CN-122022081-A - Mechanism model guided corn yield prediction method

CN122022081ACN 122022081 ACN122022081 ACN 122022081ACN-122022081-A

Abstract

The invention belongs to the technical field of remote sensing prediction, and particularly relates to a mechanism model guided corn yield prediction method. The invention is based on a Monte Carlo meteorological sequence of historical statistics, combines with LAI characteristic values of key growth periods of corns, and simulates genetic parameters, soil parameters and field management parameters corresponding to crops to be predicted. Coupling simulation is carried out through a crop growth model to be predicted, and sensitive parameters of crop yield to be predicted are constructed in advance, so that high calculation cost of traditional pixel-by-pixel simulation is avoided, and efficiency of large-area application is remarkably improved. And by means of simulation data of large-scale yield generated by the crop growth model to be predicted, the invention trains a differentiable multi-layer perceptron model, is used for approximating the internal response process of DSSAT mechanism models to yield, realizes quick and light yield prediction, and simultaneously retains physiological mechanism constraints driven by the model.

Inventors

Li Duqiang
CHEN SHENGBO
Wang Kaisi
Cao Lisai
LU PENG

Assignees

吉林大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (8)

1. A mechanism model guided corn yield prediction method is characterized by comprising the following steps: S1, in a region to be detected, collecting historical real meteorological data of crops to be predicted corresponding to a sample plot in three key growth periods, and carrying out meteorological simulation by adopting a Monte Carlo simulation method based on the historical real meteorological data to obtain meteorological sequences of the three key growth periods; S2, carrying out sensitivity analysis on sensitive parameters of the sample plot based on a Sobel method to obtain simulated sensitive parameters of the sample plot; inputting simulation sensitive parameters of a sample area and meteorological sequences of three key growth periods into a DSSAT model for processing to obtain simulation output and simulation LAI characteristic values corresponding to the three key growth periods respectively; Constructing a multi-layer perceptron model, taking simulation sensitive parameters, meteorological sequences corresponding to three key growth periods and simulation LAI characteristic values as inputs, and taking simulation output as output to train the multi-layer perceptron model to obtain a trained multi-layer perceptron model; S3, acquiring multispectral satellite data of windows corresponding to three key growth periods, inputting the multispectral satellite data of the windows corresponding to the key growth periods into a PROSAIL radiation transmission model for inversion, and obtaining real LAI characteristic values of crops to be predicted in the windows corresponding to the key growth periods; S4, calculating optimal sensitive parameters by combining real LAI characteristic values of crops to be predicted in windows corresponding to key growth periods, and inputting the optimal sensitive parameters into a trained multi-layer perceptron model for processing to obtain predicted yield; And S5, adjusting the trained multi-layer perceptron model based on the calculation result of the step S4 and the real sensitive parameters and the real yields corresponding to the sample plot to obtain a final multi-layer perceptron model, and inputting the real sensitive parameters of the region to be detected, the meteorological sequences corresponding to the three key growth periods and the real LAI characteristic values into the final multi-layer perceptron model for processing to obtain the yields of the crops to be predicted.
2. A mechanism model directed corn yield prediction method according to claim 1 wherein in step S1, the meteorological data comprises solar radiation, solar maximum temperature, solar minimum temperature and rainfall.
3. The method for predicting corn yield guided by a mechanism model according to claim 1, wherein the specific steps of adopting a Monte Carlo simulation method to perform meteorological simulation and obtaining meteorological sequences of three key growth periods are as follows: s11, determining precipitation constraint of a Monte Carlo simulation method corresponding to an ith key growth period based on historical real meteorological data of crops to be predicted in the ith key growth period, wherein the precipitation constraint comprises a wet day occurrence probability and gamma distribution parameters; S12, generating precipitation amount of crops to be predicted on the d day corresponding to the kth simulation through a Monte Carlo simulation method based on precipitation constraint of the Monte Carlo simulation method Obtaining a day-by-day precipitation sequence of the kth simulation : ={ , ,..., }; Wherein, the For simulating the precipitation amount of the corresponding crop to be predicted on the 1 st day for the kth time, For simulating the precipitation amount of the corresponding crop to be predicted on the 2 nd day for the kth time, Simulating the precipitation amount of the corresponding crop to be predicted on the D day for the kth time, wherein D is the growth period length of the crop to be predicted; S13, a daily precipitation sequence based on kth simulation Generating a kth simulated daily temperature sequence by adopting an autoregressive model with seasonal items : ; ; ; Wherein, the For the temperature on day d in the kth simulation, The month average value temperature of month m corresponding to the d day, In the form of an autoregressive coefficient