A crop yield prediction method fusing weather data and growth model
By constructing a meteorological scenario database and performing batch simulations and dynamic weight calculations, the problem of one-way interaction between meteorological data and crop growth models was solved, improving the accuracy and reliability of crop yield prediction and outputting confidence scores and risk warnings.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA AGRICULTURAL UNIVERSITY
- Filing Date
- 2026-03-05
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies, the interaction between meteorological data and crop growth models is one-way, which cannot dynamically correct future weather forecasts based on the actual growth status of crops, resulting in unstable forecast accuracy.
A meteorological scenario database was constructed and batch simulations were performed. Fractal features of real remote sensing observations and simulated leaf area index were extracted. Dynamic future meteorological data were generated through multi-scale phenological matching and dynamic weight calculation to be input into crop growth models for yield prediction.
It improves the coupling relationship between meteorological input and growth model, enhances the physiological rationality and reliability of prediction results, can dynamically adjust the prediction process to adapt to crop growth, and outputs confidence scores and risk warnings.
Smart Images
Figure CN122155020A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural remote sensing and growth simulation technology, and more specifically, to a method for predicting crop yield by integrating meteorological data and growth models. Background Technology
[0002] Crop yield forecasting is an important research area in agricultural remote sensing and crop growth simulation. Accurate yield forecasts are valuable for food security assessment, agricultural insurance loss assessment, and agricultural product market regulation. Existing yield forecasting methods mainly include estimation methods based on statistical analysis of meteorological data, empirical model methods based on remote sensing vegetation indices, and simulation methods based on crop growth mechanism models. However, single data source or single model methods cannot fully reflect the complexity of crop growth. The interaction between meteorological data and crop growth models often exhibits a unidirectional characteristic, that is, meteorological data is only used as input to the growth model, without dynamically correcting meteorological forecasts based on the actual crop growth status.
[0003] Chinese patent CN118211730A discloses a crop yield prediction method based on meteorological feature matching and crop growth models. This method simulates unknown meteorological data through meteorological data matching, combines a yield abundance / shortness index to limit yield prediction results, and assimilates remote sensing data to construct a random forest model, finally calculating the weighted yield prediction results. In this scheme, the correlation between meteorological matching and the growth model is relatively independent, and there is a lack of dynamic feedback between the selection of meteorological data and the simulation process of the crop growth model. Chinese patent CN119494448A discloses a winter wheat yield prediction method based on XLSTM and an assimilated crop growth model. This method assimilates the remote sensing leaf area index into the WOFOST model through ensemble Kalman filtering and uses the assimilated data to train an XLSTM for yield prediction. In this scheme, the assimilated data is mainly used to train the deep learning model, and the meteorological input still relies on external forecasts; the crop growth status fails to influence the selection and weighting of meteorological scenarios. The one-way interaction between meteorological data and crop growth models in existing technologies makes it difficult to maintain stable prediction accuracy when actual weather deviates from expectations. Therefore, a crop yield prediction method that integrates meteorological data and growth models is proposed to address the above problems. Summary of the Invention
[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a crop yield prediction method that integrates meteorological data and growth models. The technical problem to be solved is that the interaction between meteorological data and crop growth models in the prior art is one-way and cannot dynamically correct future meteorological forecasts according to the actual growth status of crops.
[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting crop yield by integrating meteorological data and growth models includes the following steps: S1: Acquire historical meteorological time-series data, short-term meteorological forecast data, and remote sensing observation leaf area index time-series data from sowing to the current forecast date for the target area.
[0006] S2: Based on the historical meteorological time series data and short-term meteorological forecast data, a meteorological scenario library containing multiple future meteorological time series is generated using a generative neural network.
[0007] Furthermore, the construction of the meteorological scenario database in S2 includes: using the historical meteorological time-series data as training samples, training the generative neural network with a variational autoencoder or a generative adversarial network, and extracting the temporal distribution features of the historical meteorological data; The short-term weather forecast data is used as a constraint condition and input into the trained generative neural network. The short-term weather forecast data is injected with disturbances that conform to the time series distribution characteristics and follow a preset probability distribution to generate multiple future weather time series that conform to the short-term forecast trend, thus forming the weather scenario database.
