Soybean key growth period aboveground biomass estimation method based on hyperspectral image and machine learning

Abstract

The invention discloses a soybean key growth period aboveground biomass estimation method based on a hyperspectral image and machine learning. Belongs to the technical field of electric digital data processing. According to the method, the technical problems of low information utilization rate, poor stability in the trans-growth period, insufficient model interpretation and the like in the overground biomass prediction of the key growth period of the soybeans in the existing method are solved. A feature fusion method of hyperspectral derivative transformation and wavelet decomposition is provided, noise is effectively suppressed, and the spectral difference in the growth period is highlighted; a multi-source feature coupling system is constructed, traditional spectral indexes are utilized, waveband combination and wavelet features are introduced, and feature diversity and robustness are improved; the stability of AGB estimation in the cross-growth period is improved through the Stacking integrated model, and the Stacking integrated model is superior to a single model.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing technology, specifically relating to a method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning. Background Technology

[0002] Aboveground biomass (AGB) is a crucial agronomic parameter reflecting crop growth and yield potential. Existing AGB measurement methods primarily rely on manual sampling and weighing, which is destructive, time-consuming, labor-intensive, and its accuracy is significantly affected by human intervention, making it unsuitable for efficient monitoring of large-scale farmland. With the development of hyperspectral technology, crop canopy reflectance spectra have become an important non-destructive method for monitoring AGB during key soybean growth stages. However, estimation methods relying solely on raw spectra or traditional vegetation indices suffer from information redundancy, significant noise interference, and poor model stability, resulting in insufficient accuracy in modeling across growth stages. Furthermore, most existing methods rely on single algorithms and lack feature fusion and interpretative analysis mechanisms, hindering stable prediction of crop biomass in complex growth dynamics. Therefore, in predicting aboveground biomass (AGB) during key soybean growth stages, there is an urgent need for a method that can effectively integrate multi-source spectral features, suppress noise influence, and combine machine learning algorithms to construct a high-precision estimation model. Summary of the Invention To address the technical problems of existing methods in predicting aboveground biomass (AGB) during key growth stages of soybean, such as low information utilization, poor stability across growth stages, and insufficient model interpretability, this invention proposes a method for estimating aboveground biomass during key growth stages of soybean based on hyperspectral imagery and machine learning. The method includes the following steps: S1. Data collection: Collect raw hyperspectral data of soybean canopy and corresponding measured data of aboveground biomass during key growth stages of soybean across the growth period; S2. Data preprocessing: Radiometric correction, image cropping, vegetation mask generation, and Savitzky-Golay smoothing are performed on the raw hyperspectral data to obtain preprocessed hyperspectral data. S3. Spectral Transformation: The hyperspectral data preprocessed in step S2 is further subjected to first-order and second-order derivative transformations to obtain first-order and second-order derivative spectral data. S4. Spectral Feature Construction: Construct multi-dimensional spectral features, including spectral feature parameters, spectral vegetation index, wavelet coefficient features, and custom band combination features; screen out common sensitive features at different growth stages using a recursive feature elimination method and construct a coupled feature set; S5. Construction of integrated biomass estimation model: Several machine learning algorithms are used to construct several aboveground biomass estimation models. The stacking ensemble learning strategy is used to fuse the several aboveground biomass estimation models to obtain the integrated biomass estimation model with the best performance. S6. Aboveground biomass estimation: The best-performing integrated biomass estimation model is used to predict the aboveground biomass during key growth stages of soybean.

[0003] Furthermore, the reproductive period includes the initial flowering stage, the early pod-setting stage, the peak pod-setting stage, and the full maturity stage.

[0004] Furthermore, radiometric calibration is performed on the acquired hyperspectral images. The mean values ​​of the dark and white references are calculated by band, and the original data is calibrated band by band. Then, the data is cropped, removing the left and right 1 / 5 regions. The cropped RGB image is converted into an HSV hue-saturation-brightness space image. The HSV range of the green pixels is set, and a binary mask is generated to filter out background interference areas. Nearest neighbor interpolation is used to scale the mask to the spatial resolution of the hyperspectral image to ensure pixel alignment.

