Corn leaf chlorophyll content spectral feature fusion and inversion method

Abstract

According to the corn leaf chlorophyll content spectral feature fusion and inversion method, a corn leaf chlorophyll content prediction model is constructed in combination with a machine learning algorithm, and the corn leaf chlorophyll content is accurately and quantitatively estimated. According to the method, on the basis of VIs and spectral data obtained after wavelet transform analysis optimal preprocessing, VIs sensitive to chlorophyll change are screened out through correlation analysis, and in combination with a peak value extraction method, wavelet coefficients with the optimal correlation under different wavelet basis function processing are screened out to serve as WF. An SE attention mechanism is introduced to fuse the screened WF and VIs spectral features, and weights corresponding to different spectral features are optimized; and on the basis of Vls, WF, Vls-WF and SE-Vls-WF spectral characteristics, constructing a corn leaf chlorophyll content prediction model by adopting an SVR (Support Vector Regression) algorithm and an RFR (Radio Frequency Regression) algorithm. The method belongs to the field of crop leaf nutrition information monitoring, the operation is simple, the analysis steps are simple, the corn leaf chlorophyll content can be accurately predicted by using the method, and a new thought is provided for crop nutrition information estimation.

Description

Technical Field

[0001] This invention relates to the field of crop leaf nutrient information monitoring, specifically a method for fusing and inverting the spectral characteristics of chlorophyll content in maize leaves. Background Technology

[0002] As a major food, feed, and industrial raw material crop, maize accounts for 36.6% of global food production, playing a vital role in ensuring global food security. Chlorophyll is a key pigment in maize leaves for photosynthesis, promoting light energy utilization and organic matter synthesis. Changes in chlorophyll content characterize maize stress levels, photosynthetic capacity, and nitrogen use efficiency, providing a theoretical basis for estimating maize nutrient levels, growth status, and yield. Therefore, accurate prediction of maize leaf chlorophyll content is of great significance for achieving high and stable maize yields and ensuring global food security.

[0003] Hyperspectral technology, due to its advantages of speed, efficiency, and non-destructive operation, has been widely applied in the monitoring of chlorophyll content in crop leaves. By analyzing the relationship between hyperspectral information and chlorophyll content in crop leaves, the influence of crop leaf structural parameters and physicochemical parameters on spectral emissivity can be quantitatively described, thereby estimating the chlorophyll content of crop leaves. Common methods for acquiring hyperspectral information include UAV remote sensing platforms and portable hyperspectral instruments. UAV remote sensing platforms acquire crop canopy spectral information by carrying multispectral cameras, offering advantages such as wide monitoring range and high timeliness. However, UAVs are susceptible to interference from environmental factors such as noise, light, and wind during flight, leading to errors in the monitoring data and making it difficult to obtain accurate models. Compared to UAV remote sensing platforms, portable hyperspectral instruments offer advantages such as high resolution and high flexibility, making them suitable for both practical production and scientific research.

[0004] While hyperspectral data provides a wealth of spectral information, its high correlation between adjacent bands and significant redundancy make it prone to overfitting. Therefore, before using hyperspectral data for modeling, spectral data processing and variable selection are necessary to improve model predictive capabilities. Currently, researchers use portable hyperspectral instruments to acquire spectral information and combine data processing methods such as feature band extraction, spectral data transformation, and vegetation indices with machine learning algorithms to predict chlorophyll content in crop leaves. However, many of these models are based on spectral features obtained using a single data processing method, neglecting effective spectral features from other methods, thus limiting the selection of effective feature variables. To address this issue, integrating multiple data processing methods to obtain multispectral features of crop leaves can overcome the limitations of single spectral characteristics and provide a new approach to optimizing the accuracy of crop leaf chlorophyll content prediction. Summary of the Invention

[0005] To optimize the prediction accuracy of chlorophyll content in maize leaves, this invention proposes a method for fusing and inverting spectral features of chlorophyll content in maize leaves. Spectral vegetation indices (VIs) and wavelet transform are used to analyze leaf spectral reflectance. Based on correlation analysis, VIs sensitive to changes in maize leaf chlorophyll content and the wavelet features (WF) with the best correlation under different wavelet basis function processing are selected. The SE attention mechanism is used to fuse VIs and WF. Partial least squares regression (PLSR), support vector regression (SVR), and random forest regression (RFR) are combined to estimate the chlorophyll content of maize leaves. This invention explores the feasibility of different spectral data processing and modeling methods for estimating chlorophyll content in maize leaves, providing a theoretical basis and technical reference for accurate prediction of crop leaf chlorophyll content.

