Method for estimating rice yield related traits based on spectrum-texture-dimension reduction-machine learning algorithm

Abstract

The invention relates to the technical field of rice yield related character estimation, in particular to a method for estimating rice yield related characters based on a spectrum-texture-dimension reduction-machine learning algorithm. The method comprises the following steps: acquiring canopy hyperspectral data of rice in each period by virtue of a hyperspectral imager carried by an unmanned aerial vehicle; and synchronously collecting each character data of the rice. Hyperspectral data is preprocessed through a smoothing algorithm, dimension reduction is carried out on the preprocessed spectral data, and sensitive wavebands corresponding to all characters are screened out. Sensitive wave band spectral reflectivity is used as an input value, each actually measured character index is used as an output value, each character estimation model is constructed, and an optimal combination of spectrum, dimension reduction and machine learning is obtained. And extracting texture features of the sensitive wave band to construct a high-precision estimation model. The optimization combination of spectrum + texture + dimension reduction + machine learning can significantly improve the precision of the estimation model of each rice character, and provides scientific basis and technical guidance for accurate diagnosis of rice yield related character phenotypes.

Description

Technical Field

[0001] The present invention relates to the technical field of estimating rice yield-related traits, and specifically relates to a method for estimating rice yield-related traits based on a spectrum-texture-dimension reduction-machine learning algorithm. Background Art

[0002] The leaf nitrogen concentration (LNC), leaf area index (LAI), above-ground biomass (AGB), and grain yield (GY) of rice are the most common monitoring and prediction indicators for characterizing the growth of high-yield and high-efficiency rice. LAI refers to the total green leaf area per unit area and is an important indicator for measuring the photosynthetic capacity of crops and characterizing the growth trend of crop populations. AGB is an important indicator for regulating the quality of crop populations and plays an important role in light energy utilization, dry matter accumulation, yield formation, etc. Some studies have shown that by optimizing the spatial distribution of LAI and AGB, the maximum canopy photosynthesis and grain yield of crops can be increased. Crop yield prediction is crucial for food production, and accurate prediction of crop yield provides a reference basis for national food allocation and supply. Previous studies have shown that there are obvious differences in the dimension reduction methods for different monitoring objects and trait indicators. For different crops or different trait indicators, the most suitable model algorithms are also different. In addition, single spectral data may not be able to provide comprehensive information on the growth status of crops. Spectral data mainly reflects the biochemical characteristics of crops, while texture features contain the spatial structure and morphological information of crops. Therefore, the effective combination of spectral features and texture features is of great significance for improving the accuracy of estimation models for various trait indicators of crops. In the construction of spectral estimation models for crop biochemical parameters, different data dimension reduction techniques and machine learning algorithms have their own advantages, but they still face great challenges in the selection of the best methods. So far, there have been few reports on the study of estimating rice yield-related traits based on combined models such as spectrum + texture + dimension reduction + machine learning. Therefore, conducting research on hyperspectral prediction of multiple indicator traits of rice growth is of great significance for promoting the healthy growth of rice and enhancing food security. However, there is currently a lack of a method in the existing technologies that can accurately predict rice yield-related traits for use. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for estimating rice yield-related traits based on a spectrum-texture-dimension reduction-machine learning algorithm to solve the technical problems raised in the above background art.

[0004] To solve the above problems, the present invention adopts the following technical solutions for solution:

[0005] In a first aspect, the present invention provides a method for estimating rice yield-related traits based on a spectral-texture-dimension reduction-machine learning algorithm, which uses the Pearson correlation coefficient method (PCC), the successive projections algorithm (SPA), and the competitive adaptive reweighted sampling method (CARS) to extract the spectral sensitive bands of rice leaf nitrogen concentration (LNC), leaf area index (LAI), aboveground biomass (AGB), and grain yield (GY), constructs an estimation model for rice yield-related traits based on artificial neural network (ANN), support vector machine regression (SVR), one-dimensional convolutional neural network (1DCNN), and long short-term memory network (LSTM), and obtains the optimal combination of spectrum + dimension reduction + machine learning; by extracting the texture features of the corresponding sensitive bands, a high-precision estimation model for rice yield-related traits such as spectrum + texture + dimension reduction + machine learning is constructed; the specific steps are as follows:

[0006] S1. Measure the rice LNC, LAI, and AGB at the early-season heading stage, late-season initial panicle differentiation stage, and late-season heading stage of the current year and the early-season initial panicle differentiation stage of the following year, and measure the GY at the maturity stage of the current year's early and late seasons and the early season of the following year; obtain the rice canopy hyperspectral data during the same period;

[0007] S2. Preprocess the hyperspectral image, and use the spectral average value of the region of interest in each plot as the original spectral reflectance of the plot;

[0008] S3. Use the Savitzky-Golay convolutional smoothing algorithm to smooth and denoise the original spectral data;

[0009] S4. Use a variety of dimension reduction methods to determine the sensitive bands of each trait index according to the correlation between the spectral reflectance and each trait index;

[0010] S5. Input the sensitive bands into the model of the selected machine learning algorithm to predict the trait index;

[0011] S6. For each trait index, construct a data set using the corresponding sensitive bands and measured values, and divide the data set into a training set and a validation set according to a preset ratio to verify the models of various machine learning algorithms;

[0012] S7. Select the optimal yield-related trait estimation model from the constructed models, and on this basis, combine the texture features of the sensitive bands to estimate LNC, LAI, AGB, and GY;

[0013] S8. Evaluate the verification results of the model according to the coefficient of determination R 2 and the root mean square error RMSE, and determine the selected dimension reduction method and selected machine learning algorithm corresponding to each trait index.