[0008] S3: Input each future weather time series in the meteorological scenario database into the crop growth model to simulate and obtain a set of simulated time series data of leaf area index.
[0009] S4: Extract the fractal features of the remotely sensed leaf area index time series data as the true growth feature fingerprint, and extract the fractal features of each member in the simulated leaf area index time series data set as the simulated growth feature fingerprint set.
[0010] Furthermore, the fractal feature extraction in S4 includes: performing multifractal detrending fluctuation analysis on the remotely sensed leaf area index time series data and the simulated leaf area index time series data respectively, and extracting the singular width of the multifractal singular spectrum; Rescaled range analysis was performed on the time series data to extract the Hearst exponent; Sliding window rescaled range analysis was performed on the time series data to extract the local Hurst exponent sequence; By splicing the singular width, Hearst exponent, and local Hearst exponent sequences, and using principal component analysis for dimensionality reduction, the set of real growth feature fingerprints and simulated growth feature fingerprints is generated.
[0011] S5: Divide the entire growth period of the crop into multiple phenological windows, calculate the feature distance between the real growth feature fingerprint and each member in the simulated growth feature fingerprint set within each phenological window, and convert it into a local matching degree.
[0012] Furthermore, S5 divides the entire crop growth period into multiple phenological windows, including: simulating the phenological development process for each future meteorological time series based on accumulated temperature models and crop physiological parameters; The phenological window boundary for each future meteorological time series is determined based on the phenological development process, and the phenological window boundary is dynamically divided using accumulated temperature thresholds. The phenological windows include the sowing to emergence window, the emergence to jointing window, the jointing to tasseling window, the tasseling to grain filling window, and the grain filling to maturity window.
[0013] Furthermore, in S5, calculating the feature distance and converting it into local matching degree includes: In the Within a phenological window, the true growth feature fingerprint is calculated and compared with the first... Euclidean distance between simulated growth feature fingerprints ; when Less than or equal to the preset distance threshold At that time, adopt Calculate the local matching degree, where The preset maximum tolerance distance; when Greater than At that time, adopt Calculate the local matching degree, where For the first The preset feature difference tolerance for each phenological window.
[0014] S6: The local matching degree of each phenological window is weighted and fused to generate the comprehensive matching degree of each future meteorological time series. The comprehensive matching degree is normalized to obtain the dynamic weight of each future meteorological time series.
[0015] Furthermore, the weighted fusion of the local matching degree of each phenological window in S6 includes: obtaining the preset window weights of each phenological window. and preset meteorological factor sensitivity weights The Including temperature weighting Precipitation weight and radiation weight ; The arithmetic mean of each dimension is used as the overall sensitivity weight. ,Right now ; By summing the weighted matching degrees of all phenological windows, we obtain the first... Comprehensive matching degree of future weather time series , ,in This represents the total number of phenological windows.
[0016] Furthermore, in S6, the calculation of dynamic weights based on the overall matching degree includes: normalized overall matching degree. The initial probability of each future weather time series is obtained. ,in This represents the total number of future weather events. Calculate the information entropy of the initial probability distribution , ; Calculate confidence based on information entropy , ; Preset high confidence threshold and low confidence threshold ,satisfy ; when At that time, with As dynamic weight ; when and At that time, adopt Calculate the dynamic weights, where Preset adjustment parameters; when At that time, the backtracking mechanism is triggered.
[0017] S7: Based on the dynamic weights, the future meteorological data after the current forecast date in the meteorological scenario database are weighted and fused to generate dynamic future meteorological data.
[0018] S8: Input the dynamic future meteorological data into the crop growth model to complete the simulation of the remaining growing season and obtain the yield prediction value.
[0019] Furthermore, the method also includes a backtracking mechanism: when Less than At that time, backtrack to S2; During backtracking, the noise injection strategy of the generative neural network is adjusted to increase the amplitude of random perturbations or change the probability distribution type of perturbations, thereby generating a new meteorological scenario database. Re-execute S3 through S6 until... achieve The above or the preset maximum number of backtracking attempts has been reached. .
