​​of phenotypic parameters.

[0023] The environmental correction layer is used to calculate the water stress coefficient and nutrient stress coefficient, and combined with the initial predicted values, to generate the basic predicted values ​​of phenotypic parameters.

[0024] Specifically, the steps for extracting the measured phenotypic parameter values ​​of crops in the three-dimensional virtual farmland include:

[0025] Based on the soil region in the three-dimensional virtual farmland, the soil benchmark is obtained through the mean of Gaussian Z coordinates;

[0026] Based on the crop area in the three-dimensional virtual farmland, the 95th quantile of the Gaussian Z coordinate is used as the average height of the crop canopy. Combined with the soil benchmark, the average plant height of the target farmland is obtained.

[0027] Based on the Gaussian distribution of crop leaves, the projected area of ​​each Gaussian distribution on the horizontal plane is calculated, and combined with the land area of ​​the target farmland, the leaf area index is obtained.

[0028] Based on the Gaussian of crop stems, the scale parameters of each Gaussian in the X and Y directions are extracted, the equivalent diameter of each stem is calculated, and the average stem diameter of the crop is obtained by the mean of the equivalent diameters of all stems.

[0029] The DBSCAN clustering algorithm was used to perform Gaussian clustering on all crop leaves, the number of clusters was counted, the total number of leaves was obtained, and the number of leaves per unit area was obtained by combining the land area.

[0030] Principal component analysis is performed on each cluster to obtain the principal direction vector of the corresponding blade. The angle between the principal direction vector and the horizontal plane is taken as the angle of the corresponding blade. The average blade angle is obtained by averaging all blade angles.

[0031] Specifically, optimizing the initial parameters of the growth coupling model includes:

[0032] After each update of the three-dimensional virtual farmland, phenotypic parameters are automatically extracted to generate a dynamic measured sequence.

[0033] Phenotypic weights are assigned to different phenotypic parameters based on the analytic hierarchy process.

[0034] The weighted mean square error loss function is used as the loss function of the growth coupling model to calculate the loss function value between the predicted phenotypic parameters and the measured phenotypic parameters.

[0035] Based on the dynamic measured sequence, a training set is obtained, and error backpropagation is performed using mini-batch gradient descent. The parameters of the growth coupling model are then updated, and iterative optimization is performed until the loss function converges or the maximum number of iterations is reached.

[0036] Specifically, establishing response functions of different management parameters to crop growth parameters includes:

[0037] Using irrigation amount and irrigation time as independent variables and the adjustment ratio of water stress coefficient as dependent variable, a multivariate nonlinear regression analysis was used to construct an irrigation response function.

[0038] Using nitrogen application rate, phosphorus application rate, potassium application rate and fertilization time as independent variables, and the adjustment ratio of nutrient stress coefficient as dependent variable, a fertilization response function was established by multiple nonlinear regression analysis.

[0039] Using pesticide type, application rate, and control time as independent variables, and adjusting ratio of pest and disease incidence as dependent variable, logistic regression analysis was used to establish a plant protection response function.

[0040] The irrigation response function, fertilization response function, and plant protection response function are embedded into the correction layer of the growth coupling model as additional input features of the correction layer. At the same time, the response features of management measures, including irrigation response, nitrogen response, phosphorus response, potassium response, and plant protection response, are added to the correction layer.

[0041] For the corrected layer, the parameters are retrained using field experiment data and historical production data.

[0042] Specifically, incremental updates to the three-dimensional virtual farmland include:

[0043] Construct a real-time target set and combine it with the latest growth coupling model to predict the phenotypic parameter values ​​of each grid cell for the next day;

[0044] Iterate through all stem Gaussians and obtain the predicted plant height and predicted average stem diameter for the corresponding grid based on the spatial ID;

[0045] Extract the current average plant height and stem diameter from the three-dimensional virtual farmland, and calculate the plant height growth rate and stem diameter growth rate.

[0046] For the Z direction of the stem Gaussian, the change in plant height is obtained by multiplying the current average plant height by the plant height growth rate. Combined with the corresponding Z coordinate of the stem Gaussian, the Z coordinate of the stem Gaussian is updated by summing.

[0047] For the horizontal direction of the stem Gaussian, the growth rate of stem diameter is used as the growth ratio of the X and Y direction scales to update the horizontal scale of the stem Gaussian;

[0048] Iterate through all Gaussian leaves and calculate the leaf area index growth rate and the change in leaf angle.

[0049] The planar scale of the leaf Gaussian is updated using the product of the leaf spread coefficient and the leaf area index growth rate as the growth rate, and the rotation angle of the leaf Gaussian is updated using Euler angle rotation.

[0050] Specifically, incremental updates to the three-dimensional virtual farmland also include:

[0051] The predicted number of leaves generated by the growth coupling model is obtained, and the actual number of leaves in the three-dimensional virtual farmland is read. Based on the difference between the predicted number of leaves and the actual number of leaves, the leaf change is calculated, and the updated Gaussian quantity is calculated in combination with the grid area.

[0052] When the change in leaf size is greater than 0, a new leaf Gaussian is generated based on the updated Gaussian value. The initial parameters of the new leaf are generated based on the statistical distribution of parameters of the surrounding mature leaves.

[0053] When the leaf change is less than 0, the leaf is removed from the three-dimensional virtual farmland by gradually reducing the transparency, based on the updated Gaussian value.

[0054] Local crop images collected daily by fixed-point monitoring stations are obtained and compared with rendered images of the three-dimensional virtual farmland from the same viewpoint. The structural similarity index is used to calculate local differences.

[0055] Filter out error regions where the local difference is greater than a preset difference threshold, and extract all Gaussian parameters within the error regions;

[0056] Using the constructed local optimization loss function, the Gaussian parameters in the error region are optimized by gradient descent to minimize the loss function value. After optimization, the Gaussian parameters in the error region are updated.

[0057] Specifically, the evaluation of all candidate management measures includes:

[0058] Define decision variables, including irrigation decision variables, fertilization decision variables, and plant protection decision variables. At the same time, set the value range for each decision variable to form a search space.

[0059] The NSGA-II multi-objective optimization algorithm is used to perform multi-objective optimization search in the search space to generate several candidate management measures;

[0060] Using days as the basic time step, an independent virtual copy is created for each group of candidate management measures, and the deduction of all candidate management measures is executed in parallel.

[0061] Store the three-dimensional virtual farmland snapshots, phenotypic parameter sequences, resource consumption data, and yield prediction data for each group of candidate management measures at each time step, and establish an index of the inference results;

[0062] A multi-dimensional evaluation system is formed by setting up secondary indicators under production, resource, and environmental indicators as primary indicators;

[0063] By using range standardization, all indicators are converted into standardized values, and the weighted TOPSIS method is used to calculate the comprehensive score of each candidate strategy.

[0064] All candidate management measures were ranked from highest to lowest based on their comprehensive scores, and the top-ranked measures were selected. One candidate management measure was selected as the optimal management measure; among them... It is a preset positive integer value.