[0006] The technical solution adopted in this invention is: a method for fusion and inversion of spectral characteristics of chlorophyll content in maize leaves, specifically including the following steps: S1, Plant grouping: Select a number of corn plants and group them accordingly; S2. Raw Data Acquisition and Processing: The maize plants grouped in step S1 were tested, and hyperspectral data and chlorophyll content (SPAD value) of maize leaves at different growth stages of each plant were obtained. For the obtained maize leaf spectral data and SPAD value, abnormal bands and samples of spectral data were removed by combining PLSR and Monte Carlo method. Abnormal data of SPAD value were removed by k-means clustering and Euclidean distance method to obtain the processed test data. S3. Spectral data preprocessing: Based on the detection data processed in step S2, and using commonly used spectral preprocessing methods and their combinations, the optimal spectral preprocessing method and the prediction model framework for maize leaf chlorophyll content were selected. S4. Optimal Vegetation Index (VIs) Selection: Based on the optimal preprocessed spectral curve selected in step S3, several spectral vegetation indices are extracted, and the optimal VIs data for constructing a chlorophyll content prediction model are compared and obtained. S5. Selection of Optimal Wavelet Feature (WF) Coefficients: Based on the optimal preprocessed spectral curve selected in step S3, eight wavelet basis functions were selected to decompose the spectral reflectance data at the scale of 1-1024. The wavelet coefficients with the best correlation under the processing of the eight wavelet basis functions were selected as wavelet features (WF) to construct a chlorophyll content prediction model. S6. Attention mechanism fusion of VIs-WF spectral features: The SE attention mechanism is used to compress and activate the best VIs index and best WF coefficient selected in steps S4 and S5, obtain the weight value corresponding to the spectral feature, and assign it to the spectral feature to realize weight update assignment, and finally obtain the weighted spectral feature. S7. Model Building: Based on the modeling method selected in step S3, a prediction model for maize leaf chlorophyll content at different growth stages was constructed, using WF, VIs, VIs-WF and SE attention mechanism fusion of multiple feature variables (SE-VIs-WF) as independent variables and maize leaf chlorophyll content (SPAD value) as dependent variable.

[0007] As a preferred option, in step S2, hyperspectral data and chlorophyll content (SPAD value) of maize leaves at the six-leaf stage (V6), eight-leaf stage (V8), twelve-leaf stage (V12), and flowering and silking stage (R1) are obtained for each plant. Three sampling points (upper, middle, and lower) are set at the top fully expanded leaves of maize V6, V8, and V12 and the ear-position leaves of R1. The average value of the three measurements is taken as the final spectral reflectance and SPAD value of the leaf.

[0008] As a preferred option, the specific steps in step S2, which combine PLSR and Monte Carlo methods to remove abnormal bands and samples from spectral data, are as follows: Set the maximum number of components to 20, and use 10-fold cross-validation to select the optimal number of PLSR components; set the number of iterations N for the Monte Carlo method to 1000, set the outlier threshold to 3 times the standard deviation, fit the initial PLSR model using the optimal number of components, and calculate the initial residuals; perform N random samplings, randomly select samples and fit the PLSR model each time, calculate the residuals, and mark samples with residuals exceeding the threshold as outliers.

[0009] As a preferred option, when using k-means clustering combined with Euclidean distance to remove SPAD value outliers in step S2, it is assumed that the distance from the data point to the cluster center follows a normal distribution, and data points with a distance greater than "mean distance plus twice the standard deviation" are defined as outliers.

[0010] As a preferred option, the commonly used spectral preprocessing methods and combinations in step S3 are as follows: based on the original spectral data (OR) and smoothing algorithm (SG), standard normal variable transformation (SNV), multivariate scattering correction (MSC), SG-SNV, and SG-MSC preprocessing methods, combined with partial least squares regression (PLSR), support vector regression (SVR), and random forest regression (RFR) methods, a maize leaf chlorophyll content prediction model is constructed to select the best maize leaf chlorophyll content prediction model framework.

