Paddy rice yield growing season high-precision prediction method based on general primary productivity remote sensing product

Abstract

The invention provides a rice yield growth season high-precision prediction method based on a total primary productivity remote sensing product. The method comprises the following steps: firstly, calculating rice total primary productivity GPP and a normalized difference red-edge vegetation index NDRE in a rice growth season on a remote sensing cloud platform based on a time sequence Sentinel-2 image; then, respectively carrying out correlation analysis on the yield by utilizing GPP and NDRE of monthly-scale rice; performing harmonic fitting on the time sequence GPP based on a harmonic analysis technology, and determining an optimal time window in a growth season for rice yield prediction at an interval of 10 days; the rice yield is estimated by utilizing transfer learning and GPP and NDRE of the optimal time window, so that the advantages of GPP in yield estimation are clear; and an inter-annual GPP correction method suitable for different ecological points is constructed, so that the yield estimation precision of independent years is improved. According to the method, the yield of the rice in the field can be accurately estimated in seasons in a regional range, and accurate management of the rice can be promoted by farmers and agricultural technology promoters.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural remote sensing monitoring, and is a method for accurately predicting rice yield during the growing season based on remote sensing products of gross primary productivity. Background Art

[0002] Early and reliable field-level crop yield estimation can not only help farmers make management decisions during the growing season, but also effectively cope with the impact of climate change. At the same time, early yield estimation can also help the government timely formulate strategies to address potential food market fluctuations. In crop yield estimation methods, statistical experience / data-driven methods and remote sensing and crop model assimilation methods are the main ways to achieve mid-season crop yield estimation. When the data assimilation method is applied in a regional scope, due to the complexity of parameters (crop varieties, soil parameters, etc.) and the difficulty of obtaining them, its large-scale application is restricted to a certain extent. Therefore, the data-driven method shows great application potential in mid-season crop yield estimation, especially in the 1-2 months before harvest or the heading and filling stage, which is considered the best estimation window. Theoretically, the best early estimation period is closely related to crop phenology, and there are also obvious differences in crop phenology regionally and within regions. However, these time windows are relatively rough, difficult to reflect the dynamic changes of crops in different growth stages, and cannot meet the needs of current precision agriculture management (such as irrigation and fertilization). Therefore, it is necessary to clarify a more refined time window for crop yield estimation based on time series remote sensing images and statistical experience / data-driven methods. However, no matter which estimation method is adopted, optical remote sensing images are affected by clouds, which limits the research on crop yield estimation. For this reason, filling and interpolation techniques aim to reconstruct continuous time series remote sensing images from insufficient original multi-temporal images, thereby providing important support for crop growth monitoring and yield estimation.

[0003] When estimating crop yields using statistical experience / data-driven methods, remote sensing vegetation indices are often used as intermediate variables, while GPP data, which is significantly correlated with crop biomass and yields, is less used. This may be due to the relatively low spatial resolution (≤500m) of current GPP products, which are mainly applied to global-scale ecosystems, and the underestimation phenomenon of the widely used MOD17 GPP product in agricultural ecosystems. Therefore, it is particularly important to develop high-quality and high-spatial-resolution GPP products suitable for agricultural ecosystems for mid-season crop yield estimation. Through high-resolution time-series GPP products, a more refined time window for crop yield estimation can be determined, and the differences in the capabilities of GPP and vegetation indices in crop yield estimation can be systematically evaluated, especially in terms of the stability of inter-annual yield estimation. Currently, mid-season estimation of crop yields for individual years is mainly achieved through data-driven models, which rely heavily on field-measured yield data on the ground, resulting in problems such as poor stability and low accuracy. As a machine learning method, transfer learning aims to apply the learned knowledge and experience from one task or domain to another task or domain to improve learning performance. This algorithm has great potential in mid-season estimation of individual-year yields, but still requires a large amount of field-measured yield data. Therefore, based on high-spatial-resolution GPP data, it is particularly important to develop a simple-to-operate method for mid-season estimation of inter-annual crop yields to achieve high-precision estimation of individual-year crop yields in the case of small samples. Summary of the Invention

[0004] The technical problem solved by the present invention is to provide a method for high-precision prediction of rice yield during the growing season based on remote sensing products of gross primary productivity. Using high-spatial-resolution GPP products of different years as data sources, combined with harmonic analysis technology and trigonometric function transformation, the correlation between GPP, vegetation indices and yield, as well as the differences in inter-annual yield estimation accuracy, are systematically analyzed; the optimal time window for yield estimation of individual and cumulative GPP values at 10-day intervals is analyzed; and a cross-year GPP correction (CGC) method applicable to different ecological sites is proposed, its inter-annual yield estimation accuracy in the mid-season of different ecological sites is evaluated, and yield prediction at the field scale of a large area (provincial level) and a 10-meter spatial resolution rice yield distribution product are achieved. This method reveals the performance differences and their reasons of high-resolution GPP and vegetation indices in yield estimation, determines the optimal time window for yield estimation at 10-day intervals, can achieve high-precision mid-season yield estimation, eliminate the influence of inter-annual differences in rice growth, has simple steps and is easy to operate, and can be widely applied to high-precision prediction of large rice yields in different rice planting areas.

