Winter wheat regional yield intelligent estimation method and system fusing crop mechanism model and machine learning

Abstract

The invention provides a winter wheat regional yield intelligent estimation method and system fusing a crop mechanism model and machine learning, and relates to the technical field of agricultural remote sensing and crop growth simulation. The method comprises the following steps: identifying a winter wheat range through object-oriented classification by using a multi-temporal satellite remote sensing image, and constructing a winter wheat mask and remote sensing features thereof; selecting sample points in the mask, inverting a leaf area index time sequence through a radiation transfer model and driving a crop growth model, and performing ensemble Kalman filtering assimilation to obtain single-point simulation yield to form a training sample library. Performing growth grading according to the normalized vegetation index accumulated value of the booting stage and the flowering stage, selecting representative sample points to construct a training sample set, establishing a winter wheat yield estimation model and applying the model to all pixels, obtaining yield spatial distribution and administrative unit yield, and realizing high-precision rapid estimation of the winter wheat area yield.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing and crop growth simulation technology, and in particular to a method and system for intelligent estimation of regional yield of winter wheat that integrates crop mechanism models and machine learning. Background Technology

[0002] Winter wheat is one of the world's three major food crops, and efficient and accurate monitoring of winter wheat planting area and yield estimation provide important basis for adjusting food policies. Compared with manual statistical methods, satellite remote sensing technology has advantages in regional crop identification, growth monitoring, and yield estimation, and has become an important means of obtaining regional winter wheat information.

[0003] Mechanistic crop growth models, in simulating winter wheat growth, comprehensively consider multiple factors such as soil, climate, nutrients, variety, and field management. They can depict the processes of photosynthesis, respiration, transpiration, and organogenesis on a daily basis, reflecting crop growth mechanisms more effectively than traditional statistical models and yield estimation models based solely on simple remote sensing indices. Existing studies have explored the performance of different assimilation algorithms, assimilation variables, and crop growth models in regional yield estimation by coupling crop growth models with remote sensing information assimilation methods. Among these, the method of assimilating leaf area index time-series data into the WOFOST crop growth model using ensemble Kalman filtering shows good accuracy in regional crop yield simulation. However, these models are mostly based on area assimilation, which requires significant computational resources when assimilating high spatial resolution remote sensing data.

[0004] On the one hand, forward modeling of winter wheat at a single point using crop growth models already requires significant computation. While introducing data assimilation at a regional scale and using high-resolution remote sensing data can accurately represent the spatial distribution of winter wheat yield, the computational and time requirements for area assimilation are substantial, resulting in low efficiency for large-scale applications. On the other hand, yield estimation methods based on machine learning algorithms such as multiple linear regression and random forest regression, which establish empirical relationships between remote sensing indices and winter wheat yield, have achieved good results in some study areas. However, these methods require a large number of yield samples collected in the field, and related studies typically require data collected over the past five years, making the samples highly time-sensitive. The winter wheat harvest period is generally only about one week, which also limits the number of available samples. Existing regional winter wheat yield estimation models still have shortcomings in simultaneously ensuring mechanistic constraints, high spatial resolution representation, and yield estimation efficiency, while reducing dependence on large-scale long-term yield samples. There is an urgent need for a method that utilizes crop growth models and remote sensing data for physical constraints at limited representative points, and combines the resulting single-point yield estimation results with regional remote sensing information to achieve rapid regional winter wheat yield estimation. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for intelligent estimation of winter wheat regional yield that integrates crop mechanism models and machine learning. By assimilating a crop growth model into a small number of representative sample points to generate highly reliable yield samples and combining machine learning to achieve high-resolution extrapolation at the regional scale, the accuracy and efficiency of winter wheat regional yield estimation are significantly improved.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for intelligently estimating regional yield of winter wheat by integrating crop mechanism models and machine learning includes: Multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat were acquired. Vegetation index, texture features and spectral band features were extracted. An object-oriented classification method was used to identify the planting range of winter wheat and to obtain the winter wheat mask and the multi-temporal remote sensing features corresponding to the winter wheat mask. Sample points were selected within the winter wheat mask. Based on the radiative transfer model, the multi-temporal remote sensing features of each sample point were inverted to obtain the leaf area index time series of each sample point. The leaf area index time series, together with meteorological parameters, soil parameters, and field management parameters, were used as inputs to drive the crop growth model. Based on the maximum target leaf area index, the model parameters related to the maximum leaf area index were calibrated using parameter optimization methods to obtain the calibrated crop growth model. The leaf area index time series of each sample point is used as an assimilation variable and input into the calibrated crop growth model. The ensemble Kalman filter method is used for sequential assimilation to update the state variables and yield-related parameters of the crop growth model. The single-point simulated yield of each sample point is output, and a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield is constructed. Based on the cumulative normalized vegetation index (NVI) values ​​of winter wheat during the heading stage and the flowering stage, the pixels within the winter wheat mask are classified according to their growth status. Representative sample points are selected from each growth status level, and the multi-temporal remote sensing features of the representative sample points are paired with the corresponding single-point simulated yield to obtain a machine learning training sample set. Using the machine learning training sample set as input, a regional winter wheat yield estimation model is established using the random forest regression algorithm. Multi-temporal remote sensing features are used as model input, and the single-point simulated yield is used as model output, so that the regional winter wheat yield estimation model can characterize the correspondence between remote sensing features and yield under the constraints of crop growth model mechanism. The regional winter wheat yield estimation model is applied to all pixels within the winter wheat mask. Using the multi-temporal remote sensing features of each pixel as input, the spatial distribution results of winter wheat yield with a spatial resolution of 10m are obtained. Based on the spatial distribution results of yield, area-weighted statistics are performed on administrative units to obtain the winter wheat yield estimation results at the administrative unit scale.

[0007] Preferably, the object-oriented classification method uses a segmentation algorithm based on simple non-iterative clustering to segment the multi-temporal Sentinel satellite active and passive remote sensing images. The Sentinel radar polarization features, vegetation index, and texture features calculated from the gray-level co-occurrence matrix are used as input features. A supervised classification algorithm based on random forest is used to classify winter wheat and non-winter wheat, so that the overall accuracy of winter wheat planting range identification is not less than 95%.

[0008] Preferably, in the step of acquiring multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat, multiple images with cloud cover of less than 10% are selected from the optical images during the winter wheat greening and jointing stage and are combined using time-series averages. Radar images are acquired by combining monthly averages to ensure the temporal continuity and spatial consistency of the multi-temporal remote sensing features.

[0009] Preferably, the radiative transfer model is a unified radiative transfer model for leaves and canopy, using the multispectral reflectance of each sample point on each observation date as input to perform forward modeling simulation of leaf area index and canopy structure parameters related to leaf area index; the parameter optimization method adopts a direct search optimization algorithm in the leaf area index inversion process, with the goal of minimizing the error between the forward modeled reflectance and the corresponding band reflectance in the multi-temporal remote sensing features, to solve for the time series of leaf area index of each sample point on each observation date.

