Grain yield estimation method based on land-atmosphere coupling model and machine learning algorithm

Abstract

The present invention discloses a grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm. This method relates to the field of agricultural meteorological forecasting and intelligent decision-making technology, and includes the following steps: Step S1: A dynamic corn growth model is integrated with a weather research and forecasting model through a kernel-level algorithm to construct a WRF-Crop coupling system. The WRF-Crop coupling system uses a bidirectional data channel to achieve real-time interaction between corn canopy transpiration and albedo changes and temperature, humidity, and radiation data. This grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm overcomes the limitations of the traditional dynamic corn growth model and meteorological model operating in isolation by constructing a WRF-Crop bidirectional coupling model. This method achieves dynamic feedback simulation between vegetation growth and local climate, effectively capturing the crop stress response mechanism under extreme climate events.

Description

Technical Field

[0001] The present invention relates to the field of agricultural meteorological forecasting and intelligent decision-making technology, and specifically to a grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm. Background Art

[0002] Traditional crop yield forecasting methods suffer from a fundamental bottleneck: crop growth models and atmospheric circulation models have long operated in isolation, resulting in a failure to effectively capture the dynamic interactions between vegetation and the climate system. While dynamic corn growth models can simulate physiological processes such as photosynthesis and water use, they fail to account for the feedback effects of canopy transpiration and albedo changes on local climate. Meteorological models, while capable of predicting climate factors such as temperature and precipitation, lack the ability to capture the dynamic responses of crop phenology. This one-way simulation model underestimates key mechanisms, such as crop stress responses to extreme weather events and the regulatory effects of regional climate microcirculation on corn growth. This leads to cumulative and amplified forecast errors during key growth stages, such as the tasseling and grain filling periods, over time. Furthermore, existing approaches that incorporate remote sensing observations for correction often employ static data assimilation strategies with fixed time steps, making them incapable of adapting to the sudden changes in environmental parameters during the rapid growth period. Furthermore, machine learning algorithms are used solely as standalone forecasting tools, lacking a collaborative optimization mechanism with physical models. This results in systematic deviations in forecasts when encountering climate patterns not covered by historical data. The combined effects of model isolation, data assimilation lag, and lack of algorithm interpretability have severely restricted the reliability of corn yield assessment in climate change-sensitive areas, becoming a core obstacle to the upgrading of agricultural precision management technology. Summary of the Invention

[0003] (1) Technical problems solved

[0004] In response to the shortcomings of the existing technology, the present invention provides a grain yield prediction method based on a land-atmosphere coupling model and a machine learning algorithm, which solves the problem of how to overcome the lack of a dynamic coupling mechanism between crop growth and the atmospheric environment.

[0005] (2) Technical solution

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm, comprising the following steps:

[0007] Step S1: A kernel-level algorithm fusion is performed on the dynamic corn growth model and the weather research and forecasting model to construct a WRF-Crop coupling system. The WRF-Crop coupling system realizes real-time interaction between corn canopy transpiration, albedo changes, and temperature, humidity, and radiation data through a bidirectional data channel;

[0008] Step S2: Divide the assimilation stages according to the maize phenological process, adopt a daily assimilation frequency from the seedling emergence to the jointing stage, and switch to a 6-hour assimilation frequency from the tasseling to the grain filling stage. Fusing multispectral remote sensing leaf area index, soil moisture satellite inversion data, and real-time monitoring values ​​from field sensors, a dynamic growth environment parameter set is generated; wherein, the preset assimilation frequency is a 6-hour assimilation frequency;

[0009] Step S3: construct a dual-channel physically constrained machine learning model that includes a long-short-term memory network and an attention mechanism. The model uses the meteorological-crop interaction data output by WRF-Crop as input, integrates historical climate-yield relationships and stress response characteristics through a gating unit, and imposes canopy heat flux conservation constraints and nitrogen accumulation continuity constraints during the prediction process.

[0010] In step S4, spatial block cross-validation and independent year extreme climate sample adversarial testing are used to perform Bayesian probability fusion on machine learning predicted yield, WRF-Crop mechanism simulated yield and agricultural statistical data. When the data difference exceeds the preset threshold, a parameter recalibration cycle is triggered until the error converges to the preset range.

[0011] Preferably, the construction process of the WRF-Crop coupling system in step S1 includes:

[0012] A dynamic corn growth model core is embedded within the land surface process of the Weather Research and Forecasting Model. This core integrates a dynamic library of corn physiological parameters, including root distribution and stomatal conductance thresholds. This library uses localized experiments to establish adaptive correction functions for environmental stress intensity. When constructing the WRF-Crop coupled system, the core algorithms of the dynamic corn growth model are seamlessly integrated into the land surface process of the meteorological model, enabling real-time interaction between two-way data channels. The dynamic corn growth model dynamically adjusts the vertical distribution of roots and the sensitive thresholds for stomatal opening and closing based on the characteristics of different corn growth stages. For example, during the jointing stage, it prioritizes shallow root water absorption, while enhancing deep root drought resistance after the tasseling stage. The meteorological model provides real-time feedback on local temperature, humidity, and radiation data, corrected for crop canopy transpiration. Key physiological parameters for each growth stage are calibrated through in situ field experiments, and dynamic correction rules are established to adapt these parameters to environmental stress intensity. This ensures that the model's spatial resolution during sensitive growth stages accurately matches regional climate characteristics. Environmental stress intensity includes drought index and the duration of accumulated high temperatures.