[0011] As a preferred option, there are 23 spectral vegetation indices in step S4, as shown below. Based on the VIs that satisfy the significance of p≤0.001 in periods V6, V8, V12, and R1, VIs with correlation coefficients greater than 0.6, 0.65, 0.75, and 0.75 are selected for modeling.

[0012] As a preferred option, step S5 specifically involves selecting eight wavelet basis functions: gb4, haar, db4, bior1.5, mexh, rbio3.3, sym4, and coif2. 1 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 2 10 Wavelet coefficients at 10 specific scales were selected, and the decomposition scale and spectral band corresponding to the strongest correlation under 8 wavelet basis function processing at each period were screened to obtain the optimal wavelet coefficients, i.e., WF. The wavelet transform calculation formula is as follows: (1) In the formula, CWT f (s,τ) represents the result after wavelet transform, where s and τ are the scale and translation parameters, respectively, f(t) is the spectral reflectance of maize leaves after optimal preprocessing, and Ψ*(x) is the wavelet basis function.

[0013] As a preferred approach, step S6 involves the following steps: The SE attention mechanism is used to compress and activate the optimal VIs and WF selected in steps S4 and S5. The Squeeze operation performs global average pooling on the input VIs-WF spectral features Y to generate z. The Excitation operation uses weight W to generate the required weight information, and processes the Squeeze operation vector through two fully connected layers W1 and W2 to obtain the weight values ​​corresponding to the spectral features. The Scale operation assigns the generated weight vector s to the feature Y to update the weights, ultimately obtaining the weighted spectral feature Y'. The core calculation method is as follows: (2) (3) (4) In the formula, F sq F ex F scale These represent global average pooling, adaptive recalibration, and reweighting operations, respectively. cLet z be the c-th element.

[0014] As a preferred option, the specific steps of step S7 are as follows: Based on the modeling method selected in step S3, WF, VIs, VIs-WF and SE-VIs-WF are used as independent variables, and the chlorophyll content (SPAD value) of maize leaves is used as the dependent variable. The training set and the test set are divided in an 8:2 ratio to construct the maize leaf chlorophyll content prediction model for periods V6, V8, V12 and R1.

[0015] The present invention has the following beneficial effects: Firstly, this invention utilizes hyperspectral technology to invert chlorophyll content in crop leaves, offering advantages such as fast analysis speed, high measurement accuracy, and non-destructive testing. After data filtering, sample spectral data and corresponding chlorophyll content are obtained, reducing abnormal data caused by external stray light, instrument dark current and noise, and human error. Multiple methods are employed to preprocess the raw spectral data, eliminating noise interference with spectral reflectance and chlorophyll-sensitive spectral signals, ensuring data accuracy. Furthermore, by integrating various data processing methods, multispectral characteristics of maize leaves are obtained, overcoming the limitations of single spectral characteristics and optimizing the prediction accuracy of maize leaf chlorophyll content, thereby improving model precision.

[0016] Secondly, based on correlation analysis, the method described in this invention introduces an SE attention mechanism to fuse VIs and WF spectral features, enhancing the model's response to important features and improving the sensitivity and accuracy of spectral feature extraction. Results show that the maize leaf chlorophyll content prediction model constructed using the (SE-VIs-WF)-RFR method has high accuracy, particularly in the V12 test set R... 2 With a mean value of 0.824 and an RMSE of 1.994, it can effectively and non-destructively predict the chlorophyll content of maize leaves, providing new ideas and methods for the rapid and accurate identification of maize leaves and leaves of other crops. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the SE attention mechanism used in this invention to fuse VIs-WF spectral features; Figure 2 The average spectral response curves of maize leaves at different growth stages; Figure 3 The correlation coefficient matrix of SPAD values ​​and VIs in maize leaves; Figure 4 The results of chlorophyll content prediction in maize leaves based on (SE-VIs-WF)-RFR are shown (a, b, c, and d are V6, V8, V12, and R1 periods, respectively). Detailed Implementation

[0019] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "a," "an," or "the," and similar words used in the specification and claims of this patent application do not express a limitation of quantity, but rather indicate the presence of at least one; terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, but do not exclude other elements or objects having the same function.