[0005] Furthermore, the spectral characteristic parameters include blue edge parameters, yellow edge parameters, red edge parameters, red valley parameters, green peak parameters, ratio parameters, and normalization parameters.

[0006] Furthermore, the spectral vegetation index is an indicator that can effectively characterize crop growth status and physiological parameters, obtained through a series of mathematical operations based on the reflectance of a specific spectral band.

[0007] Furthermore, the wavelet coefficient features are obtained by selecting orthogonal db4 wavelet basis functions to perform four-level discrete wavelet transform on the original hyperspectral data and the first-order and second-order derivative spectral data. The obtained cA4 approximation coefficients reflect the macroscopic low-frequency features of the spectrum, while the cD1-cD4 detail coefficients capture local information of different frequency bands in layers.

[0008] Furthermore, the method for obtaining the custom band combination features is as follows: select all bands of the original hyperspectral data and the first-order and second-order derivative spectral data for custom combination operation, and calculate the sum, difference, ratio and normalized value of the reflectance of different wavelengths of the spectrum.

[0009] Furthermore, the recursive feature elimination method selected is the RFE recursive feature elimination method.

[0010] The beneficial effects of the method described in this invention are as follows: (1) A feature fusion method of hyperspectral derivative transformation and wavelet decomposition was proposed, which effectively suppressed noise and highlighted the spectral differences during the reproductive period; (2) A multi-source feature coupling system was constructed, which not only utilized traditional spectral indices, but also introduced band combinations and wavelet features, thereby improving feature diversity and robustness; (3) The stability of AGB estimation across reproductive periods was improved by using the Stacking ensemble model, which is superior to the single model. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method described in an embodiment of the present invention; Figure 2 This is a comparison diagram of spectral SG smoothing filtering in an embodiment of the present invention; Figure 3 is a heatmap of feature engineering correlation in an embodiment of the present invention; Figure 4 shows a scatter plot of the training and test sets of spectral feature set 1 and R. 2 Charts showing the results of RMSE and NRMSE indices; Figure 5 shows a scatter plot of the training and test sets of spectral feature set 2 and R. 2 Chart showing the results of RMSE and NRMSE indicators. Detailed Implementation

[0012] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0013] Example 1 This embodiment provides a method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning, including the following steps: S1. Data collection: Collect raw hyperspectral data of soybean canopy and corresponding measured data of aboveground biomass during key growth stages of soybean across the growth period; S2. Data preprocessing: Radiometric correction, image cropping, vegetation mask generation, and Savitzky-Golay smoothing are performed on the raw hyperspectral data to obtain preprocessed hyperspectral data. S3. Spectral Transformation: The hyperspectral data preprocessed in step S2 is further subjected to first-order and second-order derivative transformations to obtain first-order and second-order derivative spectral data. S4. Spectral Feature Construction: Construct multi-dimensional spectral features, including spectral feature parameters, spectral vegetation index, wavelet coefficient features, and custom band combination features; screen out common sensitive features at different growth stages using a recursive feature elimination method and construct a coupled feature set; S5. Construction of integrated biomass estimation model: Several machine learning algorithms are used to construct several aboveground biomass estimation models. The stacking ensemble learning strategy is used to fuse the several aboveground biomass estimation models to obtain the integrated biomass estimation model with the best performance. S6. Aboveground biomass estimation: The best-performing integrated biomass estimation model is used to predict the aboveground biomass during key growth stages of soybean.

[0014] Example 2 This embodiment further defines Embodiment 1. Step S1 will be further explained.