[0005] The technical solution to achieve the object of the present invention is as follows:

[0006] A high-precision prediction method for rice yield during the growing season based on remote sensing products of gross primary productivity, comprising the following steps:

[0007] Step 1: Obtain the time series Sentinel-2 satellite images of the rice growing season in the study area, preprocess the images on the remote sensing cloud platform respectively, combine the spatio-temporal interpolation technology to obtain the daily Sentinel-2 images, calculate the daily rice gross primary productivity GPP and the normalized red-edge vegetation index NDRE, calculate the average value and cumulative value of GPP and NDRE respectively in each month of the rice growing season, and conduct Pearson correlation analysis with the rice yield;

[0008] Step 2: Based on the daily high-resolution GPP products of the time series, use the ordinary least squares regression and the discrete Fourier transform HANTS method for pixel-by-pixel fitting to eliminate the influence of high-frequency noise. Then, calculate the harmonic fitting GPP data from the original high-resolution GPP data in the study area through this method. Calculate the cumulative values of the original GPP and the harmonic fitting GPP every 10 days during the rice growing season. Use the original and harmonic fitting GPP data with a 10-day time interval during the rice growing season in the study area for linear regression with the yield to determine the optimal time window for yield estimation;

[0009] Step 3: Based on the cumulative fitting high-resolution GPP and NDRE data within the best 10-day window, use the Transformer architecture of the transfer learning algorithm for mid-season prediction of rice yield, and conduct random split method and leave-one-year independent validation accuracy evaluation to clarify the differences between high-resolution GPP and the vegetation index NDRE in rice yield prediction;

[0010] Step 4:

[0011] Step 4-1: Extract the maximum value of the harmonic fitting GPP within the rice growing season in different ecological regions;

[0012] Step 4-2: The measured dataset of rice yield and the maximum value of the harmonic fitting GPP within the provincial scope in the previous n years, take the data of 1 year as the independent validation set V; the remaining dataset D within the provincial scope;

[0013] Step 4-3: Based on the triangular transformation principle technology, conduct inter-annual correction on the harmonic fitting GPP, and construct the inter-annual transfer correction method CGC;

[0014] Step 4-4: Based on the CGC method and the training set D, realize mid-season yield estimation and conduct mid-season prediction of rice yield in different ecological regions.

[0015] Furthermore, for the high-precision prediction method for rice yield during the growing season based on remote sensing products of gross primary productivity proposed by the present invention, the specific steps in Step 1 include:

[0016] Step 1-1: Obtain the rice planting distribution and rice growth stage information in the study area;

[0017] Step 1-2: Extract the Sentinel-2 time series images during the rice growing season using the rice growth stage information in the study area;

[0018] Step 1-3: Preprocess and cloud mask the time series Sentinel-2 images, and use spatio-temporal interpolation technology to obtain daily Sentinel-2 images;

[0019] Step 1-4: Calculate the gross primary productivity GPP and the normalized red-edge vegetation index NDRE using the daily Sentinel-2 images. The specific formulas are as follows;

[0020] GPP = (ε msu × APAR su + ε msh × APAR sh ) × g(T a ) × Min(g(VPD), g(CO2))

[0021]

[0022] In the formula, ε msu represents the maximum light use efficiency of sun leaves, APAR su represents the absorbed photosynthetically active radiation of sun leaves, ε msh represents the maximum light use efficiency of shade leaves, APAR sh represents the absorbed photosynthetically active radiation of shade leaves, g(CO2) represents the carbon dioxide limitation factor, g(VPD) represents the water limitation factor, g(T a ) represents the air temperature limitation factor, ρ NIR and ρ Red-edge represent the reflectances of Sentinel-2 band 8 (785 - 900 nm) and band 6 (733 - 748 nm) respectively;

[0023] Step 1-5: Calculate the average and cumulative values of GPP and NDRE for each month during the rice growing season;

[0024] Step 1-6: Conduct Pearson correlation analysis on the average and cumulative values of GPP and NDRE with the rice yield respectively to clarify the differences of different remote sensing variables in rice yield estimation;

[0025] Step 1-7: Apply the square of the correlation coefficient r 2 to evaluate the results.