[0014] Preferably, in the step S1, for the method for estimating rice yield-related traits based on spectral-texture-dimension reduction-machine learning algorithm, it is characterized in that for the acquisition of rice canopy hyperspectral data, a multi-rotor drone M300RTK is used to carry an X20P airborne hyperspectral imager, with a spectral range of 350-1000 nm, a resolution of 4 nm, 164 effective bands, and a drone flight altitude of 50 m; the measurement time is from 10:00 to 14:00 Beijing time, and the weather is sunny.

[0015] Preferably, in the step S2, through image stitching, radiometric calibration, atmospheric correction, orthorectification, image fusion, geometric correction, band normalization, etc., the spectral average value of the region of interest (ROI) of each plot is extracted using the Region of Interest (ROI) tool as the original spectral reflectance of the plot.

[0016] Preferably, in the step S3, to overcome the influence of environmental factors to generate noise bands and improve the prediction accuracy of the estimation model, in this study, the original spectrum is processed by the Savitzky-Golay convolution smoothing algorithm; the Savitzky-Golay convolution smoothing algorithm performs polynomial least squares fitting on the data within the moving window through a polynomial, and its essence is a weighted average method that emphasizes the central role of the center point more, and can better retain the original spectral information while reducing noise; after parameter tuning and optimization, the window_length of the Savitzky-Golay convolution smoothing method is set to 9, and the polyorder is set to 2.

[0017] Preferably, in the step S4, among the sensitive bands of each trait index determined, the top 3 bands with the strongest correlation of the Pearson correlation coefficient method (PCC) and the competitive adaptive reweighted sampling method (CARS), and the top 5 bands of the successive projections algorithm (SPA) are selected as the sensitive bands; the best sensitive bands of LNC screened by the PCC method are 410 nm, 414 nm, and 418 nm; the best sensitive bands of LAI are 422 nm, 486 nm, and 678 nm; the best sensitive bands of AGB are 370 nm, 422 nm, and 678 nm; the best sensitive bands of GY are 650 nm, 654 nm, and 658 nm; the best sensitive bands of LNC screened by the SPA method are 398 nm, 410 nm, 478 nm, 498 nm, and 590 nm; the best sensitive bands of LAI are 350 nm, 378 nm, 394 nm, 486 nm, and 890 nm; the best sensitive bands of AGB are 402 nm, 422 nm, 450 nm, 718 nm, and 898 nm; the best sensitive bands of GY are 630 nm, 654 nm, 658 nm, 690 nm, and 742 nm; the best sensitive bands of LNC screened by the CARS method are 406 nm, 410 nm, and 414 nm; the best sensitive bands of LAI are 402 nm, 674 nm, and 678 nm; the best sensitive bands of AGB are 366 nm, 370 nm, and 418 nm; the best sensitive bands of GY are 350 nm, 398 nm, and 898 nm.

[0018] Preferably, in the step S5, the multiple machine learning algorithms include artificial neural networks (ANN), support vector regression (SVR), one-dimensional convolutional neural networks (1D CNN), and long short-term memory networks (LSTM);

[0019] The main parameters of the artificial neural network, namely "activation", "alpha", "hidden_layer_sizes", "learning_rate", "max_iter", "momentum", "solver", and "tol", are set to relu, 0.0001, 100, adaptive, 200, 0.7, adam, and 0.00001 respectively;

[0020] The main parameters of the support vector machine, namely "kernel", "degree", "gamma", "coef0", "tol", "C", "Epsilon", "shrinking", "cache_size", "verbose", "max_iter", are respectively set to rbf, 3, auto, 0.0, 0.001, 1.0, 0.1, True, 200, False, -1;

[0021] The main parameters of the one-dimensional convolutional neural network, namely "in_channels", "out_channels", "kernel_size", "padding", "num_epochs", are respectively set to 1, 16, 3, 1, 10;

[0022] The main parameters of the long short-term memory network LSTM, namely "input_size", "output_size", "hidden_size_temp", "num_layer_temp", "drop_temp", are respectively set to 3, 1, 64, 1, 0.4.

[0023] Preferably, in step S6, the train_test_split function in the sklearn.model_selection module in python is used to divide the training set and the test set to verify the model; the ratio of the test set to the validation set is 7:3.

[0024] Preferably, in step S7, the texture features of each sensitive band are extracted by using the gray-level co-occurrence matrix method. The gray-level co-occurrence matrix method (Gray-Level Co-occurrence Matrix, GLCM) of ENVI 5.3 is used to screen the selected sensitive bands, and the texture features of each sensitive band are screened out by using a window size of 3×3 resolution; the texture features include mean, variance, homogeneity, contrast, heterogeneity, entropy, second-order moment, and correlation; the region of interest is demarcated for each band's texture information image, and the texture value of this region can be extracted.

[0025] Preferably, in step S8, the evaluation index of the estimation model adopts the coefficient of determination R 2 and the root mean square error RMSE. The calculation formulas of each evaluation index are as follows:

[0026]

[0027] Among them, n is the number of samples, is the predicted value, is the average value, and y iis the actual value.

[0028] Preferably, the preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the leaf nitrogen concentration are the successive projections algorithm and the artificial neural network, respectively;

[0029] The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the leaf area index are the successive projections algorithm and the artificial neural network, respectively;

[0030] The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the aboveground biomass are the competitive adaptive reweighted sampling method and the artificial neural network, respectively;

[0031] The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the rice yield are the competitive adaptive reweighted sampling method and the artificial neural network, respectively.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. Based on the spectral data of rice at the key growth stages, the present invention uses different data dimensionality reduction methods to determine the sensitive bands corresponding to each trait index of rice, and uses the sensitive bands and the corresponding texture features as the feature inputs of different machine learning models to achieve high-precision prediction of each trait index.