[0015] Soybean canopy hyperspectral data acquisition: Soybean canopy hyperspectral data were acquired using the TraitDiscover field-track high-throughput plant phenotyping platform from Huinuo Ruide. Forty representative soybean lines were photographed in individual plots at four growth stages: R1 (initial flowering), R3 (early pod formation), R4 (peak pod formation), and R8 (full maturity). The specific timing was around 9 PM, with the lens pointing vertically downwards at a distance of approximately 1 meter from the top of the canopy. First, the lens was turned off to photograph the whiteboard, then turned on to scan the whiteboard for calibration. The measured hyperspectral band range was 397.01 nm–1004.52 nm, with a spectral resolution of 1.4 nm, yielding raw, jpg (rgb), and hr files.

[0016] Soybean aboveground biomass data collection: After collecting hyperspectral data, several representative soybean plants with average growth were selected from each plot as samples for harvesting. The underground parts of the samples were removed, and the remaining parts were washed and separated into stems, leaves, and pods (separation of stems, leaves, and pods at R8 stage). The fresh weight of each organ was weighed. The samples were then numbered and placed in nylon bags. The oven temperature was adjusted to 105℃ for blanching, followed by drying at 80℃ until constant weight was achieved. The dry weight of each organ was measured, and the soybean AGB value (g / m³) was calculated based on the population density and sample dry weight. 2 ).

[0017] Example 3 This embodiment further defines embodiment 1. Step S2 will be further explained.

[0018] Radiometric calibration was performed on the acquired hyperspectral images. The mean values ​​of the dark and white references were calculated for each band, and the original data was calibrated band by band. Then, the data was cropped (removing the left and right 1 / 5 regions). The cropped RGB image was converted to HSV (Hue-Saturation-Lightness) space, and the HSV range for green pixels was set (Hue H: 51-122, Saturation S≥25, Lightness V≥25). A binary mask was generated to filter out background interference areas such as soil. Nearest neighbor interpolation was used to scale the mask to the spatial resolution of the hyperspectral image, ensuring pixel alignment. Furthermore, Savitzky-Golay (SG) smoothing filtering was applied to the original hyperspectral data. Figure 2 As shown, SG smoothing filtering is a signal denoising strategy that uses a moving window to perform polynomial fitting of the data using the least squares method, preserving the detailed features of the spectrum while denoising the spectral data. The filter parameters selected are window_length=9 and polyorder=2.

[0019] Example 4 This embodiment further defines Embodiment 1. Step S3 will be further explained.

[0020] Raw hyperspectral data often contains a lot of noise and redundant information. Directly applying this data for analysis increases the computational burden on the model, thus reducing its accuracy. Performing derivative transformation on the raw spectrum can effectively eliminate baseline drift and background interference, highlight subtle features in the spectrum, and enhance the sensitivity and discriminative ability to detect crop physiological states.

[0021] The first derivative reflects the slope of the spectral curve, helping to identify absorption peaks and reflection features, especially the red edge positions. The second derivative further highlights the curvature of the spectrum, improving the accuracy of feature extraction. Therefore, we use python.numpy.diff to calculate the first and second derivative spectra.

[0022] Example 5 This embodiment further defines embodiment 1. Step S4 will be further explained.

[0023] Spectral characteristic parameters: The absorption characteristics of pigments inside crops to different bands of the spectrum form a distinctive feature that distinguishes them from other land cover. These features can be quantified to obtain a series of spectral characteristic parameters such as amplitude, location, area, kurtosis, and skewness. In this embodiment, five locations—blue edge, yellow edge, red edge, red valley, and green peak—are selected to construct spectral characteristic parameters. The selected spectral characteristic parameters are shown in Table 1.

[0024] Table 1. Spectral characteristic parameters:

[0025] Spectral vegetation index: A vegetation index (VI) is an indicator that effectively characterizes crop growth status and physiological parameters, obtained through a series of mathematical operations based on the reflectance of a specific spectral band. It is widely proven and used to retrieve a range of agronomic parameters of crops. In this embodiment, the selected spectral vegetation indices are shown in Tables 2.1-2.2.