[0010] Preferably, when calibrating crop growth model parameters related to the maximum leaf area index, the parameter optimization method first selects a set of crop growth model parameters sensitive to the maximum leaf area index based on a global sensitivity analysis method. These parameters are then used as the parameters to be optimized. A particle swarm optimization algorithm is then used to search for parameter combinations to minimize the root mean square error between the leaf area index time series simulated by the calibrated crop growth model and the leaf area index time series inverted from the radiative transfer model.

[0011] Preferably, when the ensemble Kalman filtering method inputs the leaf area index time series of each of the sample points as an assimilation variable into the calibrated crop growth model, it uses the leaf area index as an external observation, updates the state variables of the crop growth model and the crop growth model parameters related to yield at each time node with remote sensing observation, outputs the single-point simulated yield of each of the sample points, and constructs a yield sample library containing the location of the sample points, the leaf area index time series, and the single-point simulated yield.

[0012] Preferably, based on the cumulative normalized vegetation index (NDI) values ​​during the winter wheat heading stage and the flowering stage, the pixels within the winter wheat mask are classified according to their growth vigor, and representative sample points are selected from each growth vigor level, including: Using the cumulative value of the normalized vegetation index during the heading stage and the cumulative value of the normalized vegetation index during the flowering stage as growth evaluation indicators, the growth was divided into four levels: excellent growth, good growth, average growth and poor growth using the natural discontinuity grading method. Representative sample points were selected in each growth level according to the principle of spatial uniform distribution.

[0013] Preferably, the number of samples in the machine learning training sample set is not less than 625, and the representative sample points cover the winter wheat planting areas of each growth level and each year within the winter wheat mask. The number of samples of each growth level is allocated according to a preset ratio, so that the regional winter wheat yield estimation model established by the random forest regression algorithm has stable generalization ability under multiple years and multiple growth conditions.

[0014] Preferably, in the step of applying the regional winter wheat yield estimation model to all pixels within the winter wheat mask, the crop growth model assimilated by the ensemble Kalman filter is run only at the sample point to obtain the simulated yield at that single point. Within the entire pixel range of the winter wheat mask, the crop growth model and ensemble Kalman filter assimilation are not run again. Instead, the multi-temporal remote sensing features of each pixel are input into the regional winter wheat yield estimation model to obtain the spatial distribution result of the winter wheat yield, and the spatial distribution result of the winter wheat yield is output as the regional winter wheat yield estimation result.

[0015] A smart estimation system for regional winter wheat yield that integrates crop mechanism models and machine learning includes: The winter wheat mask construction unit is used to acquire multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat, extract vegetation index, texture features and spectral band features, identify the winter wheat planting range using an object-oriented classification method, and obtain the winter wheat mask and the multi-temporal remote sensing features corresponding to the winter wheat mask. The leaf area index inversion and crop growth model calibration unit is used to select sample points within the winter wheat mask, invert the multi-temporal remote sensing features of each sample point based on the radiative transfer model, obtain the leaf area index time series of each sample point, use the leaf area index time series together with meteorological parameters, soil parameters and field management parameters as input to drive the crop growth model, and calibrate the model parameters related to the maximum leaf area index using parameter optimization methods according to the target maximum leaf area index, so as to obtain the calibrated crop growth model. The leaf area index assimilation and single-point simulated yield generation unit is used to input the leaf area index time series of each sample point as an assimilation variable into the calibrated crop growth model, perform sequential assimilation using the ensemble Kalman filter method, update the state variables and yield-related parameters of the crop growth model, output the single-point simulated yield of each sample point, and construct a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield. The growth grading and representative sample point construction unit is used to grade the growth of pixels in the winter wheat mask based on the cumulative value of the normalized vegetation index during the heading stage and the cumulative value of the normalized vegetation index during the flowering stage. Representative sample points are selected from each growth grading level, and the multi-temporal remote sensing features of the representative sample points are paired with the corresponding single-point simulated yield to obtain a machine learning training sample set. The regional yield estimation model training unit is used to establish a regional winter wheat yield estimation model by taking the machine learning training sample set as input and using the random forest regression algorithm. It takes multi-temporal remote sensing features as model input and the single-point simulated yield as model output, so that the regional winter wheat yield estimation model can represent the correspondence between remote sensing features and yield under the constraints of crop growth model mechanism. The regional yield spatialization representation and statistical output unit is used to apply the regional winter wheat yield estimation model to all pixels within the winter wheat mask. Using the multi-temporal remote sensing features of each pixel as input, it obtains the spatial distribution result of winter wheat yield with a spatial resolution of 10m. Based on the yield spatial distribution result, it performs area-weighted statistics on administrative units to obtain the winter wheat yield estimation result at the administrative unit scale.

[0016] The present invention discloses the following technical effects: This invention significantly improves the accuracy and relevance of remote sensing input data in the regional winter wheat yield estimation process by first constructing a winter wheat mask and acquiring corresponding multi-temporal remote sensing features. Compared with prior art methods that rely solely on simple threshold segmentation or coarse-grained classification, this invention employs object-oriented classification technology that integrates vegetation indices, texture features, and spectral features. This greatly improves the accuracy of winter wheat spatial distribution identification, providing a high-quality data foundation for subsequent crop physiological parameter inversion, sample assimilation, and yield spatial extrapolation, thereby reducing the accumulation of yield estimation errors from the source.

[0017] This invention introduces a radiative transfer model to retrieve the leaf area index at the sample point level and combines it with meteorological, soil, and field management parameters to drive the crop growth model. By using parameter optimization methods to calibrate model parameters related to the maximum leaf area index, it effectively solves the problems of parameter estimation difficulties and unstable model accuracy in the regional application of crop growth models in the prior art. Because the model parameters are constrained to the retrieved leaf area index time series, this invention achieves local adaptation of the crop model, significantly enhancing the consistency between single-point growth simulation and actual growth, thereby improving the reliability of simulated yield.

[0018] This invention uses the time series of leaf area index at sample points as an assimilation variable input into a calibrated crop growth model. It sequentially updates state variables and yield-related parameters using an ensemble Kalman filter method, obtaining high-precision single-point simulated yield at a limited number of locations. This overcomes the prominent problem in previous techniques where "direct assimilation at the regional scale leads to enormous computational burden and low efficiency." By restricting the assimilation calculation to sample points, this invention significantly reduces the computational burden while maintaining the advantages of mechanistic constraints, making the mechanistic model-remote sensing assimilation system more adaptable to the actual needs of regional-scale yield estimation.

[0019] This invention classifies crop growth based on the cumulative normalized vegetation index (NDI) values ​​during the heading and flowering stages, and selects representative sample points from each growth level to construct a machine learning training sample set. This solves the problems of machine learning yield estimation in the background technology, which heavily relies on actual yield samples, has high sample timeliness, and is difficult to collect. Since the training samples of this invention are derived from single-point simulated yields obtained through assimilation of crop growth models, they have clear physiological significance and cross-year consistency, freeing the machine learning model from dependence on a large number of offline measured yields and significantly improving the scalability of regional yield estimation models.

