A rice yield prediction method coupling crop model and remote sensing data
By coupling remote sensing data with crop models, and utilizing Sentinel-2 and Sentinel-1 image inversion data and ensemble Kalman filtering algorithms, the rice growth status is updated in real time, solving the problems of timeliness and accuracy in rice yield prediction. This method is suitable for grain storage and scheduling and grain trade.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 柳州市农业气象试验站
- Filing Date
- 2026-05-12
- Publication Date
- 2026-07-31
AI Technical Summary
Existing rice yield prediction methods are inadequate in terms of timeliness and accuracy, especially when extreme weather or changes in management practices cause large prediction errors and lack real-time correction capabilities.
By combining crop models with remote sensing data, leaf area index and soil moisture content data are retrieved from Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images. The crop model is assimilated using an ensemble Kalman filter algorithm, and the rice growth status is updated in real time. The model parameters are then corrected using grain storage data to achieve yield prediction.
It significantly improves the timeliness and accuracy of rice yield forecasting, and can output pixel-level yield forecasts 4-60 days before harvest, making it suitable for grain storage and dispatching and grain trade.
Smart Images
Figure CN122491597A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of rice yield prediction methods, and in particular to a rice yield prediction method that couples crop models with remote sensing data. Background Technology
[0002] Rice yield forecasting is of great value for ensuring national food security, guiding grain storage and procurement, and stabilizing market prices. Traditional rice yield forecasting methods are mainly divided into two categories: one is an empirical model based on statistical regression, which relies on historical yields and meteorological factors or vegetation indices to establish empirical relationships; the other is a process model based on crop growth mechanisms, which estimates the final yield by simulating light and temperature production potential and water and nutrient stress.
[0003] However, existing methods suffer from the following technical limitations: Purely statistical methods lack mechanistic understanding, making it difficult to capture the nonlinear impact of extreme weather or changes in management practices on yield. Furthermore, their extrapolation capabilities are poor, and prediction errors increase significantly when the climate deviates from historical averages. While pure crop models have a clear physiological basis, their simulation accuracy is highly dependent on accurate meteorological inputs, soil parameters, and field management data. In practical operational applications, regional-scale soil and crop parameters often use default values or coarse estimates, leading to accumulated model biases. More importantly, these models cannot obtain real-time observational information on crop growth, hindering timely correction of simulation errors.
[0004] For the reasons mentioned above, the timeliness and accuracy of current rice yield forecasts are insufficient. Summary of the Invention
[0005] The purpose of this invention is to provide a method for predicting rice yield by coupling crop models and remote sensing data, which can improve the timeliness and accuracy of rice yield prediction.
[0006] To achieve the above objectives, this invention provides a method for predicting rice yield by coupling crop models and remote sensing data, comprising: Obtain rice planting distribution maps, daily meteorological reanalysis data, soil grid data, as well as rice variety parameters and field management parameters for the predicted area; Multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season were acquired, and leaf area index time series data and surface soil moisture content time series data were obtained by inversion, respectively. Based on the WOFOST crop model, the daily meteorological reanalysis data, soil grid data and rice variety parameters were used as inputs to simulate the daily growth status of rice pixel by pixel. The status included leaf area index, soil moisture content, aboveground biomass and storage organ dry weight. An ensemble Kalman filter algorithm was used to assimilate the WOFOST crop model pixel by pixel, using the inverted leaf area index time series data and surface soil moisture content time series data as observations, and update the model's state variables. The assimilated WOFOST crop model was run to the physiological maturity stage of rice. The memory stem weight of each ensemble member was extracted, the ensemble mean was calculated as the predicted yield of that pixel, and the ensemble standard deviation was calculated as the prediction uncertainty. Based on the boundaries of the grain storage area or the service radius of the drying plant, the pixel-level predicted output is spatially aggregated to obtain the predicted total output value and the predicted peak harvest value for the region, and to generate planning suggestions for storage resources.
[0007] The process includes spatial aggregation of pixel-level predicted yields based on the boundaries of the grain storage area or the service radius of the drying plant, to obtain the predicted total yield and peak harvest period for the region. After generating planning suggestions for grain storage resources, the process also includes the following steps: During the actual harvest period, the predicted yield was verified using grain depot weighbridge purchase data, and the crop model parameters and assimilation parameters for the next growing season were adaptively adjusted based on the prediction error.
[0008] In the steps of acquiring multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season, and then retrieving time-series data of leaf area index and topsoil moisture content respectively,... The leaf area index obtained by the inversion is specifically calculated as follows: based on the surface reflectance of Sentinel-2 optical imagery, the leaf area index is calculated pixel by pixel using the PROSAIL radiative transfer model lookup table method.