[0013] Preferably, the generation of the dynamic growth environment parameter set in step S2 further includes: dynamically optimizing the assimilation weight matrix based on the meteorological-crop mutual feedback residual output by WRF-Crop, wherein the residual includes the temperature feedback delay error and the transpiration calculation deviation, and after the weight matrix is ​​optimized, the error in capturing the environmental parameters during the tasseling period is reduced.

[0014] Preferably, the structure of the dual-channel, physically constrained machine learning model in step S3 includes: a long-short-term memory network channel for analyzing the nonlinear temporal relationship between historical climate data and corn yield, and an attention mechanism channel for extracting water stress and heat stress characteristics output by the dynamic corn growth model. The gating unit dynamically adjusts the weighting of the two channels based on the growth stage. The physically constrained machine learning model utilizes a dual-channel architecture. The long-short-term memory network channel focuses on analyzing the nonlinear temporal patterns between historical climate fluctuations and yield formation, while the attention mechanism channel captures the stress response signals output by the dynamic corn growth model in real time. The model dynamically integrates the dual-channel outputs via the gating unit and imposes energy and mass conservation constraints during the prediction process. When the difference between the predicted canopy sensible and latent heat fluxes exceeds the field measurement range, a reverse verification of the meteorological model boundary layer parameters is triggered. When the predicted nitrogen uptake rate systematically deviates from the dynamic corn growth model simulation results, the root nitrogen uptake depth parameter is preferentially adjusted, and the nitrogen allocation logic of the machine learning model is constrained. This constraint mechanism achieves coordinated optimization of prediction results and physical mechanisms through a multi-layer feedback loop.

[0015] Preferably, the canopy heat flux conservation constraint is achieved by forcing the prediction results to satisfy the sensible heat flux and latent heat flux balance equation, and the nitrogen accumulation continuity constraint is achieved by verifying the consistency of the nitrogen absorption rate in the predicted yield with the simulated value of the dynamic corn growth model.

[0016] Preferably, the operation process of spatial block cross-validation in step S4 is: divide the study area into 5km×5km grid units, randomly shield the input data of 20% of the area, use the data of the remaining 80% of the area to train the model and verify the prediction accuracy of the shielded area.

[0017] Preferably, the triggering condition for the parameter recalibration cycle in step S4 is: when the Bayesian probability fusion result shows that the difference between the machine learning predicted yield and the WRF-Crop mechanism simulated yield exceeds 8%, or the difference between both and agricultural statistical data exceeds 10%, the joint optimization of the stomatal conductance threshold of the dynamic corn growth model and the machine learning gating unit weight is initiated. The model verification system adopts a cross-scale strategy that combines spatial blocking with extreme climate confrontation testing. In spatial verification, local area data is shielded according to geographic grids, and the machine learning feature extraction weights are optimized based on the error distribution characteristics; wherein, the error distribution characteristics include the difference between irrigated areas and rain-fed areas; historical extreme climate year data are introduced in the confrontation test to limit the adjustment range of model parameters to prevent overfitting. When the verification error continues to exceed the standard, the system automatically triggers the joint calibration mechanism to simultaneously optimize crop physiological parameters, machine learning weights and data assimilation strategies until the error converges.

[0018] Preferably, the method for improving the resolution of the dynamic library of maize physiological parameters includes: using 3km×3km grid division during the tasseling and grain filling stages, dynamically correcting the stomatal conductance threshold parameters through real-time monitoring data from field sensors, and keeping the correction frequency synchronized with the assimilation stage.

[0019] Preferably, the frequency of two-way data interaction of the WRF-Crop coupling system during the grouting period is increased to once per minute, and the interactive data includes feedback correction values ​​of canopy temperature to boundary layer height.

[0020] Preferably, the yield estimation result outputted in step S4 includes the spatiotemporal resolution data generated by the WRF-Crop coupling system, and the data is specifically manifested as a correlation map between the daily canopy transpiration change curve and the temperature fluctuation of the corresponding grid. The reliability of the final yield estimation result is ensured by multi-source data fusion and reverse verification. When an abnormal correlation between canopy transpiration and temperature fluctuation is detected, the system automatically marks the problem area and starts a special calibration process; for the correlation deviation caused by the terrain, the dynamic weighted fusion output of the mechanism model and remote sensing data is adopted. All results must pass the dual verification of spatial block verification and historical parameter comparison to ensure the physical consistency of the simulation data.