[0020] This invention employs a machine learning model for regional spatial extrapolation, mapping multi-temporal remote sensing features to a 10m resolution spatial distribution of winter wheat yield. This achieves a harmonious balance between the constraints of crop growth model mechanisms and the spatial coverage capabilities of remote sensing data, overcoming the dual shortcomings of low efficiency in regional extrapolation of mechanistic models and a lack of mechanistic basis in empirical models in previous technologies. By utilizing single-point simulated yields with good mechanistic consistency to construct training samples, this invention achieves high-resolution regional yield estimation without sacrificing estimation accuracy. It provides a technical approach for multi-scale, multi-year winter wheat yield monitoring that combines accuracy, efficiency, and scalability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The purpose of this invention is to provide a method and system for intelligent estimation of regional yield of winter wheat that integrates crop mechanism model and machine learning. It utilizes the mechanism model to assimilate and construct training samples with physiological consistency, and uses multi-temporal remote sensing features to drive the machine learning model to complete the spatial expression of regional yield, thereby achieving a synergistic improvement in the accuracy, efficiency and generalizability of regional yield estimation of winter wheat.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides an intelligent estimation method for regional winter wheat yield that integrates crop mechanism models and machine learning, comprising: Step 100: Acquire multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat, extract vegetation index, texture features and spectral band features, identify the winter wheat planting range using an object-oriented classification method, and obtain the winter wheat mask and the multi-temporal remote sensing features corresponding to the winter wheat mask. Step 200: Select sample points within the winter wheat mask, invert the multi-temporal remote sensing features of each sample point based on the radiative transfer model, obtain the leaf area index time series of each sample point, and use the leaf area index time series together with meteorological parameters, soil parameters and field management parameters as input to drive the crop growth model. Based on the target maximum leaf area index, use parameter optimization methods to calibrate the model parameters related to the maximum leaf area index, and obtain the calibrated crop growth model. Step 300: Input the leaf area index time series of each sample point as an assimilation variable into the calibrated crop growth model, use the ensemble Kalman filter method for sequential assimilation, update the state variables and yield-related parameters of the crop growth model, output the single-point simulated yield of each sample point, and construct a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield. Step 400: Based on the cumulative normalized vegetation index (NVI) values ​​of winter wheat during the heading stage and the flowering stage, classify the growth status of the pixels within the winter wheat mask, select representative sample points from each growth status level, and pair the multi-temporal remote sensing features of the representative sample points with the corresponding single-point simulated yield to obtain the machine learning training sample set. Step 500: Using the machine learning training sample set as input, a regional winter wheat yield estimation model is established using the random forest regression algorithm. Multi-temporal remote sensing features are used as model input, and single-point simulated yield is used as model output, so that the regional winter wheat yield estimation model can represent the correspondence between remote sensing features and yield under the constraints of crop growth model mechanism. Step 600: Apply the regional winter wheat yield estimation model to all pixels within the winter wheat mask, use the multi-temporal remote sensing features of each pixel as input, obtain the spatial distribution results of winter wheat yield with a spatial resolution of 10m, and perform area-weighted statistics on administrative units based on the yield spatial distribution results to obtain the winter wheat yield estimation results at the administrative unit scale.

[0027] Specifically, in this embodiment, step 100 first selects multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat in the target area. The active remote sensing images are optical images in the visible and near-infrared bands, and the passive remote sensing images are synthetic aperture radar images. In this embodiment, it is preferred to screen and check the quality of multiple images with cloud cover of less than 10% in the optical images during the winter wheat greening to jointing stage, retaining only images with small cloud coverage areas and complete geometric correction. Then, the multiple images of the same stage are averaged and synthesized in the time dimension to obtain a staged optical temporal image with a spatial resolution of 10 meters. The radar images are grouped according to natural months, and the multiple radar images are averaged and synthesized in each month to form a radar temporal image covering the greening, jointing, heading, flowering, grain filling and maturity stages, so as to ensure the temporal continuity and spatial consistency of the multi-temporal remote sensing features. In this embodiment, it is preferred that the number of effective time phases of optical time-series images be no less than 3, and the number of effective time phases of radar time-series images be no less than 6, so as to ensure that the key stages of the entire growth process of winter wheat are observed.

[0028] In this embodiment, after completing the spatiotemporal registration of optical and radar time-series images, multi-temporal remote sensing features for winter wheat identification are extracted from each image period. In this embodiment, multi-temporal remote sensing features refer to a combined feature sequence of vegetation indices, radar backscattering features, and texture features extracted for the same pixel or pixel object on different observation dates. The vegetation indices preferably include the normalized vegetation index calculated based on the red and near-infrared bands, used to characterize the canopy leaf area and growth status of winter wheat; the radar backscattering features preferably include the backscattering coefficients of the vertical transmission and vertical reception channels, used to characterize canopy structure and soil moisture information; the texture features preferably involve constructing a gray-level co-occurrence matrix on a single band of each optical image period, and calculating contrast, homogeneity, and entropy indices based on a 5×5 pixel neighborhood window, used to describe the spatial texture differences of winter wheat plots. In this embodiment, it is preferable to form a multi-temporal remote sensing feature sequence containing at least 3 vegetation index features, 2 radar backscatter features and 3 texture features on each observation date, so that the feature dimension of a single pixel or pixel object is not less than 8 dimensions within a growing season.

[0029] In this embodiment, to achieve object-oriented classification of winter wheat planting areas, a segmentation algorithm based on simple non-iterative clustering is first used in step 100 to segment multi-temporal Sentinel satellite active and passive remote sensing images. In this embodiment, the simple non-iterative clustering segmentation algorithm refers to aggregating adjacent pixels into several image objects based on spectral similarity and spatial adjacency relationships between optical and radar images, without iterative optimization. Each image object is considered a ground feature unit with relatively uniform spectral and textural features. Preferably, the segmentation scale parameter is set between 30 and 50 to ensure that the generated image objects basically match the spatial scale of the winter wheat plots. The mean of the aforementioned multi-temporal remote sensing features is calculated within each image object as an object-level feature vector. Sample objects from both winter wheat and non-winter wheat plots are used as training samples to train a supervised classification algorithm based on random forest. All image objects are then classified as either winter wheat or non-winter wheat. Finally, objects classified as winter wheat are grouped together to form a winter wheat mask. In this embodiment, the number of training sample objects is preferably no less than 2,000, the number of decision trees in the random forest is set between 100 and 200, and the overall accuracy of the final winter wheat planting range identification is no less than 95%, with corresponding user accuracy and mapping accuracy both no less than 90%.

[0030] Optionally, in step 200 of this embodiment, sample points for model inversion and calibration are first selected within the winter wheat mask. In this embodiment, sample points refer to representative pixel locations within the winter wheat planting area. Each sample point corresponds to a complete multi-temporal remote sensing feature sequence and subsequent crop growth process simulation. Preferably, sample points are distributed within the study area using a combination of regular grids and typical plots, ensuring that the sample points cover different soil types and planting management methods while maintaining a uniform spatial distribution. Preferably, the number of sample points is no less than 100, and the distance between any two sample points is preferably greater than 100 meters to reduce the interference of spatial autocorrelation on the parameter inversion results.