[0020] Furthermore, S8 also includes: output production forecasts. and the corresponding confidence score, wherein the confidence score is a confidence level calculated based on information entropy. ; Based on dynamic weights Extracting weighted rankings The future weather time series and its future weather data are output as a list of scenario contribution. Ranked by weight Risk warning information is generated for extreme weather events such as high temperature damage, drought, or waterlogging in future weather time series.
[0021] Furthermore, the method also includes a dynamic update step: after the remote sensing observation data for subsequent phenological windows are acquired, S5 and S6 are re-executed to update the comprehensive matching degree. and confidence level ; If the confidence level is updated Reduced to a preset low confidence threshold The backtracking mechanism will then be executed. If the maximum number of backtracking iterations is reached... Still below If the result with the highest confidence in the current iteration is used as the final output, a low confidence warning flag will be added.
[0022] The technical effects and advantages of this invention are as follows: This invention constructs a meteorological scenario database and performs batch simulations of crop growth models for each meteorological time series to obtain a set of simulated leaf area index (LAI) trajectories corresponding to different meteorological scenarios. Based on this, the fractal features of the actual remote sensing LAI and the simulated LAI are extracted as growth feature fingerprints. This transforms the comparison of crop growth trajectories from a numerical level to a comparison at the level of intrinsic dynamic characteristics. This approach helps improve the ability to capture the essential laws of crop growth, making the selection criteria for meteorological scenarios closer to the intrinsic mechanisms of crop growth.
[0023] This invention divides the entire crop growth period into multiple phenological windows. Within each window, the feature distance between the actual growth characteristic fingerprint and the simulated growth characteristic fingerprint is calculated and converted into a local matching degree. The comprehensive matching degree is obtained by weighted fusion of the local matching degrees of each window. This multi-scale phenological matching method takes into account the differences in sensitivity of different growth stages to meteorological factors such as temperature, precipitation, and radiation, which helps to improve the physiological rationality of the matching results and enables the comprehensive matching degree to more accurately reflect the degree of consistency between meteorological scenarios and actual crop growth.
[0024] This invention obtains the initial probability based on the normalization of the comprehensive matching degree, and then calculates the information entropy and confidence level. Different weight calculation methods are selected to generate dynamic weights based on the relationship between the confidence level and a preset threshold. When the confidence level falls below the threshold, a backtracking mechanism is triggered. The meteorological scenario database is regenerated and rematched by adjusting the noise injection strategy. This adaptive mechanism helps ensure the reliability of the prediction results and avoids matching failures due to design flaws in the initial scenario database.
[0025] This invention uses dynamic weights to weight and fuse future meteorological data after the current forecast date to generate dynamic future meteorological data. This data is then input into a crop growth model to simulate the remaining growing season and obtain yield predictions. Simultaneously, it outputs confidence scores, a list of scenario contributions, and risk warnings. This fusion method improves the coupling between meteorological input and the growth model by using reverse filtering based on the actual crop growth trajectory and weighted future meteorological scenarios, enabling the prediction process to dynamically adjust as the crop grows. Attached Figure Description
[0026] Figure 1 This is a flowchart illustrating the method execution of the present invention.
[0027] Figure 2 This is a flowchart of the fractal feature extraction and phenological window matching process of the present invention. Detailed Implementation
[0028] 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.
[0029] Example 1 As attached Figures 1 to 2 This invention presents a method for predicting crop yield by integrating meteorological data and growth models. This embodiment uses spring maize in a target area as the prediction object to provide a detailed explanation of the method proposed in this invention.
[0030] Step 1: Data Acquisition and Preprocessing Historical meteorological time series data for a preset historical period in the target area is obtained. The data comes from a public meteorological database and includes meteorological elements such as daily average temperature, maximum temperature, minimum temperature, precipitation, solar radiation, wind speed, and relative humidity.
[0031] At the same time, short-term weather forecast data for the current forecast date is obtained, using forecast products issued by the meteorological forecast center, which include daily forecast values of meteorological elements for the preset number of days from the current forecast date.
