Winter wheat yield estimation method based on key growth period spectral feature index combination

Abstract

The present application belongs to the technical field of agricultural remote sensing and crop yield estimation, and specifically discloses a winter wheat yield estimation method based on a combination of spectral characteristic indexes in key growth periods. The method is based on winter wheat canopy hyperspectral reflectance data, and first constructs spectral characteristic indexes FVI and a spectral characteristic index set. Then, the core indexes related to yield estimation are obtained through progressive screening, and the index combination optimization is carried out in each growth period respectively, and further growth period combination optimization is carried out, so as to determine the target key growth period combination for yield estimation. On this basis, a cross-period dynamic characteristic is further constructed, and a two-stage XGBoost winter wheat yield estimation model oriented to key growth period information utilization is established, so as to realize accurate estimation of winter wheat yield.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing and crop yield estimation technology, and specifically relates to a method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods. Background Technology

[0002] Winter wheat yield estimation technology is a method for obtaining winter wheat yield based on crop growth information, which can provide data support for field management, agricultural production decision-making, and food security. Existing winter wheat yield estimation methods mainly include field measurement methods, crop model methods, and remote sensing estimation methods. Among them, remote sensing estimation methods have the advantages of large monitoring range, fast acquisition speed, and repeatable observation, and have been widely used in winter wheat growth monitoring, yield prediction, and agricultural management.

[0003] In recent years, with the development of hyperspectral remote sensing technology, extracting spectral features from winter wheat canopy hyperspectral reflectance data and establishing yield estimation models has become an important technical approach for winter wheat yield estimation. However, in the process of winter wheat remote sensing estimation, winter wheat yield is not a direct response of the canopy spectrum, but is formed by the combined processes of photosynthesis, dry matter accumulation, and nutrient translocation at different growth stages. Since different growth stages contribute differently to the final yield, spectral information from a single growth stage cannot fully reflect the key growth information in the yield formation process, while directly using a large amount of spectral data from multiple stages easily introduces redundant variables, leading to a decrease in model stability. At the same time, hyperspectral data has characteristics such as multiple bands, high dimensionality, and strong correlation between adjacent bands. Directly modeling with these data can easily lead to problems such as variable redundancy, model complexity, and overfitting, thus affecting the accuracy of yield estimation. Current winter wheat yield estimation techniques mainly include two categories:

[0004] 1. Estimation methods based on the entire growth period or a single growth period directly utilize spectral data of winter wheat throughout its entire growth period, or select spectral data from only one growth period to extract vegetation indices or related spectral features and establish a yield estimation model. These methods tend to introduce redundant information when using data from the entire growth period, increasing model complexity. Conversely, using data from only one growth period fails to comprehensively reflect the key information at different stages of winter wheat yield formation, thus limiting estimation stability and accuracy. The reason for these problems lies in the fact that winter wheat yield formation is influenced by multiple growth periods, with varying roles at each stage. Existing methods utilize growth period information inappropriately, failing to simultaneously ensure both completeness and specificity.

[0005] 2. Estimation methods based on key growth period screening first identify key growth periods with significant impact on yield formation from multiple growth stages of winter wheat, then extract vegetation indices or spectral features for the corresponding periods, and establish yield estimation models. This type of method can reduce redundant information and improve model specificity to some extent. However, existing schemes mostly use traditional vegetation indices or general spectral features. While these features can reflect some canopy information, they are insufficient in representing the sensitive band information of key growth periods, making it difficult to fully characterize the impact of canopy spectral changes on yield formation and further improve the yield estimation capability of winter wheat. Furthermore, some existing schemes integrate multiple input features such as LAI, SPAD, or phenological information. Although this enhances the ability to describe the crop growth process, it also increases the complexity of data acquisition and model construction, hindering the stable application of the model under different years and conditions. The reason for these problems is that winter wheat yield estimation is affected by multiple factors. Existing methods often improve accuracy by increasing input information, but this also increases the difficulty of model application.

[0006] In summary, while machine learning methods have been introduced into winter wheat yield estimation research, improving yield estimation capabilities by learning the relationship between spectral features and yield is a significant advancement, existing models generally suffer from high input feature dimensionality, model complexity, and insufficient generalization ability. Their stability remains to be improved, particularly under different years, regions, and growth conditions. Furthermore, current techniques often rely on traditional vegetation indices or general spectral features to build estimation models, failing to fully utilize sensitive spectral information from key growth stages, thus impacting the accuracy of winter wheat yield estimation.

[0007] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0008] The purpose of this invention is to propose a method for estimating winter wheat yield based on a combination of spectral characteristic indices for key growth periods. This method uses winter wheat canopy hyperspectral reflectance data as a foundation. First, it constructs a spectral characteristic index (FVI) and a set of spectral characteristic indices. Then, it obtains core indices related to yield estimation through progressive screening, and optimizes index combinations for each growth period. Further optimization of growth period combinations determines the target key growth period combination for yield estimation. Based on this, it constructs inter-period dynamic features and establishes a two-stage XGBoost winter wheat yield estimation model oriented towards utilizing key growth period information, ultimately facilitating accurate estimation of winter wheat yield.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The method for estimating winter wheat yield based on the combination of spectral characteristic indices during key growth periods includes the following steps:

[0011] Step 1. Obtain canopy spectral data during the key growth stages of winter wheat and calculate canopy reflectance data;

[0012] Step 2. Calculate the set of spectral characteristic indices for key reproductive periods based on canopy reflectance data;

[0013] Step 3. Perform data cleaning and scale matching preprocessing on the spectral feature index set obtained in Step 2;

[0014] Step 4. Perform progressive screening on the spectral characteristic indices obtained after preprocessing in Step 3 to obtain the core index set;

[0015] Step 5. Based on the core index set obtained in Step 4, index combinations are optimized at the jointing stage, booting stage, heading and flowering stage, and grain-filling stage to determine the optimal index combination for each growth stage;

[0016] Step 6. Based on the optimal index combination for each growth period determined in Step 5, further optimize the growth period combination to determine the target key growth period combination for winter wheat yield estimation.

[0017] Step 7. Based on the target key growth period combination determined in Step 6, extract the optimal spectral features of the corresponding growth period, and further construct inter-period dynamic features to form a yield estimation feature set for winter wheat yield estimation.

[0018] Step 8. Based on the yield estimation feature set formed in Step 7, establish a two-stage winter wheat yield estimation model based on the XGBoost model, which includes a first-stage basic yield prediction model and a second-stage subsequent correction model.

[0019] The first-stage basic yield prediction model uses static spectral characteristics during the booting stage to generate basic yield prediction results.

[0020] The second-stage follow-up correction model uses static spectral characteristics during the grain-filling period and dynamic characteristics across periods to correct the output of the first-stage basic yield prediction model, thus obtaining the final estimate of winter wheat yield.