[0021] To more clearly describe the specific operation of the present invention, the accompanying drawings are provided. Figure 1 To be continued Figure 4 This embodiment is described as follows: This embodiment specifically includes the following steps: S1, Plant grouping: Select a number of corn plants and group them accordingly; S2. Raw Data Acquisition and Processing: The maize plants grouped in step S1 were tested, and hyperspectral data and chlorophyll content (SPAD value) of maize leaves at different growth stages of each plant were obtained. For the obtained maize leaf spectral data and SPAD value, abnormal bands and samples of spectral data were removed by combining PLSR and Monte Carlo method. Abnormal data of SPAD value were removed by k-means clustering and Euclidean distance method to obtain the processed test data. S3. Spectral data preprocessing: Based on the detection data processed in step S2, and using common spectral preprocessing methods and their combinations, combined with machine learning models, the chlorophyll content of maize leaves is predicted, and the optimal spectral preprocessing method and the maize leaf chlorophyll content prediction model framework are selected. S4. Optimal Vegetation Index (VIs) Selection: Based on the optimal preprocessed spectral curve selected in step S3, several spectral vegetation indices are extracted, and the optimal VIs data for constructing a chlorophyll content prediction model are compared and obtained. S5. Selection of Optimal Wavelet Feature (WF) Coefficients: Based on the optimal preprocessed spectral curve selected in step S3, eight wavelet basis functions were selected to decompose the spectral reflectance data at the scale of 1-1024. The wavelet coefficients with the best correlation under the processing of the eight wavelet basis functions were selected as wavelet features (WF) to construct a chlorophyll content prediction model. S6. Attention mechanism fusion of VIs-WF spectral features: The SE attention mechanism is used to compress and activate the best VIs index and best WF coefficient selected in steps S4 and S5, obtain the weight value corresponding to the spectral feature, and assign it to the spectral feature to realize weight update assignment, and finally obtain the weighted spectral feature. S7. Model Building: Based on the S3 selection modeling method, we constructed a prediction model for maize leaf chlorophyll content at different growth stages by using WF, VIs, VIs-WF and SE attention mechanism to fuse multiple feature variables (SE-VIs-WF) as independent variables and maize leaf chlorophyll content (SPAD value) as dependent variable.

[0022] In this embodiment, in step S1, several groups of maize plants are selected, and three nitrogen treatment control groups and two variety control groups are set up for the maize plants. The nitrogen treatment is 0 kg / hm. 2 200kg / hm 2 400kg / hm 2 The varieties used were Zhengdan 958 and Yuke 918, and all treatments were treated with 150 kg / hm² of phosphate fertilizer. 2 150 kg / hm of potassium fertilizer 2 It should be applied all at once during sowing.

[0023] In this embodiment, in step S2, hyperspectral data and chlorophyll content (SPAD value) of maize leaves at the six-leaf stage (V6), eight-leaf stage (V8), and twelve-leaf stage (V12) of each plant are obtained respectively. Three sampling points (upper, middle, and lower) are set at the top fully expanded leaves of maize V6, V8, and V12 and the ear position leaf of R1. The average value of the three measurements is taken as the final spectral reflectance and SPAD value of the leaf.

[0024] In this embodiment, the specific steps in step S2, which combine PLSR and Monte Carlo methods to remove abnormal bands and samples from spectral data, are as follows: Set the maximum number of components to 20, and use 10-fold cross-validation to select the optimal number of PLSR components; set the number of iterations N for the Monte Carlo method to 1000, set the outlier threshold to 3 times the standard deviation, fit the initial PLSR model using the optimal number of components, and calculate the initial residuals; perform N random samplings, randomly select samples and fit the PLSR model each time, calculate the residuals, and mark samples with residuals exceeding the threshold as outliers.

[0025] In this embodiment, when using k-means clustering combined with Euclidean distance to remove SPAD value outliers in step S2, it is assumed that the distance from the data point to the cluster center follows a normal distribution, and data points with a distance greater than "mean distance plus twice the standard deviation" are defined as outliers.

[0026] In this embodiment, there are 23 spectral vegetation indices in step S4, as shown below. Based on the VIs that satisfy the significance of p≤0.001 in periods V6, V8, V12, and R1, VIs with correlation coefficients greater than 0.6, 0.65, 0.75, and 0.75 are selected for modeling.