[0021] (3) Beneficial effects

[0022] The present invention provides a grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm. It has the following beneficial effects:

[0023] (1) This grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm breaks through the limitations of the traditional dynamic corn growth model and meteorological model operating in isolation by constructing a WRF-Crop bidirectional coupling model, achieving dynamic feedback simulation between vegetation growth and local climate, and effectively capturing the crop stress response mechanism under extreme climate events. It combines growth-period-sensitive data assimilation with a physically constrained machine learning algorithm to improve the parameter capture accuracy of key stages such as the tasseling and grain filling periods, thereby reducing the yield estimation error compared to traditional methods. At the same time, a cross-scale verification system and a multimodal data fusion mechanism ensure the robustness of the model in climate change-sensitive areas and complex terrain areas, providing high-credibility support for agricultural disaster warning and precision irrigation decision-making.

[0024] (2) This grain yield estimation method based on the land-atmosphere coupling model and machine learning algorithm solves the problem of local application deviation of general models by deeply integrating physical mechanisms with data intelligence and adapting to regional corn characteristics through a dynamic parameter library; the machine learning prediction results are verified by energy balance and material conservation constraints to avoid the failure risk of black box models in unknown climate scenarios and enhance decision-making interpretability; high-frequency data interaction and terrain correction mechanisms further expand the application scenarios of the technology and support multi-level yield estimation needs from field scale to regional scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the overall structure / framework of the present invention;

[0026] Figure 2 It is a flow chart / control logic timing diagram of the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0028] See also Figure 1 and Figure 2 The present invention provides a technical solution: a grain yield estimation method based on a land-atmosphere coupling model and a machine learning algorithm, comprising the following steps:

[0029] Step S1: The dynamic corn growth model and the weather research and forecasting model are integrated at the kernel level to construct a WRF-Crop coupling system. The WRF-Crop coupling system realizes real-time interaction between corn canopy transpiration, albedo changes, and temperature, humidity, and radiation data through a bidirectional data channel.

[0030] Step S2: Divide the assimilation stages according to the maize phenological process. A daily assimilation frequency is used from the seedling emergence to jointing stage, and a 6-hour assimilation frequency is switched from the tasseling to the grain filling stage. Multispectral remote sensing leaf area index, soil moisture satellite inversion data, and real-time monitoring values ​​of field sensors are integrated to generate a dynamic growth environment parameter set.

[0031] Step S3: construct a dual-channel physically constrained machine learning model that includes a long-short-term memory network and an attention mechanism. The model uses the meteorological-crop interaction data output by WRF-Crop as input, integrates historical climate-yield relationships and stress response characteristics through a gating unit, and imposes canopy heat flux conservation constraints and nitrogen accumulation continuity constraints during the prediction process.

[0032] In step S4, spatial block cross-validation and independent year extreme climate sample adversarial testing are used to perform Bayesian probability fusion on machine learning predicted yield, WRF-Crop mechanism simulated yield, and agricultural statistical data. When the data difference exceeds the preset threshold, a parameter recalibration cycle is triggered until the error converges to the range of 5% to 8%.

[0033] The construction process of the WRF-Crop coupling system in step S1 includes: embedding the dynamic corn growth model kernel in the land surface process of the weather research and forecasting model. The dynamic corn growth model kernel integrates a dynamic library of corn physiological parameters including root distribution and stomatal conductance threshold. The parameter dynamic library establishes a correction function that is adaptive to environmental stress intensity through localized experiments. It should be further explained that, in the specific implementation process, during the construction of the WRF-Crop coupling system, the kernel of the dynamic corn growth model was first embedded in the land surface process of the weather research and forecasting model. The kernel established a real-time interactive interface with meteorological elements by analyzing the vertical distribution characteristics of the roots and the stomatal opening and closing threshold parameters during the corn growth period. Through localized field experiments, the physiological parameters of corn in different growth periods were obtained, including the density gradient distribution of the roots in the 0-50cm soil layer and the closing threshold of the stomatal conductance when the photosynthetic active radiation intensity was lower than 200μmol / m² / s. The parameter correction function was dynamically adjusted according to the intensity of environmental stress. Among them, the growth period includes the jointing stage, the tasseling stage, and the filling stage. The intensity of environmental stress includes a drought stress index exceeding 0.6 or a continuous high temperature exceeding 35°C.

[0034] When the model detects that the soil moisture content is lower than 60% of the field water holding capacity, it automatically triggers the downward adjustment mechanism of the stomatal conductance threshold and extends the root water absorption depth to the deep soil by 10-15 cm to simulate drought resistance. The update of the above parameter dynamic library is synchronized with the radiation balance calculation of the land surface process of the Weather Research and Forecasting Model (WRF), ensuring that during the sensitive stages of the tasseling and grain filling periods, the model spatial resolution is divided into 3km×3km grids to achieve coordinated simulation of the canopy microclimate and regional atmospheric circulation.

[0035] Generating the dynamic growth environmental parameter set in step S2 further includes dynamically optimizing an assimilation weight matrix based on the meteorological-crop feedback residuals output by WRF-Crop. The residuals include temperature feedback delay errors and transpiration calculation biases. The optimized weight matrix reduces the error in capturing environmental parameters during the tasseling period. It should be further explained that, in specific implementations, when the WRF-Crop coupled system detects a phenological phase transition signal from the tasseling period to the grain filling period, the data assimilation frequency is automatically increased from once daily to once every six hours. Furthermore, environmental parameters for priority assimilation are dynamically selected through real-time analysis of leaf area index mutation signatures from multispectral remote sensing data and water deficit gradients from satellite-derived soil moisture data. Leaf area index mutation signatures include daily growth exceeding 0.5, and water deficit gradients include soil moisture content in the 0-20 cm layer below 18%.