[0009] In the steps of acquiring multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season, and then retrieving time-series data of leaf area index and topsoil moisture content respectively,... The inversion of surface soil moisture content specifically involves: using the dual-polarization backscattering coefficient of Sentinel-1 radar imagery and subtracting vegetation influence using a water cloud model to obtain the soil volumetric moisture content at a depth of 0-10 cm.
[0010] In this process, based on the WOFOST crop model, daily meteorological reanalysis data, soil grid data, and rice variety parameters are used as inputs to simulate the daily growth status of rice pixel by pixel. This simulation includes steps for leaf area index, soil moisture content, aboveground biomass, and storage organ dry weight. The WOFOST crop model is implemented using the PCSE framework, and the rice variety parameters include photoperiod sensitivity coefficient, grain filling rate, accumulated temperature during the growth period, and distribution coefficient.
[0011] In the process of employing the ensemble Kalman filter algorithm, using the retrieved leaf area index time-series data and surface soil moisture content time-series data as observations, the WOFOST crop model is assimilated pixel-by-pixel, and the model's state variables are updated. Each assimilation uses both leaf area index and topsoil moisture content observations, and the observation error covariance matrix is set according to the prior accuracy of remote sensing inversion.
[0012] Among the steps, spatial aggregation of pixel-level predicted yields based on the boundaries of grain storage areas or the service radius of drying plants is used to obtain the predicted total yield and peak harvest period values for the region, and to generate planning recommendations for grain storage resources. The proposed resource acquisition and storage plan includes the area of temporary storage, the number of drying lines in operation, and the peak daily acquisition funds.
[0013] This invention provides a rice yield prediction method that couples a crop model with remote sensing data. By using ensemble Kalman filtering to simultaneously assimilate the leaf area index retrieved by Sentinel2 optical inversion and the soil moisture retrieved by Sentinel1 radar inversion, the state variables of the crop model are corrected in real time. Ultimately, pixel-level yield predictions can be output 4560 days before harvest, which can significantly improve the timeliness and accuracy of rice yield prediction. This invention solves the problems of existing rice yield prediction methods, such as the difficulty in achieving both mechanistic and timeliness, insufficient expression of spatial heterogeneity, and lack of operational decision indicators. It is particularly suitable for practical scenarios such as grain storage and scheduling, grain trade, and yield monitoring. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0015] Figure 1 This is a flowchart of a rice yield prediction method that couples crop models with remote sensing data according to the present invention. Detailed Implementation
[0016] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.
[0017] Please see Figure 1 This invention provides a method for predicting rice yield by coupling crop models and remote sensing data, comprising: S1 acquires the rice planting distribution map, daily meteorological reanalysis data, soil grid data, rice variety parameters, and field management parameters for the predicted area; In this implementation, firstly, a rice planting distribution map of the predicted area is obtained. Paddy fields are extracted using random forest classification based on the previous year's Sentinel2 imagery, and planting boundaries are updated using the current year's transplanting period imagery, outputting a 100m resolution raster mask. Then, ERA5Land is downloaded and meteorological data is reanalyzed, aggregating temperature, precipitation, and radiation into daily values, and downscaling to a 100m grid using bilinear interpolation. Next, soil texture, field water holding capacity, wilting coefficient, and other attributes are obtained from SoilGrids or HWSD platforms and resampled to the same grid. Finally, rice variety parameters, including photoperiod sensitivity coefficient, grain-filling rate, accumulated temperature during the growth period, distribution coefficient, etc., as well as management parameters such as average sowing date and planting density, are collected from county-level statistical data and organized according to the WOFOST model format.
[0018] S2 acquires multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season, and retrieves time-series data of leaf area index and topsoil moisture content, respectively. In this step, the leaf area index is retrieved by calculating it pixel-by-pixel using the PROSAIL radiative transfer model lookup method based on the surface reflectance of Sentinel-2 optical imagery. The surface soil moisture content is retrieved by subtracting vegetation influence from the dual-polarization backscattering coefficient of Sentinel-1 radar imagery using a water cloud model to obtain the soil volumetric moisture content at a depth of 0–10 cm.
[0019] In this embodiment, firstly, Sentinel2 L2A imagery is automatically retrieved using Google Earth Engine during the growing season. The leaf area index (LAI) is then retrieved pixel-by-pixel using the PROSAIL radiative transfer model lookup table method to generate LAI time-series data at 5-day intervals, which is then resampled to a 100m grid. Next, Sentinel1 GRD data is acquired synchronously. After preprocessing with SNAP software, the influence of vegetation is subtracted using a water cloud model, and the 0-10cm topsoil volumetric water content is retrieved to generate a soil moisture time-series aligned with the LAI time.