[0032] Remote sensing images from the current sowing period to the current forecast date are acquired from remote sensing satellites. The remote sensing images are then preprocessed by radiometric calibration, atmospheric correction, and geometric fine correction. Based on the reflectance of each band of the preprocessed images, the leaf area index is retrieved using a radiative transfer model. The model parameters are calibrated according to the physiological characteristics of crops in the target area.
[0033] The inversion yielded daily remote sensing observation time-series data of leaf area index, denoted as... ,in It is the number of days from the sowing date. , This represents the total number of days from the sowing date to the current forecast date.
[0034] Step 2: Constructing a meteorological scenario database Historical meteorological time series data for a preset historical period are used as training samples, and a generative neural network is trained using a variational autoencoder or a generative adversarial network.
[0035] The network consists of an encoder and a decoder. The encoder maps the input historical meteorological time-series data to latent space distribution parameters, and the decoder reconstructs the meteorological time-series data after sampling from the latent space.
[0036] After training, the generative neural network learned and extracted the temporal distribution features of historical meteorological data.
[0037] The short-term weather forecast data for the current forecast date is used as a constraint and input into a trained generative neural network. By injecting random perturbations that conform to the aforementioned time series distribution characteristics and follow a preset probability distribution into the short-term weather forecast data, multiple independent samplings are performed to generate multiple future weather time series that are generally consistent with the short-term forecast trend but have reasonable variability in specific values.
[0038] Each future weather time series contains daily weather data from the current forecast date to the harvest period, forming a weather scenario database. ,in This represents the total number of future weather events. Indicates the first A future weather time series, .
[0039] Step 3: Batch Simulation and Growth Trajectory Generation Each future weather time series in the weather scenario database As driving data, it is input into the crop growth model. The parameters of the crop growth model are calibrated based on the physiological characteristics of the main crop varieties in the target area. These parameters include at least the accumulated temperature required for seedling emergence, photosynthetic parameters, and dry matter partition coefficient.
[0040] For each The crop growth model simulates the daily growth process from sowing to harvest, and outputs the corresponding simulated time-series data of leaf area index. ,in This refers to the number of days from the sowing date, with a value range similar to that in the first step. of Consistent.
[0041] Based on the above simulation process, a set of simulated time-series data for leaf area index was obtained. .
[0042] Step 4: Fractal Feature Extraction Time series data of leaf area index observed by remote sensing and each simulation time series data Multifractal detrending fluctuation analysis was performed separately. For ease of description, the time series data to be analyzed are uniformly denoted as... The analysis process includes the following sub-steps: First, calculate the mean of the time series data. Then, the cumulative deviation sequence is calculated as the contour sequence. Secondly, the contour sequence is divided into segments of length [missing information]. of There are several non-overlapping intervals. Then, a polynomial fit is performed on each interval to remove local trends, and the residual variance is calculated. .
[0043] Next, calculate First-order wave function: ,in This is a pre-defined sequence of orders. Based on this, analysis is performed. and The linear relationship is used to obtain the scaling index through linear fitting. .
[0044] Finally, the multifractal singular spectrum is obtained through the Legendre transform. Then extract the singular width. ,in and Each of the following is a singular spectrum The maximum and minimum values.
[0045] Rescaled range analysis was performed on time series data to extract the Hearst exponent. Regarding timing length Calculate the mean Then calculate the cumulative deviation sequence. .
[0046] The range was calculated based on this sequence. and standard deviation For multiple Value Calculation , fitting and The linear relationship yields the slope, which is the Hearst exponent. .
[0047] A sliding window recalibrated range analysis was performed on the time series data, with the window length set to a fixed value based on the average duration of the crop growth stages. The step size is 1 day. The rescaled range analysis is repeated within each window to calculate the Hearst exponent, resulting in a local Hearst exponent sequence. ,in .
[0048] singular width Hearst index and local Hearst exponent sequences Vector concatenation is performed to form the original feature vector. Principal component analysis is then used to reduce the dimensionality of the original feature vector, retaining those with a cumulative contribution rate greater than a preset threshold. The former One principal component.
[0049] Based on this, the final true growth feature fingerprint is generated. and simulated growth feature fingerprint set Each fingerprint is dimensional vector, The first fingerprint representing the true growth characteristics One portion, Indicates the first The first simulated growth feature fingerprint One portion, .