[0031] In this embodiment, for each sample point, a unified radiative transfer model for leaves and canopy is used to invert multi-temporal remote sensing features to obtain the leaf area index time series for each observation date. The unified radiative transfer model for leaves and canopy in this embodiment refers to a physical model that simultaneously describes the optical properties of leaves and the structural features of the canopy. This model takes the multispectral reflectance of each observation date as input and maps variables such as leaf pigment content, leaf structure, canopy leaf area, and canopy vertical distribution to surface reflectance. The leaf area index inversion process employs a direct search optimization method. That is, within a pre-defined range of leaf area index and the range of canopy structural parameters related to the leaf area index, different parameter combinations are progressively searched. The error between the simulated reflectance output by the radiative transfer model and the corresponding band reflectance in the multi-temporal remote sensing features is calculated. The leaf area index corresponding to the parameter combination with the smallest error is taken as the inversion result for that observation date. In this embodiment, the search range of the leaf area index is preferably set to 0 to 7, and the search range of the canopy structure related parameters covers the common values ​​of the actual winter wheat canopy height and leaf tilt angle, so that the minimum inversion error of a single observation date is preferably controlled between 0.02 and 0.05.

[0032] In this embodiment, the leaf area index time series obtained through radiative transfer model inversion, along with meteorological parameters, soil parameters, and field management parameters, are used as inputs to drive the crop growth model to simulate the growth process. Meteorological parameters in this embodiment refer to daily maximum and minimum temperatures, total solar radiation, precipitation, and wind speed, preferably derived from reanalysis data or station observation data covering the study area, and uniformly interpolated to the sample point locations. Soil parameters in this embodiment include soil texture, soil bulk density, field water holding capacity, and soil organic matter content, used to characterize the water storage and supply capacity of the root zone. Field management parameters include sowing date, sowing amount, fertilization time, and fertilization amount, used to describe the artificially regulated conditions for crop growth. In this embodiment, the preferred meteorological data time step is 1 day, covering a continuous period from 10 days before winter wheat sowing to 10 days after maturity and harvest. The soil profile depth is preferably set to at least 1 meter, and values ​​are assigned based on experimental site data or regional soil survey data.

[0033] In this embodiment, to calibrate crop growth model parameters related to the maximum leaf area index (LAI), a global sensitivity analysis method is first employed to select a set of parameters sensitive to the maximum LAI from all candidate parameters of the crop growth model. In this embodiment, global sensitivity analysis refers to performing multiple joint samplings of parameters within a given parameter value range. By comparing the degree of change in the simulated maximum LAI under different parameter combinations, the influence strength of each parameter on the maximum LAI is quantitatively assessed, and the set of parameters requiring focused calibration is determined based on the magnitude of the influence. After selecting this set of sensitive parameters as the parameters to be optimized, a particle swarm optimization method is used to search for parameter combinations to be optimized. The goal is to minimize the root mean square error between the LAI time series simulated by the crop growth model and the LAI time series inverted from the radiative transfer model. The values ​​of the parameters to be optimized are iteratively updated until the error convergence condition is met. In this embodiment, the preferred number of sensitive parameters is 4, the parameter search range is set to be 20% above and below the commonly used values ​​based on existing literature and regional experience, the particle swarm size is preferably set to 30 to 50 candidate solutions, the number of iterations is preferably between 80 and 150, and the root mean square error is preferably controlled below 0.3, thereby obtaining the calibrated crop growth model.

[0034] In a further specific application of this embodiment, winter wheat in Hebi City, Henan Province, is taken as the research object. In step 200, the inversion results of the radiative transfer model are first verified by combining the measured leaf area index (LAI) data. This embodiment preferably conducts multiple field observations within a single growing season. The measured LAI of sample points is obtained using a vegetation canopy analyzer at the greening, jointing, booting, and grain-filling stages, and the remote sensing image observation time for the corresponding dates is recorded. Subsequently, using the multispectral reflectance of these dates as input, the LAI is inverted using the SUBPLEX algorithm in conjunction with a unified leaf and canopy radiative transfer model. The inverted LAI is then compared with the measured LAI. The results show that the coefficient of determination of the inversion method based on SUBPLEX and the radiative transfer model at the sample points in Hebi City is approximately 0.86, and the root mean square error of the LAI is approximately 0.72. This indicates that the inversion method can obtain the time series of LAI at sample points with high accuracy, providing reliable input for subsequent crop growth model driving. As an example, the main crop parameters of winter wheat in the crop growth model (WOFOST model) of this embodiment are shown in Table 1.

[0035] Table 1. Main crop parameters of winter wheat in the WOFOST model

[0036] In this embodiment, the meteorological driving data for the crop growth model preferably uses daily temperature, radiation, and precipitation data from the NASA Power dataset, and time-series meteorological data for the corresponding grid is extracted based on the latitude and longitude coordinates of the sample points. Soil parameters are preferably obtained from the national-scale soil database, including information such as wilting humidity, field water holding capacity, and soil organic matter content, and are uniformly converted into the soil hydrological parameter format required by the crop growth model. Field management parameters are determined based on the actual winter wheat planting system in the study area, including sowing date, harvest date, and main fertilization stage. Within the Hebi City area, sowing time is preferably from early to mid-October each year, and harvest time is preferably from early June of the following year. Through the combination of the above meteorological parameters, soil parameters, and field management parameters, this embodiment ensures that the daily driving data of the crop growth model is continuous and has clear physical meaning throughout a complete growing season. As an example, the WOFOST model management and main growing season parameters of this embodiment are shown in Table 3.

[0037] Table 3 WOFOST Model Management and Key Fertility Period Parameters

[0038] In this embodiment, to identify crop growth model parameters that significantly affect the maximum leaf area index (LAI) under NASA Power meteorological data conditions, the Sobol global sensitivity analysis method is introduced to evaluate the contribution of candidate parameters to the crop growth process. In this embodiment, Sobol global sensitivity analysis involves jointly perturbing multiple parameters within a given parameter value range and calculating the first-order sensitivity and total sensitivity indices based on the degree of change in model output under different parameter combinations. By performing sensitivity analyses on the two target variables, potential dry matter and maximum LAI, this embodiment found that some parameters related to initial aboveground matter, temperature response, and dry matter distribution have significantly higher sensitivities than other parameters. Four crop growth model parameters most sensitive to the maximum LAI were selected as the set of parameters to be optimized, thus compressing the search space for subsequent parameter calibration.