[0026] Table 2.1 Spectral Vegetation Indices: (in the table) represent (Average spectral reflectance at wavelength)

[0027] Table 2.2 Spectral Vegetation Indices: (in the table) represent (Average spectral reflectance at wavelength)

[0028] Wavelet coefficient characteristics: Over the past 20 years, wavelet transform has been widely applied in signal processing, data compression, human vision, and pattern recognition. It has also demonstrated great potential in denoising and dimensionality reduction of hyperspectral data. Among these applications, wavelet transform based on the Daubechies4 (db4) basis function has been extensively used in spectral data processing. Therefore, orthogonal db4 wavelet basis functions were selected to perform a four-level discrete wavelet transform (DWT) on the original spectrum and first and second derivative spectral data. The resulting cA4 approximation coefficients reflect the macroscopic low-frequency characteristics of the spectrum, while the cD1-cD4 detail coefficients capture local information in different frequency bands (cD4 reflects medium-scale features such as wide absorption valleys, while cD1 focuses on high-frequency details such as narrow absorption peaks and noise). Their orthogonality ensures a non-redundant representation of the cA4 and cD coefficients. The specific decomposition and reconstruction calculation process is as follows: ; ; ; In the formula, The wavelet decomposition level is denoted as . This is the position index of the wavelet coefficients. For the spatial index of the input signal, These are approximate coefficients. For detail coefficients, It is a low-pass filter. For high-pass filters, To reconstruct the low-pass synthesized filter, To reconstruct the high-pass synthesized filter.

[0029] Custom band combination features: A custom combination operation was performed on all bands of the original spectrum and the first and second derivative spectra. The sum (Sub), difference (Di), ratio (Ri), and normalized (Vi) values ​​of reflectance at different wavelengths were calculated to compensate for the limitations of the AGB estimation model relying solely on spectral characteristic parameters or vegetation indices under complex practical application conditions. The specific calculation and representation methods are shown in Table 3 below (in the table...). represent Average spectral reflectance at wavelength).

[0030] Table 3: Features of Custom Band Combinations

[0031] The common sensitive features at different reproductive stages were screened out by the recursive feature elimination method and a coupled feature set was constructed.

[0032] Recursive Feature Elimination (RFE) is a data dimensionality reduction and feature selection method. It couples a classifier or regression model, repeatedly trains it, and gradually eliminates the least important features. Finally, it selects the feature subset that contributes the most to the prediction target. At the same time, it sets a threshold to stipulate that the correlation coefficient between any two features in the final subset should not be greater than 0.7 / 0.8, so as to avoid the impact of feature collinearity on modeling.

[0033] Example 6 This embodiment further defines embodiment 1. Step S5 will be further explained.

[0034] Several machine learning algorithms include: Partial Least Squares Regression (PLS) was proposed by Swedish statistician Herman Wold in 1966. Its core idea is to build a predictive model by extracting the common principal components of independent and dependent variables and maximizing the covariance. It is a supervised dimensionality reduction regression method that is suitable for modeling high-dimensional, collinear, and small sample data.

[0035] Ridge Regression (RR), proposed by Hoerl and Kennard in 1970, is a regularization method specifically designed to address multicollinearity in linear regression. It introduces an L2 regularization term (penalty term) into the objective function of Ordinary Least Squares (OLS), effectively controlling model complexity by adding the square of the coefficient vector.

[0036] Random Forest (RF) was proposed by Leo Breiman and Adele Cutler in 2001. It combines multiple decision trees, using a bootstrap method to randomly sample subsets from the original dataset for training. Furthermore, during the construction of the decision trees, the algorithm randomly selects some features for node splitting, thereby reducing the correlation between trees. Random Forest ultimately yields predictions through either voting (classification) or averaging (regression).