[0036] For the meteorological-crop feedback residuals output by WRF-Crop, if the residual type is a temperature feedback delay error (i.e., the time series offset between the model-predicted temperature and the field sensor measured value exceeds 2 hours), the assimilation weight of the remotely sensed leaf area index is increased to 0.7. If the residual type is a transpiration calculation bias (i.e., the difference between the simulated value and the sensor measured value exceeds 15%), the assimilation weight of the soil moisture data is increased to 0.6. In the event of regional high temperature events, such as a maximum temperature exceeding 35°C for three consecutive days, or extreme precipitation, such as a single-day rainfall exceeding 50 mm, an emergency optimization mode of bidirectional residual feedback is activated, reducing the update frequency of the assimilation weight matrix to once per hour. The correction amplitude of the albedo parameter in the WRF land surface process is automatically adjusted based on the direction of the residual (i.e., positive or negative deviation). This reduces the capture error of key environmental parameters during the tasseling period by 40% compared to the traditional static assimilation strategy. Key environmental parameters during the tasseling period include canopy temperature and soil available moisture.

[0037] The structure of the dual-channel physical constraint machine learning model in step S3 includes: a long short-term memory network channel is used to analyze the nonlinear time series relationship between historical climate data and corn yield, an attention mechanism channel is used to extract the water stress and heat stress characteristics output by the dynamic corn growth model, and a gating unit dynamically adjusts the weight ratio of the dual channels according to the growth stage. It should be further explained that during the specific implementation process, when running the dual-channel physically constrained machine learning model, the long-short-term memory network channel receives daily temperature, precipitation, and radiation data from the past five years as input. By analyzing the nonlinear relationship between the trend of accumulated temperature changes in the 30 days before the tasseling period and the cumulative sunshine hours during the grain filling period, a mapping relationship between climate fluctuations and historical yield is established. At the same time, the attention mechanism channel analyzes the water stress index and heat stress intensity output by WRF-Crop in real time to identify the occurrence time and spatial distribution patterns of key stress events. The water stress index includes a daily evaporation deficit exceeding 2 mm, and the heat stress intensity includes a canopy temperature exceeding 38°C for three consecutive hours. When the model detects that the current growth stage is the jointing stage, the gated unit sets the LSTM channel weight to 0.7 to strengthen the learning of historical climate patterns. If a continuous high temperature warning is issued after entering the tasseling period, such as a daily maximum temperature exceeding 35°C for more than three days, the attention channel weight is dynamically increased to 0.6 to enhance the real-time stress response.

[0038] During the prediction process, when the daily transpiration prediction value output by machine learning exceeds 15% of the WRF-Crop mechanism simulation value, the canopy heat flux conservation verification is forced to be triggered. If the difference between the sensible heat flux and the latent heat flux exceeds 20% of the actual field measured value, the hidden layer parameters of the LSTM channel are recalibrated until the energy balance error converges to within 5%; for the nitrogen accumulation continuity constraint, when the nitrogen absorption rate in the predicted yield deviates from the dynamic corn growth model simulation value by more than 10% in the late grain filling period, the feature extraction intensity of the attention mechanism for soil nitrate nitrogen concentration data is automatically increased, and the weight of historical climate data is reduced through the gating unit to ensure that the material conservation constraint is met first; among them, the late grain filling period is 40 to 50 days after flowering.

[0039] The canopy heat flux conservation constraint is implemented by forcing the predicted results to satisfy the sensible heat flux and latent heat flux balance equations. The nitrogen accumulation continuity constraint is implemented by verifying the consistency of the nitrogen absorption rate in the predicted yield with the value simulated by the dynamic corn growth model. It should be further explained that in the specific implementation process, when the canopy heat flux conservation constraint is applied, if the difference between the sensible heat flux and the latent heat flux predicted by the machine learning model exceeds 20% of the range of the field measured data, the system automatically triggers the boundary layer parameter feedback verification of the WRF-Crop coupling system. By comparing the time series offset direction of the measured canopy temperature and the simulated value, the activation function threshold of the hidden layer in the LSTM channel is dynamically adjusted. The sensible heat flux includes the turbulent exchange value between the canopy and the air, and the latent heat flux includes the energy consumption value of transpiration.

[0040] The process of dynamically adjusting the activation function threshold of the hidden layer in the LSTM channel includes: if the measured value is continuously higher than the predicted value, the sensible heat flux weight coefficient is reduced by 0.1-0.3; if the measured value is lower than the predicted value and is accompanied by an abnormal decrease in soil moisture content, the latent heat flux compensation factor is increased to 1.2 times to enhance the transpiration energy consumption representation, among which the abnormal decrease in moisture content includes a daily decrease of more than 5%.

[0041] For the continuity constraint of nitrogen accumulation, the nitrogen absorption rate predicted by machine learning is compared daily with the simulated value of the dynamic corn growth model in the middle grain filling period. When the deviation exceeds 10% for three consecutive days, the two-level correction mechanism is activated. Among them, the middle grain filling period is 30-40 days after flowering.