[0050] Step 5: Phenological window division and local matching degree calculation The entire crop growth period is divided into multiple phenological windows. Based on the accumulated temperature model and crop physiological parameters, a plan is developed for each future meteorological time series. Simulate the corresponding phenological development process and determine the boundaries of each phenological window based on the accumulated temperature threshold required for each phenological stage of the target crop.
[0051] In this embodiment, the phenological windows are divided into: sowing to emergence window, emergence to jointing window, jointing to tasseling window, tasseling to grain filling window, and grain filling to maturity window. The total number of windows is denoted as follows: , This indicates the window number. The phenological window in which the current forecast date falls is determined based on the actual accumulated temperature, and the boundaries of subsequent phenological windows are determined based on each... The accumulated temperature dynamics are determined.
[0052] Based on this, the feature distance is calculated and converted into local matching degree, which includes the following sub-steps: In the Within a single phenological window, calculate the fingerprint of true growth characteristics. With the Simulated growth feature fingerprint Euclidean distance between them: ; Preset distance threshold Based on historical data statistics. Preset maximum tolerance distance. Take the preset quantile of the feature distance of historical years. When At that time, a linear transformation function is used to calculate the local matching degree: ; when At that time, a nonlinear transformation function is used to calculate the local matching degree: ; in For the first The preset characteristic difference tolerance for each phenological window is set based on the standard deviation of the characteristic distance of historical years within that window.
[0053] The local matching degree calculation for the completed phenological windows has been completed. The phenological windows in progress are partially included in the calculation. The local matching degree of subsequent windows has not yet been calculated and will be dynamically updated after the remote sensing observation data of the corresponding windows are acquired.
[0054] Step 6: Calculation of Overall Matching Degree and Dynamic Weights Based on the aforementioned local matching degree, the local matching degree of each phenological window is weighted and fused, specifically including the following sub-steps: Obtain the preset window weights for each phenological window. and preset meteorological factor sensitivity weights Window weight The parameters are set according to the contribution of each phenological window to the final yield, to meet the following requirements. .
[0055] Sensitivity weight of meteorological factors It is a three-dimensional vector, including temperature weights. Precipitation weight and radiation weight Based on crop physiological characteristics, the weights of each dimension are set to satisfy... .
[0056] The arithmetic mean of each dimension is used as the overall sensitivity weight. Participating in the calculation, that is The weighted matching degree of all phenological windows is summed to obtain the first... Comprehensive matching degree of future weather time series : ; Then, dynamic weights are calculated based on the overall matching degree, which specifically includes the following sub-steps: Normalized overall matching degree The initial probability of each future weather time series is obtained: ; Calculate the information entropy of the initial probability distribution : ; Confidence level calculated based on information entropy : ; Preset high confidence threshold and low confidence threshold ,satisfy Choose the weighting calculation method based on the confidence level interval: when At that time, directly with As dynamic weight ; when and At that time, the dynamic weights are calculated using a weighted adjustment formula: ; in For preset adjustment parameters, ; when At that time, the backtracking mechanism is triggered.
[0057] The resulting dynamic weights satisfy .
[0058] Step 7: Generation of Dynamic Future Weather Data Based on dynamic weights We perform weighted fusion of future meteorological data after the current forecast date in the meteorological scenario database. For the future... The weighted fusion values for each meteorological element of the day are calculated separately: temperature: ; precipitation: ; radiation: ; Wind speed: ; Relative humidity: ; in , , , , The first The future weather sequence in the first Forecast values for temperature, precipitation, radiation, wind speed, and relative humidity for the day. Weighted and fused data are then used to generate complete dynamic future meteorological data. This data includes daily meteorological element values from the day after the current forecast date to the harvest period.
[0059] Step 8: Production Forecasting and Output Dynamic future weather data Input a crop growth model, drive the model to complete the simulation of the remaining growing season from the current forecast date to the harvest date, and obtain the final yield forecast. .