matrix, As a random disturbance vector on day d in the kth simulation, As a covariance matrix of the random disturbance, For the highest temperature on day d in the kth simulation, The lowest temperature on day d in the kth simulation; s14, a daily precipitation sequence based on the kth simulation Synchronous simulation solar radiation sequence Day-by-day precipitation sequence based on kth simulation Day-by-day temperature sequence of kth simulation And the kth simulated solar radiation sequence Generating a simulated weather sequence for the kth time : ; Wherein k is an index of the simulation times; s15, repeating the steps S11-S14 until K times of simulation are completed, wherein K=1, 2, & gt, K is the same as the weather sequence of the ith key growth period; s16, replacing the weather data of the ith key growth period with the weather data of the (i+1) th key growth period, and repeating the steps S11-S15 until weather sequences of three key growth periods are obtained.
4. The method for predicting corn yield guided by a mechanism model according to claim 3, wherein the step S11 comprises the following steps: S111, setting a date that the single-day precipitation amount in the historical real meteorological data of the ith key growth period is larger than a precipitation amount threshold value as a wet day; s112, calculating the occurrence probability of the wet days of the ith key growth period by the following formula: ; ; Wherein F is the total precipitation of crops to be predicted in the ith key growth period, For the precipitation of crops to be predicted in the ith key growth period, D is the growth period length of the ith key growth period of the crops to be predicted, P min is the precipitation threshold value for the probability of occurrence of wet days; s113, gamma distribution parameter estimation is carried out based on the precipitation amount of wet days: ; Wherein, the As the shape parameter estimation value, a method for estimating the shape parameter, As an estimate of the scale parameter, And s is the mean value of the simulated precipitation, and s is the variance of the simulated precipitation.
5. The method for predicting corn yield guided by a mechanism model according to claim 1, wherein in the step S2, the sensitive parameters comprise genetic parameters, field management parameters and soil parameters, wherein the genetic parameters comprise radiation utilization rate, photosynthetic effective radiation canopy extinction coefficient and leaf development heat accumulation coefficient, the field management parameters comprise planting date, planting density and nitrogen application amount, and the soil parameters comprise effective soil depth and soil organic matter content.
6. The method for predicting corn yield by using a mechanism model according to claim 1 wherein when the crop to be predicted is corn, the three key growth phases include a jointing phase, a male-pulling and silking phase and a middle-late grouting phase.
7. The method for predicting corn yield guided by a mechanism model according to claim 1, wherein the step S4 comprises the following steps: s41, calculating optimal parameters based on the real LAI characteristic values of the windows corresponding to the key growth periods and the real meteorological data of the windows corresponding to the key growth periods by the following formula: ; ; Wherein, the As a matching function for calculating the parameter theta, The weight of the i-th key fertility period, i is the i-th key fertility period, i=1, 2,3, To adjust the weights of the real LAI eigenvalues, Is the simulated LAI characteristic value corresponding to the ith key growth period under the parameter theta output by DSSAT model, In order to adjust the weight of the real weather data, For the simulated meteorological values at the ith critical growth period, As a result of the optimal parameters, Is the average value of the LAI characteristic values under the ith key growth period, As the average meteorological data at the ith critical growth period, For the true LAI eigenvalues inverted by the PROSAIL model under the i-th key growth period corresponding window, Real meteorological data of a window corresponding to the ith key growth period; S42, inputting optimal parameters, real LAI characteristic values of windows corresponding to the key growth periods and real meteorological data of windows corresponding to the key growth periods into a trained four-layer perceptron model to obtain yield prediction of crops to be predicted under the condition of no sample participation: ; Wherein, the () In order to train the four-layer perceptron model, The predicted yield of the crop to be predicted under the condition of no sample participation.
8. The method for mechanism model guided corn yield prediction of claim 7 wherein the mixing loss function used to adjust the trained multi-layer perceptron model is: ; ; = ; ; Wherein, the In order to mix the loss function, For the small batch sample set extracted from the simulation result of DSSAT model, the number of samples contained in the small batch sample set is 10-20, The mean square error of the predicted output and the real output of the multi-layer perceptron model, Is a model of the multi-layer perceptron, For the true yield of the plot, As the weight of the material to be weighed, A small batch truth value set corresponding to the real yield of the sample area, the small batch sample set and the small batch truth value set have the same sample quantity, For the weight importance protection term(s), For the regularized intensity super-parameter, For Fisher information approximation parameters, W k is the parameters of the trained four-layer perceptron model, Is a parameter of the original four-layer perceptron model.