[0042] The process of activating the two-level correction mechanism includes: if the deviation is caused by abnormal soil nitrate nitrogen concentration monitoring data, the frequency of feature extraction of soil moisture satellite inversion data by the enhanced attention mechanism is increased to once per hour. Among them, the abnormal soil nitrate nitrogen concentration monitoring data includes a single-day fluctuation of soil nitrate nitrogen concentration exceeding 15%; if the deviation is caused by the error in the root nitrogen uptake depth simulated by WRF-Crop, the correction function of the stomatal conductance threshold is linked to expand the root distribution parameters to the deep soil and limit the shallow nitrogen uptake ratio to below 70%, until the nitrogen accumulation curve and the trend error output by the dynamic corn growth model converge to within 8%. Among them, the error in the root nitrogen uptake depth includes the predicted value showing that the roots are mainly distributed in the 0-30cm layer, while the sensor detects accelerated nitrate nitrogen consumption in the 40-50cm soil layer.

[0043] In the event of regional drought, for example, when the soil moisture content is lower than 50% of the field water holding capacity for five consecutive days, the strict matching requirement of the absorption rate in the nitrogen constraint will be temporarily lifted, and the deviation threshold will be allowed to be expanded to 15%. However, the verification frequency of canopy heat flux conservation will be increased to 3 times a day at the same time to ensure that energy balance is met first.

[0044] The spatial block cross-validation process in step S4 is as follows: divide the study area into 5km×5km grid cells, randomly block the input data in 20% of the area, use the data from the remaining 80% of the area to train the model, and verify the prediction accuracy of the blocked area. It should be further explained that in the specific implementation process, when performing spatial block cross-validation, the study area is first divided into 5km×5km grid cells based on longitude and latitude. When the system detects that there are grids in the target area with more than 20% missing historical production data, such grids are automatically marked as priority blocking areas and included in the random blocking pool.

[0045] During the training phase, 20% of the total number of grids are randomly selected for shielding in each iteration. The shielding rules follow the spatial continuity constraint, including: if more than 30% of the adjacent grids around the shielded grid are already in a shielded state, reselection is performed to ensure that the data blank area does not form a large continuous block; after using the WRF-Crop output data of the remaining 80% of the grids and the corresponding field sensor monitoring values ​​to train the machine learning model, the yield of the shielded area is predicted. When the root mean square error between the predicted value and the actual statistical data exceeds 15%, the local parameter optimization mechanism is triggered. Specifically, if the error is mainly concentrated in the irrigation area, the weight of the water stress feature extraction by the attention mechanism is enhanced, where the irrigation area is the soil moisture standard deviation less than 0.3; if the error is concentrated in the rain-fed area, the water stress feature extraction weight is enhanced, where the standard deviation of the soil moisture content in the irrigation area is ... In agricultural areas, the memory strength of the LSTM channel for historical drought events is enhanced. Among them, the precipitation dependence in rain-fed agricultural areas exceeds 80%. In the adversarial testing phase, if the input independent year extreme climate sample is a regional drought year, such as: the precipitation in the growing season is less than 60% of the annual average, the system automatically activates the emergency verification mode, increases the shielding ratio to 30%, and prioritizes retaining grids with irrigation facility records as training data. At the same time, the adjustment range of the gated unit weight of the machine learning model is limited to no more than plus or minus 20% of the initial value to prevent overfitting. When the verification results show that the deviation of the predicted yield in the tasseling period is continuously higher than 10%, the linkage parameters are recalibrated, and the simultaneous optimization of the stomatal conductance threshold and the root distribution parameters is forced to start until the verification error of the grid falls within the 8% threshold.

[0046] The triggering condition for the parameter recalibration loop in step S4 is: when the Bayesian probability fusion results show that the difference between the machine learning predicted yield and the WRF-Crop mechanism simulated yield exceeds 8%, or the difference between both and agricultural statistical data exceeds 10%, the joint optimization of the stomatal conductance threshold of the dynamic corn growth model and the machine learning gating unit weight is initiated. It should be further explained that, in the specific implementation process, in the triggering mechanism of the parameter recalibration cycle, when the Bayesian probability fusion detects that the difference between the yield predicted by machine learning and the yield simulated by the WRF-Crop mechanism exceeds 8%, the system first determines the type of error source: if the difference is mainly caused by the canopy temperature simulation deviation, such as the simulated value is more than 2°C higher than the measured value for three consecutive days, the stomatal conductance threshold in the dynamic corn growth model is optimized first, and the threshold is lowered to the stomatal opening and closing critical point corresponding to the photosynthetic active radiation intensity of 200μmol / m² / s by linking the transpiration rate data monitored in real time by field sensors; if the difference is due to the insufficient response of the machine learning model to historical drought events, such as the predicted value still maintains normal fluctuations in years when precipitation is less than 50% of the annual average, the LSTM channel weight is reduced to below 0.4 and the attention mechanism's parsing strength of real-time soil moisture data is increased.