[0026] Furthermore, for the high-precision prediction method of rice yield during the growing season based on the remote sensing product of gross primary productivity proposed by the present invention, the specific steps of step 2 include:

[0027] Step 2-1: Obtain the daily GPP data during the rice growing season according to the growth process information of the study area;

[0028] Step 2-2: Use ordinary least squares regression and the discrete Fourier transform HANTS method to perform pixel-by-pixel fitting to obtain the harmonic fitting GPP data. The specific calculation formula is as follows;

[0029]

[0030] Phase = arctan(A i / B i )

[0031] where f(t) is the GPP varying with time, Amplitude represents the amplitude, and Phase represents the phase; t is the observation date, expressed as the fraction of the calendar year, ranging from 0 to 1. n is the number of harmonic terms, ω is the frequency, A i and B i are the cosine coefficient and sine coefficient of the i-th harmonic term respectively. C is the intercept coefficient.

[0032] Step 2-3: Calculate the original and harmonic fitting GPP at 10-day intervals during the rice growing season in the study area to obtain the original GPP values and cumulative values every 10 days, as well as the harmonic fitting GPP values and cumulative values every 10 days;

[0033] Step 2-4: Perform linear regression on the original GPP and the fitted GPP data at each 10-day interval with the rice yield respectively;

[0034] Step 2-5: Apply the square of the correlation coefficient r 2 to evaluate the regression results at each 10-day interval and determine the optimal 10-day interval window during the rice growing season.

[0035] Furthermore, for the high-precision prediction method of rice yield during the growing season based on the total primary productivity remote sensing product proposed by the present invention, the specific steps in Step 3 include:

[0036] Step 3-1: Perform harmonic fitting calculation on the daily normalized red-edge vegetation index NDRE during the rice growing season to obtain the fitted NDRE;

[0037] Step 3-2: Extract the cumulative values of GPP and NDRE obtained in Step 2 into the transfer learning Transformer framework at the optimal time window, and use random splitting 7:3 and independent year verification to obtain the mid-season yield prediction accuracy and compare and evaluate;

[0038] Step 3-3: Input the above two types of data into the input transfer learning Transformer framework respectively, clarify the differences between the gross primary productivity GPP and the vegetation index NDRE in rice yield prediction, and determine the optimal indicators for interannual yield prediction;

[0039] Step 3-4: From the measured rice yield dataset within the provincial scope in the previous n years, take the measured data of 1 year as the independent validation set V; and the remaining measured rice yield dataset D within the provincial scope;

[0040] Step 3-5: Evaluate the verification results respectively based on the test set T and the independent validation set V by using the coefficient of determination R2, root mean square error RMSE, and relative root mean square error RRMSE.

[0041] Furthermore, the high-precision prediction method for rice yield during the growing season based on the remote sensing product of gross primary productivity proposed by the present invention performs interannual correction on the harmonic fitting GPP based on the triangular transformation principle technology, and constructs an interannual migration correction method CGC. The specific steps are as follows;

[0042] First, calculate the amplitudes of the reference and target fitting GPP curves at any time period during the two-year rice growing season. According to the reference and target fitting GPP curves, use the formula to calculate the target-corrected fitting GPP curve.

[0043] Second, determine the scale factor (sf), that is, the ratio of the target GPP to the reference GPP at a single or cumulative value within any two phenological periods. For example, use the formula to calculate the ratio of the maximum value of the target GPP to the maximum value of the reference GPP for quick implementation.

[0044] Third, use the formula to construct a reference annual yield estimation model, where the dependent variable is the single or cumulative GPP at any phenological period, which is determined by two scenarios. One scenario is to construct a yield estimation model applicable to similar years using only the yield data of one year. Another scenario is to construct a model using the yield data of multiple years for yield estimation in the remaining years. For simplicity of processing, this study selects the maximum value of GPP during the rice growing season as the dependent variable of the yield estimation model.

[0045] Fourth, use the formula to predict the yield of the independent year.

[0046] GPP Target_corrected =GPP Target ÷Amplitude Target ×Amplitude Reference (4-4)

[0047]

[0048] Y Reference =A×X Reference+B; (4 - 7)

[0049] Y Target = a × X Target_corrected +b; a = A × sf; b = B × sf (4 - 8)

[0050] In the formula, GPP Target and GPP Target_corrected respectively represent the target curve GPP and its corrected value; Amplitude Target and Amplitude Reference respectively represent the amplitudes of the target curve and the reference curve; GPP Target_max and GPP Reference_max respectively represent the maximum GPP values of the target curve and the reference curve; A, B, a, and b are model coefficients respectively; X Reference , Y Reference , X Target_corrected and Y Target respectively represent the modeling variables and yields in the reference year, and the modeling variables and yields in the target year.