[0060] Output production forecast and the corresponding confidence score, wherein the confidence score is the confidence level calculated in step six. Based on dynamic weights Extracting weighted rankings The following are output as a list of scenario contribution values: [List of future weather time series and their corresponding future weather data]. This is a preset quantity. Ranked by weight. Risk warnings are generated for extreme weather events in future weather series, including heatstroke, drought, and waterlogging, which are identified based on preset judgment thresholds. High-temperature heat damage: Daily maximum temperature exceeding the threshold And the duration exceeds ; Drought: Continuous There was no effective precipitation and the soil relative humidity was below the threshold. ; Waterlogging: continuous Rainfall exceeding ,or The cumulative rainfall exceeded .
[0061] Step 9: Dynamic Updates and Backtracking Once the remote sensing data for subsequent phenological windows is acquired, the relevant calculations in steps five and six are re-executed to update the overall matching degree. and confidence level If the confidence level is updated Reduced to a preset low confidence threshold The following will trigger the backtracking mechanism.
[0062] The backtracking mechanism executes as follows: First, it backtracks to step two, adjusting the noise injection strategy of the generative neural network. This adjustment includes increasing the perturbation amplitude or changing the probability distribution type of the perturbation. Then, based on the adjusted parameters, a new meteorological scenario database is generated, and steps three through six are re-executed.
[0063] Repeat the above process until the confidence level is reached. achieve The above, or the preset maximum number of backtracking attempts has been reached. If the maximum number of backtracking iterations is reached... Still below If the result with the highest confidence in the current iteration is used as the final output, a low confidence warning flag will be added.
[0064] Through the above steps, a multi-scale weather-growth coupling based on the fractal characteristics of growth trajectory is realized, which can predict yield before crop harvest. The prediction results include yield value, confidence score, scenario contribution and risk warning information.
[0065] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for predicting crop yield by integrating meteorological data and growth models, characterized in that, Includes the following steps: S1: Acquire historical meteorological time-series data, short-term meteorological forecast data, and remote sensing observation leaf area index time-series data from sowing to the current forecast date for the target area; S2: Based on the historical meteorological time series data and short-term meteorological forecast data, a meteorological scenario library containing multiple future meteorological time series is generated using a generative neural network; S3: Input each future weather time series in the meteorological scenario database into the crop growth model to simulate and obtain a set of simulated time series data of leaf area index; S4: Extract the fractal features of the remotely sensed leaf area index time series data as the real growth feature fingerprint, and extract the fractal features of each member in the simulated leaf area index time series data set as the simulated growth feature fingerprint set. S5: Divide the entire crop growth period into multiple phenological windows, calculate the feature distance between the real growth feature fingerprint and each member in the simulated growth feature fingerprint set within each phenological window, and convert it into a local matching degree; S6: The local matching degree of each phenological window is weighted and fused to generate the comprehensive matching degree of each future meteorological time series. The comprehensive matching degree is normalized to obtain the dynamic weight of each future meteorological time series. S7: Based on the dynamic weights, the future meteorological data after the current forecast date in the meteorological scenario database are weighted and fused to generate dynamic future meteorological data; S8: Input the dynamic future meteorological data into the crop growth model to complete the simulation of the remaining growing season and obtain the yield prediction value.
2. The crop yield prediction method based on the fusion of meteorological data and growth models according to claim 1, characterized in that, The construction of the weather scenario database in S2 includes the following sub-steps: S21: Using the historical meteorological time-series data as training samples, the generative neural network is trained by a variational autoencoder or a generative adversarial network to extract the temporal distribution features of the historical meteorological data. S22: Input the short-term weather forecast data as a constraint into the trained generative neural network; S23: Inject disturbances that conform to the time series distribution characteristics and follow a preset probability distribution into the short-term weather forecast data to generate multiple future weather time series that conform to the short-term forecast trend, thus forming the weather scenario database.
3. The crop yield prediction method integrating meteorological data and growth models according to claim 1, characterized in that, Fractal feature extraction in S4 includes the following sub-steps: S41: Perform multifractal detrending fluctuation analysis on the remotely sensed leaf area index time series data and the simulated leaf area index time series data respectively, and extract the singular width of the multifractal singular spectrum; S42: Perform rescaled range analysis on the time series data and extract the Hearst exponent; S43: Perform sliding window rescaled range analysis on the time series data to extract the local Hurst exponent sequence; S44: Concatenate the singular width, Hearst exponent, and local Hearst exponent sequences, and use principal component analysis to reduce dimensionality, generating the set of real growth feature fingerprints and simulated growth feature fingerprints.