[0047] When the difference between the two and agricultural statistical data exceeds 10% at the same time, the double verification mode is activated, including: first, looking back at the two-way data interaction records of WRF-Crop during the tasseling period, if a systematic deviation is found in the canopy transpiration feedback value, such as the simulated value at 2:00 p.m. every day is continuously 15% lower than the sensor data, the root water absorption depth parameter of the dynamic corn growth model and the climate data weight distribution ratio of the machine learning gating unit are adjusted synchronously; if the deviation is concentrated in the late filling period and is accompanied by abnormal fluctuations in the nitrogen absorption rate, the nitrogen accumulation continuity constraint is forcibly enabled to limit the upper limit of nitrogen allocation of the machine learning prediction value, and the nitrate nitrogen migration coefficient in the dynamic corn growth model is recalibrated.

[0048] In a regional drought year scenario, for example, if soil moisture is less than 40% of field capacity for 10 consecutive days, if the error cannot be converged to the threshold after parameter recalibration, the manual intervention protocol will be triggered, calling the optimized parameter combination of similar historical drought years and temporarily freezing the automatic update function of the albedo parameter in WRF-Crop until three complete growth period simulation verification cycles are completed.

[0049] The resolution improvement method for the dynamic library of maize physiological parameters includes: using a 3km×3km grid during the tasseling and grain filling stages, dynamically correcting the stomatal conductance threshold parameter using real-time monitoring data from field sensors, with the correction frequency synchronized with the assimilation stage. It should be further explained that during the specific implementation process, when the system recognizes the entry into the tasseling or grain filling stages, it automatically activates the 3km×3km high-precision grid mode. The stomatal conductance threshold parameter is dynamically corrected based on real-time data collected by field leaf temperature sensors and multi-layer soil moisture probes deployed within the grid. The multi-layer soil moisture probes include soil moisture probes covering depths of 0-20cm, 20-40cm, and 40-60cm.

[0050] The process of dynamically correcting the stomatal conductance threshold parameters includes the following: if the canopy temperature in a certain grid exceeds 35°C for three consecutive hours and the soil moisture content in the 0-20 cm layer is less than 18%, the stomatal conductance closure threshold of the grid is lowered to a photosynthetically active radiation intensity of 150 μmol / m² / s, and the root distribution parameters are triggered to migrate to the 40-60 cm deep soil layer. When the daily drop in stem moisture content during the grain filling period is detected in the grid, the stomatal conductance threshold emergency correction mechanism is activated, temporarily raising the midday threshold by 20% to reduce unnecessary transpiration water consumption. At the same time, by linking the assimilation weight matrix, the frequency of collecting satellite-retrieved soil moisture data for the grid is increased to once per hour, with the midday period being 10:00-14:00.

[0051] When correcting the parameters, if it is found that the fluctuation range of the stomatal conductance threshold after two consecutive corrections for the same grid exceeds 15%, the optimal parameter interpolation algorithm under the environmental conditions of similar historical growth periods will be automatically activated, and the threshold setting value that is closest to the combination of canopy temperature and soil moisture content in the same period of the past three years will be preferentially adopted; for grid cells with terrain undulation exceeding 5%, an additional slope orientation correction factor will be superimposed. When the grid is on a sunny slope, that is, when the sunshine duration is 20% higher than the regional average, the stomatal conductance threshold is limited to a reduction of no more than 30% of the basic value. If it is on a shady slope and the soil moisture content is continuously higher than 80% of the field water holding capacity, the threshold is allowed to expand its dynamic floating range during the day to ±25%.

[0052] All parameter correction operations are strictly synchronized with the dynamic data assimilation phase. Parameter updates are performed immediately after the end of every 6-hour assimilation cycle during the tasseling period, and are bound to every 6-hour assimilation window during the high-frequency assimilation phase of the grouting period. This ensures that the matching error between the model resolution improvement and real-time environmental changes is controlled within 3%.

[0053] The dynamic data assimilation process regulates the assimilation frequency and data source priority in stages based on the maize growth process. When the dynamic maize growth model detects the start of the tasseling phase, such as the completion of tassel differentiation, the assimilation system automatically switches to high-frequency mode, prioritizing the integration of real-time remotely sensed leaf area index and soil moisture data. Simultaneously, the assimilation weight allocation strategy is dynamically adjusted based on the direction of the temperature feedback residual output by the meteorological model, which includes both positive and negative offsets.

[0054] In response to extreme climate events, the system activates emergency optimization protocols, shortens the assimilation window and strengthens the verification function of satellite inversion data, significantly reducing the deviation in capturing environmental parameters during key growth periods.

[0055] The frequency of bidirectional data exchange within the WRF-Crop coupled system during the grain-filling period has been increased to once per minute. This interaction includes feedback corrections to the boundary layer height from canopy temperature. It should be noted that, during implementation, when bidirectional data exchange is initiated once per minute during the grain-filling period, the system first monitors the temporal synchronization between the real-time data from the canopy temperature sensor and the WRF-Crop simulation. If a deviation exceeding 1.5°C is detected within 10 minutes, the boundary layer height feedback correction mechanism is automatically activated. The measured canopy temperature is fed into the atmospheric boundary layer of the Weather Research and Forecasting Model at minute intervals, dynamically adjusting the turbulent diffusion coefficient and sensible heat flux parameters. Simultaneously, the corrected boundary layer height value is fed back in real time to the stomatal conductance calculation unit of the dynamic corn growth model.