[0034] 2. The present invention conducts practical research on the influence of different dimensionality reduction methods and different machine learning algorithms on the prediction accuracy, and determines the "spectrum + texture + dimensionality reduction + machine learning" model combination with the best prediction effect corresponding to each trait index of rice based on the determination coefficient R 2 and the root mean square error RMSE.

[0035] 3. On the basis of the optimal selection of the preferred dimensionality reduction method and the preferred machine learning algorithm, the present invention combines texture features to further improve the prediction accuracy of the model, providing a scientific basis for the non-destructive real-time prediction of the phenotypic precision diagnosis of rice yield-related traits of indica rice in South China. Description of the Drawings

[0036] Figure 1 is a general view of the test site in the present invention;

[0037] Figure 2 is a schematic diagram of the construction of the model of the machine learning algorithm in the present invention;

[0038] Figure 3 is the spectroreflectance map after SG smoothing in the present invention;

[0039] Figure 4 is the measurement result map of each trait index of rice in the present invention;

[0040] Figure 5 It is the correlation analysis diagram of each trait index of rice in the present invention;

[0041] Figure 6 It is the correlation coefficient curve diagram of spectral reflectance and trait indexes (biological indexes) of rice in the present invention;

[0042] Figure 7 It is the model verification result diagram of each machine learning algorithm in the present invention;

[0043] Figure 8 It is the model verification result diagram of combining texture features on the basis of optimizing machine learning algorithms in the present invention. Detailed implementation manners

[0044] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation to the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise stated, the meaning of "a plurality" is two or more. In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installation", "connection", and "connection" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the internal communication of two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood through specific situations.

[0045] The present invention will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0046] Embodiment 1

[0047] This embodiment provides a method for estimating rice yield-related traits based on spectral-texture-dimension reduction-machine learning algorithms, specifically including:

[0048] Step S1: At the critical growth stages of rice (the early-season HD stage in 2022, the late-season PI stage, the late-season HD stage in 2022, and the early-season PI stage in 2023, where PI and HD represent the initial stage of panicle initiation and the heading stage respectively), use an unmanned aerial vehicle (UAV) equipped with a hyperspectral imager to collect spectral data of each plot, and randomly select plants within each plot to measure trait indicators, where the trait indicators include leaf nitrogen concentration (Leaf Nitrogen Concentration, LNC, g / kg), leaf area index (Leaf Area Index, LAI), above-ground biomass (Above Ground Biomass, AGB, g / m2), and grain yield (Grain yield, GY, t / ha);

[0049] Regarding the said Step S1, in a specific embodiment, the tested rice varieties used for the rice in Step S1 are the high-quality indica rice varieties "Meixiangzhan 2 (V1)" and "Nanjingxiangzhan (V2)" with a relatively large planting area in South China. The experiment in this embodiment was carried out at the experimental base of the Guangdong Academy of Agricultural Sciences in Zhongluotan Town, Baiyun District, Guangzhou City, Guangdong Province in the early and late seasons of 2022 and the early season of 2023 ( Figure 1 ). The physical and chemical properties of the experimental field soil are as follows: PH 5.95, organic matter 22.48 g / kg, total nitrogen 1.29 g / kg, total phosphorus 0.42 g / kg, total potassium 8.43 g / kg, available nitrogen 58.03 mg / kg, available phosphorus 6.49 mg / kg, and available potassium 47.00 mg / kg.

[0050] Figure 1 Among them, N0, N1, N2, N3, N4 represent treatments of 0 kg / ha, 60 kg / ha, 120 kg / ha, 180 kg / ha, and 240 kg / ha respectively; V1, V2 represent Meixiangzhan 2 and Nanjingxiangzhan respectively.

[0051] The experiment adopted a split-plot design, with nitrogen level as the main plot and variety as the sub-plot. Five nitrogen fertilizer levels were set, namely 0, 60, 120, 180, and 240 kg N / ha. Transplanting was carried out by the method of manual wire-pulling. The transplanting density was 20 cm × 20 cm, with 2 seedlings per hill, and there were 3 replicates. Nitrogen fertilizer was added in the form of urea, and phosphate fertilizer and potassium fertilizer were applied in the forms of superphosphate and potassium chloride respectively. Nitrogen fertilizer was applied according to the ratio of basal fertilizer: tillering fertilizer: panicle fertilizer = 5:2:3. The basal fertilizer was applied one day before transplanting, the tillering fertilizer was applied 15 days after transplanting, and the panicle fertilizer was applied at the PI stage of rice. Phosphate fertilizer and potassium fertilizer were applied according to the standards of P2O5 54 kg / ha and K2O 144 kg / ha. All phosphate fertilizer was used as basal fertilizer, and half of the potassium fertilizer was used as basal fertilizer and the other half as panicle fertilizer. The plots were wrapped with plastic film (PVP) around the ridges to prevent cross-fertilization between plots. Field management was carried out according to the "Three-Control" fertilization technical regulations for rice, and pests and diseases were strictly controlled. Hyperspectral data of plot canopy height and data of four agronomic trait indicators (LNC, LAI, AGB, and GY) were collected at the early-season HD stage in 2022, the late-season PI stage, the late-season HD stage, and the early-season PI stage in 2023 respectively. Among them, PI and HD represent the initial stage of young panicle differentiation and the heading stage respectively.

[0052] In a specific embodiment, the steps for measuring the trait indicators are specifically as follows: After spectral measurement, randomly select 12 representative rice plants in the experimental plot, remove the roots, separate the stems and leaves at the PI stage, and separate the stems, leaves, and panicles at the HD stage. Measure the leaf area (S) with a leaf area meter. Place the stems, leaves, and panicles in an oven at 105 °C for 30 min for deactivation, and then dry them at a constant temperature of 75 °C until constant, and record the dry weights (w1, w2, w3) respectively.