Description

Mechanism model guided corn yield prediction method Technical Field The invention belongs to the technical field of remote sensing prediction, and particularly relates to a mechanism model guided corn yield prediction method. Background Corn is taken as one of the most important grains and forage crops to be predicted in China, and the space-time distribution of the yield is jointly influenced by multiple factors such as climate change, soil conditions, variety difference, field management and the like. The main current remote sensing crop to be predicted production estimation method at home and abroad mainly comprises four major categories, namely a remote sensing vegetation index statistical model, a light energy utilization rate model, a remote sensing-crop to be predicted growth coupling model and a deep learning regression model. However, these methods still have significant limitations in the wide range, high precision, and robust task of predicting crop yield to be predicted. The remote sensing vegetation index statistical model is mainly based on experience and statistical relation, and the yield is directly deduced from satellite remote sensing indexes. The model does not explicitly consider the crop growth process to be predicted and meteorological driving factors, can only reflect the overall trend of yield, has limited precision, and is difficult to stably migrate in different regions and years. The light energy utilization rate model is an important way to estimate vegetation productivity, and is built by theoretical simplification of the photosynthesis process. Although the physiological mechanism of the crops to be predicted is partially involved, key parameters of the crops often depend on a large amount of field observation to be accurately acquired, and meanwhile, human factors such as management measures and the like are difficult to effectively incorporate, so that the crop physiological mechanism has high parameterization cost and weak popularization. The coupling method of the remote sensing and the crop model to be predicted obtains the yield by assimilating the remote sensing information into the crop growth model to be predicted, and has stronger capability for expanding time and space. However, pixel-by-pixel assimilation is often computationally intensive, resulting in inefficiency and high cost when applied in large areas. The deep learning regression method relies on the yield measurement sample to learn the highly nonlinear yield mapping relation from the multi-source characteristics, and has higher precision. However, the model relies on a large amount of measured data, and has weak interpretation, and the generalization capability is insufficient under the condition of data distribution deviation. In summary, the existing method still has difficulty in considering mechanism interpretability, calculation efficiency, space-time generalization capability and prediction accuracy, and a new technical paradigm is needed to break through the key bottleneck of corn remote sensing estimation. Disclosure of Invention In view of the above, the invention aims to provide a mechanism model guided corn yield prediction method so as to solve the problem that the mechanism interpretability, the calculation efficiency, the time-space generalization capability and the prediction precision are difficult to be considered in the prior art. In order to achieve the above purpose, the technical scheme of the invention is realized as follows: a mechanism model guided corn yield prediction method specifically comprises the following steps: S1, in a region to be detected, collecting historical real meteorological data of crops to be predicted corresponding to a sample plot in three key growth periods, and carrying out meteorological simulation by adopting a Monte Carlo simulation method based on the historical real meteorological data to obtain meteorological sequences of the three key growth periods; S2, carrying out sensitivity analysis on sensitive parameters of the sample plot based on a Sobel method to obtain simulated sensitive parameters of the sample plot; inputting simulation sensitive parameters of a sample area and meteorological sequences of three key growth periods into a DSSAT model for processing to obtain simulation output and simulation LAI characteristic values corresponding to the three key growth periods respectively; Constructing a multi-layer perceptron model, taking simulation sensitive parameters, meteorological sequences corresponding to three key growth periods and simulation LAI characteristic values as inputs, and taking simulation output as output to train the multi-layer perceptron model to obtain a trained multi-layer perceptron model; S3, acquiring multispectral satellite data of windows corresponding to three key growth periods, inputting the multispectral satellite data of the windows corresponding to the key growth periods into a PROSAIL radiation transmission model fo