[0039] In this embodiment, the four crop growth model parameters sensitive to the maximum leaf area index (LAI) are jointly calibrated using particle swarm optimization (PSO). PSO in this embodiment refers to setting several candidate solutions in a multidimensional parameter space, treating each candidate solution as a particle, and iteratively updating the particle positions within the parameter space according to both individual optimality and swarm optimality criteria, so that the objective function gradually decreases. The objective function in this embodiment is defined as the root mean square error (RMSE) between the LAI time series obtained from the crop growth model simulation and the LAI time series obtained from the radiative transfer model inversion. The parameter search range is extended by approximately 20% both above and below the recommended values ​​in the literature and based on regional experience. In this embodiment, the preferred particle number is set between 30 and 50, and the number of iterations is set between 80 and 150. The RMSE of the calibrated crop growth model at multiple sample points is preferably controlled below 0.3. Based on subsequent single-point yield accuracy evaluation, the coefficient of determination between the single-point yield estimated using the calibrated model and assimilation method and the measured yield is close to 0.86, and the RMSE of the yield is approximately 570 kg / ha.

[0040] In this embodiment, the particle swarm optimization method used to calibrate crop growth model parameters related to the maximum leaf area index is further improved with sensitivity constraints, directly embedding the sensitivity results obtained from global sensitivity analysis into the particle update formula. Specifically, in the parameter space, the position of each particle represents a set of crop growth model parameter values. In the t-th iteration, the velocity update for the j-th parameter to be optimized is: in, For the first The particle velocity of the j-th parameter to be optimized in the next iteration; Let be the particle velocity of the j-th parameter to be optimized in the t-th iteration; Let be the inertia weight at the t-th iteration; For individual learning factors; Let be the global sensitivity coefficient of the j-th parameter to be optimized; A uniformly random number between 0 and 1; Let be the optimal position of the individual particle at the j-th parameter to be optimized during the t-th iteration; Let be the value of the particle's current position on the j-th parameter to be optimized at the t-th iteration; For group learning factors; A uniformly random number between 0 and 1; Let be the value of the j-th parameter to be optimized at the global optimal position of all particles in the t-th iteration. For error feedback weights; is the dimensionless objective function value calculated based on the time-series root mean square error of the leaf area index at the t-th iteration; This is the moving average of all historical objective function values ​​up to the t-th iteration. In this embodiment, it is preferred that... and The value ranges from 1.0 to 2.0. The sensitivity coefficient ranges from 0.1 to 0.5. After normalization, the values ​​are limited to between 0 and 1, allowing sensitive parameters to receive a larger update range during the search process, while non-sensitive parameters automatically shrink their update step size.

[0041] In this embodiment, to adaptively balance the global search and local convergence of the parameter space during particle swarm optimization iterations, an error-driven exponential decay form is introduced for the inertia weight. The exploration capability of the particle swarm is automatically adjusted based on the current iteration error, and the inertia weight is updated as follows: in, The inertia weight at the i-th iteration; This represents the lower limit of the inertia weight. This represents the upper limit of the inertia weight; This is the error sensitivity coefficient; This is the dimensionless objective function value obtained after normalizing the time-series root mean square error based on the leaf area index at the t-th iteration. In this embodiment, it is preferred that... Take a value between 0.2 and 0.4. Take a value between 0.8 and 0.95. The value is between 1.0 and 3.0. When the objective function value is large, the inertial weight approaches the upper limit to enhance the global search capability. When the objective function value gradually decreases and approaches convergence, the inertial weight automatically decays and approaches the lower limit to promote the aggregation of parameters near the optimal solution and improve the stability of crop growth model parameter calibration.

[0042] In this embodiment, to ensure that the crop growth model parameters obtained based on particle swarm optimization not only fit well at the maximum leaf area index, but also maintain the consistency of the leaf area index over time throughout the entire growth period, this embodiment sets the objective function value... Defined as a normalized comprehensive index of the time-series root mean square error of leaf area index at multiple key growth stages for each sample point, and using an error convergence threshold as the stopping condition during the iteration process. When the error convergence threshold is reached in several consecutive iterations... The iteration terminates and outputs the optimal particle position when the position remains below a preset threshold. By combining the above-mentioned sensitivity-constrained particle swarm update formula with the error adaptive inertia weight update formula, this embodiment can stably control the root mean square error of the leaf area index time series below 0.3 within a relatively small number of iterations, under the conditions of 30 to 50 particles and a maximum number of iterations of 80 to 150. This significantly reduces invalid searches and improves the convergence speed and reliability of crop growth model parameter calibration.

[0043] Further, in step 300 of this embodiment, after completing the crop growth model parameter calibration in step 200, the leaf area index time series of each sample point is input as an assimilation variable into the calibrated crop growth model, and sequential assimilation is performed using the ensemble Kalman filter method. In this embodiment, ensemble Kalman filtering refers to using a set of independent model ensemble members to simultaneously simulate the crop growth process. On dates with remote sensing observations, the leaf area index inversion result is used as an external observation and compared with the leaf area index predicted by each ensemble member. Based on the difference between the observation and the prediction, the state variables and some yield-related parameters of the crop growth model are jointly updated, making the updated ensemble statistically closer to the actual winter wheat growth state. In this embodiment, it is preferable that the leaf area index observation time points within a growing season are no less than 4 and no more than 8, the number of ensemble members is preferably between 30 and 80, and the time step of the crop growth model is 1 day, covering a continuous growth cycle from 10 days before sowing to 10 days after harvest.

[0044] In this embodiment, the state variables of the crop growth model refer to the internal variables describing the growth status of winter wheat on any given date, including aboveground dry matter, leaf dry matter, stem dry matter, sink organ dry matter, and root zone soil moisture content. Yield-related parameters refer to crop physiological parameters that control the proportion of dry matter allocated to sink organs, sink organ moisture content at maturity, and photosynthetic efficiency. During the ensemble Kalman filter sequential assimilation process, this embodiment first uses the calibrated crop growth model to predict the leaf area index (LAI) of all ensemble members daily at each time point with remote sensing observations of LAI. Then, based on the deviation between the observed LAI and the predicted LAI, the state variables and yield-related parameters of each ensemble member are corrected according to the update principle of the ensemble Kalman filter, making the updated model more closely resemble the observed time series in subsequent growth simulations. Through this "prediction-update-repredict" cycle, this embodiment achieves multiple LAI assimilations within the same growing season, preferably keeping the root mean square error of the assimilated LAI time series within 0.3.

[0045] In this embodiment, after completing the ensemble Kalman filter sequential assimilation for the entire growing season, the updated crop growth model is used to calculate the single-point simulated yield for each sample point. In this embodiment, the single-point simulated yield refers to the yield per unit area calculated at harvest time by combining the dry matter content of the sink organs output by the model with the conventional harvest moisture content of winter wheat in the region; the preferred unit is kilograms per hectare. This embodiment performs multi-year simulations for each sample point, repeating the assimilation process in years with different sowing dates and meteorological conditions to obtain a single-point simulated yield sequence covering multiple growing seasons. The spatial location of the sample points, the temporal sequence of leaf area index, and the corresponding single-point simulated yield are recorded in a unified data table to construct a yield sample library. Preferably, this embodiment obtains no fewer than 100 effective sample points within the study area, and after covering three growing seasons, the number of yield sample entries is no fewer than 300. The correlation index between these samples and the measured yield is preferably around 0.86, and the root mean square error of the single-point yield is preferably controlled within 600 kilograms per hectare, providing a reliable training sample basis for the subsequent construction of a regional winter wheat yield estimation model.