[0053] Leaf Nitrogen Concentration (LNC, g / kg): After grinding, the nitrogen concentration of rice leaves was determined by the Kjeldahl method.

[0054] The calculation method of Leaf Area Index (LAI) is as shown in Equation (1):

[0055]

[0056] In Equation (1), C is the number of sampled hills, and D is the planting density, which is expressed as the number of rice hills per unit area.

[0057] Above Ground Biomass (AGB, g / m 2 ): It was calculated according to the planting density at the sampling point and the dry weight of rice sampling. The calculation method is as shown in Equation (2):

[0058]

[0059] In formula (2), C is the number of sampling holes, and D is the planting density.

[0060] Grain yield (GY, t / ha): At maturity, 125 rice plants in each plot were actually harvested (5 m 2 ), and the grain was air-dried. About 100 g was taken and dried at 105 °C for 48 h to measure the moisture content, and the grain was converted into the grain yield with a moisture content of 14%.

[0061] In a specific embodiment, in the step of using a drone to carry a hyperspectral imager to collect spectral data of each plot, for the collection of hyperspectral data of the rice canopy, a multi-rotor drone M300RTK is used to carry an X20P airborne hyperspectral imager. The spectral range is 350 - 1000 nm, the resolution is 4 nm, the number of effective bands is 164, and the flight altitude of the drone is 50 m. The measurement time is from 10:00 to 14:00 Beijing time, and the weather is clear.

[0062] Step S2: The obtained hyperspectral images are subjected to image stitching, radiometric calibration, atmospheric correction, orthorectification, image fusion, geometric correction, and band normalization, etc. The spectral average value of the region of interest (ROI) of each plot is extracted using the Region of Interest (ROI) tool as the original spectral reflectance of the plot.

[0063] Step S3: The Savitzky-Golay convolution smoothing algorithm is used to perform smoothing and noise reduction processing on the original spectral data;

[0064] Regarding the above step S3, in order to overcome the influence of environmental factors to generate noise bands and improve the prediction accuracy of the estimation model, in this embodiment, the original spectrum is processed by the Savitzky-Golay convolution smoothing algorithm. The Savitzky-Golay convolution smoothing algorithm performs polynomial least squares fitting on the data within the moving window through a polynomial. Its essence is a weighted average method that emphasizes the central role of the center point more, and can better retain the original spectral information while reducing noise. After parameter tuning and optimization, the window_length of the Savitzky-Golay convolution smoothing method is set to 9, and the polyorder is set to 2.

[0065] Step S4: Using a variety of dimensionality reduction methods, according to the correlation between the spectral reflectance and each trait index, determine the sensitive bands of each trait index;

[0066] In a specific embodiment, through comparative analysis, three most representative dimensionality reduction methods are adopted in this embodiment for analysis, namely Pearson Correlation Coefficient (PCC), Successive Projections Algorithm (SPA), and Competitive Adaptive Reweighted Sampling (CARS). PCC is a method based on the partial least squares regression model for analysis. It mainly calculates and analyzes the correlation between the spectral data corresponding to each band and the rice biological index data, so as to screen out the characteristic band combinations with relatively large correlations. SPA uses vector projection to select the spectral variables with the least amount of information, solves problems such as information cross-overlap and collinearity in the spectrum, reduces the number of variables in modeling, and thus speeds up the modeling efficiency. CARS simulates Darwin's "survival of the fittest" evolutionary theory. Through the adaptive reweighted sampling technique, it selects the wavelengths with relatively large absolute values of the regression coefficients calculated by the partial least squares method, and uses cross-validation to select the combination set with the lowest root mean square error value.

[0067] Step S5: Input the sensitive bands combined with texture features into the preferred machine learning algorithm to predict the trait indicators.

[0068] Regarding the models of multiple machine learning algorithms in step S5, in this embodiment, four algorithms, namely Artificial Neural Networks (ANN), Support Vector Regression (SVR), 1D Convolutional Neural Networks (1D CNN), and Long Short-Term Memory (LSTM), are used to establish hyperspectral prediction models for 4 rice trait indicators.

[0069] Among them, ANN was first proposed by psychologist McCulloch and mathematician Pitts in 1943. By constructing the M-P model, they combined the working principle of neurons with logical operations, laying a theoretical foundation for the development of artificial neural networks. ANN is a computational model that mimics the structure and function of the human brain neuron network. It consists of a large number of nodes (or called "neurons"), which are usually arranged in layers, such as the input layer, hidden layer, and output layer. Each node receives inputs from the nodes in the previous layer, performs a weighted sum, and then generates an output through a non-linear activation function and passes it to the next layer. ANN learns the complex relationships and patterns between input data by adjusting the connection weights between neurons.

[0070] The main parameters of the artificial neural network are set as follows:

[0071] activation='relu',

[0072] alpha = 0.0001,

[0073] hidden_layer_sizes = 100,

[0074] learning_rate='adaptive',

[0075] max_iter = 200,

[0076] momentum = 0.7,

[0077] solver='adam', tol = 1e-05

[0078] SVR is widely used in the fields of machine learning, artificial intelligence, big data, etc. It itself solves binary classification problems. SVR finds a regression plane to make the distance from all data in a set to this plane the closest. This algorithm is set with a kernel function and can flexibly solve various non-linear regression problems. The SVR model formula is as follows:

[0079]

[0080] In the formula, i represents the i-th sample point, and ai is the Lagrange coefficient; represents the mean of the random variable, the predicted value; M represents the sequence of data vectors; b is the threshold, representing the kernel function of the SVM model, and T represents the state;

[0081] The main parameters of the SVR are set as follows:

[0082] kernel='rbf'

[0083] degree = 3,

[0084] gamma='auto',

[0085] coef0 = 0.0,

[0086] tol = 1e-3,

[0087] C = 1.0,

[0088] Epsilon = 0.1,

[0089] shrinking = True,

[0090] cache_size = 200,

[0091] verbose = False,

[0092] max_iter = -1

[0093] The prototype of the 1D CNN, Le Net-5, was first proposed in 1998. This model is based on the convolutional and pooling network structures and is trained in combination with the BP algorithm. The convolutional neural network can automatically extract data features without manual intervention, has strong robustness and fault tolerance, and is convenient for training and optimization. At the same time, it also has characteristics such as local perception, parameter sharing, and multi-level feature abstraction.