[0037] The Support Vector Machine (SVR) regression algorithm was proposed by Cortes et al. in 1995. SVM is a classifier that addresses nonlinear problems by introducing a kernel function. Subsequently, this algorithm was also applied to regression problems, where it fits the data by constructing an interval band that tolerates ε-bias, ensuring that most sample points fall within this band while minimizing model complexity.

[0038] Extreme Gradient Boosting (XGBoost) is actually a type of ensemble learning (EL) algorithm, belonging to the reinforcement learning category. It was proposed by Tianqi Chen and Carlos Guestrin in 2016. This algorithm improves prediction performance by integrating multiple weak learners (decision trees). Its core idea is gradient boosting, where the training objective of each new tree is to minimize the residual of the previous prediction, gradually reducing the value of the loss function through an additive model. XGBoost's objective function includes a loss function and L1 and L2 regularization terms to control model complexity and prevent overfitting. Furthermore, the algorithm introduces a weighted quantile sketch to efficiently find feature split points, a sparsity-aware algorithm to automatically handle missing values ​​and sparse features, and employs techniques such as block structure, cache optimization, and an approximate greedy algorithm to improve computational efficiency.

[0039] By employing a stacking ensemble learning strategy to fuse several aboveground biomass estimation models, a top-performing ensemble biomass estimation model can be obtained. Stacking is a type of ensemble learning strategy that improves model performance by combining the outputs of multiple individual learners. Its process consists of two levels: first, multiple training subsets are generated on the training dataset through resampling or cross-validation, and a set of base classifiers are trained. Then, the outputs of these base classifiers are used as input to train a meta-classifier to generate the final prediction.

[0040] A grid search-based cross-validation strategy is adopted: a detailed hyperparameter search process is set up for the base model, and the performance of each parameter group is evaluated through 5-fold cross-validation to select the optimal combination. At the same time, all possible base model-meta-model combinations are traversed to construct the optimal stacking ensemble model.

[0041] Using the coefficient of determination (R) 2 Ri, Root Mean Square Error (RMSE), and Standard Root Mean Square Error (NRMSE) are used as evaluation metrics for regression models, while P is used as an evaluation metric for classification models. 2 The closer a value is to 1, the stronger the model's explanatory power and the better its fit. RMSE reflects the degree of deviation between predicted and measured values; the smaller the value, the higher the model's accuracy. NRMSE eliminates the influence of dimensions through normalization and is used to quantify the relative level of the model's prediction error. A value less than 10% is excellent, 10%-20% is good, 20%-30% is moderate, and a value exceeding 30% indicates poor model performance.

[0042] Simultaneously, the KNN algorithm is employed to determine the growth stage, and a multi-base learner is fused using a Stacking ensemble learning strategy to improve the robustness of cross-stage prediction. Regarding model interpretation, this invention introduces the SHAP method to analyze the interpretability of the contribution of input features, clarifying the mechanism by which different spectral features affect AGB estimation. Finally, the obtained method for estimating aboveground biomass during key soybean growth stages is as follows: Figure 1 As shown.

[0043] Example 7 This embodiment further defines embodiments 1-6. The above method was experimentally verified.

[0044] In this embodiment, a comparative experiment is set up. First, a spectral feature set 1 is constructed using the existing feature engineering construction method. Then, a spectral coupling feature set 2 is constructed using the feature engineering construction method of the present invention. Biomass is estimated using the spectral feature set 1 and the spectral coupling feature set 2 respectively through the biomass estimation model, and the obtained data are compared.

[0045] Spectral feature set 1: Visualization scatter plots of training and test sets and R 2 The results of the RMSE and NRMSE indices are shown in Figure 4af: The Stacking ensemble model combines RF, SVR, and XGBoost-ElasticNet. Observing Figure 4 (af), we find that the Stacking ensemble learning model performs best on the test set R. 2 The RMSE and NRMSE were 0.694 and 281.865 g / m, respectively. 2 And 16.1%, besides, the RF and XGB models performed better on the test set R2 The RMSE and NRMSE were 0.679 and 289.003 g / m, respectively. 2 And 16.5% and 0.641, 305.691 g / m 2 and 17.5%. Test set R for SVR and PLS models 2 All are above 0.6, RR of the RR model 2 The accuracy was 0.599, indicating that the model was generally accurate and stable, with no models showing low accuracy or overfitting.