[0021] The present invention has the following advantages:

[0022] As described above, this invention discloses a method for estimating winter wheat yield based on a combination of spectral characteristic indices for key growth stages. This method first constructs a spectral characteristic index (FVI) based on winter wheat canopy reflectance data to characterize the red-far-red difference in fluorescence-sensitive spectral regions. While retaining traditional spectral characteristic information, it improves the ability to characterize information in key sensitive bands, thus enhancing the model's responsiveness to physiological information related to winter wheat yield formation. This invention constructs a set of spectral characteristic indices and obtains core indices for winter wheat yield estimation through progressive screening steps such as spatial discrimination tests, observation reliability tests, yield correlation tests, and collinearity redundancy removal. This reduces the interference of invalid and redundant features on the model input, improving the effectiveness and consistency of the model's input features. Furthermore, this invention does not directly use single growth stage or full growth stage features for modeling. Instead, it first optimizes index combinations within each growth stage, and then further optimizes growth stage combinations to determine the target key growth stage combination for yield estimation. This technique reduces redundant information caused by directly using full growth stage data for modeling and avoids the problem of insufficient information from a single growth stage, thus improving the targeting of winter wheat yield estimation. Furthermore, this invention constructs intertemporal dynamic features based on representative values ​​of the indices corresponding to the target key growth stages, used to characterize the changes in canopy spectral response between different key growth stages. Compared with methods using only static spectral features, this technique can supplement the information reflecting the stage changes of winter wheat from the booting stage to the grain-filling stage, which is beneficial to improving the characterization ability of winter wheat yield formation process. Finally, based on the constructed yield estimation feature set, this invention establishes a two-stage winter wheat yield estimation model based on the XGBoost model, which includes a first-stage basic yield prediction model and a second-stage subsequent correction model. The first-stage basic yield prediction model uses static spectral features of the booting stage to form the basic yield prediction result; the second-stage subsequent correction model uses static spectral features of the grain-filling stage and intertemporal dynamic features to correct the output result of the first-stage basic yield prediction model, obtaining the final winter wheat yield estimation result. By constructing a two-stage modeling approach, this invention not only corresponds to the staged process of winter wheat yield formation, but also makes the information of different key growth stages have a clearer division of labor in the model, which is beneficial to improving the targeting of key growth stage information utilization. In summary, this invention forms a complete technical solution from canopy spectral data acquisition, spectral feature index construction, core index screening, indices combination optimization within the growth period, growth period combination optimization, cross-period dynamic feature construction to two-stage yield estimation output, which effectively achieves accurate estimation of winter wheat yield. Attached Figure Description

[0023] Figure 1 This is a flowchart of the winter wheat yield estimation method based on the combination of spectral characteristic indices of key growth periods in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the coefficient of variation (CV) results in an embodiment of the present invention;

[0025] Figure 3 This is a graph showing the observation reliability test results in an embodiment of the present invention; wherein... Figure 3 In the figure, (a) and (b) represent the intragroup correlation coefficients of the spectral characteristic indices, respectively. and signal-to-noise ratio ;

[0026] Figure 4 This is a graph showing the partial correlation analysis results in an embodiment of the present invention;

[0027] Figure 5 This is a diagram showing the collinearity redundancy results in an embodiment of the present invention. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0029] Example 1

[0030] To address the problems in existing winter wheat yield estimation methods, such as insufficient representation of key sensitive band information by traditional vegetation indices, large input redundancy in modeling throughout the entire growth period, insufficient utilization of features at different growth stages, and the difficulty of existing models in simultaneously considering the effectiveness of input features and estimation stability, this invention proposes a winter wheat yield estimation method based on a combination of spectral feature indices for key growth stages. This method first constructs a spectral characteristic index (FVI) based on winter wheat canopy reflectance data to characterize the red-far-red difference in the fluorescence-sensitive spectral region. Then, it constructs a set of spectral characteristic indices and obtains core indices relevant to yield estimation through progressive screening. Based on this, it optimizes index combinations for each growth stage to determine the optimal combination for each growth stage, and further optimizes growth stage combinations to determine the target key growth stage combination for yield estimation. Subsequently, it constructs inter-period dynamic features based on the determined target key growth stage combination to characterize the stage-specific changes between key growth stages. Finally, based on the static spectral characteristics and inter-period dynamic features of the target key growth stages, it establishes a two-stage XGBoost winter wheat yield estimation model for utilizing key growth stage information. This forms a complete technical solution from canopy spectral acquisition, feature construction, index screening, index combination optimization, growth stage combination optimization, inter-period dynamic feature construction to yield estimation output, ultimately improving the accuracy, stability, and operability of winter wheat yield estimation in the field.

[0031] like Figure 1 As shown, the method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods includes the following steps:

[0032] Step 1. Obtain canopy spectral data during the key growth stages of winter wheat and calculate the canopy reflectance data.

[0033] During the key growth stages of winter wheat, canopy spectral data were collected from various plots in the target area, and the measured yields of the corresponding plots were obtained. The key growth stages of winter wheat include the jointing stage, booting stage, heading and flowering stage, and grain-filling stage.

[0034] Canopy spectral data were acquired using a portable ground object spectrometer (e.g., RS-8800).

[0035] During the spectral acquisition process, the spectrometer probe is placed at a preset height (1.5m) vertically above the sample plot. After the spectral measurement of each sample plot is completed, reflectance correction is performed using a whiteboard.

[0036] Multiple canopy spectral data points were collected for each sample plot. After quality screening, the mean value was calculated to obtain the canopy reflectance data for that sample plot. In this embodiment, for example, nine canopy spectral data points were collected for each sample plot; however, this is merely an example.

[0037] Step 2. Calculate the set of spectral characteristic indices for key reproductive periods based on canopy reflectance data.

[0038] Based on the canopy reflectance data obtained in step 1, a set of spectral characteristic indices is constructed. This set comprises three main categories: physiological indices, structural indices, and greenness indices, all based on canopy reflectance. Physiological indices include FVI; structural indices include Ratio Vegetation Index (RVI), Soil-Adjusted Vegetation Index (SAVI), Fraction of Absorbed Photosynthetically Active Radiation (FPAR), and Near-Infrared Reflectance of Vegetation (NIRv); and greenness indices include Normalized Difference Vegetation Index (NDVI), Red-edge Normalized Difference Vegetation Index (RENDVI), and Two-band Enhanced Vegetation Index (EVI2).

[0039] The calculation formulas for RVI, SAVI, FPAR, NIRv, NDVI, RENDVI, and EVI2 are as follows:

[0040] , , , ;

[0041] , , .

[0042] Represents the reflectivity of the red light band at 660nm; Represents the reflectivity at 800nm ​​in the near-infrared band; Represents the reflectivity of the 710nm red-edge band; and This represents the specific narrowband reflectance at 760 nm and 685 nm.