[0051] Compared with the prior art, the present invention adopts the above technical solutions and has the following technical effects:

[0052] 1. The high-precision prediction method for rice yield during the growing season based on the total primary productivity remote sensing product of the present invention breaks through the limitations of current vegetation index yield estimation, eliminates the influence of inter-annual differences in rice growth, and is simple to operate and highly universal;

[0053] 2. The high-precision prediction method for rice yield during the growing season based on the total primary productivity remote sensing product of the present invention can more accurately track the growth status of rice and its response to the environment, with high stability, high precision, and relatively low dependence on samples;

[0054] 3. The high-precision prediction method for rice yield during the growing season based on the total primary productivity remote sensing product of the present invention only requires the yield model of the reference year, has simple steps, is convenient and efficient, and can be automatically applied to rice yield prediction in different planting areas. Brief Description of the Drawings

[0055] Figure 1 is the technical roadmap of the high-precision prediction method for rice yield during the growing season based on the total primary productivity remote sensing product of the present invention;

[0056] Figure 2 is the schematic diagram of the inter-annual GPP correction method of the high-precision prediction method for rice yield during the growing season based on the total primary productivity remote sensing product of the present invention;

[0057] Figure 3It is a graph of the square of the Pearson correlation coefficient between GPP and NDRE and yield on a monthly basis for the high-precision prediction method of rice yield during the growing season based on the remote sensing product of gross primary productivity of the present invention, where A and B represent GPP and NDRE respectively;

[0058] Figure 4 It is a graph of the optimal fine time window for yield estimation of the high-precision prediction method of rice yield during the growing season based on the remote sensing product of gross primary productivity of the present invention, where (A), (B), (C), and (D) represent the square of the correlation of the 10-day step of the original GPP, the 10-day step of the harmonic-fitted GPP, the cumulative value of the 10-day step of the original GPP, and the cumulative value of the 10-day step of the harmonic-fitted GPP respectively;

[0059] Figure 5 It is the yield verification accuracy of different years with 2022 as the reference year for the high-precision prediction method of rice yield during the growing season based on the remote sensing product of gross primary productivity of the present invention, where (A), (B), (C), (D), (E), and (F) represent the scatter plots of yield prediction for 2019 to 2021 without using and using the CGC method with 2022 as the reference year;

[0060] Figure 6 It is a spatial distribution map of the rice yield in Jiangsu Province in 2021 estimated by the model based on the reference year 2022 for the high-precision prediction method of rice yield during the growing season based on the remote sensing product of gross primary productivity. (A) uses the CGC method, and (D) does not use the CGC method. (C) and (F) show the latitude profiles with the shaded area representing the standard error, and (B) and (E) show the longitude profiles with the shaded area representing the standard error;

[0061] Figure 7 It is a spatial distribution map of the rice yield in Wuchang City, Heilongjiang Province in 2021 estimated by the model based on the reference year 2022 for the high-precision prediction method of rice yield during the growing season based on the remote sensing product of gross primary productivity. (A) uses the CGC method, and (B) does not use the CGC method. (C) and (F) show the latitude profiles with the shaded area representing the standard error, and (D) and (E) show the longitude profiles with the shaded area representing the standard error. Detailed implementation mode

[0063] The following details the implementation mode of the present invention. Examples of the implementation mode are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The implementation mode described by referring to the drawings below is exemplary and is only used to explain the present invention and cannot be construed as a limitation of the present invention.

[0064] The present invention is implemented based on time series Sentinel-2 satellite images and ground-measured yield data. The specific usage characteristics are shown in Table 1.

[0065] Table 1 Sentnel-2 characteristic variables

[0066]

[0067] This embodiment is carried out in Jiangsu Province and Wuchang City, Heilongjiang Province, China's major rice-growing provinces. The rice planting area in Jiangsu Province exceeds 33 million mu, and the output exceeds 19.8 million tons, ranking fourth in the country. It is mainly distributed in Yancheng City, Huai'an City, Nantong City, Taizhou City, etc. Wuchang City is a county-level city in the southern part of Heilongjiang Province in Northeast China. It is famous for its suitable climate, soil and water conditions and is an ideal area for rice growth. Moreover, the training and testing of the field-scale rice yield prediction model were repeated in Jiangsu Province from 2019 to 2022 and Wuchang City from 2019 to 2021, respectively (in Jiangsu Province, from June 1 to October 31, and in Wuchang City, Heilongjiang Province, from May 1 to September 30, the maximum values of the original and fitted GPP were extracted, respectively; based on the principle of trigonometric function transformation, an interannual GPP correction method was proposed to estimate the yield of independent years that were not used for model training; in the case of limited samples, a yield estimation model for the reference year was constructed by using a single or cumulative GPP of any phenological period, thereby realizing cross-year yield estimation), which can effectively verify the advancement, universality and robustness of the present invention.