[0046] Furthermore, in this embodiment, step 400 first constructs a growth evaluation index based on the normalized vegetation index (NVI) of key growth stages of winter wheat. To this end, after completing steps 100 and 300, this embodiment extracts the NVI sequences of each pixel within the winter wheat mask for the corresponding dates of the booting and flowering stages, and accumulates the daily NVI values ​​within their respective time windows to obtain the cumulative values ​​of the NVI during the booting and flowering stages. The cumulative NVI value during the booting stage is used in this embodiment to reflect the population growth vigor and leaf area accumulation of winter wheat from jointing to booting, while the cumulative NVI value during flowering reflects the canopy retention and population photosynthetic capacity from flowering to the initial grain-filling stage. Preferably, the booting stage time window covers approximately 30 days, and the flowering stage time window covers approximately 20 days, with the cumulative results preferably controlled within the typical range of 0.5 to 6.0 in winter wheat planting areas.

[0047] In this embodiment, to convert the cumulative normalized vegetation index (NVR) values ​​of the two stages into discrete growth levels, the cumulative NVR values ​​of all pixels within the winter wheat mask during the heading and flowering stages are combined into a two-dimensional feature space. Each pixel is represented as a feature point with the cumulative value during the heading stage as the first dimension and the cumulative value during the flowering stage as the second dimension. In this feature space, the natural discontinuity grading method is used to segment the feature point cloud, dividing the overall sample into four levels: excellent growth, good growth, average growth, and poor growth. In this embodiment, the natural discontinuity grading method involves traversing different segmentation schemes and selecting a preferred combination of boundary values. This ensures that the cumulative NVR values ​​of pixels within the same level during the heading and flowering stages are relatively concentrated, while maximizing the differences between different levels. This guarantees that pixels with excellent growth have simultaneously high cumulative values ​​in both stages, while pixels with poor growth have significantly lower cumulative values ​​in at least one stage. In this embodiment, the preferred proportions of each growth level in the whole area are approximately 20%, 30%, 30%, and 20%, respectively.

[0048] In this embodiment, after completing the growth classification, to ensure that the machine learning training samples are representative of different growth conditions and spatial locations, representative sample points are selected in each growth level according to the principle of spatial uniform distribution. The principle of spatial uniform distribution in this embodiment means that within each growth level, the winter wheat mask is first divided into regular grids. Within each grid, pixels that are close to the grid center and possess complete multi-temporal remote sensing features and single-point simulated yield records are preferentially selected as candidate samples. Then, a repeatability check is performed on the candidate samples, and pixels that are too close to the selected samples are removed to avoid spatial clustering. Preferably, this embodiment selects no less than 150 representative sample points in each growth level, ensuring that the total number of representative sample points for the four growth levels is no less than 600, and ensuring that the distance between any two representative sample points is preferably greater than 50 meters to reduce the influence of spatial autocorrelation.

[0049] In this embodiment, the aforementioned representative sample points are matched with the yield sample library constructed in step 300 to form a machine learning training sample set. Specifically, for each representative sample point, the corresponding multi-temporal remote sensing feature sequence is extracted, including the vegetation index time series for the entire growing season, radar backscattering features for key time phases, and texture feature indicators defined in step 100. These are then paired one-to-one with the single-point simulated yield obtained after ensemble Kalman filtering assimilation of the sample point to form a single training sample record. In this embodiment, the above-mentioned growth grading and representative sample point selection process is repeated over three growing seasons, and samples from different years are merged to finally form a machine learning training sample set with no fewer than 625 samples. The proportion of samples for each growth grading level is preferably basically consistent with the proportion of the area of ​​the growth grading level. When constructing the training sample set, approximately 30% of the samples are reserved as independent validation data for subsequent evaluation of the generalization ability of the regional winter wheat yield estimation model based on the random forest regression algorithm under multi-year and multi-growth conditions.

[0050] Furthermore, in this embodiment, step 500 first constructs a regional winter wheat yield estimation model based on the machine learning training sample set obtained in step 400. In this embodiment, the machine learning training sample set refers to a set of samples formed by representative sample points. Each sample record consists of two parts: input features and output target. The input features are the multi-temporal remote sensing feature sequences of the representative sample point throughout the entire growing season, specifically including vegetation index time series, key temporal radar backscattering features, and texture features. The output target is the simulated yield of a single point obtained after ensemble Kalman filtering assimilation in step 300. To balance model stability and computational efficiency, this embodiment preferably divides the training sample set of no less than 625 samples into a training set and a validation set. The training set preferably accounts for 70%, and the validation set preferably accounts for 30%, ensuring that the training data covers different years, different growth levels, and different spatial locations, while retaining a sufficient number of independent samples for performance evaluation.

[0051] In this embodiment, the regional winter wheat yield estimation model is constructed using the random forest regression algorithm. The random forest regression algorithm in this embodiment refers to an ensemble learning method that uses multiple regression decision trees as base learners. During training, multiple subsets of samples are generated by sampling with replacement from the training samples. At each node split, a subset of features is randomly selected from all input features as candidate split variables, thereby constructing a set of regression decision trees with different structures and low error correlation. The average output of each regression decision tree is used as the final yield estimation result. In this embodiment, the number of regression decision trees in the random forest regression model is preferably set to 200 to 500, the maximum depth of a single regression decision tree is preferably limited to between 10 and 20 layers, and the number of candidate features at each node split is preferably set to be close to the square root of the total number of input features. This ensures the model's nonlinear fitting ability while avoiding overfitting and excessively long training time.

[0052] In this embodiment, to ensure that the regional winter wheat yield estimation model can stably represent the correspondence between remote sensing features and yield under the constraints of the crop growth model mechanism, this embodiment comprehensively evaluates the fitting accuracy and generalization ability of the random forest regression model after training. Evaluation indicators include the coefficient of determination, root mean square error, and mean absolute error on the training and validation sets. Based on the feature importance measurement method within the random forest, the contribution of vegetation indices, radar backscattering features, and texture features at different time phases to the yield estimation results is calculated, thereby verifying whether the model fully utilizes remote sensing information from key growth stages. In this embodiment, the coefficient of determination obtained on the training set is preferably not less than 0.90, and the coefficient of determination obtained on the validation set is preferably not less than 0.85. The corresponding root mean square error of yield is preferably controlled at around 600 kg / ha. Meanwhile, the feature importance ranking results show that the vegetation indices at the heading and flowering stages and the radar backscattering features in the later stages of growth are significantly more important than other time phase features, indicating that the regional winter wheat yield estimation model constructed in this embodiment can reasonably learn the mapping relationship between remote sensing features and yield under the constraints of the yield samples output by the crop growth model.