[0043] The soil background adjustment parameter is set to 0.5 in this embodiment; and This is the FPAR empirical conversion factor, used to convert the corresponding vegetation index into the photosynthetically active radiation absorption ratio. , .

[0044] In addition, this invention also constructs a spectral characteristic index for characterizing key sensitive spectral regions of winter wheat, namely the Fluorescence-sensitive Vegetation Index (FVI).

[0045] The spectral characteristic index FVI, a newly added index in this invention, is used to characterize the red-far-red difference in the fluorescence-sensitive spectral region. FVI is included in the spectral characteristic index set as a physiologically sensitive characteristic during the key growth stages of winter wheat; its calculation formula is as follows:

[0046] .

[0047] In the formula The canopy reflectance at 760 nm The canopy reflectance is at 685 nm; 685 nm is located in the red fluorescence sensitive region, and 760 nm is located in the far-red fluorescence sensitive region and is close to the O2-A absorption band.

[0048] Based on winter wheat canopy reflectance data, this invention selects two fluorescence-sensitive bands, 685nm and 760nm, and innovatively constructs the spectral characteristic index FVI to characterize the red-far-red difference in the fluorescence-sensitive spectral region.

[0049] The spectral characteristic index FVI improves the ability to characterize sensitive band information related to winter wheat yield formation.

[0050] Step 3. Perform data cleaning and scale matching preprocessing on the spectral feature index set obtained in Step 2.

[0051] The spectral characteristic index data obtained in step 2 are preprocessed to improve the quality of the modeling data.

[0052] The preprocessing includes: removing outliers by combining physical boundary screening with Hampel filtering; filling in missing data points using linear interpolation; and averaging the spectral feature index time series data on a daily scale.

[0053] Hampel filtering is a filtering method that identifies outlier observations based on the sliding window midpoint and absolute deviation.

[0054] Step 4. To obtain effective input features for winter wheat yield estimation, the spectral feature indices obtained after preprocessing in Step 3 are progressively screened to obtain a core index set. In this embodiment, the spectral feature indices include eight indices: FVI, RVI, SAVI, FPAR, NIRv, NDVI, RENDVI, and EVI2.

[0055] The process of progressively screening spectral characteristic indices includes the following steps:

[0056] I. Spatial Discrimination Test: This test calculates the degree of variability among plots in spectral characteristic indices over the entire growth period, eliminating spectral characteristic indices with weak spatial discrimination capabilities. The formula is as follows:

[0057] .

[0058] In the formula, The standard deviation of a specific feature across all observed plots; This is the arithmetic mean of the plots with this characteristic. Step I can eliminate characteristics with extremely low coefficients of variation (CV), retaining only variables with spatial differentiation potential.

[0059] like Figure 2 As shown, after comparing the degree of variation among plots throughout the entire growth period of the eight spectral characteristic indices, it was found that the NDVI had the lowest coefficient of variation and relatively weak spatial discrimination, indicating that its ability to respond to yield differences among different plots was relatively insufficient.

[0060] Therefore, NDVI is removed in step I. After the spatial discrimination test in step I, seven spectral characteristic indices, namely FVI, RVI, SAVI, FPAR, NIRv, RENDVI and EVI2, are retained and proceed to the next step of reliability test.

[0061] II. Verification of observation reliability: Calculate the intraclass correlation coefficient and signal-to-noise ratio of the spectral characteristic indices. Indices significantly affected by environmental noise and random observation errors are excluded; the calculation formula is: , .

[0062] ICC stands for Intra-group Correlation Coefficient, which characterizes the proportion of inter-plot differences in the total variation; SNR stands for Signal-to-Noise Ratio, which characterizes the strength of effective inter-plot differences relative to random observation noise.

[0063] In the formula The correlation coefficient within a single observation group. Characterize the effective signal variance caused by differences in crop biology among plots; It characterizes the variance of random noise caused by short-term environmental fluctuations and observation errors.

[0064] The intra-group correlation coefficient is used to see the true differences of the same indicator between different plots, how much of it is a "valid signal" and how much is overwhelmed by daily fluctuations and observation errors within the plot.

[0065] This step eliminates variables dominated by environmental noise, retaining those with high observational stability.

[0066] Figure 3 The following diagram illustrates the results of the observation reliability test in an embodiment of the present invention; wherein... Figure 3 In the figure, (a) and (b) represent the intragroup correlation coefficients of the spectral characteristic indices of locations A and B, and the two locations after pooling, respectively. and signal-to-noise ratio contrast.

[0067] The results show that, after reliability testing, FVI, RVI, SAVI, FPAR, NIRv, RENDVI, and EVI2 all meet the reliability requirements, indicating that the above indices have good observation stability over time series.

[0068] Therefore, no further indexes were removed in step II, and the above 7 indices continued to the next step of partial correlation analysis.

[0069] III. Partial Correlation Analysis: Analyze the correlation between spectral characteristic indices and measured yield, and preferentially retain spectral characteristic indices that are significantly correlated with yield; the calculation formula is:

[0070] .

[0071] In the formula, and The first The residuals of characteristic X and output Y of a sample after removing the influence of control variable Z; and is the mean of the corresponding residuals, and n represents the number of samples participating in the partial correlation analysis. The partial correlation coefficient between characteristic X and output Y after controlling variable Z is expressed. It reflects the strength and direction of the correlation.

[0072] Feature X refers to a specific spectral characteristic index, such as one of SIF, FVI, RVI, RENDVI, etc. Yield Y refers to the measured yield of winter wheat. Control variable Z refers to the variable that needs to be controlled in the partial correlation analysis.

[0073] Z is not a single variable, but a matrix of control variables, specifically including the region and year. Z is not directly written in the correlation coefficient formula on the right, but is implicit in the residuals. and In other words, we first perform regression on X against Z to obtain the residuals. Then perform regression on Y against Z to obtain the residuals. Finally, the Pearson correlation coefficient between the two sets of residuals was calculated.

[0074] Step III introduces the Benjamini-Hochberg false discovery rate (FDR) for multiple test correction, retaining only the characteristic of significant correlation (q<0.05) after correction. q represents the significance level of the p-value of the multiple test after Benjamini-Hochberg correction; it reflects statistical significance, not the strength of the correlation.

[0075] like Figure 4 As shown in this embodiment, after comparing the correlation of the spectral characteristic indices retained in step II, it was found that the spectral characteristic indices and the measured yield of winter wheat both maintained a high correlation.

[0076] Among them, RVI has the highest correlation with yield, followed by RENDVI, and NIRv, FVI, FPAR, EVI2, and SAVI decrease in that order. This indicates that the above spectral characteristic indices all have a certain ability to characterize yield. Therefore, this step does not further eliminate indices, but uses the strength of the correlation as an important basis for determining representative indices in the next step of collinearity removal and redundancy screening.