[0068] A method for high-precision prediction of rice yield during the growing season based on remote sensing products of gross primary productivity, such as Figure 1 As shown in the figure, it mainly includes remote sensing satellite preprocessing and spatiotemporal interpolation, field-scale rice yield collection, systematic analysis of the correlation between GPP and vegetation index and yield, as well as the difference in inter-annual yield estimation accuracy, and analysis of the optimal time window for yield estimation of single and cumulative GPP values with a time interval of 10 days; and proposes a cross-year GPP correction (CGC) method suitable for different ecological points. Figure 2 As shown in the figure, the accuracy of annual yield estimation in different ecological point seasons is evaluated, and a high spatial resolution spatial distribution map of rice yield in different planting areas is drawn, which specifically includes the following steps:

[0069] Step 1: Obtain the time-series Sentinel-2 satellite images during the rice growing season in Jiangsu Province, and preprocess and perform spatio-temporal interpolation on the Sentinel-2 images on the remote sensing cloud platform. The processing steps of Sentinel-2 include cloud removal, spatio-temporal interpolation, vegetation index calculation, gross primary productivity calculation, maximum value synthesis, and cumulative value synthesis. The Sentinel-2 cloud probability dataset can characterize the probability of each pixel in the Sentinel-2 image being contaminated by clouds. In the present invention, the original Sentinel-2 surface reflectance image is first de-clouded based on this dataset to obtain a cloud-free Sentinel-2 image; then, to address the data missing caused by cloud contamination, the linear time interpolation method is used to fill in the missing values on the Google Cloud platform. The present invention calculates the daily rice gross primary productivity GPP and the normalized red-edge vegetation index NDRE based on the original Sentinel-2 bands, and generates the average and cumulative values of GPP and NDRE of Sentinel-2 for each month during the rice growing season; finally, the Pearson correlation between the measured rice yields in Jiangsu Province from 2019 to 2022 and GPP and NDRE at the monthly scale is analyzed (as Figure 3 shown) to clarify the differences between GPP and the vegetation index NDRE in yield estimation. Specifically:

[0070] GPP = (ε msu ×APAR su +ε msh ×APAR sh )×g(T a )×Min(g(VPD), g(CO2))

[0071]

[0072] In the formula, ε msu represents the maximum light use efficiency of sunlit leaves, APAR su represents the absorbed photosynthetically active radiation by sunlit leaves, ε msh represents the maximum light use efficiency of shaded leaves, APAR sh represents the absorbed photosynthetically active radiation by shaded leaves, g(CO2) represents the carbon dioxide limitation factor, g(VPD) represents the water limitation factor, g(T a ) represents the air temperature limitation factor, ρ NIR and ρ Red-edge represent the reflectances of Sentinel-2 band 8 (785 - 900 nm) and band 6 (733 - 748 nm), respectively.

[0073] From Figure 3It can be seen that rice yield is significantly correlated with GPP and NDRE of a single month and cumulative months (p<0.05). Generally speaking, whether it is GPP or NDRE, the correlation between cumulative monthly variables and yield is significantly higher than that between single-month variables and yield. Compared with single-month data, the correlation between cumulative monthly data and yield increases significantly during the reproductive growth period, manifested as the change of r 2 (Δr 2 )(GPP: Δr 2 is 0.22 in September and Δr 2 is 0.58 in October; NDRE: Δr 2 is 0.11 in September and Δr 2 is 0.21 in October). Specifically, the correlation between cumulative GPP in August and yield is significantly higher than that of single-month GPP in August, while NDRE shows the opposite trend. In addition, both single-month (Δr 2 =0.04 - 0.29) and cumulative-month (Δr 2 =0.22 - 0.41) GPP are better than single-month and cumulative-month NDRE. It should be noted that the Δr 2 of cumulative-month GPP relative to NDRE during the four-year period is significantly higher than that of single-month GPP relative to NDRE's Δr 2 .

[0074] Step 2: Determine the optimal fine time window for yield prediction based on high-resolution total primary productivity of the time series, specifically including:

[0075] Step 2-1: Obtain daily GPP data of the rice growing season according to the growth process information of the study area;

[0076] Step 2-2: Use ordinary least squares regression and the discrete Fourier transform HANTS method for pixel-by-pixel fitting to obtain harmonic fitting GPP data. The specific calculation formula is as follows;

[0077]

[0078] Phase = arctan(A i / B i )

[0079] where f(t) is the GPP changing with time, Amplitude represents the amplitude, and Phase represents the phase; t is the observation date, expressed as the fraction of the calendar year, ranging from 0 to 1. n is the number of harmonic terms, ω is the frequency, A i and B i are the cosine coefficient and sine coefficient of the i-th harmonic term respectively. C is the intercept coefficient.