[0027] In this embodiment, the commonly used spectral preprocessing methods and their combinations in step S3 are as follows: based on the original spectral data (OR) and smoothing algorithm (SG), standard normal variable transformation (SNV), multivariate scattering correction (MSC), SG-SNV, and SG-MSC preprocessing methods, combined with partial least squares regression (PLSR), support vector regression (SVR), and random forest regression (RFR) methods, a maize leaf chlorophyll content prediction model is constructed to select the best preprocessing and modeling method.

[0028] In this embodiment, step S5 specifically involves selecting eight wavelet basis functions: gb4, haar, db4, bior1.5, mexh, rbio3.3, sym4, and coif2. 1 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 2 10 Wavelet coefficients at 10 specific scales were selected, and the decomposition scale and spectral band corresponding to the strongest correlation under 8 wavelet basis function processing at each period were screened to obtain the optimal wavelet coefficients, i.e., WF. The wavelet transform calculation formula is as follows: (1) In the formula, CWT f (s,τ) represents the result after wavelet transform, where s and τ are the scale and translation parameters, respectively, f(t) is the spectral reflectance of maize leaves after optimal preprocessing, and Ψ*(x) is the wavelet basis function.

[0029] As a preferred approach, step S6 involves the following steps: The SE attention mechanism is used to compress and activate the optimal VIs and WF selected in steps S4 and S5. The Squeeze operation performs global average pooling on the input VIs-WF spectral features Y to generate z. The Excitation operation uses weight W to generate the required weight information, and processes the Squeeze operation vector through two fully connected layers W1 and W2 to obtain the weight values ​​corresponding to the spectral features. The Scale operation assigns the generated weight vector s to the feature Y to update the weights, ultimately obtaining the weighted spectral feature Y'. The core calculation method is as follows: (2) (3) (4) In the formula, F sq F ex F scale These represent global average pooling, adaptive recalibration, and reweighting operations, respectively. c Let z be the c-th element.

[0030] In this embodiment, the specific steps of step S7 are as follows: based on the modeling method selected in step S3, WF, VIs, VIs-WF and SE-VIs-WF are used as independent variables, and the chlorophyll content (SPAD value) of maize leaves is used as the dependent variable. The training set and the test set are divided in an 8:2 ratio to construct the maize leaf chlorophyll content prediction model for periods V6, V8, V12 and R1.

[0031] In this embodiment, the detection device mainly includes a Field Spec Hand Held 2™ portable ground object spectrometer and a SPAD-502Plus chlorophyll meter. The data processing and modeling methods are all MATLAB-R2022a, and the data analysis software is OriginPro-2024b.

[0032] Specific steps: Figure 1This diagram illustrates the SE attention mechanism fusion of VIs-WF spectral features used in this invention. The spectral reflectance of maize leaves was measured using a FieldSpecHandHeld2™ portable ground-based spectrometer from ASD (Americas). Spectral measurements were conducted on clear, windless days from 9:00 AM to 5:00 PM. Healthy, disease-free maize plants were selected from each experimental area, and hyperspectral data were obtained from the fully expanded top leaves at stages V6, V8, and V12, and the ear-position leaves at stage R1. Standard white board calibration was performed before each sampling. Measurements avoided the maize leaf veins. Each leaf was measured three times (upper, middle, and lower) and marked. The average value was taken as the original spectral reflectance of the sample. The chlorophyll content of maize leaves, i.e., the SPAD value, was obtained using a SPAD-502Plus chlorophyll meter. Measurement calibration was performed before each sampling. Measurements were taken simultaneously with hyperspectral data acquisition, at the same location as the hyperspectral data. Each sample point was measured three times, and the average value was taken as the leaf SPAD value.

[0033] Based on maize leaf spectral data and SPAD values, abnormal bands and samples in the spectral data were removed using PLSR and Monte Carlo methods. Abnormal SPAD values ​​were further removed using k-means clustering and Euclidean distance methods. The spectral data corresponding to abnormal SPAD values ​​and the SPAD values ​​corresponding to abnormal spectral data were simultaneously removed to obtain the original experimental data. The table below shows the original SPAD values ​​of maize leaves at different time periods. The maximum value ranged from 60.400 to 67.500, and the minimum value ranged from 41.900 to 47.900. The coefficients of variation were all less than 10%, which is consistent with statistical characteristics.