[0077] IV. Collinearity Removal: To avoid high overlap of information between spectral feature indices leading to model input redundancy, the spectral feature indices retained in step III are further screened for collinearity removal.

[0078] The correlation between spectral characteristic indices is preferably evaluated using the Pearson correlation coefficient matrix, as shown in the formula:

[0079] .

[0080] In the formula, and The first Spectral characteristic index of each sample With spectral characteristic index Observed values; and This is the mean of the corresponding index. The Pearson correlation coefficient represents the relationship between spectral characteristic indices i and j, reflecting the degree of information overlap between the two candidate indices. This is particularly relevant for indices exhibiting extreme collinearity (e.g., ...). The index cluster ("index cluster" refers to a set of spectral feature indices that are highly correlated and have highly overlapping information) follows the principle of information redundancy removal and physical meaning complementarity, retaining only the single representative index that is most strongly correlated with output.

[0081] like Figure 5 As shown in this embodiment, there is a clear collinearity among the spectral characteristic indices, with FVI and FPAR being completely collinear, and their correlation coefficient reaching r=1.00, indicating that the information they represent is highly overlapping.

[0082] Based on the partial correlation analysis results in step III, FVI is more correlated with yield than FPAR. In addition, FVI is directly constructed from the two sensitive bands of 685nm and 760nm, and its physical meaning is clearer. Therefore, FVI is retained and FPAR is removed in this collinear group.

[0083] NIRv, EVI2, and SAVI form a highly collinear cluster, with pairwise correlation coefficients reaching or approaching r=0.99-1.00. Meanwhile, RENDVI also maintains a high correlation with each index in this collinear cluster, with a correlation coefficient of approximately r=0.98-0.99, indicating that these indices have significant information overlap in characterizing canopy structure and greenness changes.

[0084] Based on the results of the correlation test with yield in step III, RENDVI has a higher correlation with yield than NIRv, EVI2, and SAVI. Therefore, RENDVI is retained in this group, while NIRv, EVI2, and SAVI are removed.

[0085] Although RVI shows a high correlation with the other spectral characteristic indices, its correlation coefficient is generally lower than that of the extremely collinear clusters mentioned above, indicating that it still has a certain degree of independence. Combining the results of the correlation test with yield in step III, RVI has the highest correlation with yield, and as a structural index, it has good complementarity with FVI and RENDVI in terms of physical meaning, therefore it is retained.

[0086] After collinearity removal and redundancy screening in step IV, FVI, RVI, and RENDVI were finally determined as the core index set for winter wheat yield estimation, and used for subsequent selection of index combinations within the growth period and selection of growth period combinations.

[0087] This invention establishes a progressive core index screening method based on a spectral feature index set. The spectral feature index set is constructed and sequentially screened through spatial discrimination test, observation reliability test, yield correlation test, and collinearity redundancy removal to obtain core indices for winter wheat yield estimation, thereby reducing the interference of invalid and redundant features on the model input.

[0088] Step 5. Based on the core index set obtained in Step 4, index combinations are optimized at the jointing stage, booting stage, heading and flowering stage, and grain-filling stage to determine the optimal index combination for each growth stage.

[0089] In this embodiment, the core indices retained in step 4 are FVI, RVI, and RENDVI.

[0090] Therefore, within each growth period, different index combination schemes were constructed using FVI, RVI, and RENDVI as candidate input indices. Each combination scheme was then used as an input variable in a ridge regression model, with the measured yield of winter wheat as the output variable, to evaluate the yield representation ability of different combination schemes.

[0091] The candidate combinations include: (1) single-index schemes: FVI, RVI, and RENDVI; (2) dual-index combination schemes: RVI and FVI combination, RVI and RENDVI combination, and RENDVI and FVI combination; (3) triple-index combination schemes: RVI, RENDVI, and FVI combined. That is, a total of 7 candidate index combination schemes are constructed in each reproductive period, and modeling and evaluation are carried out respectively.

[0092] Preferably, ridge regression, by introducing an L2 regularization term into the loss function, can mitigate the impact of the correlation between input exponents on model stability under small sample conditions. Its objective function is:

[0093] .

[0094] In the formula For loss function, For the sample size, For the first The measured yield of a sample For the first Each sample corresponds to an input vector of candidate index combinations. For the first The regression coefficients of each input feature, where p is the dimension of the input feature. This is the regularization parameter.

[0095] For each reproductive period, ridge regression models were established for the seven candidate index combinations mentioned above, and the correlation coefficient r, significance level p, and determination coefficient R were used to determine the appropriate regression results. 2 And a comprehensive comparison of different combination schemes based on the input feature dimensions.

[0096] Preferably, when the model fitting results are similar, the combination with fewer input features and a simpler model structure is selected as the optimal index combination for the reproductive period to reduce model complexity and input redundancy.

[0097] Table 1 presents a comparison of the performance of index combinations during the single-fertility period.

[0098] Table 1 Performance of Index Combinations During the Single-Fertility Period

[0099]

[0100] As shown in Table 1, in this embodiment, the optimal index combination determined after selecting the combination of each reproductive period index is as follows:

[0101] (1) During the jointing stage, the combination of three exponential RVI, RENDVI and FVI has the best fitting effect, but the combination of double exponential RENDVI and FVI has a similar fitting effect, while the number of input features is less.

[0102] Therefore, the optimal index combination for the jointing stage is determined by the combination of the dual-index RENDVI and FVI.

[0103] (2) During the spikelet incubation period, the combination of three-index RVI, RENDVI and FVI has the best fitting effect, but the combination of two-index RENDVI and FVI has a similar fitting effect and a simpler model structure.

[0104] Therefore, the optimal index combination during the booting stage is determined to be the combination of the two indices RENDVI and FVI.

[0105] (3) During the heading and flowering stage, the fitting effect of the combination of three-exponential RVI, RENDVI, and FVI is close to that of the combination of two-exponential RVI and RENDVI. This invention comprehensively considers the influence of model accuracy and input feature dimension.

[0106] Therefore, the optimal index combination for the heading and flowering period is determined by the combination of the dual-index RVI and RENDVI.

[0107] (4) During the grouting period, the fitting effect of the combination of three exponential RVI, RENDVI, and FVI is close to that of the combination of two exponential RVI and FVI. This invention comprehensively considers model accuracy and input feature dimensions.

[0108] Therefore, the optimal index combination during the grouting period is determined by the combination of the dual-index RVI and FVI.

[0109] After step 5, the optimal index combinations for each growth stage are as follows: jointing stage: RENDVI and FVI combination; booting stage: RENDVI and FVI combination; heading and flowering stage: RVI and RENDVI combination; grain filling stage: RVI and FVI combination.