[0080] Step 2-3: Calculate the original harmonic fitting GPP of the rice growing season in the study area at 10-day intervals, and obtain the original GPP values and cumulative values every 10 days, as well as the harmonic fitting GPP values and cumulative values every 10 days;

[0081] Step 2-4: Perform linear regression on the original GPP and fitted GPP data of every 10 days per day with the rice yield respectively;

[0082] Step 2-5: Apply the square of the correlation coefficient r 2 to evaluate the results and determine the optimal time window for yield prediction ( Figure 4 ). For the original GPP data, the performance during the vigorous growth period of rice (DOY 213-232: r 2 = 0.45 - 0.59) is always better than other time periods. The cumulative value of data every 10 days is significantly better than that of individual 10-day data in terms of yield estimation performance. Specifically, during the period from 2019 to 2022, the r 2 of DOY 183-232 increased by 0.03 - 0.11, while the r 2 of DOY 183-302 increased by 0.41 - 0.66. Compared with the original GPP data, the harmonic fitting GPP is more superior in both single (Δr 2 = 0.03 - 0.48) and cumulative periods (Δr 2 = 0.01 - 0.47), and also advances the optimal time window by at least 10 days.

[0083] Step 3: Use the transfer learning Transformer framework to predict the mid-season rice yield for different remote sensing variables within the provincial scope, specifically including:

[0084] Step 3-1: Perform harmonic fitting calculation on the daily normalized red-edge vegetation index NDRE during the rice growing season to obtain the fitted NDRE;

[0085] Step 3-2: Extract the cumulative values of GPP and NDRE at the optimal time window obtained in Step 2 and input them into the transfer learning Transformer framework respectively, and use random splitting 7:3 and independent year verification to obtain the mid-season yield prediction accuracy and compare and evaluate;

[0086] Step 3-3: Input the above two types of data into the transfer learning Transformer framework respectively, clarify the differences between the gross primary productivity GPP and the vegetation index NDRE in rice yield prediction, and determine the optimal indicators for inter-annual yield prediction;

[0087] Step 3-4: From the measured rice yield dataset within the provincial region in the previous n years, select the measured data of 1 year as the independent validation set V; divide the remaining measured rice yield dataset D within the provincial region into two mutually exclusive sets according to the holdout method, where one set is the training set S and the other is the test set T, that is, D = S ∪ T, S ∩ T = ∅;

[0088] Step 3-5: Evaluate the validation results based on the test set T and the independent validation set V using the coefficient of determination R 2 , root mean square error RMSE, and relative root mean square error RRMSE.

[0089] Step 4: Construct an inter-annual transfer GPP correction CGC method that can predict rice yields in different years during the growing season. The specific steps are as follows:

[0090] Step 4-1: Extract the maximum value of the harmonic fitting GPP during the rice growing season in different ecological regions;

[0091] Step 4-2: From the measured rice yield dataset within the provincial region in the previous n years and the maximum value of the harmonic fitting GPP, select the data of 1 year as the independent validation set V; for the remaining dataset D within the provincial region;

[0092] Step 4-3: Based on the principle of triangular transformation technology, perform inter-annual correction on the harmonic fitting GPP and construct an inter-annual transfer correction method CGC (as Figure 2 shown), and the specific steps are as follows;

[0093] First, calculate the amplitudes of the reference and target fitting GPP curves at any time during the two-year rice growing seasons. According to the reference and target fitting GPP curves, use the formula to calculate the target-corrected fitting GPP curve.

[0094] Second, determine the scale factor (sf), that is, the ratio of the target GPP to the reference GPP at the single or cumulative value within any two phenological periods. For example, use the formula to calculate the ratio of the maximum value of the target GPP to the maximum value of the reference GPP for quick implementation.

[0095] Third, use the formula to construct a reference annual yield estimation model, where the dependent variable is the single or cumulative GPP at any phenological period, which is determined by two scenarios. One scenario is to construct a yield estimation model applicable to similar years using only the yield data of one year. The other scenario is to construct a model using the yield data of multiple years for yield estimation of the remaining years. For simplicity of processing, this study selects the maximum value of GPP during the rice growing season as the dependent variable of the yield estimation model.

[0096] Fourth, use the formula to predict the yield of the independent year.

[0097] GPP Target_corrected = GPPTarget ÷Amplitude Target ×Amplitude Reference (4-4)

[0098]

[0099] Y Reference = A × X Reference + B; (4-7)

[0100] Y Target = a × X Target_corrected + b; a = A × sf; b = B × sf (4-8)

[0101] In the formula, GPP Target and GPP Target_corrected respectively represent the target curve GPP and its corrected value; Amplitude Target and Amplitude Reference respectively represent the amplitudes of the target curve and the reference curve; GPP Target_max and GPP Reference_max respectively represent the maximum GPP values of the target curve and the reference curve; A, B, a, and b are model coefficients respectively; X Reference , Y Reference , X Target_corrected and Y Target respectively represent the modeling variables and yields in the reference year, and the modeling variables and yields in the target year.