[0065] The 3DGS-based farmland sand table simulation system includes: a data acquisition module, a characterization and prediction module, and a candidate simulation module;

[0066] The data acquisition module is used to acquire the attribute dataset of the target farmland, assign a spatial ID to each Gaussian in the three-dimensional virtual farmland, establish a mapping relationship between the spatial ID and the attribute dataset, and acquire multi-view images of the target farmland. Using 3DGS technology, a three-dimensional virtual farmland is generated.

[0067] The characterization prediction module is used to construct a growth coupling model, combine the measured phenotypic parameter values ​​in the three-dimensional virtual farmland, optimize the initial parameters of the growth coupling model, use regression analysis to establish the response function of different management measures parameters to crop growth parameters, embed it into the growth coupling model, and incrementally update the three-dimensional virtual farmland according to the phenotypic parameter prediction values ​​generated by the growth coupling model.

[0068] The candidate simulation module generates several sets of candidate management measures using the NSGA-II multi-objective optimization algorithm, and simulates the future growth process of crops for each candidate management measure. The optimal management measure is then selected by combining the constructed multi-dimensional evaluation system.

[0069] The beneficial effects of this invention are:

[0070] This invention constructs a three-dimensional virtual farmland using 3DGS technology and establishes a one-to-one mapping relationship between each Gaussian and multi-source attribute data. This achieves deep integration of the geometric model with soil, meteorological, and crop status data, solving the pain point of separation between geometry and attributes in traditional digital twins. The constructed growth coupling model retains the physiological interpretability of crop growth and performs field calibration using measured phenotypic parameters extracted from the three-dimensional virtual farmland, significantly improving the model's adaptability to specific farmland environments. By embedding management measure response functions based on field experiments and historical data, it achieves accurate quantitative simulation of the impact of different irrigation, fertilization, and plant protection measures on crop growth. The incremental dynamic update based on the growth coupling model effectively solves the simulation problem of non-rigid, large-time-span dynamic growth of crops, ensuring long-term spatiotemporal consistency between the digital twin and the real farmland. Finally, by generating candidate management measures through multi-objective optimization and combining parallel inference and a multi-dimensional evaluation system, it can comprehensively quantify the combined benefits of different strategies, select the optimal management plan that meets the planting objectives, and provide intuitive and scientific decision support for agricultural production, realizing a complete closed loop from static three-dimensional modeling to dynamic intelligent decision-making. Attached Figure Description

[0071] Figure 1 The flowchart shows the farmland sand table simulation method based on 3DGS.

[0072] Figure 2 This is a flowchart for extracting measured phenotypic parameters in this invention;

[0073] Figure 3 This is a flowchart illustrating the evaluation of all candidate management measures in this invention;

[0074] Figure 4 This is a structural diagram of a farmland sand table simulation system based on 3DGS. Detailed Implementation

[0075] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0076] Example 1

[0077] refer to Figures 1 to 3 As shown in the figure, this embodiment introduces a farmland sand table simulation method based on 3DGS, including the following steps:

[0078] Step S1: Using the RGB camera and positioning sensor mounted on the drone, collect multi-view images of the target farmland, use 3DGS technology to quickly create a 3D model, process the multi-view images to generate a 3D virtual farmland, and simultaneously collect soil data, meteorological data, basic crop data, and real-time growth data of the target farmland to construct an attribute dataset of the target farmland. Assign a spatial ID to each Gaussian in the 3D virtual farmland and establish a mapping relationship between the spatial ID and the attribute dataset. This allows for quick querying of attribute data at the corresponding location of any Gaussian through the spatial ID, realizing dynamic binding of geometric location, soil fertility, meteorological conditions, and crop status.

[0079] The soil data includes soil moisture in the 0-20cm and 20-40cm layers, soil organic matter content, total nitrogen content, available phosphorus content, available potassium content, and pH value, collected using a portable soil analyzer with a grid-based sampling method (e.g., one sampling point per 100 square meters). Meteorological data includes air temperature, relative humidity, photosynthetically active radiation intensity, precipitation, wind speed, and wind direction, continuously collected from small field weather stations. Crop basic data includes crop variety, sowing date, emergence rate, planting density, historical irrigation records, historical fertilization records, and historical plant protection records, obtained through farmer ledgers and agricultural management terminals. Real-time growth data includes the current growth stage of the crop, relative chlorophyll content in leaves, and pest and disease occurrence, obtained through daily crop images collected from fixed monitoring stations and measurements using a portable chlorophyll meter.

[0080] Step S2: Using the WOFOST crop growth model and the Logistic growth equation, a growth coupling model is constructed to predict the crop growth status under different environmental conditions and management parameters. Based on the statistics of crop geometric parameters in the three-dimensional virtual farmland, the measured phenotypic parameter values ​​of the current crop are extracted, including average plant height, leaf area index, average stem diameter, leaf number distribution, and leaf angle distribution. These measured phenotypic parameter values ​​are used as labels. The initial parameters of the growth coupling model are optimized using the gradient descent method to adapt to specific crop varieties and local growth environments. Field experimental data and historical production data of the target farmland are obtained. Regression analysis is used to establish the response function of different management parameters to crop growth parameters. The response function is embedded in the growth coupling model to realize the quantitative simulation of the impact of management measures on crop growth.

[0081] Step S3: Using days as the basic time step, incrementally update the three-dimensional virtual farmland based on the predicted crop phenotypic parameters for the next time step generated by the growth coupling model.

[0082] Step S4: Based on the mapping relationship between spatial ID and attribute dataset, call the current crop growth status and soil data, and simultaneously obtain weather forecast data and planting targets input by the user as the data source for inference. Combined with the NSGA-II multi-objective optimization algorithm, generate several sets of candidate management measures covering three dimensions: irrigation, fertilization, and plant protection. For each candidate management measure, input the candidate management measure and weather forecast data into the growth coupling model to simulate the future growth process of the current crop. Combined with the constructed multi-dimensional evaluation system, evaluate all candidate management measures and select the optimal management measure.

[0083] Furthermore, establishing the mapping relationship between spatial IDs and attribute datasets includes:

[0084] The target farmland is divided into a 10m×10m grid. The lower left corner of the target farmland is taken as the origin of the coordinates. Each grid is assigned a unique grid code in the order from left to right and from top to bottom, and in the form of row number and column number.

[0085] All Gaussians within each grid are assigned incremental local numbers, and a spatial ID is generated for each Gaussian based on the grid number. At the same time, a correspondence table between the spatial ID and the three-dimensional coordinates of the Gaussian is established.

[0086] For discrete soil data, Kriging interpolation is used to spatially interpolate each soil attribute to generate a soil raster map, and the coordinate system of the soil raster map is converted into a local coordinate system consistent with the three-dimensional virtual farmland.

[0087] Since meteorological data has spatial uniformity at the farmland scale, the regional assignment method is used to construct the coverage area based on the effective coverage radius of the meteorological station. All 10m×10m grids within the coverage area are assigned the measured meteorological attribute values ​​of the meteorological station to generate a meteorological raster map. If there are multiple meteorological stations in the target farmland, the inverse distance weighted interpolation method is used to generate a meteorological raster map and align it with the grid division.