[0056] When the corrected boundary layer height value fluctuates by more than 15% compared with the original simulated value, the double verification protocol is triggered, that is, the canopy temperature gradient change trends of the three adjacent grids are compared. If the correction direction of the current grid is opposite to the simulated trend of the adjacent grid, such as the boundary layer height of the current grid increases while the surrounding grid decreases, the feedback correction of the grid is suspended and the historical data interpolation replacement is enabled until the difference in the interactive data falls within 5% for three consecutive times; in the event of strong wind events, the interaction frequency is temporarily reduced to once every 5 minutes to avoid turbulence parameter overload, and the satellite-retrieved surface temperature data is prioritized for assimilation through the linkage assimilation weight matrix to compensate for sensor noise. Among them, strong wind events include wind speeds exceeding 5m / s for more than 30 minutes.

[0057] For grid cells with complex terrain, such as those with a slope exceeding 8°, an additional canopy temperature altitude correction factor is introduced. When the altitude increases by 100 meters and the simulated canopy temperature is 2°C lower than the measured value, the boundary layer sensible heat flux compensation value is increased by a gradient of 0.3°C per 100 meters, and the single correction amplitude is limited to no more than 20% of the original value; the high-frequency interaction data are strictly synchronized with the correction operation of the stomatal conductance threshold, and the photosynthesis calculation parameters of the dynamic corn growth model are immediately updated after the end of each minute interaction cycle to ensure that the coupling error between the canopy microclimate feedback and the regional atmospheric circulation is controlled below the 3% threshold throughout the filling period.

[0058] The yield estimate output in step S4 includes spatiotemporal resolution data generated by the WRF-Crop coupling system. Specifically, this data is presented as a correlation graph between daily canopy transpiration changes and temperature fluctuations for the corresponding grid. It should be further explained that, in the specific implementation process, when generating the final yield estimate, the system integrates the minute-by-minute canopy transpiration data collected by the WRF-Crop coupling system with the temperature fluctuation data for the corresponding grid, generating a daily variation curve at an hourly granularity. If a daily increase in canopy transpiration from the tasseling to grain filling period exceeds 30% and is accompanied by a grid temperature fluctuation standard deviation exceeding 2°C, the grid is automatically marked as an abnormal mutual feedback zone, triggering a dynamic optimization assimilation weight matrix and prioritizing parameter updates for that region.

[0059] For the construction of the correlation map, if the overlap rate between the peak period of canopy transpiration and the trough period of temperature in the same grid exceeds 60%, it will be judged as a water stress sensitive area. The correction mechanism of the stomatal conductance threshold will be linked to force the midday threshold to be lowered to 70% of the basic value and freeze the weight of the historical climate data of the grid in the machine learning model. Among them, the peak period of canopy transpiration is 13:00-15:00 every day.

[0060] During the data fusion phase, if the correlation map shows a divergence between temperature fluctuations and evapotranspiration trends for a given grid for three consecutive days—for example, if temperature rises while evapotranspiration decreases by more than 15%—a parameter recalibration cycle is initiated, focusing on optimizing the root distribution parameters and the WRF boundary layer sensible heat flux coefficient for that grid. For areas with complex terrain, i.e., slopes exceeding 5°, an aspect correction layer is added to the correlation map. When the evapotranspiration-temperature correlation strength for a grid on a sunny slope is 40% lower than that for a grid on a shady slope, direct use of machine learning predictions is suspended. Instead, a weighted fusion of WRF-Crop simulation data and satellite inversions is used. The weighting is dynamically adjusted based on real-time soil moisture content: for every 5% decrease in soil moisture, the WRF-Crop weight increases by 10%. The output data are back-checked using spatial block cross-validation. If a grid's evapotranspiration curve and temperature map still exhibit a time series offset of more than 10% after Bayesian fusion, a manual intervention protocol is triggered, replacing the current values ​​with the optimal historical parameter combination. Three rounds of simulation verification are completed within 24 hours to ensure that the physical consistency error of the resolution data is less than 3% throughout.

[0061] To address complex terrain and microclimates, the system incorporates terrain correction factors and altitude gradient compensation mechanisms. In areas with significant slope gradients, the stomatal conductance threshold range is dynamically adjusted based on slope-to-slope insolation differences. In high-altitude grids, gradient compensation for boundary layer sensible heat flux is applied based on measured canopy temperatures. All terrain adaptation operations are strictly synchronized with high-frequency data interaction to ensure accurate model simulations in heterogeneous environments.