[0102] Step 4-4: Based on the CGC method and the training set D, realize mid-season yield estimation. Conduct mid-season prediction of rice yield in different ecological regions, and compare with the yield prediction without using this method and the prediction results of the Transformer architecture( Figure 5 ). It can be seen from the figure that when using the CGC method, the yield estimation model with Jiangsu Province in 2022 as the reference year obtained the highest accuracy (in 2019: R 2 = 0.66, in 2020: R 2 = 0.65, and in 2021: R 2 = 0.74), which is higher than the models of other reference years. For Wuchang City, Heilongjiang Province, the model with 2020 as the reference year (in 2019: R 2 = 0.56, in 2021: R 2 ) has a higher yield estimation accuracy in the independent year than the models with 2019 (in 2020: R 2 = 0.48, in 2021: R 2 = 0.55) and 2021 (in 2019: R 2 = 0.53, in 2020: R2 = 0.47) as the yield estimation model for the reference year;

[0104] Step 4-5: Apply the coefficient of determination R 2 , root mean square error RMSE, and relative root mean square error RRMSE to evaluate the verification results;

[0105] Step 4-6: Draw the yield estimation model for Jiangsu Province with 2022 as the reference year, and the spatial distribution map of rice yields at the field level in 2021 estimated with and without the CGC method ( Figure 6 ). Generally speaking, the spatial pattern of the 2021 yield estimated by the CGC method is basically the same as that of the yield estimated without the CGC method within the scope of Jiangsu Province. In addition, the spatial distribution map of rice yields in Wuchang City in 2021 (6.0 - 8.8 t / ha) estimated by the CGC method usually shows higher yields than the spatial distribution map of the 2021 yield (6.0 - 8.3 t / ha) estimated without the CGC method ( Figure 7 ).

[0106] The above are only some embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements can be made, and these improvements should be regarded as the protection scope of the present invention.

Claims

1. A high-precision prediction method for rice yield during the growing season based on remote sensing products of gross primary productivity, characterized in that It includes the following steps: Step 1: Obtain the time-series Sentinel-2 satellite images of the rice growing season in the study area, preprocess the images on the remote sensing cloud platform respectively, combine the spatio-temporal interpolation technology to obtain the daily Sentinel-2 images, calculate the daily total primary productivity GPP and the normalized red-edge vegetation index NDRE of rice, calculate the average and cumulative values of GPP and NDRE respectively for each month in the rice growing season, and conduct Pearson correlation analysis with the rice yield; Step 2: Based on the daily high-resolution GPP products of the time series, use the ordinary least squares regression and the discrete Fourier transform HANTS method to perform pixel-by-pixel fitting to eliminate the influence of high-frequency noise; Then, calculate the harmonic fitting GPP data from the original high-resolution GPP data in the study area through this method; calculate the cumulative values of the original GPP and the harmonic fitting GPP every 10 days respectively during the rice growing season; use the original and harmonic fitting GPP data with a 10-day time interval during the rice growing season in the study area to conduct linear regression with the yield to determine the optimal time window for yield estimation; Step 3: Based on the cumulative fitting high-resolution GPP and NDRE data within the best 10-day window, use the Transformer architecture of the transfer learning algorithm to predict the mid-season rice yield, and conduct the random split method and the leave-one-year independent validation accuracy evaluation to clarify the differences between the high-resolution GPP and the vegetation index NDRE in rice yield prediction; Step 4: Propose an inter-annual corrected GPP method based on the triangular transformation principle technology to achieve inter-annual yield estimation: Step 4-1: Extract the maximum value of the harmonic fitting GPP within the rice growing season of different ecological regions; Step 4-2: For the measured rice yield dataset and the maximum value of the harmonic fitting GPP within the provincial scope in the previous n years, take 1 year of data as the independent validation set V; the remaining dataset D within the provincial scope; Step 4-3: Based on the triangular transformation principle technology, perform inter-annual correction on the harmonic fitting GPP and construct the inter-annual migration correction method CGC; Step 4-4: Based on the CGC method and the training set D, achieve mid-season yield estimation and conduct mid-season rice yield prediction in different ecological regions.