[0053] Optionally, in this embodiment, after training and validating the regional winter wheat yield estimation model in step 600, the model is applied to all pixels within the winter wheat mask to generate the spatial distribution results of winter wheat yield. Specifically, firstly, within the winter wheat mask area, multi-temporal remote sensing features corresponding to each pixel throughout the entire growing season are extracted, consistent with the feature types used when constructing the machine learning training sample set in step 400. These features include the vegetation index time series covering the greening stage, jointing stage, heading stage, flowering stage, grain filling stage, and maturity stage, radar backscattering features of key observation dates, and texture feature indices. In this embodiment, the spatial resolution of the winter wheat mask pixels is preferably 10 meters. The regional winter wheat yield estimation model is called one by one for all winter wheat pixels in the study area to obtain the estimated yield per unit area for each pixel in a single growing season, thereby forming raster data of the spatial distribution of winter wheat yield with 10-meter grids as the basic unit.

[0054] In this embodiment, to improve overall computational efficiency and avoid repeatedly running the crop growth model and ensemble Kalman filter assimilation process, this embodiment only runs the assimilated crop growth model on the selected sample points in step 300 to obtain single-point simulated yields, and uses these single-point simulated yields as output targets to train the regional winter wheat yield estimation model in step 500. In step 600, the crop growth model and ensemble Kalman filter assimilation are no longer run on other pixels within the winter wheat mask; instead, the multi-temporal remote sensing features of these pixels are directly input into the already trained and parameterized regional winter wheat yield estimation model, and the model outputs the corresponding yield values. Since the regional winter wheat yield estimation model has been trained and validated on samples from multiple years, multiple growth stages, and multiple spatial locations, this embodiment can obtain the spatial distribution results of the yield of the entire region at a computational cost far lower than that of area assimilation while maintaining estimation accuracy. Under the condition that the study area is approximately several thousand square kilometers and the number of winter wheat raster pixels reaches millions of pixels, the time required to complete the full-area yield estimation in this embodiment is preferably controlled to the order of several hours.

[0055] In this embodiment, after obtaining the spatial distribution raster data of winter wheat yield, to meet the needs of grain yield statistics and management, this embodiment further performs area-weighted statistics on winter wheat yield at the administrative unit scale. Specifically, based on the administrative division boundary vector data, the spatial distribution raster data of winter wheat yield is overlaid and analyzed within the administrative unit. The yield estimates of all winter wheat rasters falling within the same administrative unit are accumulated according to the raster area to obtain the total yield estimate of that administrative unit. At the same time, the number of winter wheat rasters and the total planting area within the administrative unit are recorded, and the average yield per unit area can be further calculated. This embodiment preferably uses county-level administrative units as the statistical objects. When the number of county-level administrative units in the study area is between ten and twenty, the number of winter wheat rasters in a single administrative unit is preferably not less than several thousand to ensure that the area-weighted statistical results are spatially representative.

[0056] In this embodiment, to facilitate comparison of spatial differences in winter wheat yield under different years, meteorological conditions, and planting management measures, the above steps are repeated over multiple growing seasons in the same region. Spatial distribution results and administrative unit-level yield estimates are generated for each year, and these results are organized and archived according to year and administrative unit, thus forming a regional winter wheat yield database with time-series characteristics. This database can be used in this embodiment to analyze yield change trends and spatial stability over three consecutive years or more. Preferably, the relative error between the interannual differences in the total regional winter wheat yield estimates at the administrative unit level for consecutive years and the yield data published by the statistical department is controlled within 10%, providing reliable data support for subsequent food security assessments and planting structure optimization.

[0057] Corresponding to the above methods, such as Figure 2 As shown in the figure, this embodiment also provides an intelligent estimation system for regional winter wheat yield that integrates crop mechanism models and machine learning, including: The winter wheat mask construction unit is used to acquire multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat, extract vegetation index, texture features and spectral band features, identify the winter wheat planting range using an object-oriented classification method, and obtain the winter wheat mask and the multi-temporal remote sensing features corresponding to the winter wheat mask. The leaf area index inversion and crop growth model calibration unit is used to select sample points within the winter wheat mask, invert the multi-temporal remote sensing features of each sample point based on the radiative transfer model, obtain the leaf area index time series of each sample point, use the leaf area index time series together with meteorological parameters, soil parameters and field management parameters as input to drive the crop growth model, and calibrate the model parameters related to the maximum leaf area index using parameter optimization methods according to the target maximum leaf area index, so as to obtain the calibrated crop growth model. The leaf area index assimilation and single-point simulated yield generation unit is used to input the leaf area index time series of each sample point as an assimilation variable into the calibrated crop growth model, perform sequential assimilation using the ensemble Kalman filter method, update the state variables and yield-related parameters of the crop growth model, output the single-point simulated yield of each sample point, and construct a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield. The growth grading and representative sample point construction unit is used to grade the growth of pixels in the winter wheat mask based on the cumulative value of the normalized vegetation index during the heading stage and the cumulative value of the normalized vegetation index during the flowering stage. Representative sample points are selected from each growth grading level, and the multi-temporal remote sensing features of the representative sample points are paired with the corresponding single-point simulated yield to obtain a machine learning training sample set. The regional yield estimation model training unit is used to establish a regional winter wheat yield estimation model by taking the machine learning training sample set as input and using the random forest regression algorithm. It takes multi-temporal remote sensing features as model input and the single-point simulated yield as model output, so that the regional winter wheat yield estimation model can represent the correspondence between remote sensing features and yield under the constraints of crop growth model mechanism. The regional yield spatialization representation and statistical output unit is used to apply the regional winter wheat yield estimation model to all pixels within the winter wheat mask. Using the multi-temporal remote sensing features of each pixel as input, it obtains the spatial distribution result of winter wheat yield with a spatial resolution of 10m. Based on the yield spatial distribution result, it performs area-weighted statistics on administrative units to obtain the winter wheat yield estimation result at the administrative unit scale.

[0058] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0059] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligently estimating regional yield of winter wheat by integrating crop mechanism models and machine learning, characterized in that, include: Multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat were acquired. Vegetation index, texture features and spectral band features were extracted. An object-oriented classification method was used to identify the planting range of winter wheat and to obtain the winter wheat mask and the multi-temporal remote sensing features corresponding to the winter wheat mask. Sample points were selected within the winter wheat mask. Based on the radiative transfer model, the multi-temporal remote sensing features of each sample point were inverted to obtain the leaf area index time series of each sample point. The leaf area index time series, together with meteorological parameters, soil parameters, and field management parameters, were used as inputs to drive the crop growth model. Based on the maximum target leaf area index, the model parameters related to the maximum leaf area index were calibrated using parameter optimization methods to obtain the calibrated crop growth model. The leaf area index time series of each sample point is used as an assimilation variable and input into the calibrated crop growth model. The ensemble Kalman filter method is used for sequential assimilation to update the state variables and yield-related parameters of the crop growth model. The single-point simulated yield of each sample point is output, and a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield is constructed. Based on the cumulative normalized vegetation index (NVI) values ​​of winter wheat during the heading stage and the flowering stage, the pixels within the winter wheat mask are classified according to their growth status. Representative sample points are selected from each growth status level, and the multi-temporal remote sensing features of the representative sample points are paired with the corresponding single-point simulated yield to obtain a machine learning training sample set. Using the machine learning training sample set as input, a regional winter wheat yield estimation model is established using the random forest regression algorithm. Multi-temporal remote sensing features are used as model input, and the single-point simulated yield is used as model output, so that the regional winter wheat yield estimation model can characterize the correspondence between remote sensing features and yield under the constraints of crop growth model mechanism. The regional winter wheat yield estimation model is applied to all pixels within the winter wheat mask. Using the multi-temporal remote sensing features of each pixel as input, the spatial distribution results of winter wheat yield with a spatial resolution of 10m are obtained. Based on the spatial distribution results of yield, area-weighted statistics are performed on administrative units to obtain the winter wheat yield estimation results at the administrative unit scale.

2. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning as described in claim 1, characterized in that, The object-oriented classification method uses a segmentation algorithm based on simple non-iterative clustering to segment the multi-temporal Sentinel satellite active and passive remote sensing images. It uses Sentinel radar polarization features, vegetation index, and texture features calculated from the gray-level co-occurrence matrix as input features. It uses a supervised classification algorithm based on random forest to classify winter wheat and non-winter wheat, so that the overall accuracy of winter wheat planting range identification is not less than 95%.

3. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning as described in claim 1, characterized in that, In the process of acquiring multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat, multiple images with cloud cover of less than 10% are selected from the optical images during the winter wheat greening and jointing stage and then combined using time-series averages. Radar images are also acquired by combining monthly averages to ensure the temporal continuity and spatial consistency of the multi-temporal remote sensing features.

4. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning as described in claim 1, characterized in that, The radiative transfer model is a unified radiative transfer model for leaves and canopy. It uses the multispectral reflectance of each sample point on each observation date as input to perform forward modeling simulation of leaf area index and canopy structure parameters related to leaf area index. The parameter optimization method adopts a direct search optimization algorithm in the leaf area index inversion process. The objective is to minimize the error between the forward modeled reflectance and the corresponding band reflectance in the multi-temporal remote sensing features, and solve for the time series of leaf area index of each sample point on each observation date.

5. The intelligent estimation method for regional winter wheat yield based on the integration of crop mechanism model and machine learning as described in claim 1, characterized in that, When calibrating crop growth model parameters related to the maximum leaf area index (LAI), the parameter optimization method first selects a set of crop growth model parameters sensitive to the maximum LAI based on a global sensitivity analysis method. These parameters are then used as the parameters to be optimized. A particle swarm optimization algorithm is then used to search for parameter combinations that minimize the root mean square error between the LAI time series simulated by the calibrated crop growth model and the LAI time series inverted from the radiative transfer model.

6. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning according to claim 1, characterized in that, The ensemble Kalman filtering method, when inputting the leaf area index time series of each sample point as an assimilation variable into the calibrated crop growth model, uses the leaf area index as an external observation, updates the state variables of the crop growth model and the crop growth model parameters related to yield at each time node with remote sensing observation, outputs the single-point simulated yield of each sample point, and constructs a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield.

7. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning according to claim 1, characterized in that, Based on the cumulative normalized vegetation index (NVI) values ​​of winter wheat during the heading and flowering stages, the pixels within the winter wheat mask were classified according to their growth vigor. Representative sample points were selected from each growth vigor level, including: Using the cumulative value of the normalized vegetation index during the heading stage and the cumulative value of the normalized vegetation index during the flowering stage as growth evaluation indicators, the growth was divided into four levels: excellent growth, good growth, average growth and poor growth using the natural discontinuity grading method. Representative sample points were selected in each growth level according to the principle of spatial uniform distribution.

8. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning according to claim 1, characterized in that, The number of samples in the machine learning training sample set is no less than 625. The representative sample points cover the winter wheat planting areas of each growth level and each year within the winter wheat mask. The number of samples of each growth level is allocated according to a preset ratio, so that the regional winter wheat yield estimation model established by the random forest regression algorithm has stable generalization ability under multi-year and multi-growth conditions.

9. The intelligent estimation method for regional yield of winter wheat integrating crop mechanism model and machine learning according to claim 1, characterized in that, In the step of applying the regional winter wheat yield estimation model to all pixels within the winter wheat mask, the crop growth model assimilated by the ensemble Kalman filter is run only at the sample point to obtain the simulated yield at that single point. Within the entire pixel range of the winter wheat mask, the crop growth model and ensemble Kalman filter assimilation are not run again. Instead, the multi-temporal remote sensing features of each pixel are input into the regional winter wheat yield estimation model to obtain the spatial distribution result of winter wheat yield. The spatial distribution result of winter wheat yield is then output as the regional winter wheat yield estimation result.

10. A smart estimation system for regional winter wheat yield that integrates crop mechanism models and machine learning, characterized in that, include: The winter wheat mask construction unit is used to acquire multi-temporal Sentinel satellite active and passive remote sensing images covering the entire growth period of winter wheat, extract vegetation index, texture features and spectral band features, identify the winter wheat planting range using an object-oriented classification method, and obtain the winter wheat mask and the multi-temporal remote sensing features corresponding to the winter wheat mask. The leaf area index inversion and crop growth model calibration unit is used to select sample points within the winter wheat mask, invert the multi-temporal remote sensing features of each sample point based on the radiative transfer model, obtain the leaf area index time series of each sample point, use the leaf area index time series together with meteorological parameters, soil parameters and field management parameters as input to drive the crop growth model, and calibrate the model parameters related to the maximum leaf area index using parameter optimization methods according to the target maximum leaf area index, so as to obtain the calibrated crop growth model. The leaf area index assimilation and single-point simulated yield generation unit is used to input the leaf area index time series of each sample point as an assimilation variable into the calibrated crop growth model, perform sequential assimilation using the ensemble Kalman filter method, update the state variables and yield-related parameters of the crop growth model, output the single-point simulated yield of each sample point, and construct a yield sample library containing the location of the sample point, the leaf area index time series, and the single-point simulated yield. The growth grading and representative sample point construction unit is used to grade the growth of pixels in the winter wheat mask based on the cumulative value of the normalized vegetation index during the heading stage and the cumulative value of the normalized vegetation index during the flowering stage. Representative sample points are selected from each growth grading level, and the multi-temporal remote sensing features of the representative sample points are paired with the corresponding single-point simulated yield to obtain a machine learning training sample set. The regional yield estimation model training unit is used to establish a regional winter wheat yield estimation model by taking the machine learning training sample set as input and using the random forest regression algorithm. It takes multi-temporal remote sensing features as model input and the single-point simulated yield as model output, so that the regional winter wheat yield estimation model can represent the correspondence between remote sensing features and yield under the constraints of crop growth model mechanism. The regional yield spatialization representation and statistical output unit is used to apply the regional winter wheat yield estimation model to all pixels within the winter wheat mask. Using the multi-temporal remote sensing features of each pixel as input, it obtains the spatial distribution result of winter wheat yield with a spatial resolution of 10m. Based on the yield spatial distribution result, it performs area-weighted statistics on administrative units to obtain the winter wheat yield estimation result at the administrative unit scale.

Images