[0046] Spectral coupling feature set 2: Visualization scatter plot of training and test sets and R 2 The results of the RMSE and NRMSE indices are shown in Figure 5af: The Stacking ensemble combines primitive-meta-models such as RF and XGB-Lasso. The RF model performs best, followed by the Stacking model and the XGB model. The RF model is trained on set R. 2 The RMSE and NRMSE were 0.885 and 171.851 g / m, respectively. 2 And 7.2%, test set R 2 The RMSE and NRMSE were 0.767 and 246.274 g / m, respectively. 2 And 14.1%. Stacking generally performed similarly to RF, with training set R 2 The RMSE and NRMSE were 0.864 and 186.752 g / m, respectively. 2 And 7.8%, test set R 2 The RMSE and NRMSE were 0.767 and 246.344 g / m, respectively. 2 And 14.1%. Secondly, the XGB model test set R... 2 With a score of 0.730, all three demonstrated high accuracy and stability. The test set R of the other three models was [missing value]. 2 They are all above 0.6.

[0047] Overall, the soybean AGB model based on coupling features performs significantly better than the AGB estimation model based on spectral features.

[0048] A soybean AGB estimation model spanning four growth stages. Data from four growth stages are combined to construct spectral features based on 31 spectral vegetation indices and 28 spectral characteristic parameters. Simultaneously, based on the Pearson correlation coefficient, three commonly highly correlated spectral bands across the four growth stages and four types of combined features between them are searched to construct common sensitive features. Coupled features are then constructed based on the RFE algorithm, building upon the above two steps.

[0049] For each feature category, the 12 features with the highest absolute values ​​of the AGB correlation coefficient were selected for visualization. It was stipulated that the autocorrelation coefficients between the selected features could not exceed 0.8. Since the autocorrelation coefficients between spectral features were generally high, to ensure a sufficient number of features were selected, the autocorrelation coefficients between spectral features could not exceed 0.95. The correlation heatmap is shown in Figure 3.

[0050] Analysis of Figure 3 shows that the blue edge amplitude D in the spectral characteristics b The correlation coefficient between MCARI and biomass was the highest, with r = -0.76, followed by MCARI and R. g / R r The autocorrelation between highly correlated features and IPVI is relatively high. Furthermore, apart from MCARI, IPVI, and GNDVI, which are vegetation index features, the other highly correlated features are spectral parameters, mainly calculated as combinations of green and red light bands. In the selection of common sensitive features, the highly correlated feature is Sub (fd683.47, fd807.92), which is the sum of the reflectance of two first-derivative wavelengths located in the red and near-infrared regions. Apart from this, the other highly correlated common sensitive features selected, except for one difference combination feature, are all ratio and normalized combination features, distributed in the blue, green, red, and near-infrared regions, and are all combinations of first- or second-derivative spectral wavelengths, without single-band or combination features based on the original spectral wavelengths. In the correlation heatmap of coupled features, except for ExG, which is a vegetation index feature, SDR / SD... b For spectral characteristic parameters, other highly correlated features are all band combination features based on the three spectra. ExG has the strongest correlation with AGB, with r = -0.72. The highly correlated band combination features are also mainly based on the ratio of the first and second derivative spectral bands and the normalized combination.

[0051] A soybean fertility stage discrimination model based on multi-source features was developed. Models were constructed based on spectral features and common sensitive features, respectively, without considering AGB (Adaptive Global Gradient) analysis, to determine the fertility stage of soybeans from a feature-based perspective. The classification algorithm used was the K-Nearest Neighbors (KNN) algorithm, with Manhattan distance, which is more stable in high-dimensional space, replacing the default Euclidean distance. Adversarial validation was used to evaluate data partitioning, and t-SNE (Thinker Neighbors) was chosen as the data dimensionality reduction and visualization strategy.