[0094] The main parameter settings of the 1D CNN model are as follows:

[0095] in_channels = 1,

[0096] out_channels = 16,

[0097] kernel_size = 3,

[0098] padding = 1

[0099] num_epochs = 10

[0100] LSTM is a recurrent neural network (RNN) with a memory function and specifically designed for processing time series data, first proposed by Hochreiter et al. It solves the problems of gradient disappearance and gradient explosion caused by the long-term dependence mechanism in traditional recurrent neural networks (RNNs) when processing long time series, especially the problem of gradient disappearance.

[0101] The key parameter settings of the LSTM are as follows:

[0102] input_size = 3,

[0103] output_size = 1,

[0104] hidden_size_temp = 64,

[0105] num_layer_temp = 1,

[0106] drop_temp = 0.4,

[0107] The model construction process is as Figure 2 shown. When modeling, 70% of the samples are randomly selected for modeling, and 30% of the samples are used for model accuracy testing. Figure 2Among them, LNC, LAI, AGB, and GY represent leaf nitrogen concentration, leaf area index, aboveground biomass, and yield, respectively; PCC, SPA, and CARS represent Pearson correlation coefficient method, successive projections algorithm, and competitive adaptive reweighted sampling method, respectively; ANN, SVR, 1DCNN, and LSTM represent artificial neural network, support vector regression, one-dimensional convolutional neural network, and long short-term memory network, respectively.

[0108] Step S6: For each trait index, construct a data set using the corresponding sensitive band and measured value, and divide the data set into a training set and a validation set according to a preset ratio to verify the models of multiple machine learning algorithms.

[0109] Step S7: Select the optimal yield-related trait estimation model from the constructed models, and on this basis, estimate LNC, LAI, AGB, and GY by combining the texture features of the sensitive bands.

[0110] Regarding the texture features in step S7, in this embodiment, the gray-level co-occurrence matrix method (GLCM) of ENVI 5.3 is used to screen the texture features of each sensitive band with a window size of 3×3 resolution for the selected sensitive bands. They are: mean (Mean), variance (Var), homogeneity (Hom), contrast (Con), dissimilarity (Dis), entropy (Ent), second moment (Sec), and correlation (Cor). By delimiting the region of interest for each band's texture information image respectively, the texture values of this region can be extracted.

[0111] S8. According to the coefficient of determination R 2 and the root mean square error RMSE, evaluate the verification results of the model, and determine the preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to each trait index.

[0112] All parameter estimation models use the coefficient of determination (R 2 ) and the root mean square error (Root Mean Square Error, RMSE) to test the model accuracy. Among them, R 2 is used to evaluate the model fitting degree, and the closer R 2 is to 1, the better the model fitting degree; RMSE is used to evaluate the model stability, and the smaller the RMSE, the more stable the model.

[0113]

[0114] Among them, n is the number of samples, is the predicted value, is the average value, y i is the actual value

[0115] Example 2

[0116] Based on the specific method of Example 1, this example will be described in combination with the actual test results.

[0117] Regarding the effect of spectral reflectance smoothing:

[0118] Figure 3 It represents the spectral reflectance of the rice canopy after SG smoothing at each growth stage of Meixiangzhan 2 (A, C, E, G) and Nanjingxiangzhan (B, D, F, H). Figure 3 In it, PI and HD respectively represent the initial stage of young panicle differentiation and the heading stage; N0, N1, N2, N3, N4 respectively represent the treatments of 0, 60, 120, 180, 240 kg / ha.

[0119] is the spectral reflectance after SG smoothing. After SG smoothing, it can effectively reduce the "burr" phenomenon of the spectral reflectance. The canopy spectral reflectances of the two varieties at the four sampling times all show a consistent change trend. At 350 nm - 720 nm, the canopy spectral reflectance decreases with the increase of nitrogen application rate, and at 720 nm - 898 nm, the spectral reflectance increases with the increase of nitrogen application rate.

[0120] Regarding the measurement results of each trait index:

[0121] Figure 4 It shows the effects of different nitrogen fertilizer levels on leaf nitrogen concentration, leaf area index, aboveground biomass, and yield at each growth stage of Meixiangzhan 2 (A, C, E, G) and Nanjingxiangzhan (B, D, F, H). Figure 4 In it, HD in 2022E, PI in2022L, HD in 2022L, and PI in 2023E respectively represent the heading stage in the early season of 2022, the initial stage of young panicle differentiation in the late season of 2022, the heading stage in the late season of 2022, and the initial stage of young panicle differentiation in the early season of 2023; N0, N1, N2, N3, N4 respectively represent the treatments of 0, 60, 120, 180, 240 kg / ha.

[0122] From Figure 4 it can be seen that with the increase of nitrogen application rate, the rice LNC, LAI, AGB, and GY all show a gradually increasing change trend. Whether it is the rice variety Meixiangzhan 2 or Nanjingxiangzhan, when the nitrogen application rate reaches 120 kg / ha, there is no significant difference in the rice GY under the N2, N3, and N4 treatments (P < 0.05).