[0088] For basic crop data, values ​​are assigned to plots to generate basic raster maps. For real-time growth data, inverse distance weighted interpolation is used to spatially interpolate each real-time growth attribute to generate growth raster maps. Plots are pre-defined natural planting zones in the target farmland, defined by those skilled in the art.

[0089] The spatial extent of all raster maps is consistent with the spatial extent of the three-dimensional virtual farmland, forming a multi-source raster map set;

[0090] Traverse all Gaussians carrying spatial IDs, locate the corresponding grid pixels in the base raster and meteorological raster based on the corresponding grid number, and obtain the crop basic attribute and meteorological attribute values ​​of the corresponding grid. For soil raster and growth raster, obtain the soil attribute and real-time growth attribute values ​​corresponding to the corresponding Gaussian position through bilinear interpolation, and establish a Gaussian mapping table between geometry and attributes.

[0091] Furthermore, the steps for constructing a growth coupling model include:

[0092] Soil and meteorological parameters are used as environmental parameters, while management parameters and crop status parameters are set as model input parameters. Among them, management parameters include irrigation amount, irrigation time, nitrogen application rate, phosphorus application rate and potassium application rate, and crop status parameters are the current growth stage, expressed as effective accumulated temperature days.

[0093] Set phenotypic parameters, including plant height, leaf area index, average stem diameter, number of leaves, and average leaf angle;

[0094] The system utilizes a two-layer cascaded architecture, including a basic mechanism layer and a correction layer. The basic mechanism layer receives input parameters, and different sub-layers select specific parameters to use based on functional requirements.

[0095] Based on the WOFOST and Logistic growth equations in the agricultural field, a basic mechanism layer is set up to describe the basic physiological laws of crop growth. The basic mechanism layer is configured with an accumulated temperature layer, a distribution layer, and an environmental correction layer. Among them, regarding the specific parameters of different sublayers, the accumulated temperature layer only uses the daily average temperature from meteorological parameters, the distribution layer uses meteorological parameters and crop state parameters, and the environmental correction layer uses soil parameters.

[0096] The accumulated temperature layer utilizes three cardinal temperature points to construct a dependence function between crop growth rate and temperature. Based on the daily average temperature in meteorological parameters, the daily growth rate is calculated. The temperature difference is calculated by subtracting the average temperature from the minimum growth temperature. The effective accumulated temperature for that day is calculated by multiplying the temperature difference by the daily growth rate. The cumulative effective accumulated temperature since sowing is obtained by summing these values. The three cardinal temperature points include the minimum growth temperature. Optimal growth temperature Maximum growth temperature , which are inherent genetic parameters of the crop variety, obtained from agricultural industry standards; the dependency function expression is shown below:

[0097]

[0098] In the formula, This represents the relative growth rate for that day. The average daily temperature;

[0099] The allocation layer utilizes the WOFOST crop growth model, taking the daily effective accumulated temperature, cumulative effective accumulated temperature, photosynthetically active radiation (PADR) from meteorological parameters, and the current leaf area index as dynamic time-varying inputs to generate initial predicted values ​​of phenotypic parameters under conditions without any environmental stress. In this embodiment, daily average temperature, daily precipitation, specific leaf area, initial biomass, and maintenance respiration coefficient are used as fixed attribute parameters for the day. These fixed attribute parameters are essential basic configuration parameters for model operation and must be pre-configured in the model, remaining constant throughout a single crop growth simulation cycle. The fixed attribute parameters are determined by the genetic characteristics of the crop variety. The daily average temperature was calculated by collecting air temperature hourly using a micro-weather station in the field and calculating the 24-hour average. Daily precipitation was recorded by accumulating the total precipitation daily using a field rain gauge. Five to ten representative plants were selected at each growth stage of the crop, and fully mature and expanded leaves were cut off. The total leaf area was measured using a leaf area meter. The leaves were then blanched at 105℃ for 30 minutes and dried at 80℃ to constant weight. The dry weight of the leaves was measured, and the ratio of total leaf area to leaf dry weight was the specific leaf area for the corresponding growth stage. Initial biomass refers to the basic dry matter content per unit area of ​​the aboveground part after crop emergence. After emergence stabilized, five groups of 1m² plants were randomly selected. 2 Representative plots were prepared by isolating the aboveground parts of all plants, blanching them at 105℃, and drying them at 80℃ to constant weight. The average weight was taken as the initial biomass. A dark environment gradient temperature experiment was conducted, with four temperature gradients set at 15℃, 20℃, 25℃, and 30℃. The dark respiration rate of crop leaves was measured using a portable photosynthesis meter. Using temperature gradient as the independent variable and dark respiration rate as the dependent variable, the temperature response function was used to... The maintenance respiration coefficient and respiration temperature coefficient were obtained by fitting, where, To maintain the respiratory rate, The respiratory temperature coefficient, The reference temperature is 25℃;

[0100] The process of generating initial predictions using the WOFOST crop growth model includes:

[0101] The daily total photosynthetic production and respiration consumption are calculated. The net dry matter accumulation for the day is obtained by calculating the difference between the daily total photosynthetic production and respiration consumption. The daily total photosynthetic production is obtained by using the daily photosynthetically active radiation and the current leaf area index as inputs and by using WOFOST's canopy light distribution and photosynthesis calculations. Based on the daily average temperature and the maintenance respiratory coefficient, the respiratory consumption is calculated. The expression is as follows:

[0102]

[0103]

[0104] In the formula, The maximum photosynthetic rate saturated by light is the inherent physiological parameter of the crop variety. Mature and fully expanded leaves at various growth stages of the crop were selected. Portable photosynthesis measurement was used. The leaves were first light-induced with saturated light intensity for 30 minutes. Multiple gradients of photosynthetically active radiation were set. With photosynthetically active radiation as the independent variable and the measured net photosynthetic rate as the dependent variable, a non-rectangular hyperbola was used for fitting to obtain the maximum photosynthetic rate saturated by light. , is an integral variable, representing the cumulative leaf area index at the position calculated from the top downwards within the canopy; The total leaf area index of the current crop canopy is obtained directly from the Gaussian parameters of the leaves in the three-dimensional virtual farmland. The initial light energy utilization rate is used to characterize the light energy conversion efficiency of the leaf under low light conditions, and the initial slope is used to obtain the non-right-angle hyperbola fitting in the maximum photosynthetic rate under light saturation. The extinction coefficient of the canopy is determined by measuring the photosynthetically active radiation at the top of the canopy. With photosynthetically active radiation at the bottom It is obtained by combining the total leaf area index with the back calculation, that is ; The canopy reflectance is obtained by measuring the canopy reflectance using a canopy hyperspectral analyzer. The total daily photosynthetically active radiation incident at the top of the canopy is the cumulative daily value of photosynthetically active radiation reaching the top of the crop canopy, which is continuously collected daily by a photosynthetically active radiation sensor mounted on a field micro-weather station. The total dry matter of the aboveground part of the crop is obtained by adding the net dry matter accumulation from the initial biomass at sowing to the daily net dry matter accumulation since sowing.