[0110] The optimal index combination for each of the above reproductive periods serves as the input basis for the next step of optimizing the reproductive period combination.

[0111] Step 6. Based on the optimal index combination for each growth period determined in Step 5, further optimize the growth period combination to determine the target key growth period combination for winter wheat yield estimation.

[0112] This invention uses the optimal index combination for each growth period as a fixed input unit to construct different growth period combination schemes and compares the ability of each scheme to represent the measured yield of winter wheat.

[0113] The reproductive period combination schemes include single reproductive period combinations, double reproductive period combinations, triple reproductive period combinations, and quadruple reproductive period combinations. The input variables for each combination scheme are composed of the optimal index combination for the corresponding reproductive period.

[0114] Preferably, ridge regression models are established for all combinations of the jointing stage, booting stage, heading and flowering stage, and grain-filling stage, and the results are calculated based on the correlation coefficient r, significance level p, and coefficient of determination R. 2 The root mean square error (RMSE) and input feature dimensions are used to comprehensively compare different fertility period combinations. Here, r represents the correlation coefficient between predicted and measured yields, p represents the probability value of the correlation significance test, R² represents the coefficient of determination, used to characterize the model's explanatory power for yield variation, and RMSE represents the root mean square error (RMSE), used to characterize the average deviation level between model predicted and measured yields.

[0115] In the process of optimal combination, priority is given to the balance between model fitting ability and input dimension.

[0116] When the fitting effects of different combinations of reproductive periods are similar, the preferred option is the one with fewer input features, fewer reproductive periods, and a simpler model structure, in order to reduce input redundancy and improve the targeting of subsequent model construction.

[0117] As shown in Table 2, in this embodiment, the multi-fertility period combination is generally superior to the single-fertility period combination.

[0118] Among them, the four-stage growth combination (a combination of jointing stage, booting stage, heading and flowering stage, and grain filling stage) and some three-stage growth combinations have high fitting accuracy, but their input dimensions are relatively high.

[0119] In contrast, the combination of the booting and grain-filling stages during the dual-growth period results in higher r and R values. 2 It is close to the three- and four-fertility combinations in terms of metrics such as RMSE, while having lower input dimensionality and a simpler model structure.

[0120] After comprehensively considering model accuracy, error level, and input feature dimensions, this invention ultimately determines that the booting stage and grain-filling stage are combined as the target key growth period combination for winter wheat yield estimation.

[0121] After step 6, the target key growth period combination is determined to be the booting stage and the grain-filling stage. The corresponding optimal index combination serves as the input basis for the next step of constructing inter-period dynamic characteristics and establishing a yield estimation model.

[0122] Preferably, the core index screening, index combination optimization, and reproductive period combination optimization in steps 4 to 6 are all completed independently on the training dataset, and the verification dataset does not participate in the above screening and optimization process to avoid information leakage.

[0123] Table 2 Performance of Combined Products During the Reproductive Period

[0124]

[0125] This invention establishes a two-level optimization method combining index combination optimization within the growth period and growth period combination optimization through steps 5 and 6 above. First, different candidate index combinations composed of core indices are optimized within each growth period to determine the optimal index combination for each growth period. Then, using the optimal index combination for each growth period as a fixed input unit, different growth period combination schemes are optimized to determine the target key growth period combination for winter wheat yield estimation, thereby improving the targeting of key growth period information utilization and providing a clear input basis for subsequent model construction.

[0126] Step 7. Based on the target key growth period combination determined in Step 6, extract the optimal spectral features of the corresponding growth period, and further construct inter-period dynamic features to form a yield estimation feature set for winter wheat yield estimation.

[0127] Based on the target key growth period combination determined in step 6, the optimal spectral features of the corresponding growth period are extracted, and cross-period dynamic features are further constructed to form a yield estimation feature set for winter wheat yield estimation.

[0128] In this embodiment, based on the results of steps 5 and 6, step 7 extracts the RENDVI, FVI, RVI and FVI of the heading stage, the grain-filling stage as static spectral features.

[0129] To improve the ability to utilize information on changes in key growth stages of winter wheat in yield estimation, a cross-period dynamic feature is further constructed based on the extraction of the above-mentioned static spectral features.

[0130] Since FVI appears in both the optimal index combination during the booting stage and the optimal index combination during the grain-filling stage, it is preferable to construct a cross-period dynamic feature based on the representative values ​​of FVI during the booting and grain-filling stages.

[0131] Among them, FVI b This represents the representative value of FVI during the fertile growth period at the booting stage. g This represents the representative value of the FVI during the grouting period.

[0132] Preferably, the representative value for the reproductive period is obtained by averaging the effective observations of FVI within the corresponding reproductive period.

[0133] Based on this, the following intertemporal dynamic features are constructed:

[0134] (1) Ratio characteristics:

[0135] ;

[0136] (2) Characteristics of relative rate of change:

[0137] .

[0138] Among them, Q FVI P is used to characterize the relative ratio of FVI during the heading and grain-filling stages. FVI Used to characterize the relative variation of FVI from the heading stage to the grain-filling stage.

[0139] After processing in step 7, the resulting set of asset valuation characteristics includes:

[0140] (1) Static spectral characteristics during the booting stage: RENDVI and FVI during the booting stage; (2) Static spectral characteristics during the grain-filling stage: RVI and FVI during the grain-filling stage; (3) Dynamic characteristics across stages: Q FVI P FVI .

[0141] In constructing the training dataset, a single plot of land was used as a sample unit. For each sample, the RENDVI, FVI during the heading stage, RVI during the heading stage, and FVI during the grain-filling stage were extracted, and Q was further calculated. FVI and P FVI This forms the yield estimation feature set for the sample, and the corresponding output variable is the measured yield Y of winter wheat for the sample.

[0142] This invention establishes a method for constructing dynamic features across target critical growth periods. Based on the representative values ​​of the growth period corresponding to the target critical growth period index, ratio features and relative change rate features are constructed to characterize the change process of canopy spectral response between target critical growth periods, thereby improving the ability to characterize the change information of winter wheat yield formation stages.

[0143] Step 8. Based on the yield estimation feature set formed in Step 7, establish a two-stage winter wheat yield estimation model based on the XGBoost model to achieve winter wheat yield estimation. The two-stage XGBoost winter wheat yield estimation model established in this embodiment includes a first-stage basic yield prediction model and a second-stage subsequent correction model.

[0144] The first-stage basic yield prediction model uses static spectral characteristics during the booting stage to generate basic yield prediction results.

[0145] The second-stage follow-up correction model uses static spectral characteristics during the grain-filling period and dynamic characteristics across periods to correct the output of the first-stage basic yield prediction model, thus obtaining the final estimate of winter wheat yield.