[0020] S3 is based on the WOFOST crop model. It takes daily meteorological reanalysis data, soil grid data and rice variety parameters as inputs to simulate the daily growth status of rice pixel by pixel. The status includes leaf area index, soil moisture content, aboveground biomass and storage organ dry weight. In this step, the WOFOST crop model is implemented using the PCSE framework, where rice variety parameters include photoperiod sensitivity coefficient, grain-filling rate, accumulated temperature during the growth period, and distribution coefficient.
[0021] In this implementation, firstly, a WOFOST model (using the PCSE framework) is independently initialized for each paddy field pixel, inputting the meteorological, soil, and crop parameters from step 1. Then, the model simulates light and temperature production potential, water balance, and dry matter distribution daily in daily increments, outputting daily LAI, soil moisture content, aboveground biomass, and storage organ dry weight. The simulation continues until the set maturity date (or prediction date).
[0022] S4 uses the ensemble Kalman filter algorithm, taking the inverted leaf area index time series data and surface soil moisture content time series data as observations, to assimilate the WOFOST crop model pixel by pixel and update the model's state variables; In this step, leaf area index and surface soil moisture content observations are used simultaneously for each assimilation, and the observation error covariance matrix is set according to the prior accuracy of remote sensing inversion.
[0023] In this implementation, firstly, 50 ensemble members are generated for each pixel, and a 10% Gaussian perturbation is applied to the emergence date, initial available soil water, and grain filling rate. Then, a forecast step is performed: all members are run daily from the previous assimilation time to the next remote sensing observation time, resulting in an ensemble forecast of the state vector. Next, when new LAI and soil moisture observations are available simultaneously, an analysis step is performed: the ensemble mean and covariance are calculated, Kalman gain is constructed, and the state (LAI, soil moisture, biomass, organ weight) of each member is updated using the perturbed observations. The observation error covariance is set according to the prior accuracy of remote sensing inversion. Forecasting and analysis are performed iteratively until the physiological maturity of rice. For meteorological inputs for future unobserved periods, historical average climate data or ECMWF SEAS5 seasonal forecast data are used as substitutes. Finally, the memory organ weight ensemble of each member at maturity is obtained.
[0024] S5 runs the assimilated WOFOST crop model to the physiological maturity stage of rice, extracts the memory stem weight of each set member, calculates the set mean as the predicted yield of that pixel, and calculates the set standard deviation as the prediction uncertainty. In this embodiment, firstly, the ensemble mean of all member memory stem weights on each cell is calculated as the final predicted yield for that cell. Then, the ensemble standard deviation is calculated as the prediction uncertainty, and a yield spatial raster map is output.
[0025] S6 spatially aggregates pixel-level predicted yields based on the boundaries of the grain storage area or the service radius of the drying plant to obtain the predicted total yield and peak harvest period for the region, and generates planning suggestions for storage resources. In this step, the proposed resource planning recommendations include temporary storage area, number of drying lines in operation, and daily peak acquisition funds.
[0026] In this embodiment, firstly, a radiation zone with a radius of 30 km is delineated centered on the coordinates of the grain depot, and paddy field pixels are superimposed on it. Then, the predicted yield of all pixels within the zone is multiplied by the pixel area and summed to obtain the predicted total yield for the region, and the standard deviation of the total yield is calculated. Next, the peak harvest period (the 10-day interval with the highest density) is statistically determined based on the maturity dates output by each pixel. Finally, a resource planning recommendation for grain storage is generated according to the formula given in claim 8. Temporary storage area (m) 2 = Forecast total yield (t) × Rice bulk density (0.58t / m³) 3 Stacking height (4m) × safety factor (0.7); Number of drying lines in operation = Predicted peak daily purchase volume (t / day) / Single line daily processing capacity (t / day), rounded up; Peak daily acquisition funds (ten thousand yuan) = Predicted peak daily acquisition volume (t / day) × Acquisition unit price (yuan / t) / 10000.
[0027] The system automatically generates a report that includes forecast charts and the above recommendations.
[0028] During the actual harvest season, S7 uses grain depot weighbridge purchase data to verify the predicted yield and adaptively adjusts the crop model parameters and assimilation parameters for the next growing season based on the prediction error.
[0029] In this embodiment, during the actual harvest period, the predicted yield is verified plot by plot using the purchase data recorded by the grain depot weighbridge. The relative error is calculated, and if the error exceeds a preset threshold, the error for that year is fed back to the model parameter library. The crop model parameters (such as grain filling rate and photoperiod coefficient) and assimilation parameters (such as observation error covariance) for the next growing season are adjusted using Bayesian update or Kalman smoothing methods to achieve adaptive calibration.