[0105] Combining the Logistic growth equation, the current growth stage is determined based on the cumulative effective accumulated temperature and the stage-specific accumulated temperature thresholds for the crop variety. The corresponding dry matter allocation coefficient is then matched according to the current growth stage. The stage-specific accumulated temperature threshold is an inherent genetic parameter of the corresponding crop variety. The sowing date, the start date of each key growth stage (emergence, tillering, jointing, heading, grain-filling, and maturity) of the target crop variety over a 5-year period, and the daily average temperature data within the corresponding experimental period are obtained to generate a historical experimental set. For each year's experimental data, the effective accumulated temperature is accumulated daily starting from the sowing date. The cumulative effective accumulated temperature value corresponding to the start date of each growth stage is labeled, resulting in accumulated temperature samples for each growth stage in a single year. Accumulated temperature samples from the same growth stage over multiple years are compiled, and extreme outliers are removed using the 3σ criterion to form an accumulated temperature sample set for each growth stage. A normality fit test, such as the Shapiro-Wilk test, is performed on the sample set. The arithmetic mean of the samples is used as the stage-specific accumulated temperature threshold for that growth stage. The threshold range between two adjacent growth stages is defined. The interval refers to the accumulated temperature range corresponding to the growth stage. For example, the correspondence between the growth stages and the accumulated temperature thresholds for winter wheat is as follows: seedling stage (0℃~200℃), tillering stage (200℃~500℃), jointing stage (500℃~1000℃), heading stage (1000℃~1500℃), grain-filling stage (1500℃~2000℃), and maturity stage (≥2000℃). Representative plants are selected at each growth stage, and the stems, leaves, and ears are separated, dried to constant weight, and weighed to calculate the accumulated temperature thresholds for each stage. The proportion of organ dry weight to the total dry weight of the upper part is the dry matter allocation coefficient for the corresponding growth stage, with a value ranging from 0 to 1. The sum of the allocation coefficients of all organs at the same growth stage is 1. For example, in winter wheat, the leaf allocation coefficient is 0.65 and the stem allocation coefficient is 0.35 at the seedling stage, 0.55 and 0.45 at the tillering stage, 0.40 and 0.60 at the jointing stage, and 0.25, 0.35 and 0.40 at the heading stage.

[0106] The net dry matter accumulation is allocated to vegetative organs, such as stems and leaves, according to the dry matter distribution coefficient. Based on the dry matter mass of each vegetative organ and combined with the organ growth law described by the Logistic growth equation, the initial predicted values ​​of phenotypic parameters under no environmental stress conditions are generated.

[0107] The environmental correction layer receives the initial predicted values ​​of phenotypic parameters. Based on the crop water stress coefficient calculation method published in the FAO's guidelines for calculating crop water requirements, it calculates the crop water stress coefficient. Based on the content of total nitrogen, available phosphorus, and available potassium in the soil parameters, combined with the optimal nitrogen, phosphorus, and potassium contents required for crop growth, it calculates the nutrient stress coefficient through ratio calculation and minimum value selection. Finally, it multiplies the water stress coefficient and nutrient stress coefficient by the initial predicted values ​​to obtain the basic predicted values ​​of phenotypic parameters corrected by soil environment and basic management practices. The expression is shown below:

[0108]

[0109] In the formula, This is the water stress coefficient. This represents the actual soil moisture content. Field holding capacity The wilting coefficient is the soil moisture content at which crops begin to permanently wilt. It is an inherent physical parameter of the soil, with a value ranging from 0.05 to 0.25. It can be obtained by looking up a table of soil texture or by measuring it using the pressure membrane method.

[0110] At the basic mechanism layer, a correction layer is configured using a spatiotemporal graph convolutional network, such as a graph attention network. The correction layer receives the basic predicted values ​​of phenotypic parameters, extracts spatial features between different plots through spatial convolution, and extracts the temporal features of crop growth through temporal convolution. The extracted spatiotemporal features are input into two fully connected layers, and the output is a correction coefficient with the same dimension as the phenotypic parameters. The basic predicted value of the phenotypic parameters is multiplied by the corresponding correction coefficient to obtain the predicted value of the phenotypic parameters. Specifically, the spatial convolution uses two graph attention layers, each with eight attention heads, a single head output dimension of 16, an attention dropout rate of 0.6, and LeakyReLU activation function. Each attention layer is followed by a LayerNorm normalization layer. The temporal convolution uses two one-dimensional convolution layers with a kernel size of 3, a stride of 1, and padding of 1. The ReLU activation function is used, and each convolution is followed by a BatchNorm normalization layer.

[0111] The training steps for the growth coupling model include:

[0112] Based on input and phenotypic parameters, a historical dataset containing different growth stages of the target area over the past three years was obtained. For example, the historical dataset contains a total of 12,000 samples, including 2,500 from the seedling stage, 3,500 from the tillering stage, 3,000 from the jointing stage, 2,000 from the heading stage, and 1,000 from the grain-filling stage. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio, using a time-series partitioning method. Data annotation is achieved through a combination of manual field measurements and 3DGS model extraction, with an annotation accuracy of ±5%. The historical dataset covers three soil types (sandy loam, loam, and clay), two climate years (normal and high water years), and four management modes (conventional irrigation, water-saving irrigation, conventional fertilization, and soil testing-based fertilizer application).

[0113] The mean squared error loss function is used as the loss function of the growth coupling model to calculate the error between the predicted phenotypic parameters and the measured phenotypic parameters.

[0114] Using the Adam optimizer, the initial learning rate was set to 0.001, the batch size to 32, and the number of training epochs to 100. Training was stopped early when the validation set loss no longer decreased for 10 consecutive epochs.

[0115] Save the parameters of the basic mechanism layer and the correction layer after training as the initial parameters of the growth coupling model.

[0116] Furthermore, optimizing the initial parameters of the growth coupling model includes:

[0117] Based on a 3D virtual farmland, soil regions are screened using color threshold segmentation and height threshold methods. Soil baselines are obtained by averaging the Gaussian Z-coordinates within each soil region. The color threshold segmentation method includes: statistically analyzing the RGB color values ​​of all Gaussians, setting a soil color threshold range, and extracting Gaussians whose colors fall within the threshold range as candidate soil Gaussians. The height threshold method includes: calculating the average and standard deviation of the Z-coordinates of all candidate soil Gaussians, using the 3σ criterion to remove outlier coordinate values, and generating soil regions. The soil color threshold range is set based on the typical color characteristics of brown soil in the North China Plain farmland, with a range of R: 100-180, G: 80-150, and B: 50-120.

[0118] The crop region is screened using the height difference method and the color threshold segmentation method. The 95th percentile of the Z coordinates of all Gaussians in the crop region is extracted to obtain the average height of the crop canopy. The average plant height of the target farmland is obtained by the difference between the average height of the crop canopy and the soil reference. The height difference method includes: calculating the difference between the Z coordinates of all Gaussians and the soil reference, retaining Gaussians with a difference greater than 0.1m as candidate crop Gaussians, setting a crop threshold range, extracting Gaussians whose color falls within the crop threshold range, and generating crop regions. The crop threshold range is set based on the green spectral characteristics of the corresponding crop, such as the range R:50-150, G:100-200, B:30-100.