4. The crop yield prediction method integrating meteorological data and growth models according to claim 1, characterized in that, S5 divides the entire crop growth period into multiple phenological windows, including: Based on accumulated temperature models and crop physiological parameters, the phenological development process is simulated for each future meteorological time series. The phenological window boundary for each future meteorological time series is determined based on the phenological development process, and the phenological window boundary is dynamically divided using accumulated temperature thresholds. The phenological windows include the sowing to emergence window, the emergence to jointing window, the jointing to tasseling window, the tasseling to grain filling window, and the grain filling to maturity window.
5. The crop yield prediction method integrating meteorological data and growth models according to claim 1, characterized in that, Calculating feature distance and converting it to local matching degree in S5 includes the following sub-steps: S51: In the Within a phenological window, the true growth feature fingerprint is calculated and compared with the first... Euclidean distance between simulated growth feature fingerprints ; S52: When Less than or equal to the preset distance threshold At that time, adopt Calculate the local matching degree, where The preset maximum tolerance distance; S53: When Greater than At that time, adopt Calculate the local matching degree, where For the first The preset feature difference tolerance for each phenological window.
6. The crop yield prediction method based on the fusion of meteorological data and growth models according to claim 5, characterized in that, The weighted fusion of local matching degrees for each phenological window in S6 includes the following sub-steps: S61: Obtain the preset window weights for each phenological window. and preset meteorological factor sensitivity weights The Including temperature weighting Precipitation weight and radiation weight ; S62: Take the arithmetic mean of each dimension as the overall sensitivity weight. ,Right now ; S63: Accumulate the weighted matching degree of all phenological windows to obtain the first... Comprehensive matching degree of future weather time series , ,in This represents the total number of phenological windows.
7. The crop yield prediction method integrating meteorological data and growth models according to claim 1, characterized in that, The calculation of dynamic weights based on comprehensive matching degree in S6 includes the following sub-steps: S71: Normalized Overall Matching Degree The initial probability of each future weather time series is obtained. ,in This represents the total number of future weather events. S72: Calculate the information entropy of the initial probability distribution , ; S73: Calculating Confidence Based on Information Entropy , ; S74: Preset high confidence threshold and low confidence threshold ,satisfy ; S75: When At that time, with As dynamic weight ; S76: When and At that time, adopt Calculate the dynamic weights, where Preset adjustment parameters; S77: When At that time, the backtracking mechanism is triggered.
8. The crop yield prediction method based on the fusion of meteorological data and growth models according to claim 7, characterized in that, The backtracking mechanism includes: when Less than At that time, backtrack to S2; During backtracking, the noise injection strategy of the generative neural network is adjusted to increase the amplitude of random perturbations or change the probability distribution type of perturbations, thereby generating a new meteorological scenario database. Re-execute S3 through S6 until... achieve The above or the preset maximum number of backtracking attempts has been reached. .
9. The crop yield prediction method based on the fusion of meteorological data and growth models according to claim 1, characterized in that, S8 is followed by the following steps: S91: Output Production Forecast and the corresponding confidence score, wherein the confidence score is a confidence level calculated based on information entropy. ; S92: Based on dynamic weights Extracting weighted rankings The future weather time series and its future weather data are output as a list of scenario contribution. S93: Ranking based on weight Risk warning information is generated for extreme weather events such as high temperature damage, drought, or waterlogging in future weather time series.
10. The crop yield prediction method based on the fusion of meteorological data and growth models according to claim 1, characterized in that, It also includes a dynamic update step: Once the remote sensing data for subsequent phenological windows is acquired, S5 and S6 are executed again to update the overall matching degree. and confidence level ; If the confidence level is updated Reduced to a preset low confidence threshold The backtracking mechanism described in claim 8 shall then be executed. If the maximum number of backtracking iterations is reached... Still below If the result with the highest confidence in the current iteration is used as the final output, a low confidence warning flag will be added.