[0034] Figure 2 The spectral reflectance curves of maize leaves at different growth stages show that the shapes of the spectral reflectance curves of maize leaves at different growth stages are basically similar, which is consistent with the spectral characteristics of general healthy green vegetation.

[0035] Based on the original spectral data (OR) and preprocessing methods such as SG, SNV, MSC, SG-SNV, and SG-MSC, a prediction model for maize leaf chlorophyll content was constructed using PLSR, SVR, and RFR machine learning algorithms. The table below shows the results of the PLSR, SVR, and RFR models built based on spectral data after different preprocessing steps: Among them, the PLSR, SVR, and RFR models constructed from maize leaf spectral data preprocessed with MSC showed the best performance. In period V6, the prediction set R of the PLSR, SVR, and RFR models constructed from maize spectral data preprocessed with MSC was the highest. 2The values ​​were 0.483, 0.511, and 0.534 respectively, with the RFR model showing the best prediction performance. During the V8 period, the prediction sets R of the PLSR, SVR, and RFR models constructed based on MSC-preprocessed maize spectral data were... 2 The values ​​were 0.608, 0.697, and 0.671, respectively, with the SVR model showing the best prediction performance. In period V12, the prediction sets R of the PLSR, SVR, and RFR models constructed based on MSC-preprocessed maize spectral data were... 2 The values ​​were 0.698, 0.699, and 0.683, respectively, with the SVR model showing the best prediction performance. In the R1 period, the prediction sets R of the PLSR, SVR, and RFR models constructed based on MSC-preprocessed maize spectral data were... 2 The values ​​were 0.672, 0.676, and 0.684, respectively, with the RFR model showing the best predictive performance. Therefore, this embodiment further analyzes the spectral data after MSC preprocessing and uses the SVR and RFR methods to construct predictive models for chlorophyll content in maize leaves at different stages.

[0036] Correlation analysis was performed between the SPAD values ​​of maize leaves and 23 selected VIs, and the Pearson correlation coefficient matrix was obtained as follows: Figure 3 As shown, there are 17, 20, 22 and 22 VIs that satisfy the significance of p≤0.001 in periods V6, V8, V12 and R1, respectively. The maximum correlation coefficients are 0.620, 0.692, 0.784 and -0.575, respectively, and the average correlation coefficients are 0.373, 0.487, 0.663 and 0.637, respectively. To improve model accuracy, this invention selects VIs with a significance p≤0.001 for periods V6, V8, V12, and R1, respectively, and models them with correlation coefficients greater than 0.6, 0.65, 0.75, and 0.75, namely CIrededge (0.617), CVI (0.620), NDRE (0.655), LCI (0.651), CIrededge (0.692), CVI (0.690), GDVI (0.759), CCI (0.777), CIrededge (0.760), CVI (0.784), and ENDVI (-0.757), CIrededge (0.762).

[0037] Eight wavelet basis functions (gb4, haar, db4, bior1.5, mexh, rbio3.3, sym4, and coif2) were used to decompose the spectral data after MSC processing into a 1-1024 scale. However, the 1-1024 scale decomposition resulted in excessive data volume, so two were selected. 1 2 2 2 3 2 4 2 5 26 2 7 2 8 2 9 2 10 Ten specific scales (labeled S1-S10) were used to analyze the correlation between wavelet coefficients and maize leaf SPAD values ​​at different scales using the Pearson method. The strongest correlations were found at V6, V8, V12, and R1, reaching 0.644, 0.741, 0.819, and 0.790, respectively. Compared with VIs, WF showed a significant improvement in the correlation between WF and maize leaf SPAD values. Since wavelet coefficients under the same wavelet basis function treatment exhibited strong collinearity after wavelet transformation reconstruction of spectral data, peak extraction was used to screen the decomposition scales and spectral bands corresponding to the strongest correlations under eight wavelet basis function treatments at each period, obtaining the optimal wavelet coefficient, WF, as shown in the table below. Based on the above VIs, WF, and fused spectral characteristic variables, SVR and RFR machine learning regression algorithms were used to establish prediction models for maize leaf chlorophyll content. The model results are shown in the table below: Compared to the SVR method, the maize leaf chlorophyll content prediction model based on RFR performs better. Specifically, the prediction set R of the maize leaf chlorophyll content prediction model at different time stages constructed based on VIs is superior. 2 The values ​​were 0.453, 0.539, 0.642, and 0.559, respectively; the prediction set R of the maize leaf chlorophyll content prediction model constructed based on WFs at different stages was... 2 The values ​​were 0.530, 0.768, 0.744, and 0.672, respectively; the prediction set R of the maize leaf chlorophyll content prediction model constructed based on VIs-WF at different stages was... 2 The values ​​are 0.543, 0.694, 0.784, and 0.682, respectively; the model prediction set R based on SE-VIs-WF is... 2 The values ​​were 0.570, 0.760, 0.824, and 0.742, respectively. It can be seen that the performance of the models built based on different spectral features is V12>V8>R1>V6, SE-VIs-WF>VIs-WF>WF>VIs. The V12 period is more suitable for predicting the chlorophyll content of maize leaves. Multispectral feature fusion is more advantageous than single spectral feature modeling.