[0119] The leaves of all crops are screened using anisotropic feature analysis and scale thresholding. By iterating through all Gaussian covariance matrices, the projected area of ​​each Gaussian on the horizontal plane is calculated. The leaf area index is obtained by comparing the sum of the projected areas of all leaves with the land area of ​​the target farmland. The anisotropic feature analysis includes calculating the anisotropy coefficient of each crop's Gaussian covariance matrix by comparing the ratio of the largest to the smallest eigenvalue. Since leaves are thin and exhibit significant anisotropy, Gaussian covariance coefficients greater than 5 are retained. The scale thresholding method includes retaining Gaussian covariance matrices with scales in the X and Y directions within the range of 0.01-0.1m and a scale in the Z direction less than 0.01m to obtain the crop leaves.

[0120] The directional feature method and scale threshold method are used to screen the stems of all crops. The Gaussian stalks of all stems are iterated, and the scale parameters of each Gaussian stalk in the X and Y directions are extracted. The equivalent diameter of each stem is calculated by multiplying the product of the scale parameters by 2 and then performing a square root operation. The average stem diameter of the crop is obtained by averaging the equivalent diameters of all stems. The directional feature method and scale threshold method include: calculating the principal direction vector of each crop Gaussian stalk; since the stems grow vertically, Gaussian stalks with a principal direction vector angle less than 15° to the Z-axis are retained; Gaussian stalks with a Z-axis scale greater than 0.1m and X and Y-axis scales within the range of 0.005-0.02m are retained to obtain the crop stems.

[0121] The DBSCAN clustering algorithm is used to perform Gaussian clustering on all crop leaves. Each cluster corresponds to a complete leaf. The number of clusters is counted to obtain the total number of leaves. Combined with the land area of ​​the target farmland, the number of leaves per unit area is obtained.

[0122] Principal component analysis is performed on each cluster to obtain the principal direction vector of the corresponding blade. The angle between the principal direction vector and the horizontal plane is taken as the angle of the corresponding blade. The average blade angle is obtained by averaging all blade angles.

[0123] For the continuously updated 3D virtual farmland, after each update, phenotypic parameters are automatically extracted, the phenotypic parameter values ​​after each update are integrated, and a dynamic measured sequence with timestamps is generated.

[0124] Based on the analytic hierarchy process (AHP), phenotypic weights are assigned to different phenotypic parameters. Specifically, agricultural experts compare the importance of phenotypic parameters pairwise to construct a judgment matrix and calculate the weight of each parameter.

[0125] The weighted mean square error loss function is used as the loss function of the growth coupling model to calculate the loss function value between the predicted phenotypic parameters and the measured phenotypic parameters.

[0126] Based on the dynamic measured sequence, data from the past 30 days is selected as the training set. At the same time, environmental data and crop status data for the corresponding dates are obtained. Through mini-batch gradient descent, the error is backpropagated and the parameters of the growth coupling model are updated. The optimization is iteratively performed until the loss function converges or the maximum number of iterations is reached. The maximum number of iterations is set by those skilled in the art. In this embodiment, the maximum number of iterations is set to 1000.

[0127] Furthermore, the response functions of different management measures to crop growth parameters are established, including:

[0128] Collect field trial data from the target area over the past three years, including comparative crop growth data under different irrigation amounts, fertilizer amounts, and plant protection measures. Also collect historical production data from the target area over the past five years, including historical management records for different plots (irrigation time, irrigation amount, fertilizer type, fertilizer amount, plant protection time, pesticide application) and corresponding crop yield and growth data.

[0129] Using irrigation amount and irrigation time as independent variables, and the adjustment ratio of irrigation amount and irrigation time to water stress coefficient as dependent variable, a multivariate nonlinear regression analysis was used to construct an irrigation response function. The adjustment ratio is the ratio of the actual change parameter of the adjusted irrigation amount or irrigation time to the change parameter under no intervention. The change parameters include water stress coefficient, nutrient stress coefficient, and disease and pest incidence rate.

[0130] Using nitrogen application rate, phosphorus application rate, potassium application rate and fertilization time as independent variables, and the adjustment ratio of nutrient stress coefficient as dependent variable, a fertilization response function was established by multiple nonlinear regression analysis.

[0131] Using pesticide type, application rate, and control time as independent variables, and adjusting proportion of pest and disease incidence as the dependent variable, logistic regression analysis was used to establish a plant protection response function; among which, pesticide type... For categorical variables, dummy variable encoding is used. Indicates insecticide, Indicates a bactericide. This indicates the herbicide, with all others taking a value of 0. The application rate is the amount of active ingredient, the control time is the number of days of effective accumulated temperature, and the ratio of the actual incidence of crop diseases and pests after the implementation of plant protection measures to the initial incidence of diseases and pests under conditions without plant protection is used as the adjustment ratio for the incidence of diseases and pests.

[0132] The irrigation response function, fertilization response function, and plant protection response function are embedded into the correction layer of the growth coupling model as additional input features of the correction layer. At the same time, the response features of management measures, including irrigation response, nitrogen response, phosphorus response, potassium response, and plant protection response, are added to the correction layer. The parameters are then retrained for the correction layer.

[0133] Furthermore, incremental updates to the 3D virtual farmland include:

[0134] Based on the latest environmental parameters, implemented management measures parameters, and current crop status parameters, a real-time target set is constructed. Combined with the latest growth coupling model, and utilizing the parallel computing capabilities of GPUs, the phenotypic parameter values ​​for each grid cell on the next day are predicted in batches. The prediction results are stored as a whole-field prediction table according to spatial ID.

[0135] Traverse all stem Gaussians, extract the predicted plant height and predicted average stem diameter of the corresponding grid from the whole field prediction table according to the spatial ID, and extract the current average plant height and stem diameter from the three-dimensional virtual farmland to calculate the plant height growth rate and stem diameter growth rate.

[0136] For the Z direction of the stem Gaussian, the change in plant height is obtained by multiplying the current average plant height by the plant height growth rate. Combined with the corresponding Z coordinate of the stem Gaussian, the Z coordinate of the stem Gaussian is updated by summing.

[0137] For the horizontal direction of the stem Gaussian, the growth rate of stem diameter is used as the growth ratio of the X and Y direction scales to update the horizontal scale of the stem Gaussian;

[0138] Iterate through all leaf Gaussians, obtain the predicted leaf area index and predicted average leaf angle of the corresponding grid from the whole field prediction table according to the spatial ID, and extract the current leaf area index and current average leaf angle from the three-dimensional virtual farmland to calculate the leaf area index growth rate and leaf angle change.

[0139] For the planar orientation of the leaf Gaussian, the growth rate is the product of the leaf spread coefficient and the leaf area index growth rate, and the planar orientation scale of the leaf Gaussian is updated. The spread coefficient is dynamically determined by those skilled in the art based on the current size and growth stage of the leaf, and its value ranges from 0.1 to 1. For example, the spread coefficient for newly curled leaves (spread < 30%) is 0.1-0.3, for mid-growth leaves (spread 30%-70%) is 0.4-0.7, and for fully expanded mature leaves (spread > 70%) is... The coefficient ranges from 0.8 to 1.0. The unfolding degree is the ratio of the current actual unfolded area of ​​the leaf to the maximum unfolded area when the leaf is fully mature. The value ranges from 0 to 1. It is used to quantitatively characterize the growth and maturity of the leaf. The actual unfolded area is directly calculated by Gaussian from the leaves in the three-dimensional virtual farmland, which is the product of the X and Y direction scales. The maximum unfolded area is the product of the average X and Y direction scales of mature leaves of the same variety, which is obtained by statistical analysis of the Gaussian parameters of mature leaves in the three-dimensional virtual farmland. The unfolding degree coefficient and the unfolding degree are linearly positively correlated and correspond one-to-one using equal interval mapping.