[0123] Regarding the correlation analysis between each trait index:

[0124] Figure 5The correlations between LNC (A), leaf area index (B), above-ground biomass (C) and yield at different growth stages of rice, as well as the correlations between leaf nitrogen concentration and leaf area index (D), leaf nitrogen concentration and above-ground biomass (E), and leaf area index and above-ground biomass (F) are shown. Figure 5 In Figure 5 , PI and HD represent the PI stage and heading stage respectively. "*" and "**" indicate significant correlations at the P<0.05 and P<0.01 levels respectively.

[0125] Combined with Figure 5 , significant positive correlations were found between LNC, LAI, AGB and GY. Among the correlations between AGB, LNC, LAI and GY, the correlation between AGB and GY was relatively high, with R 2 ≥0.60, followed by LAI, with R 2 ≥0.44. Among the correlations between AGB, LNC and LAI, the correlation between AGB and LAI was the highest, with R 2 ≥0.75.

[0126] Regarding the screening of characteristic bands:

[0127] The correlation analysis was carried out between the spectral reflectance of the canopy and each trait index. When the absolute value of the correlation coefficient was greater than 0.5, it was considered to have a strong correlation. As can be seen from Figure 6 , using the PCC dimensionality reduction method, it was found that the bands with strong correlations with LNC were 406nm, 410nm, 414nm, 418nm, 422nm. Among them, 410nm had the strongest correlation with LNC, and the correlation coefficient was 0.508. AGB had strong correlations in the range of 366nm - 694nm, and the strongest correlation was at 422nm, with the correlation coefficient reaching 0.845. LAI had strong correlations at 454nm - 502nm, 425nm and 426nm, and the strongest correlation was at 486nm, with the correlation coefficient reaching 0.686. AGB had strong correlations in the range of 366nm - 694nm, and the strongest correlation was at 422nm, with the correlation coefficient reaching 0.845. GY had strong correlations in the range of 642nm - 670nm, and the strongest correlation was at 654nm, with the correlation coefficient reaching 0.523. In addition, the successive projection algorithm (SPA) and competitive adaptative reweighted sampling (CARS) were also used to screen sensitive bands. Finally, the top 3 bands with the strongest correlations of PCC and CARS and the top 5 bands with the strongest correlations of SPA were selected as sensitive bands, and the results are shown in Table 1.

[0128] Table 1 Sensitive bands of rice yield-related traits after dimensionality reduction methods

[0129]

[0130] Regarding the determination of the optimal dimensionality reduction method and the optimal machine learning algorithm:

[0131] Three dimensionality reduction methods, namely PCC, SPA, and CARS, are used for sensitive band screening. The sensitive bands are used as features for input. Four machine learning algorithms, namely ANN, SVR, 1DCNN, and LSTM, are used to construct estimation models for four indicators, namely LNC, LAI, AGB, and GY. The R of the obtained models 2 and RMSE are shown in Table 2. Using PCC to screen sensitive bands, the effect of estimating AGB is the best. Among them, the model accuracy and stability of the model using the ANN algorithm are the highest, and R 2 is 0.89; RMSE is 86.25. Using SPA to screen sensitive bands, the effect of estimating LNC, LAI, and AGB is the best. Among them, the model accuracy and stability of the model using the ANN algorithm are the highest, and R 2 are 0.82, 0.75, and 0.90 respectively, and RMSE are 3.68, 0.47, and 81.29 respectively. Using CARS to screen sensitive bands, the effect of estimating LAI, AGB, and GY is the best. Among them, for LAI, the model accuracy and stability of the model using the SVR algorithm are the highest, and R 2 and RMSE are 0.62 and 0.54 respectively; for AGB and GY, the model accuracy and stability of the model using the ANN algorithm are the highest, and R 2 are 0.90 and 0.63 respectively; RMSE are 79.05 and 0.59 respectively. In summary, among the 3 data dimensionality reduction methods and 4 modeling methods, when estimating LNC and LAI, the model constructed by SPA-ANN has high accuracy and good stability. When estimating AGB and GY, the model constructed by CARS-ANN is the best.

[0132] Table 2 Model accuracy

[0133]

[0134] Regarding the model verification of the machine learning algorithm:

[0135] Select the optimal dimensionality reduction method and machine learning algorithm under 4 indicators. Use 70% as the training set and 30% as the validation set for verification. The model verification results are as Figure 7 shown. The SPA-ANN model has the best effect in estimating LNC and LAI, and R 2 are 0.81 and 0.76 respectively; RMSE are 4.82 and 0.43 respectively. The CARS-ANN model has the best effect in estimating AGB and GY, and R 2They are 0.90 and 0.61 respectively; the RMSEs are 94.93 and 0.57 respectively. The above results show that compared with the PCC method, the sensitive bands screened by SPA and CARS are more representative, and compared with SVR, 1DCNN and LSTM, the ANN algorithm has more advantages in estimating the LNC, LAI, AGB, and GY of rice.