[0030] This invention provides a rice yield prediction method that couples a crop model with remote sensing data. By using ensemble Kalman filtering to simultaneously assimilate the leaf area index retrieved by Sentinel2 optical inversion and the soil moisture retrieved by Sentinel1 radar inversion, the state variables of the crop model are corrected in real time. Ultimately, pixel-level yield predictions can be output 4560 days before harvest, which can significantly improve the timeliness and accuracy of rice yield prediction. This invention solves the problems of existing rice yield prediction methods, such as the difficulty in achieving both mechanistic and timeliness, insufficient expression of spatial heterogeneity, and lack of operational decision indicators. It is particularly suitable for practical scenarios such as grain storage and scheduling, grain trade, and yield monitoring.
[0031] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.
Claims
1. A rice yield prediction method of coupling a crop model with remote sensing data, characterized in that, include: Obtain rice planting distribution maps, daily meteorological reanalysis data, soil grid data, as well as rice variety parameters and field management parameters for the predicted area; Multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season were acquired, and leaf area index time series data and surface soil moisture content time series data were obtained by inversion, respectively. Based on the WOFOST crop model, the daily meteorological reanalysis data, soil grid data and rice variety parameters were used as inputs to simulate the daily growth status of rice pixel by pixel. The status included leaf area index, soil moisture content, aboveground biomass and storage organ dry weight. An ensemble Kalman filter algorithm was used to assimilate the WOFOST crop model pixel by pixel, using the inverted leaf area index time series data and surface soil moisture content time series data as observations, and update the model's state variables. The assimilated WOFOST crop model was run to the physiological maturity stage of rice. The memory stem weight of each ensemble member was extracted, the ensemble mean was calculated as the predicted yield of that pixel, and the ensemble standard deviation was calculated as the prediction uncertainty. Based on the boundaries of the grain storage area or the service radius of the drying plant, the pixel-level predicted output is spatially aggregated to obtain the predicted total output value and the predicted peak harvest value for the region, and to generate planning suggestions for storage resources. 2.The rice yield prediction method of coupling a crop model with remote sensing data according to claim 1, wherein, Based on the boundaries of the grain storage area or the service radius of the drying plant, the pixel-level predicted output is spatially aggregated to obtain the total regional output forecast and the peak harvest forecast. After generating storage resource planning suggestions, the process also includes the following steps: During the actual harvest period, the predicted yield was verified using grain depot weighbridge purchase data, and the crop model parameters and assimilation parameters for the next growing season were adaptively adjusted based on the prediction error.
3. The method of claim 2, wherein the crop model is coupled with the remote sensing data to predict the rice yield. In the steps of acquiring multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season, and then retrieving time-series data of leaf area index and topsoil moisture content respectively,... The leaf area index obtained by the inversion is specifically calculated as follows: based on the surface reflectance of Sentinel-2 optical imagery, the leaf area index is calculated pixel by pixel using the PROSAIL radiative transfer model lookup table method.
4. The method of claim 3, wherein the crop model is coupled with the remote sensing data. In the steps of acquiring multi-temporal Sentinel-2 optical remote sensing images and Sentinel-1 radar remote sensing images of the predicted area during the rice growing season, and then retrieving time-series data of leaf area index and topsoil moisture content respectively,... The inversion of surface soil moisture content specifically involves: using the dual-polarization backscattering coefficient of Sentinel-1 radar imagery and subtracting vegetation influence using a water cloud model to obtain the soil volumetric moisture content at a depth of 0-10 cm.
5. The method of claim 4, wherein the crop model is coupled with the remote sensing data. Based on the WOFOST crop model, using daily meteorological reanalysis data, soil grid data, and rice variety parameters as input, the daily growth status of rice is simulated pixel by pixel. This simulation includes steps for leaf area index, soil moisture content, aboveground biomass, and storage organ dry weight. The WOFOST crop model is implemented using the PCSE framework, and the rice variety parameters include photoperiod sensitivity coefficient, grain filling rate, accumulated temperature during the growth period, and distribution coefficient.
6. The method of claim 5, wherein the crop model is coupled with the remote sensing data to predict the rice yield. In the step of employing the ensemble Kalman filter algorithm, using the retrieved leaf area index time-series data and surface soil moisture content time-series data as observations, the WOFOST crop model is assimilated pixel-by-pixel, and the model's state variables are updated. Each assimilation uses both leaf area index and surface soil moisture content observations, and the observation error covariance matrix is set according to the prior accuracy of remote sensing inversion.
7. The method of claim 6, wherein the crop model is coupled with the remote sensing data to predict the rice yield. Based on the boundaries of the grain storage area or the service radius of the drying plant, the pixel-level predicted output is spatially aggregated to obtain the total regional output forecast and the peak harvest forecast, and suggestions for grain storage resource planning are generated. The proposed resource acquisition and storage plan includes the area of temporary storage, the number of drying lines in operation, and the peak daily acquisition funds.