2. The method according to claim 1, characterized in that Step 1 specifically includes: Step 1-1: Obtain the rice planting distribution and rice growth process information in the study area; Step 1-2: Use the rice growth process information in the study area to extract the Sentinel-2 time-series images of the rice growing season; Step 1-3: Preprocess and cloud mask the time-series Sentinel-2 images, and use the spatio-temporal interpolation technology to obtain the daily Sentinel-2 images; Step 1-4: Use the daily Sentinel-2 images to calculate the total primary productivity GPP and the normalized red-edge vegetation index NDRE, and the specific formulas are as follows; GPP = (ε msu × APAR su + ε msh × APAR sh ) × g(T a ) × Min(g(VPD), g(CO2)) where ε msu represents the maximum light use efficiency of sunlit leaves, APAR su represents the absorbed photosynthetically active radiation of sunlit leaves, ε msh represents the maximum light use efficiency of shaded leaves, APAR sh represents the absorbed photosynthetically active radiation of shaded leaves, g(CO2) represents the carbon dioxide limitation factor, g(VPD) represents the water limitation factor, g(T a ) represents the air temperature limitation factor, ρ NIR and ρ Red-edge represent the reflectances of Sentinel-2 bands 8 and 6, respectively, where band 8 is 785 - 900 nm and band 6 is 733 - 748 nm; Step 1-5: Calculate the average and cumulative values of GPP and NDRE for each month in the rice growing season; Step 1-6: Conduct Pearson correlation analysis on the average and cumulative values of GPP and NDRE with the rice yield respectively to clarify the differences of different remote sensing variables in rice yield estimation.

3. The method according to claim 1, wherein Step 2 specifically includes: Step 2-1: Obtain the daily GPP data during the rice growing season according to the growth process information of the study area; Step 2-2: Use ordinary least squares regression and the discrete Fourier transform HANTS method for pixel-by-pixel fitting to obtain the harmonic fitting GPP data. The specific calculation formula is as follows; Phase=arctan(A i / B i ) where f(t) is the GPP varying with time, Amplitude represents the amplitude, and Phase represents the phase; t is the observation date, expressed as the fraction of the calendar year, ranging from 0 to 1; n is the number of harmonic terms, ω is the frequency, A i and B i are the cosine coefficient and sine coefficient of the i-th harmonic term respectively; C is the intercept coefficient; Step 2-3: Calculate the original and harmonic fitting GPP at 10-day intervals during the rice growing season in the study area to obtain the original GPP values and cumulative values every 10 days, as well as the harmonic fitting GPP values and cumulative values every 10 days; Step 2-4: Perform linear regression on the original GPP and the fitted GPP data at 10-day intervals respectively with the rice yield; Step 2-5: Apply the square of the correlation coefficient r 2 Perform a comparative analysis on the regression results for each 10-day interval to determine the optimal 10-day interval window within the rice growing season.

4. The method according to claim 1, characterized in that, The specific steps in Step 3 include: Step 3-1: Perform harmonic fitting calculation on the daily normalized red-edge vegetation index NDRE during the rice growing season to obtain the fitted NDRE; Step 3-2: Extract the cumulative values of GPP and NDRE obtained in Step 2 into the transfer learning Transformer framework respectively in the optimal time window obtained. Use random splitting of 7:3 and independent year verification to obtain the mid-season yield prediction accuracy and compare and evaluate; Step 3-3: Input the above two types of data into the transfer learning Transformer framework respectively to clarify the differences between the gross primary productivity GPP and the vegetation index NDRE in rice yield prediction, and determine the optimal indicators for inter-annual yield prediction; Step 3-4: In the measured rice yield dataset within the provincial scope in the previous n years, take the measured data of 1 year as the independent verification set V; the remaining measured rice yield dataset within the provincial scope D; Step 3-5: Evaluate the verification results respectively based on the test set T and the independent validation set V using the coefficient of determination R 2 , root mean square error RMSE, and relative root mean square error RRMSE.

5. The method according to claim 1, characterized in that, The specific steps of Step 4-3 are as follows; First, calculate the amplitudes of the reference and target fitting GPP curves at any time during the two-year rice growing season. According to the reference and target fitting GPP curves, use the formula to calculate the target-corrected fitting GPP curve; Second, determine the scale factor sf, that is, the ratio of the target GPP to the reference GPP for the single or cumulative value within any two phenological periods; Third, use the formula to construct a reference annual yield estimation model, and select the maximum value of GPP during the rice growing season as the dependent variable of the yield estimation model; Fourth, use the formula to predict the yield of the independent year: GPP Target_corrected = GPP Target ÷ Amplitude Target × Amplitude Reference (4 - 4) Y Reference = A × X Reference + B; (4 - 7) Y Target = a × X Target_corrected + b; a = A × sf; b = B × sf (4-8) In the formula, GPP Target and GPP Target_corrected represent the target curve GPP and its corrected value respectively; Amplitude Target and Amplitude Reference represent the amplitudes of the target curve and the reference curve respectively; GPP Target_max and GPP Reference_max represent the maximum GPP values of the target curve and the reference curve respectively; A, B, a, and b are model coefficients; X Reference , Y Reference , X Target_corrected and Y Target represent the modeling variables and yields of the reference year, as well as the modeling variables and yields of the target year respectively.

Images