[0146] This invention establishes a two-stage XGBoost winter wheat yield estimation model based on the utilization of information from key growth stages. The model is grounded in the fact that winter wheat yield formation exhibits distinct stages, with different growth stages having varying impacts on the final yield. The booting stage is a crucial period for ear development and canopy productivity formation; the spectral characteristics of this stage effectively reflect crop growth, canopy status, and the foundation for yield formation, making it suitable for generating baseline yield predictions. The grain-filling stage involves the accumulation, translocation, and filling of assimilated products; the spectral characteristics of this stage further reflect the degree to which the previous yield baseline has been achieved. Furthermore, the inter-stage dynamic characteristics between the booting and grain-filling stages characterize the phased changes between key growth stages, making it suitable for subsequent correction of baseline yield predictions.

[0147] Based on the above mechanism, this invention adopts a phased modeling approach: the first phase utilizes the optimal spectral characteristics of the heading stage to establish a basic yield prediction model, which characterizes the foundation for winter wheat yield formation; the second phase utilizes the optimal spectral characteristics of the grain-filling stage and cross-phase dynamic characteristics to establish a subsequent correction model, which characterizes the impact of the later grain filling process on the final yield and corrects the results of the first phase. Compared with inputting different growth stage characteristics into a single model all at once, this approach can utilize the early basic information and the later realization information separately, reducing the information ambiguity caused by simultaneous input of features from different stages, thereby improving the model's ability to characterize the winter wheat yield formation process and further improving the accuracy and stability of winter wheat yield estimation. Therefore, the two-stage modeling approach of this invention not only corresponds to the phased process of winter wheat yield formation but also allows for a clearer division of labor among different key growth stage information in the model, which is beneficial for improving the targeted utilization of key growth stage information.

[0148] I. Establishment of the basic output prediction model in the first stage.

[0149] In the first stage, the RENDVI and FVI of the booting stage in the yield estimation feature set formed in step 7 are used as input variables, and the measured yield of winter wheat is used as the output variable to establish the basic yield prediction model for the first stage.

[0150] Let the input variables for the first stage be denoted as:

[0151] ;

[0152] In the formula, RENDVI b Indicates RENDVI and FVI during the spikelet incubation period. b Let FVI represent the booting stage. Let Y be the measured yield of winter wheat. Then the output of the first-stage basic yield prediction model is the basic yield prediction result, denoted as:

[0153] ;

[0154] In the formula This represents the first stage of the XGBoost model. This indicates the results of the first-stage basic production forecast.

[0155] II. Establishment of the second-stage subsequent correction model.

[0156] In the second stage, the grouting period RVI, grouting period FVI, and Q are included in the production estimation feature set formed in step 7. FVI and P FVI Using the correction between the first-stage baseline yield forecast and the measured yield as input variables, the second-stage subsequent correction model is established.

[0157] In the formula, Indicates the RVI during the grouting period. Indicates the grouting period FVI, Indicates the characteristics of the FVI ratio. This indicates the relative rate of change of FVI.

[0158] During the model training phase, the output variable of the second phase is the correction variable, denoted as ΔY, which is defined as:

[0159] In the formula, Y represents the measured yield of winter wheat, Y1 represents the predicted basic yield in the first stage, and ΔY represents the difference between the measured yield and the predicted basic yield in the first stage.

[0160] Using the correction variable ΔY as the training output variable, a second-stage XGBoost model is established, denoted as:

[0161] ;

[0162] In the formula This indicates the second-stage XGBoost model. This represents the amount of prediction correction output by the second-stage model.

[0163] III. Obtaining the final production estimate.

[0164] Based on the first-stage basic yield forecast results and the prediction corrections from the second-stage model output, the final winter wheat yield estimate is obtained, denoted as:

[0165] In the formula This indicates the final estimate of winter wheat yield.

[0166] In this embodiment, both models in the two-stage XGBoost winter wheat yield estimation model need to be trained separately:

[0167] The first-stage model was used to learn the relationship between static spectral characteristics during the booting stage and winter wheat yield; the second-stage model was used to learn the relationship between static spectral characteristics during the grain-filling stage, inter-stage dynamic characteristics, and yield correction.

[0168] This invention employs a two-stage XGBoost modeling approach, rather than inputting all features into a single model at once. The aim is to utilize different information from the target critical reproductive period in stages.

[0169] The first stage uses static spectral characteristics of the booting stage to form basic yield prediction results. The second stage further uses static spectral characteristics of the grain-filling stage and dynamic characteristics across stages to correct the basic prediction results. In this way, the static characteristic information and stage change information of the target key growth stages are used respectively to improve the pertinence and accuracy of winter wheat yield estimation.

[0170] The first-stage basic yield prediction model and the second-stage subsequent correction model are both trained on the training dataset. The validation dataset is only used to independently validate the performance of the completed model. After processing in step 8, the completed two-stage XGBoost winter wheat yield estimation model can be used for winter wheat yield estimation in step 9.

[0171] After the model training is completed, steps 1 to 7 are repeated for the winter wheat in the test area to obtain the yield estimation feature set of the test plots, and the yield estimation feature set is input into the two-stage XGBoost winter wheat yield estimation model established in step 8.

[0172] The two-stage XGBoost winter wheat yield estimation model outputs the winter wheat yield estimation results for the corresponding sample plots.

[0173] In one specific implementation, a plot of land within the target area is used as the estimation object.

[0174] First, during the identified key growth periods, namely the heading stage and the grain-filling stage, canopy spectral data of the plot were collected, and canopy reflectance and corresponding spectral characteristic values ​​were obtained according to steps 1 to 3.

[0175] Then, based on the optimal index combination and target key growth period combination results determined in this invention, the RENDVI of the booting stage, FVI of the booting stage, RVI of the grain filling stage, and FVI of the grain filling stage of the plot are extracted.

[0176] The intertemporal dynamic characteristic Q is calculated according to step 7. FVI and P FVI This forms the set of valuation characteristics corresponding to the land parcel.

[0177] Subsequently, the RENDVI and FVI of the heading stage of this plot were input into the first-stage basal yield prediction model to obtain the basal yield prediction result Y1 for this plot; then the RVI, FVI, and Q of the grain-filling stage of this plot were input into the first-stage basal yield prediction model to obtain the basal yield prediction result Y1 for this plot; FVI and P FVI Input the second-stage subsequent correction model to obtain the predicted correction amount. .

[0178] Finally, the estimated winter wheat yield for this plot of land was obtained using the following formula. The formula is as follows: .

[0179] Example 2

[0180] This embodiment 2 describes a winter wheat yield estimation system based on a combination of spectral characteristic indices of key growth periods. This system is based on the same inventive concept as the method in embodiment 1 above.