[0140] For the rotation angle of the blade Gaussian, the rotation matrix constructed based on Euler angle rotation is called. The rotation matrix after rotating the blade around the rotation axis by the change in the blade angle is used as the adjustment coefficient. The rotation angle of the blade Gaussian is updated through the product operation. At the same time, the Z coordinate of the blade Gaussian is updated synchronously with the Z coordinate of the corresponding stem Gaussian.

[0141] The predicted number of leaves generated by the growth coupling model is obtained, and the actual number of leaves in the three-dimensional virtual farmland is read. Based on the difference between the predicted number of leaves and the actual number of leaves, the change in leaf quantity is calculated. When the change in leaf quantity is 0, it indicates that there is no change in the number of leaves, so the quantity update is skipped. When the change in leaf quantity is greater than 0, the product of the grid area and the change in leaf quantity is used as the update Gaussian quantity, so that new leaf Gaussians are randomly generated in the upper part of the canopy of the corresponding grid. The initial parameters of the new leaves are generated based on the statistical distribution of the parameters of the surrounding mature leaves, and the initial spread coefficient is set to 0.2 to simulate newly curled leaves.

[0142] When the leaf change is less than 0, based on the updated Gaussian value, the Gaussian leaf with the highest transparency is selected, and a gradual transparency decay strategy is adopted to increase the transparency by 0.05 every day until the transparency reaches 0.9, at which point it is removed from the 3D virtual farmland to simulate the natural withering and falling of leaves; where the transparency value ranges from 0 to 1, and the larger the value, the higher the transparency.

[0143] The local crop images collected daily by fixed monitoring stations are compared with the rendered images of a 3D virtual farmland from the same perspective. The structural similarity index is used to calculate the local differences, thereby generating an error heat map.

[0144] Error regions with local differences exceeding a preset threshold are selected from the error heatmap, and all Gaussian parameters within these regions are extracted. A local optimization loss function is then constructed using the mean square error and L2 regularization term of the rendered image and the actual acquired crop local image. The expression is as follows:

[0145]

[0146] In the formula, To render the image Compared with actual collected crop local images The mean square error term, represents the change in Gaussian parameters, and represents the vector difference in Gaussian position coordinates, 3D scale, rotation angle, color attributes, and transparency before and after optimization. For the L2 regularization term of the Gaussian parameter variation, For the regularization term weight coefficient, local real-shot images of crops at different growth stages, under different light conditions, and in different plot types of the target farmland are collected. Candidate coefficients are set with a coefficient range of 0.001-0.01 and a step size of 0.003. Local optimization tests are performed on the candidate coefficients respectively, and the candidate coefficient corresponding to the minimum local optimization loss is selected as the regularization term weight coefficient.

[0147] The Gaussian parameters in the error region are optimized using the gradient descent method to minimize the loss function value. After optimization, the Gaussian parameters in the error region are updated. The difference threshold is set by those skilled in the art; in this embodiment, the difference threshold is set to 0.85.

[0148] Furthermore, the evaluation of all candidate management measures includes:

[0149] Decision variables are defined, including irrigation decision variables, fertilization decision variables, and plant protection decision variables. At the same time, those skilled in the art, based on agricultural production experience and local agronomic standards, set the value range for each decision variable to form a search space. Among them, irrigation decision variables include irrigation time, single irrigation amount, and irrigation method; fertilization decision variables include fertilization time, nitrogen, phosphorus and potassium application amount, and fertilization method; and plant protection decision variables include control time, pesticide type, and application amount.

[0150] The NSGA-II multi-objective optimization algorithm is used to perform multi-objective optimization search in the search space. The optimization objectives include maximizing the expected output, minimizing the production cost, minimizing water resource consumption, and minimizing the environmental impact. The algorithm parameters are set as follows: population size 100, crossover probability 0.8, mutation probability 0.1, and maximum number of iterations 200 generations.

[0151] After the operation is completed, based on the Pareto optimal solution set, solutions covering different objective preferences are selected, including the solution with the highest output, the lowest cost, the lowest water consumption, and the solution with the least environmental impact, and candidate management measures are generated to cover different objective preferences.

[0152] Using days as the basic time step, an independent virtual copy is created for each group of candidate management measures. The simulation of all candidate management measures is performed in parallel, including: for each virtual copy, inputting the weather forecast data of the day and the management measure parameters of the day in the corresponding candidate strategy, calling the trained growth coupling model, calculating the predicted values ​​of crop phenotypic parameters for the day, reusing the Gaussian parameter update mechanism, updating the three-dimensional virtual farmland parameters in the virtual copy according to the predicted values ​​of phenotypic parameters, recording the crop growth status, resource consumption, and occurrence of pests and diseases for the day. The simulation process continues until the simulation end time set by the user is reached.

[0153] Store the three-dimensional virtual farmland snapshot, phenotypic parameter sequence, resource consumption data, and yield prediction data for each group of candidate management measures at each time step, establish an index of inference results, and support quick query of relevant data by candidate management measures, time, and region;

[0154] A multi-dimensional evaluation system is constructed using production, resource, and environmental indicators as primary indicators, with secondary indicators under each primary indicator. Production indicators include projected yield, crop maturity period, lodging risk, pest and disease incidence rate, and fruit quality score. Resource indicators include water consumption, nitrogen fertilizer utilization rate, phosphorus fertilizer utilization rate, and potassium fertilizer utilization rate. Environmental indicators include nitrogen loss, phosphorus loss, pesticide residue risk, and greenhouse gas emissions. Projected yield is the expected harvest yield per unit area after crop maturity. Crop maturity period is the total number of days required for the crop to reach physiological maturity from sowing. Lodging risk represents the probability of stem lodging, ranging from 0 to 1. Pest and disease incidence rate is the percentage of plants affected by pests and diseases per unit area. Fruit quality score represents the overall quality grade of the harvested crop, ranging from 0 to 100 points. The quality parameters output by the growth coupling model are calculated by weighting according to the quality grading rules of agricultural industry standards. Water consumption is the total water input per unit area during the entire crop growth cycle. Nitrogen fertilizer utilization rate is the proportion of nitrogen absorbed by the crop to the total nitrogen applied, with a value range of 0-1. Phosphorus fertilizer utilization rate is the proportion of phosphorus absorbed by the crop to the total phosphorus applied, with a value range of 0-1. Potassium fertilizer utilization rate is the proportion of potassium absorbed by the crop to the total potassium applied, with a value range of 0-1. Nitrogen loss is the total amount of nitrogen lost per unit area through runoff and leaching during the entire crop growth cycle. Phosphorus loss is the total amount of phosphorus lost per unit area through runoff and leaching during the entire crop growth cycle. Pesticide residue risk is the probability of pesticide residue exceeding the standard at crop harvest, with a value range of 0-1. Greenhouse gas emissions are the total greenhouse gas emissions per unit area during the entire crop production cycle.