[0052] The results show that the adversarial validation AUC based on the spectral feature dataset is 0.485 (±0.086), the data segmentation is reasonable, the training set accuracy is P=0.9905, and the test set accuracy is P=0.9783. The adversarial validation AUC based on the public sensitive feature dataset is 0.476 (±0.110), the data segmentation is reasonable, and the KNN model training set accuracy is P=0.9619, and the test set accuracy is P=0.9348. A satisfactory and stable model for determining fertility stage has been obtained.

Claims

1. A method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning, characterized in that... The method includes the following steps: S1. Data collection: Collect raw hyperspectral data of soybean canopy and corresponding measured data of aboveground biomass during key growth stages of soybean across the growth period; S2. Data preprocessing: Radiometric correction, image cropping, vegetation mask generation, and Savitzky-Golay smoothing are performed on the raw hyperspectral data to obtain preprocessed hyperspectral data. S3. Spectral Transformation: The hyperspectral data preprocessed in step S2 is further subjected to first-order and second-order derivative transformations to obtain first-order and second-order derivative spectral data. S4. Spectral Feature Construction: Construct multi-dimensional spectral features, including spectral feature parameters, spectral vegetation index, wavelet coefficient features, and custom band combination features; screen out common sensitive features at different growth stages using a recursive feature elimination method and construct a coupled feature set; S5. Construction of integrated biomass estimation model: Several machine learning algorithms are used to construct several aboveground biomass estimation models. The stacking ensemble learning strategy is used to fuse the several aboveground biomass estimation models to obtain the integrated biomass estimation model with the best performance. S6. Aboveground biomass estimation: The best-performing integrated biomass estimation model is used to predict the aboveground biomass during key growth stages of soybean.

2. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 1, characterized in that, The reproductive period includes the initial flowering stage, the early pod-forming stage, the peak pod-forming stage, and the full maturity stage.

3. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 2, characterized in that, Radiometric calibration is performed on the acquired hyperspectral images. The mean values ​​of the dark and white references are calculated by band, and the original data is calibrated band by band. Then, the data is cropped, removing the left and right 1 / 5 regions. The cropped RGB image is converted into an HSV hue-saturation-lightness space image. The HSV range of the green pixels is set, and a binary mask is generated to filter out background interference areas. Nearest neighbor interpolation is used to scale the mask to the spatial resolution of the hyperspectral image to ensure pixel alignment.

4. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 3, characterized in that, Spectral characteristic parameters include blue edge parameter, yellow edge parameter, red edge parameter, red valley parameter, green peak parameter, ratio parameter, and normalization parameter.

5. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 4, characterized in that, The spectral vegetation index is an indicator that effectively characterizes crop growth status and physiological parameters, obtained through a series of mathematical operations based on the reflectance of a specific spectral band.

6. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 5, characterized in that, The wavelet coefficient features are obtained by selecting the orthogonal db4 wavelet basis function to perform a four-level discrete wavelet transform on the original hyperspectral data and the first-order and second-order derivative spectral data. The obtained cA4 approximation coefficients reflect the macroscopic low-frequency features of the spectrum, while the cD1-cD4 detail coefficients capture the local information of different frequency bands in layers.

7. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 6, characterized in that, The method for obtaining custom band combination features is as follows: select all bands of the original hyperspectral data and the first-order and second-order derivative spectral data, perform custom combination operations, and calculate the sum, difference, ratio and normalized value of the reflectance of different wavelengths of the spectrum.

8. The method for estimating aboveground biomass of soybean during key growth stages based on hyperspectral imagery and machine learning according to claim 7, characterized in that, The recursive feature elimination method selected is the RFE recursive feature elimination method.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-8.

Images