[0038] Based on the SE-VIs-WF-constructed RFR model of maize leaf chlorophyll content at various stages, the relationship between measured and predicted values ​​is as follows: Figure 4 Compared to the performance of the RFR model built based on VIs-WF, the prediction set R of the model in periods V6, V8, V12, and R1 is significantly better. 2The improvements of 4.972%, 9.510%, 5.102%, and 9.090% respectively indicate that the SE attention mechanism can effectively optimize the weights of each spectral feature and increase the effective utilization of each spectral feature in the model. Compared with the RFR model constructed from maize leaf spectral data at various stages after MSC preprocessing, the model prediction sets R for stages V6, V8, V12, and R1... 2 The improvements were 6.742%, 13.264%, 20.644%, and 8.480% respectively, while the number of spectral feature variables was reduced from 551 to 10, 12, 12, and 12 respectively, effectively reducing model complexity and improving model computation efficiency.

[0039] It should be noted that although the present invention has been described through the above embodiments, the present invention may have many other embodiments. Without departing from the spirit and scope of the present invention, those skilled in the art can obviously make various corresponding changes and modifications to the present invention, but all such changes and modifications should fall within the scope of protection of the appended claims and their equivalents.

Claims

1. A method for fusing and inverting spectral features of chlorophyll content of corn leaves, characterized in that: The method comprises the following steps: S1, grouping plants: Selecting a plurality of corn plants and grouping them; S2, original data acquisition and processing: Detecting the corn plants grouped in step S1 and acquiring corn leaf hyperspectral data and chlorophyll content (SPAD value) of each plant at different growth stages, in combination with PLSR and Monte Carlo method, removing abnormal spectral data bands and samples, removing abnormal SPAD value data by using k-means clustering and Euclidean distance method, and obtaining processed detection data; S3, spectral data preprocessing: According to the processed detection data in step S2, based on the commonly used spectral preprocessing methods and their combinations, the best spectral preprocessing method and the corn leaf chlorophyll content prediction model framework are selected; S4, selection of the best vegetation index (VI): Based on the best preprocessed spectral curve selected in step S3, a plurality of spectral vegetation indexes are extracted, compared and the best VI data for constructing the chlorophyll content prediction model are obtained; S5, selection of the best wavelet feature (WF) coefficient: Based on the best preprocessed spectral curve selected in step S3, the spectral reflectance data is decomposed under 1-1024 scales by selecting 8 wavelet basis functions, and the best wavelet coefficient with the best correlation under the processing of the 8 wavelet basis functions is selected as the wavelet feature (WF) for constructing the chlorophyll content prediction model; S6, attention mechanism fusion VIs-WF spectral feature: The SE attention mechanism is used to compress and stimulate the best VIs index and the best WF coefficient selected in steps S4 and S5, the weight value corresponding to the spectral feature is obtained, and the weight value is assigned to the spectral feature to realize weight update assignment, and finally the weighted spectral feature is obtained; S7, model construction: Based on the corn leaf chlorophyll content prediction model framework selected in step S3, the WF, VIs, VIs-WF and SE attention mechanism fusion multi-feature variable (SE-VIs-WF) are respectively taken as the independent variable, and the corn leaf chlorophyll content (SPAD value) is taken as the dependent variable, and the corn leaf chlorophyll content prediction model at different growth stages is constructed to realize the prediction of the corn leaf chlorophyll content.