[0155] By using range standardization, all indicators are converted into standardized values ​​to eliminate dimensional differences. At the same time, the weight values ​​of each primary and secondary indicator are calculated based on the analytic hierarchy process. The weighted TOPSIS method is used to calculate the comprehensive score of each group of candidate management measures, including constructing a standardized decision matrix, determining the positive and negative ideal solutions, calculating the distance of each strategy to the positive and negative ideal solutions, and calculating the relative closeness of each strategy as the comprehensive score. All candidate management measures are ranked from high to low according to the comprehensive score.

[0156] Based on the ranking results of the comprehensive scores, the top 3 candidate management measures are selected as the optimal management measures. At the same time, an implementation plan is generated for each optimal management measure, including irrigation method, fertilization plan, plant protection plan, and expected effect.

[0157] Example 2

[0158] Please see Figure 4 Another embodiment of the present invention provides a farmland sand table simulation system based on 3DGS, comprising: a data acquisition module, a characterization prediction module, and a candidate simulation module;

[0159] The data acquisition module is used to collect soil data, meteorological data, basic crop data, and real-time growth data of the target farmland, construct the attribute dataset of the target farmland, assign spatial IDs to each Gaussian in the 3D virtual farmland, establish the mapping relationship between spatial IDs and attribute datasets, and at the same time, acquire multi-view images of the target farmland through the RGB camera and positioning sensor carried by the UAV, use 3DGS technology to quickly perform 3D modeling, process the multi-view images, and generate a 3D virtual farmland.

[0160] The characterization and prediction module is used to construct a growth coupling model to predict the growth status of crops under different environmental conditions and management measures. Based on the measured phenotypic parameter values ​​in the 3D virtual farmland, the initial parameters of the growth coupling model are optimized to adapt to specific crop varieties and local growth environments. Field trial data and historical production data of the target farmland are obtained. Regression analysis is used to establish the response function of different management measures to crop growth parameters. The response function is embedded in the growth coupling model to realize the quantitative simulation of the impact of management measures on crop growth. At the same time, with days as the basic time step, the 3D virtual farmland is incrementally updated according to the predicted values ​​of crop phenotypic parameters at the next time step generated by the growth coupling model.

[0161] The candidate simulation module generates several sets of candidate management measures using the NSGA-II multi-objective optimization algorithm, simulates the future growth process of crops for each candidate management measure, and evaluates all candidate management measures in conjunction with the constructed multi-dimensional evaluation system to select the optimal management measure.

[0162] Working principle and effects:

[0163] By establishing a one-to-one mapping between Gaussian and soil, meteorological, and crop attributes, the pain point of separation between geometry and attributes in traditional digital twins is solved, providing a unified foundation for point-by-point refined simulation. A growth coupling model is adopted, and the model is calibrated using measured phenotypic parameters extracted from 3DGS. The response function of management measures is embedded to achieve accurate quantitative prediction of crop growth and the effects of different irrigation, fertilization, and plant protection measures. Based on incremental updates and multi-scale error correction driven by the growth model, the problem of non-rigid and large-time-span dynamic growth simulation of crops is solved, ensuring that the digital twin and real farmland are synchronized in the long term. Finally, candidate strategies are generated through multi-objective optimization, and the optimal management scheme is selected by combining parallel inference and multi-dimensional evaluation, forming a complete closed loop from high-precision modeling to intelligent decision-making, providing scientific decision support for agricultural production.

[0164] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for farmland sand table simulation based on 3DGS, characterized in that, include: Acquire multi-view images of the target farmland and construct a three-dimensional virtual farmland using 3DGS technology; Obtain the attribute dataset of the target farmland, and construct the mapping relationship between each Gaussian in the three-dimensional virtual farmland and the attribute dataset; A growth coupling model was constructed to predict the growth status of crops under different environmental conditions and management measures. The measured phenotypic parameter values ​​of crops in the three-dimensional virtual farmland are extracted, and the initial parameters of the growth coupling model are optimized by gradient descent. Obtain field experiment data and historical production data of the target farmland, use regression analysis to establish response functions of different management measures parameters to crop growth parameters, and embed the response functions into the growth coupling model; Based on the predicted phenotypic parameters generated by the growth coupling model, the three-dimensional virtual farmland is incrementally updated. Configure the simulation data source, combine the NSGA-II multi-objective optimization algorithm to generate several sets of candidate management measures. For each candidate management measure, combine the meteorological forecast data with the growth coupling model to simulate the future growth process of the current crop. Use the constructed multi-dimensional evaluation system to evaluate all candidate management measures and select the optimal management measure. The steps for constructing the growth coupling model include: Soil parameters and meteorological parameters in the attribute dataset are used as environmental parameters, while management measure parameters and crop status parameters are set as input parameters of the growth coupling model, and phenotypic parameters are set as output parameters of the growth coupling model. It adopts a two-layer cascaded architecture, including a basic mechanism layer and a correction layer; The basic mechanism layer obtains crop growth patterns through the WOFOST and Logistic growth equations, and obtains basic predicted values ​​of phenotypic parameters through effective accumulated temperature calculation, dry matter production and distribution, and environmental stress correction. The correction layer is used to extract spatiotemporal features. The correction coefficient is calculated through the fully connected layer and multiplied with the basic prediction value to obtain the phenotypic parameter prediction value. Establishing response functions for crop growth parameters under different management measures includes: Using irrigation amount and irrigation time as independent variables and the adjustment ratio of water stress coefficient as dependent variable, a multivariate nonlinear regression analysis was used to construct an irrigation response function. Using nitrogen application rate, phosphorus application rate, potassium application rate and fertilization time as independent variables, and the adjustment ratio of nutrient stress coefficient as dependent variable, a fertilization response function was established by multiple nonlinear regression analysis. Using pesticide type, application rate, and control time as independent variables, and the adjustment ratio of pest and disease incidence rate as the dependent variable, logistic regression analysis was used to establish a plant protection response function. The irrigation response function, fertilization response function, and plant protection response function are embedded into the correction layer of the growth coupling model as additional input features of the correction layer. At the same time, the response features of management measures, including irrigation response, nitrogen response, phosphorus response, potassium response, and plant protection response, are added to the correction layer. For the corrected layer, the parameters are retrained using field experiment data and historical production data.

2. The farmland sand table simulation method based on 3DGS according to claim 1, characterized in that: The basic mechanism layer is configured with a temperature accumulation layer, a distribution layer, and an environmental correction layer; The accumulated temperature layer is used to calculate the daily growth rate using the three cardinal temperature points and the daily average temperature, and to calculate the daily effective accumulated temperature by combining the lowest growth temperature among the three cardinal temperature points. The cumulative effective accumulated temperature is obtained by accumulating these values. The allocation layer is used to calculate the total daily photosynthetic production and respiration consumption using WOFOST, thereby obtaining the daily net dry matter accumulation. Combined with the Logistic growth equation, the current growth stage is determined based on the cumulative effective accumulated temperature, and the net dry matter is allocated to generate the initial predicted values ​​of phenotypic parameters. The environmental correction layer is used to calculate the water stress coefficient and nutrient stress coefficient, and combined with the initial predicted values, to generate the basic predicted values ​​of phenotypic parameters.