[0062] By constructing the WRF-Crop bidirectional coupling model, the limitations of the isolated operation of traditional dynamic corn growth models and meteorological models are overcome, dynamic mutual feedback simulation of vegetation growth and local climate is realized, and the crop stress response mechanism under extreme climate events is effectively captured; combining the growth period sensitive data assimilation and physical constraint machine learning algorithm, the parameter capture accuracy of key stages such as the tasseling period and the filling period is improved, so that the yield estimation error is reduced compared with traditional methods; at the same time, the cross-scale verification system and multimodal data fusion mechanism ensure the robustness of the model in climate change sensitive areas and complex terrain areas, providing high-credibility support for agricultural disaster warning and precision irrigation decision-making.

[0063] By deeply integrating physical mechanisms with data intelligence, the regionalized corn characteristics are adapted through a dynamic parameter library to solve the problem of local application deviation of general models; the machine learning prediction results are verified by energy balance and material conservation constraints to avoid the failure risk of black box models in unknown climate scenarios and enhance the explainability of decisions; high-frequency data interaction and terrain correction mechanisms further expand the application scenarios of the technology and support multi-level yield estimation needs from field scale to regional scale.

[0064] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0065] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A grain yield prediction method based on a land-atmosphere coupling model and a machine learning algorithm, characterized by: The steps include: Step S1: A kernel-level algorithm fusion is performed on the dynamic corn growth model and the weather research and forecasting model to construct a WRF-Crop coupling system. The WRF-Crop coupling system realizes real-time interaction between corn canopy transpiration, albedo changes, and temperature, humidity, and radiation data through a bidirectional data channel; Step S2: Divide the assimilation stages according to the maize phenological process, adopt a daily assimilation frequency from the seedling emergence to the jointing stage, switch to a preset assimilation frequency from the tasseling to the grain filling stage, and integrate multispectral remote sensing leaf area index, soil moisture satellite inversion data, and real-time monitoring values ​​of field sensors to generate a dynamic growth environment parameter set; Step S3: construct a dual-channel physically constrained machine learning model that includes a long-short-term memory network and an attention mechanism. The model uses the meteorological-crop interaction data output by WRF-Crop as input, integrates historical climate-yield relationships and stress response characteristics through a gating unit, and imposes canopy heat flux conservation constraints and nitrogen accumulation continuity constraints during the prediction process. In step S4, spatial block cross-validation and independent year extreme climate sample adversarial testing are used to perform Bayesian probability fusion on machine learning predicted yield, WRF-Crop mechanism simulated yield and agricultural statistical data. When the data difference exceeds the preset threshold, a parameter recalibration cycle is triggered until the error converges to the preset range.

2. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The construction process of the WRF-Crop coupling system in step S1 includes: embedding a dynamic corn growth model kernel in the land surface process of the weather research and forecasting model, wherein the dynamic corn growth model kernel integrates a dynamic library of corn physiological parameters including root distribution and stomatal conductance threshold, and the parameter dynamic library establishes a correction function that is adaptive to environmental stress intensity through localized experiments.

3. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The generation of the dynamic growth environment parameter set in step S2 further includes: dynamically optimizing the assimilation weight matrix based on the meteorological-crop mutual feedback residual output by WRF-Crop, wherein the residual includes a temperature feedback delay error and a transpiration calculation deviation.

4. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The structure of the dual-channel physical constraint machine learning model in step S3 includes: a long short-term memory network channel for analyzing the nonlinear temporal relationship between historical climate data and corn yield, an attention mechanism channel for extracting water stress and heat stress characteristics output by a dynamic corn growth model, and the gating unit dynamically adjusts the weight ratio of the dual channels according to the growth stage.

5. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 4, characterized in that: The canopy heat flux conservation constraint is achieved by forcing the prediction results to satisfy the sensible heat flux and latent heat flux balance equation, and the nitrogen accumulation continuity constraint is achieved by verifying the consistency of the nitrogen uptake rate in the predicted yield with the simulated value of the dynamic corn growth model.

6. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The operation process of spatial block cross-validation in step S4 is as follows: divide the study area into 5km×5km grid units, randomly block the input data of 20% of the area, use the data of the remaining 80% area to train the model and verify the prediction accuracy of the blocked area.

7. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The triggering condition for the parameter recalibration cycle in step S4 is: when the Bayesian probability fusion result shows that the difference between the machine learning predicted yield and the WRF-Crop mechanism simulated yield exceeds 8%, or the difference between both and agricultural statistical data exceeds 10%, the joint optimization of the stomatal conductance threshold of the dynamic corn growth model and the machine learning gating unit weight is initiated.

8. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 2, characterized in that: The method for improving the resolution of the dynamic library of maize physiological parameters includes: using a 3km×3km grid during the tasseling and grain filling stages, dynamically correcting the stomatal conductance threshold parameters through real-time monitoring data from field sensors, and keeping the correction frequency synchronized with the assimilation stage.

9. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The frequency of two-way data interaction of the WRF-Crop coupling system during the grouting period was increased to once per minute, and the interaction data included feedback correction values ​​of canopy temperature to boundary layer height.

10. The method for estimating grain yield based on a land-atmosphere coupling model and a machine learning algorithm according to claim 1, characterized in that: The yield estimation result outputted in step S4 includes the spatiotemporal resolution data generated by the WRF-Crop coupling system, which is specifically represented by a correlation graph of daily canopy transpiration change curves and temperature fluctuations of the corresponding grid.

Images