[0181] The winter wheat yield estimation system based on the combination of spectral characteristic indices of key growth stages includes the following modules:

[0182] The canopy reflectance data acquisition module is used to acquire canopy spectral data during key growth stages of winter wheat and calculate canopy reflectance data.

[0183] The spectral characteristic index set calculation module is used to calculate the spectral characteristic index set for key growth periods based on canopy reflectance data;

[0184] The preprocessing module is used to perform data cleaning and scale matching preprocessing on the calculated set of spectral feature indices.

[0185] The core index set construction module is used to progressively filter the spectral feature indices obtained from preprocessing to obtain the core index set.

[0186] The optimal index combination determination module is used to select the optimal index combination for each growth stage based on the obtained core index set, specifically during the jointing stage, booting stage, heading and flowering stage, and grain filling stage.

[0187] The target critical growth period combination determination module is used to further optimize the growth period combination based on the determined optimal index combination of each growth period, so as to determine the target critical growth period combination for winter wheat yield estimation.

[0188] The yield estimation feature set construction module is used to extract the optimal spectral features of the corresponding growth period based on the determined target key growth period combination, and further construct inter-period dynamic features to form a yield estimation feature set for winter wheat yield estimation.

[0189] The prediction module is used to establish a two-stage winter wheat yield estimation model based on the XGBoost model, which includes a first-stage basic yield prediction model and a second-stage subsequent correction model, based on the formed yield estimation feature set.

[0190] The first-stage basic yield prediction model uses static spectral characteristics during the booting stage to generate basic yield prediction results.

[0191] The second-stage follow-up correction model uses static spectral characteristics during the grain-filling period and dynamic characteristics across periods to correct the output of the first-stage basic yield prediction model, thus obtaining the final estimate of winter wheat yield.

[0192] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0193] Example 3

[0194] This embodiment 3 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the winter wheat yield estimation method based on the combination of spectral characteristic indices of key growth periods described in embodiment 1 above.

[0195] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0196] Example 4

[0197] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the winter wheat yield estimation method based on the combination of spectral characteristic indices of key growth periods in embodiment 1.

[0198] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0199] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth stages, characterized in that, Includes the following steps: Step 1. Obtain canopy spectral data during the key growth stages of winter wheat and calculate canopy reflectance data; Step 2. Calculate the set of spectral characteristic indices for key reproductive periods based on canopy reflectance data; The spectral characteristic index (FVI) is used to characterize the red-far-red difference in the fluorescence-sensitive spectral region and is included in the spectral characteristic index set as a physiologically sensitive characteristic of key growth stages of winter wheat; its calculation formula is as follows: ； In the formula The canopy reflectance at 760 nm The canopy reflectance at 685 nm; 685 nm is located in the red fluorescence sensitive region, and 760 nm is located in the far-red fluorescence sensitive region and is close to the O2-A absorption band; Step 3. Perform data cleaning and scale matching preprocessing on the spectral feature index set obtained in Step 2; Step 4. Perform progressive screening on the spectral characteristic indices obtained after preprocessing in Step 3 to obtain the core index set; Step 5. Based on the core index set obtained in Step 4, index combinations are optimized at the jointing stage, booting stage, heading and flowering stage, and grain-filling stage to determine the optimal index combination for each growth stage; Step 6. Based on the optimal index combination for each growth period determined in Step 5, further optimize the growth period combination to determine the target key growth period combination for winter wheat yield estimation. Step 7. Based on the target key growth period combination determined in Step 6, extract the optimal spectral features of the corresponding growth period, and further construct inter-period dynamic features to form a yield estimation feature set for winter wheat yield estimation. RENDVI, FVI during the booting stage, RVI during the booting stage, and FVI during the grain-filling stage were extracted as static spectral features; based on the above static spectral features, dynamic features across different stages were further constructed. Since FVI appears in both the optimal index combination during the booting stage and the optimal index combination during the grain-filling stage, a cross-period dynamic feature is constructed based on the representative values ​​of FVI during the booting and grain-filling stages. Define FVI b This represents the representative value of FVI during the fertile growth period at the booting stage. g This represents the representative value of the FVI during the grouting period; The representative value for the reproductive period is obtained by the average of the valid FVI observations within the corresponding reproductive period; Based on this, the following intertemporal dynamic features are constructed: I. Ratio characteristics: ； Q FVI Used to characterize the relative ratio of FVI during the heading and grain-filling stages; II. Characteristics of relative rate of change: ； Where P FVI Used to characterize the relative change in FVI from the heading stage to the grain-filling stage; After processing in step 7, the resulting set of asset valuation characteristics includes: I. Static spectral characteristics during the booting stage: RENDVI and FVI during the booting stage; II. Static spectral characteristics during the grain-filling stage: RVI and FVI during the grain-filling stage; III. Dynamic characteristics across stages: Q FVI P FVI ; In the process of constructing the training dataset, a single land parcel is used as a sample unit; For each sample, the RENDVI, FVI during the booting stage, RVI during the grain-filling stage, and FVI during the grain-filling stage were extracted, and Q was calculated. FVI and P FVI This forms the yield estimation feature set for the sample, and the corresponding output variable is the measured winter wheat yield Y of the sample. Step 8. Based on the yield estimation feature set formed in Step 7, establish a two-stage winter wheat yield estimation model based on the XGBoost model, which includes a first-stage basic yield prediction model and a second-stage subsequent correction model. The first-stage basic yield prediction model uses static spectral characteristics during the booting stage to generate basic yield prediction results. The second-stage follow-up correction model uses static spectral characteristics during the grain-filling period and dynamic characteristics across periods to correct the output of the first-stage basic yield prediction model, thus obtaining the final estimate of winter wheat yield.

2. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 1 specifically involves: Canopy spectral data were collected from various plots in the target area during the key growth stages of winter wheat, and the measured yields of the corresponding plots were obtained. The key growth stages of winter wheat include the jointing stage, booting stage, heading and flowering stage, and grain-filling stage. Canopy spectral data were collected using a portable ground object spectrometer. During the spectral acquisition process, the spectrometer probe was placed at a preset height vertically above the sample plot. After the spectral measurement of each sample plot was completed, reflectance correction was performed using a whiteboard. Multiple canopy spectral data were collected from each sample plot, and the average value was calculated after quality screening to obtain the canopy reflectance data of that sample plot.

3. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 2 specifically involves: Based on the canopy reflectance data obtained in step 1, a set of spectral characteristic indices is constructed; the spectral characteristic indices in the set include RVI, SAVI, FPAR, NIRv, NDVI, RENDVI, EVI2, and FVI.

4. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 3 specifically involves: The spectral characteristic index data obtained in step 2 are preprocessed to improve the quality of the modeling data; Preprocessing includes: removing outliers by combining physical boundary screening with Hampel filtering; filling in missing data points using linear interpolation; and performing daily-scale averaging on the spectral feature index time series data.

5. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 4 specifically involves: The process of progressively screening spectral characteristic indices includes the following steps: I. Spatial Discrimination Test: Used to calculate the degree of variation of spectral characteristic indices among plots over the entire growth period, in order to eliminate spectral characteristic indices with weak spatial discrimination ability; II. Observation reliability verification: Intra-group correlation coefficient and signal-to-noise ratio used to calculate spectral characteristic indices. In order to eliminate spectral characteristic indices that are greatly affected by environmental noise and random observation errors; III. Partial Correlation Analysis: Used to analyze the correlation between spectral characteristic indices and measured yield, retaining spectral characteristic indices that are significantly correlated with yield; IV. Collinearity Removal: The spectral feature indices retained in step III are further screened for collinearity removal to avoid high overlap of information between spectral feature indices, which would lead to redundancy in the model input. After collinearity removal and redundancy screening in step IV, FVI, RVI, and RENDVI were finally determined as the core index set for winter wheat yield estimation, and used for subsequent steps of index combination optimization within the growth period and growth period combination optimization.

6. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 5 specifically involves: During each reproductive period, different index combination schemes were constructed using the spectral characteristic indices FVI, RVI, and RENDVI of the core index set as candidate input indices, and each combination scheme was used as an input variable to input the ridge regression model. Using the measured yield of winter wheat as the output variable, the yield representation ability of different combination schemes is evaluated. Specifically, within each reproductive period, a total of 7 candidate index combination schemes were constructed; the candidate combinations include: (1) single index scheme: FVI, RVI, RENDVI; (2) double index combination scheme: RVI and FVI combination, RVI and RENDVI combination, RENDVI and FVI combination; (3) triple index combination scheme: RVI, RENDVI, and FVI combined. For each reproductive period, ridge regression models were established for the seven candidate index combinations mentioned above, and the correlation coefficient r, significance level p, and determination coefficient R were used to determine the appropriate models. 2 And to comprehensively compare different combination schemes based on the input feature dimensions; After selecting the optimal combination of indices for each reproductive period, the following optimal combination of indices was determined: During the jointing stage, the combination of the dual-index RENDVI and FVI was determined as the optimal index combination for the jointing stage. During the booting stage, the combination of the dual-index RENDVI and FVI was determined to be the optimal index combination for the booting stage. During the heading and flowering stage, the combination of the double-index RVI and RENDVI was determined as the optimal index combination for the heading and flowering stage. During the grouting period, the combination of the dual-index RVI and FVI was determined as the optimal index combination for the grouting period.

7. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 6 specifically involves: Using the optimal index combination for each growth period as a fixed input unit, different growth period combination schemes were constructed, and the ability of each scheme to represent the measured yield of winter wheat was compared. The reproductive period combination schemes include single reproductive period combination, double reproductive period combination, triple reproductive period combination, and quadruple reproductive period combination; the input variables of each combination scheme are composed of the optimal index combination of the corresponding reproductive period. Ridge regression models were established for all combinations of schemes during the jointing stage, booting stage, heading and flowering stage, and grain-filling stage. Based on the correlation coefficient r, significance level p, and determination coefficient R... 2 The root mean square error (RMSE) and input feature dimensions are used to comprehensively compare different reproductive period combinations. Ultimately, the combination of the booting stage and the grain-filling stage was determined as the target key growth period combination for winter wheat yield estimation.

8. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 1, is characterized in that... Step 8 specifically involves: I. Establish the first-stage basic output forecasting model; In the first stage, the RENDVI and FVI of the booting stage in the yield estimation feature set formed in step 7 are used as input variables, and the measured yield of winter wheat is used as the output variable to establish the basic yield prediction model for the first stage. Let the input variables for the first stage be... Recorded as: ; In the formula RENDVI b Indicates RENDVI and FVI during the spikelet incubation period. b Let FVI represent the booting stage; let Y be the measured yield of winter wheat, then the output of the first-stage basic yield prediction model is the basic yield prediction result, denoted as: ; in This represents the first stage of the XGBoost model. This indicates the results of the first-stage basic production forecast; II. Establish the second-stage follow-up correction model; In the second stage, the grouting period RVI, grouting period FVI, and Q are included in the production estimation feature set formed in step 7. FVI and P FVI As input variables, the correction amount between the first-stage basic yield prediction result and the actual yield is used as the training output variable to establish the second-stage subsequent correction model. Let the input variables for the second stage be denoted as: In the formula Indicates the RVI during the grouting period. Indicates the grouting period FVI, Indicates the characteristics of the FVI ratio. Indicates the relative rate of change of FVI; During the model training phase, the output variable of the second phase is the correction variable, denoted as ΔY, which is defined as: In the formula, Y represents the measured yield of winter wheat, Y1 represents the predicted yield of the first stage, and ΔY represents the difference between the measured yield and the predicted yield of the first stage. Using the correction variable ΔY as the training output variable, a second-stage XGBoost model is established, denoted as: In the formula This indicates the second-stage XGBoost model. This represents the amount of prediction correction output by the second-stage model; III. Obtaining the final yield estimate: Based on the first-stage baseline yield forecast and the forecast corrections from the second-stage model output, the final winter wheat yield estimate is obtained. The formula is as follows: .

9. The method for estimating winter wheat yield based on a combination of spectral characteristic indices during key growth periods, as described in claim 8, is characterized in that... In step 8, after the model training is completed, steps 1 to 7 are repeated for the winter wheat in the area to be tested to obtain the yield estimation feature set of the sample plots to be tested, and the yield estimation feature set is input into the two-stage XGBoost winter wheat yield estimation model established and trained in step 8, so as to output the winter wheat yield estimation result of the corresponding sample plots. Specifically, a specific plot of land within the target area is selected as the object of estimation, and the following operations are performed sequentially: First, during the identified key growth periods, namely the heading stage and the grain-filling stage, canopy spectral data of the plot were collected, and canopy reflectance and corresponding spectral characteristic values ​​were obtained according to steps 1 to 3. Then, based on the determined optimal index combination and target key growth period combination results, the RENDVI, FVI, RVI and FVI of the heading stage, and the grain filling stage of the plot are extracted. The intertemporal dynamic characteristic Q is calculated according to step 7. FVI and P FVI This forms the asset valuation feature set corresponding to the land parcel; Subsequently, the RENDVI and FVI of the heading stage of this plot were input into the first-stage basal yield prediction model to obtain the basal yield prediction result Y1 for this plot; then the RVI, FVI, and Q of the grain-filling stage of this plot were input into the first-stage basal yield prediction model to obtain the basal yield prediction result Y1 for this plot; FVI and P FVI Input the second-stage subsequent correction model to obtain the predicted correction amount. ; Finally, the estimated winter wheat yield for this plot of land was obtained using the following formula. The formula is as follows: .

Images