3. The farmland sand table simulation method based on 3DGS according to claim 2, characterized in that, The steps for extracting the measured phenotypic parameter values ​​of crops in the three-dimensional virtual farmland include: Based on the soil region in the three-dimensional virtual farmland, the soil benchmark is obtained through the mean of Gaussian Z coordinates; Based on the crop area in the three-dimensional virtual farmland, the 95th quantile of the Gaussian Z coordinate is used as the average height of the crop canopy. Combined with the soil benchmark, the average plant height of the target farmland is obtained. Based on the Gaussian distribution of crop leaves, the projected area of ​​each Gaussian distribution on the horizontal plane is calculated, and combined with the land area of ​​the target farmland, the leaf area index is obtained. Based on the Gaussian of crop stems, the scale parameters of each Gaussian in the X and Y directions are extracted, the equivalent diameter of each stem is calculated, and the average stem diameter of the crop is obtained by the mean of the equivalent diameters of all stems. The DBSCAN clustering algorithm was used to perform Gaussian clustering on all crop leaves, the number of clusters was counted, the total number of leaves was obtained, and the number of leaves per unit area was obtained by combining the land area. Principal component analysis is performed on each cluster to obtain the principal direction vector of the corresponding blade. The angle between the principal direction vector and the horizontal plane is taken as the angle of the corresponding blade. The average blade angle is obtained by averaging all blade angles.

4. The farmland sand table simulation method based on 3DGS according to claim 3, characterized in that, The initial parameters for optimizing the growth coupling model include: After each update of the three-dimensional virtual farmland, phenotypic parameters are automatically extracted to generate a dynamic measured sequence. Phenotypic weights are assigned to different phenotypic parameters based on the analytic hierarchy process. The weighted mean square error loss function is used as the loss function of the growth coupling model to calculate the loss function value between the predicted phenotypic parameters and the measured phenotypic parameters. Based on the dynamic measured sequence, a training set is obtained, and error backpropagation is performed using mini-batch gradient descent. The parameters of the growth coupling model are then updated, and iterative optimization is performed until the loss function converges or the maximum number of iterations is reached.

5. The farmland sand table simulation method based on 3DGS according to claim 4, characterized in that, Incremental updates to the three-dimensional virtual farmland include: Construct a real-time target set and combine it with the latest growth coupling model to predict the phenotypic parameter values ​​of each grid cell for the next day; Iterate through all stem Gaussians and obtain the predicted plant height and predicted average stem diameter for the corresponding grid based on the spatial ID; Extract the current average plant height and stem diameter from the three-dimensional virtual farmland, and calculate the plant height growth rate and stem diameter growth rate. For the Z direction of the stem Gaussian, the change in plant height is obtained by multiplying the current average plant height by the plant height growth rate. Combined with the corresponding Z coordinate of the stem Gaussian, the Z coordinate of the stem Gaussian is updated by summing. For the horizontal direction of the stem Gaussian, the growth rate of stem diameter is used as the growth ratio of the X and Y direction scales to update the horizontal scale of the stem Gaussian; Iterate through all Gaussian leaves and calculate the leaf area index growth rate and the change in leaf angle. The planar scale of the leaf Gaussian is updated using the product of the leaf spread coefficient and the leaf area index growth rate as the growth rate, and the rotation angle of the leaf Gaussian is updated using Euler angle rotation.

6. The farmland sand table simulation method based on 3DGS according to claim 5, characterized in that, Incremental updates to the three-dimensional virtual farmland also include: The predicted number of leaves generated by the growth coupling model is obtained, and the actual number of leaves in the three-dimensional virtual farmland is read. Based on the difference between the predicted number of leaves and the actual number of leaves, the leaf change is calculated, and the updated Gaussian quantity is calculated in combination with the grid area. When the change in leaf size is greater than 0, a new leaf Gaussian is generated based on the updated Gaussian value. The initial parameters of the new leaf are generated based on the statistical distribution of parameters of the surrounding mature leaves. When the leaf change is less than 0, the leaf is removed from the three-dimensional virtual farmland by gradually reducing the transparency, based on the updated Gaussian value. Local crop images collected daily by fixed-point monitoring stations are obtained and compared with rendered images of the three-dimensional virtual farmland from the same viewpoint. The structural similarity index is used to calculate local differences. Filter out error regions where the local difference is greater than a preset difference threshold, and extract all Gaussian parameters within the error regions; Using the constructed local optimization loss function, the Gaussian parameters in the error region are optimized by gradient descent to minimize the loss function value. After optimization, the Gaussian parameters in the error region are updated.

7. The farmland sand table simulation method based on 3DGS according to claim 6, characterized in that, The evaluation of all candidate management measures includes: Define decision variables, including irrigation decision variables, fertilization decision variables, and plant protection decision variables. At the same time, set the value range for each decision variable to form a search space. The NSGA-II multi-objective optimization algorithm is used to perform multi-objective optimization search in the search space to generate several candidate management measures; Using days as the basic time step, an independent virtual copy is created for each group of candidate management measures, and the deduction of all candidate management measures is executed in parallel. Store the three-dimensional virtual farmland snapshots, phenotypic parameter sequences, resource consumption data, and yield prediction data for each group of candidate management measures at each time step, and establish an index of the inference results; A multi-dimensional evaluation system is formed by setting up secondary indicators under production, resource, and environmental indicators as primary indicators; By using range standardization, all indicators are converted into standardized values, and the weighted TOPSIS method is used to calculate the comprehensive score of each candidate strategy. All candidate management measures were ranked from highest to lowest based on their comprehensive scores, and the top-ranked measures were selected. One candidate management measure was selected as the optimal management measure; among them... It is a preset positive integer value.

8. A 3DGS-based farmland sand table simulation system, used to implement the 3DGS-based farmland sand table simulation method as described in any one of claims 1-7, characterized in that, include: Data acquisition module, characterization and prediction module, and candidate inference module; The data acquisition module is used to acquire the attribute dataset of the target farmland, assign a spatial ID to each Gaussian in the three-dimensional virtual farmland, establish a mapping relationship between the spatial ID and the attribute dataset, and acquire multi-view images of the target farmland to generate a three-dimensional virtual farmland using 3DGS technology. The characterization prediction module is used to construct a growth coupling model, combine the measured phenotypic parameter values ​​in the three-dimensional virtual farmland, optimize the initial parameters of the growth coupling model, use regression analysis to establish the response function of different management measures parameters to crop growth parameters, embed it into the growth coupling model, and incrementally update the three-dimensional virtual farmland according to the phenotypic parameter prediction values ​​generated by the growth coupling model. The candidate simulation module generates several sets of candidate management measures using the NSGA-II multi-objective optimization algorithm, and simulates the future growth process of crops for each candidate management measure. The optimal management measure is then selected by combining the constructed multi-dimensional evaluation system.

Images