2. The method according to claim 1, characterized in that: In step S2, the corn leaf hyperspectral data and chlorophyll content (SPAD value) of each plant at six-leaf stage (V6), eight-leaf stage (V8) and twelve-leaf stage (V12) are acquired, and three sampling points of upper, middle and lower are set on the fully expanded leaves at the top of the corn V6, V8 and V12 and the R1 ear leaves, and the average value of the three points is taken as the final spectral reflectance and SPAD value of the leaf.

3. The method according to claim 1, characterized in that: In step S2, the specific steps of removing spectral data abnormal bands and samples by combining PLSR and Monte Carlo method are as follows: setting the maximum component number as 20, using 10-fold cross validation to select the best PLSR component number; setting the iteration number N of Monte Carlo method as 1000, setting the threshold value of abnormal value as 3 times of standard deviation, using the best component number to fit the initial PLSR model, calculating the initial residual; performing N times of random sampling, randomly selecting samples each time and fitting the PLSR model, calculating the residual, and marking the samples with residual exceeding the threshold value as abnormal values.

4. The method according to claim 1, characterized in that: In step S2, when the three methods of k-means clustering combined with Euclidean distance are used to remove abnormal data of SPAD value, it is assumed that the distance of data points to the cluster center obeys normal distribution, and the data points with distance greater than "average distance plus twice standard deviation" are defined as outliers.

5. The method according to claim 1, characterized in that: In step S3, the commonly used spectral pretreatment methods and their combinations are based on original spectral data (OR) and smoothing algorithm (SG), standard normal variable transformation (SNV), multivariate scatter correction (MSC), SG-SNV, SG-MSC pretreatment method, combined with partial least squares regression (PLSR), support vector regression (SVR) and random forest regression (RFR) method to construct the prediction model of corn leaf chlorophyll content to select the best prediction model framework of corn leaf chlorophyll content.

6. The method according to claim 1, characterized in that: In step S4, the specific spectral vegetation index is 23, and based on V6, V8, V12 and R1 period VIs meeting p≤0.001 significance, VIs with correlation coefficient greater than 0.6, 0.65, 0.75 and 0.75 are selected for modeling.

7. The method according to claim 1, characterized in that: The specific steps of step S5 are as follows: Select eight wavelet basis functions—gb4, haar, db4, bior1.5, mexh, rbio3.3, sym4, and coif2—in 2 1 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 2 10 Wavelet coefficients at 10 specific scales were selected, and the decomposition scale and spectral band corresponding to the strongest correlation under 8 wavelet basis function processing at each period were screened to obtain the optimal wavelet coefficients, i.e., WF. The wavelet transform calculation formula is as follows: （1） In the formula, CWT f (s, τ) represents the result of wavelet transform, wherein s and τ are scale and translation parameters, respectively, f(t) is the optimal pretreated corn leaf spectral reflectance, and Ψ*(x) is a wavelet base function.

8. The method according to claim 1, characterized in that: The specific steps of step S6 are: using SE attention mechanism to compress and excite the best VIs and WF selected in step S4 and step S5, Squeeze operation performs global average pooling processing on input VI-WF spectral features Y to generate z; Excitation operation generates the required weight information by using weight W, processes the Squeeze operation vector through two fully connected layers W1 and W2 to obtain the weight value corresponding to the spectral features; Scale operation assigns the generated weight vector s to the feature Y to realize weight update assignment, and finally obtains the weighted spectral feature Y', and the core calculation method is as follows: （2） （3） （4） where F sq , F ex , F scale denote global average pooling, adaptive re-normalization and re-weighting operations, respectively, z c denotes the c-th element of z.

9. The method according to claim 1, characterized in that: The specific steps of step S7 are: based on the modeling method selected in step S3, taking WF, VIs, VIs-WF and SE-VIs-WF as independent variables, and corn leaf chlorophyll content (SPAD value) as dependent variable, the training set and test set are divided according to 8:2, and the prediction model of corn leaf chlorophyll content at V6, V8, V12 and R1 stage is constructed.

Images