[0136] Furthermore, regarding the combination of texture features on top of the preferred machine learning algorithms:

[0137] According to Figure 7 the optimal dimensionality reduction methods and machine learning algorithms for the 4 indicators in Figure 8 and Table 3, the texture features of the corresponding sensitive bands are introduced, and the results are as 2 shown. The models constructed after combining the texture features have all been improved to a certain extent. Among them, when estimating LNC, compared with only spectral modeling, the training set R of the model with texture features introduced 2 increased by 8.5%, and the RMSE decreased by 10.9%; the test set R of the test set 2 increased by 9.9%, and the RMSE decreased by 27.2%. When estimating GY, the test set R of the model with texture features introduced

[0138] Table 3 Spectral models and spectral + texture models

[0139]

[0140] Specifically, the analysis of the influence of dimensionality reduction methods, machine learning algorithms, and texture features on the model accuracy is as follows:

[0141] Regarding the influence of dimensionality reduction methods on the model accuracy: Reducing the dimensionality of spectral data to reduce data redundancy is a key step in establishing a crop growth parameter estimation model and an important link to improve the model accuracy and stability. The three dimensionality reduction methods of PCC, SPA, and CARS can all effectively reduce the dimensionality of spectral data and screen out the sensitive bands of specific growth parameters (Table 1). These methods have been widely used in many studies to extract key spectral features. Therefore, there are also obvious differences in the sensitive bands and model prediction performance obtained by different dimensionality reduction methods. This is because these dimensionality reduction methods have different focuses in feature selection. In this experiment, when estimating LNC and LAI, the dimensionality reduction of the SPA algorithm is significantly better than that of PCC and CARS (Table 2 and Figure 7)。As is well known, changes in LNC can affect the chlorophyll content of plant leaves, thereby affecting the spectral reflectance of leaves. LAI refers to the leaf area per unit land area. An increase in LAI means more leaf area, which is conducive to increasing the scattering and reflection of the spectrum. Some studies have shown that an increase in LAI will cause a significant increase in spectral reflectance in the near-infrared band (760nm - 1315nm). Therefore, both LAI and LNC have a direct connection with the spectral characteristics of leaves. When estimating AGB and GY, the sensitive bands screened by CARS dimensionality reduction are clearly superior to SPA and PCC (Table 2 and Figure 7 ). This is because AGB and GY are not only related to spectral data but also related to texture features and spatial structures, etc. There are complex interactions among these factors, and this complexity requires that the estimation model must be able to handle and interpret these complex multivariate relationships. The CARS algorithm is just suitable for dealing with complex multivariate relationships and hyperspectral data with a large amount of redundant information and multicollinearity, and can effectively screen out the key feature bands with the greatest contribution from complex data, reduce the data dimension, and thus obtain higher model accuracy and stability.

[0142] Regarding the influence of machine learning algorithms on model accuracy: The experimental results of this embodiment show that the ANN model performs better than the SVR, 1DCNN, and LSTM models, and the optimal models for the 4 trait indicators (biological indicators) are all ANN models (Table 2 and Figure 7 ). ANN is mainly applied to prediction and classification problems and is one of the widely used machine learning algorithms at present, with relatively good accuracy and stability among many machine learning algorithms. As a general function approximator, ANN has the non-linear characteristics introduced by the activation function and can learn any complex non-linear relationship between the input and output. This ability makes it perform excellently in dealing with non-linear problems. In contrast, SVR and 1DCNN are not as flexible as ANN in dealing with non-linear data sets, and although LSTM is suitable for dealing with sequence data, it has disadvantages in terms of computational complexity and parallelization ability, resulting in poor performance in the estimation of these trait indicators (biological indicators).

[0143] Regarding the influence of texture features on model accuracy: Texture features are added to the optimal combined model to optimize the model. The estimation model accuracy of LNC is improved the most, and the accuracy R of the model test set 2 can be increased by up to 9.9%, and the RMSE can be reduced by up to 27.2% (Table 3). Texture is an inherent property of the object surface, which does not change depending on color and brightness and can inhibit the occurrence of the phenomena of different objects with the same spectrum and the same object with different spectra. This is mainly because texture features provide additional spatial information, which helps to capture the microstructures and patterns of ground objects. Under the experimental conditions, the model accuracy of LAI, AGB, and GY all increases after adding texture features, but the improvement is limited (Table 3 andFigure 8 )。 This is because an optimal dimensionality reduction + machine learning combined model has been carried out during the research process, and the accuracy of the model itself is already very high, leaving limited room for accuracy improvement. In addition, among the four indicators in this experiment, the AGB estimation model has the highest accuracy (Table 2-3 and Figure 7-8 ), which is because when AGB is larger, it will absorb more light energy for photosynthesis, resulting in lower reflectance in these bands, making it easier to screen out sensitive bands, and thus a higher model accuracy can be obtained when constructing the AGB estimation model.

[0144] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art of this technology, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm, characterized in that: The spectral sensitive bands of rice leaf nitrogen concentration (LNC), leaf area index (LAI), aboveground biomass (AGB) and rice yield (GY) were extracted using the Pearson correlation coefficient method (PCC), successive projection method (SPA) and competitive adaptive reweighted sampling method (CARS). An estimation model of rice yield-related traits was constructed based on artificial neural network (ANN), support vector machine regression (SVR), one-dimensional convolutional neural network (1DCNN) and long short-term memory network (LSTM), and the optimal combination of spectrum + dimensionality reduction + machine learning was obtained. By extracting the texture features of the corresponding sensitive bands, a high-precision estimation model of rice yield-related traits such as spectrum + texture + dimensionality reduction + machine learning was constructed. The specific steps include: S1. Measure rice LNC, LAI, and AGB at the early heading stage, the beginning of panicle differentiation, and the late heading stage of the same year, and the beginning of panicle differentiation of the next year. Measure GY at the early and late seasons of the same year and the early season maturity of the next year. Acquire rice canopy hyperspectral data at the same time. S2, preprocessing the hyperspectral image, and taking the spectral average value of the area of ​​interest of each cell as the original spectral reflectance of the cell; S3, using Savitzky-Golay convolution smoothing algorithm to smooth and reduce noise of the original spectral data; S4, using a variety of dimensionality reduction methods, according to the correlation between the spectral reflectance and each property index, determining the sensitive band of each property index; S5, inputting the sensitive band into the model of the preferred machine learning algorithm to predict the trait index; S6. For each trait index, a data set is constructed using the corresponding sensitive band and the measured value, and the data set is divided into a training set and a validation set according to a preset ratio to validate the models of multiple machine learning algorithms; S7. Select the best yield-related trait estimation model from the constructed models, and on this basis, estimate LNC, LAI, AGB and GY in combination with the texture characteristics of the sensitive bands; S8, according to the determination coefficient R 2 The validation results of the model were evaluated by the root mean square error (RMSE), and the optimal dimensionality reduction method and optimal machine learning algorithm corresponding to each trait index were determined.

2. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In the step S1, the method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm is characterized in that the rice canopy hyperspectral data is collected using a multi-rotor UAV M300 RTK equipped with an X20P airborne hyperspectral imager, with a spectral range of 350-1000nm, a resolution of 4nm, and 164 effective bands. The UAV flight altitude is 50m; the measurement time is 10:00-14:00 Beijing time, and the weather is clear.

3. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In step S2, the spectral average value of the region of interest (ROI) of each cell is extracted as the original spectral reflectance of the cell by using the Region of Interest (ROI) tool through image stitching, radiation calibration, atmospheric correction, orthorectification, image fusion, geometric correction and band normalization.

4. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In step S3, in order to overcome the influence of environmental factors on the noise band and improve the prediction accuracy of the estimation model, this study processes the original spectrum through the Savitzky-Golay convolution smoothing algorithm; the Savitzky-Golay convolution smoothing algorithm uses a polynomial to perform polynomial least squares fitting on the data in the moving window. Its essence is a weighted average method that emphasizes the central role of the center point, which can better retain the original information of the spectrum while reducing noise; after parameter optimization, the Savitzky-Golay convolution smoothing method window_legth is set to 9 and polyorder is set to 2.

5. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In the step S4, among the sensitive bands of each trait index, the first three of the Pearson correlation coefficient method (PCC) and the competitive adaptive reweighted sampling method (CARS), and the first five of the continuous projection algorithm (SPA) are selected as sensitive bands; the best sensitive bands of LNC screened by the PCC method are 410nm, 414nm, and 418nm; the best sensitive bands of LAI are 422nm, 486nm, and 678nm; the best sensitive bands of AGB are 370nm, 422nm, and 678nm; the best sensitive bands of GY are 650nm, 654nm, and 658nm; the best sensitive bands of LNC screened by the SPA method are 398nm, 410nm, 478nm, The best sensitive bands of LAI are 350nm, 378nm, 394nm, 486nm, and 890nm; the best sensitive bands of AGB are 402nm, 422nm, 450nm, 718nm, and 898nm; the best sensitive bands of GY are 630nm, 654nm, 658nm, 690nm, and 742nm; the best sensitive bands of LNC screened by the CARS method are 406nm, 410nm, and 414nm; the best sensitive bands of LAI are 402nm, 674nm, and 678nm; the best sensitive bands of AGB are 366nm, 370nm, and 418nm; the best sensitive bands of GY are 350nm, 398nm, and 898nm.

6. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In step S5, the plurality of machine learning algorithms include artificial neural networks (ANN), support vector regression (SVR), one-dimensional convolutional neural networks (1D CNN) and long short-term memory networks (LSTM). The main parameters of the artificial neural network, "activation", "alpha", "hidden_layer_sizes", "learning_rate", "max_iter", "momentum", "solver", and "tol", are set to relu, 0.0001, 100, adaptive, 200, 0.7, adam, and 0.00001, respectively; The main parameters of the support vector machine, "kernel", "degree", "gamma", "coef0", "tol", "C", "Epsilon", "shrinking", "cache_size", "verbose", and "max_iter", are set to rbf, 3, auto, 0.0, 0.001, 1.0, 0.1, True, 200, False, and -1, respectively; The main parameters of the one-dimensional convolutional neural network, "in_channels", "out_channels", "kernel_size", "padding", and "num_epochs", are set to 1, 16, 3, 1, and 10, respectively; The main parameters of the long short-term memory network LSTM, "input_size", "output_size", "hidden_size_temp", "num_layer_temp", and "drop_temp", are set to 3, 1, 64, 1, and 0.4, respectively.

7. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In step S6, the train_test_split function in the sklearn.model_selection module in Python is used to divide the training set and the test set to verify the model; the ratio of the test set to the verification set is 7:

3.

8. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In step S7, the texture features of each sensitive band are extracted using the gray-level co-occurrence matrix method. The gray-level co-occurrence matrix (GLCM) method of ENVI 5.3 is used to screen the texture features of each sensitive band using a window size of 3×3 resolution. The texture features include mean, variance, homogeneity, contrast, heterogeneity, entropy, second-order moment and correlation. The region of interest is delineated for the texture information image of each band, and the texture value of the region can be extracted.

9. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: In step S8, the evaluation index of the estimation model adopts the determination coefficient R 2 , root mean square error RMSE, the calculation formulas of each evaluation index are as follows: Where n is the number of samples, is the predicted value, is the mean value, y i is the actual value.

10. The method for estimating rice yield-related traits based on spectrum-texture-dimensionality reduction-machine learning algorithm according to claim 1, characterized in that: The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the leaf nitrogen concentration are respectively a continuous projection algorithm and an artificial neural network; The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the leaf area index are a continuous projection algorithm and an artificial neural network, respectively; The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the aboveground biomass are competitive adaptive reweighted sampling method and artificial neural network, respectively; The preferred dimensionality reduction method and the preferred machine learning algorithm corresponding to the rice yield are a competitive adaptive reweighted sampling method and an artificial neural network, respectively.

Images