A method for detecting the nutritional status of lonicera japonica plant based on near infrared spectroscopy

Abstract

This invention belongs to the field of spectral detection technology, specifically relating to a method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy. It aims to solve the problems of strong specular reflection interference, moisture background shielding of nutrient signals, and insufficient applicability throughout the entire growth period in in-situ detection of live honeysuckle. The method suppresses specular reflection on the leaf surface through orthogonal extinction and collects deep diffuse reflectance spectra; it obtains standardized reflectance using real-time background difference and dark current compensation; it achieves moisture background decoupling through orthogonal signal correction and external parameter orthogonalization; it extracts feature wavelength vectors through multi-scale continuous wavelet transform; it inputs a deep residual network model with a self-attention mechanism and combines it with a phenological adaptive compensation algorithm to invert nutrient quality fractions; finally, it compares the results with a nutrient critical value database for the entire growth period to calculate the surplus / deficit index and complete nutrient diagnosis. This significantly improves the spectral signal-to-noise ratio and detection accuracy, achieving rapid and stable detection of the nutritional status of live honeysuckle throughout the entire growth period.

Description

Technical Field

[0001] This invention belongs to the field of spectral detection technology, and in particular relates to a method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy. Background Technology

[0002] With increasingly stringent global requirements for the quality, safety, and standardized cultivation of Chinese medicinal herbs, the precise nutritional regulation of honeysuckle, a plant with extremely high medicinal and economic value and considered both food and medicine, has become a key research focus in the interdisciplinary field of modern smart agriculture and traditional Chinese medicine resource science. The nutritional status of honeysuckle plants directly determines their biomass accumulation and is closely related to the synthesis and distribution of key secondary metabolites such as chlorogenic acid and luteolin. Therefore, real-time and precise monitoring of the nutritional status of honeysuckle plants is crucial for optimizing fertilization strategies, improving the quality of medicinal materials, and achieving ecological cultivation.

[0003] Traditional methods for determining plant nutrient status primarily rely on laboratory physicochemical analysis, such as the Kjeldahl method for total nitrogen content or inductively coupled plasma mass spectrometry (ICP-MS) for measuring various trace elements. While these methods are widely recognized for their accuracy, their inherently destructive sampling process, lengthy sample pretreatment procedures, and high testing costs mean they can only provide delayed, single-point data, making them unsuitable for the real-time monitoring of nutrient dynamics in large-scale plantations during critical phenological stages. To overcome these limitations, near-infrared spectroscopy (NIRS) has been increasingly adopted in crop nutrient diagnosis due to its non-destructive nature, lack of chemical contamination, and ability to simultaneously analyze multiple components. This technology acquires molecular vibrational absorption information in the near-infrared band and uses chemometric models to rapidly invert the content of target components.

[0004] However, although near-infrared spectroscopy performs well in the analysis of in vitro samples under controlled laboratory conditions, a series of deep-seated technical contradictions gradually emerge when it is applied to the in situ detection of a specific perennial woody vine, honeysuckle.

[0005] Specifically, the physiological structural characteristics of honeysuckle and its complex growth environment exert a significant modulating effect on spectral signals. From a physical perspective, honeysuckle leaves typically possess a cuticle of a certain thickness, and some populations are covered with dense pubescence on the underside of leaves or young branches. These microstructures cause near-infrared incident light to undergo intense diffuse reflection and multiple scattering at the interface before entering the plant tissue. This uncontrolled physical scattering phenomenon manifests as severe baseline drift and nonlinear noise in the spectral spectrum, significantly reducing the signal-to-noise ratio of weak absorption signals representing nutrient components (nitrogen- and phosphorus-containing groups, etc.).

[0006] A deeper problem lies in the dynamic and complex coupling of the nutrient metabolism system of honeysuckle. Under living growth conditions, the distribution of water within the plant, the transport of photosynthetic products, and the synergistic or competitive effects between elements mean that the spectral characteristics of nutrients are not simply a linear superposition. For example, water has a very strong absorption band in the near-infrared band, and the background signal it generates can easily mask the characteristic peaks of key mineral nutrients. Simultaneously, different nutrient states such as nitrogen, phosphorus, and potassium exhibit high frequency overlap in their spectral responses, causing conventional linear correction models to often show poor generalization ability and large fluctuations in prediction accuracy when processing such high-dimensional, strongly coupled, and nonlinear data. Furthermore, the physiological functions of honeysuckle differ significantly at different growth stages (such as the bud stage, flowering stage, and growth stage). This physiological rhythm-induced spectral characteristic shift makes it difficult for existing general-purpose detection models to maintain stable robustness throughout the entire growth period.

[0007] Existing detection methods often overlook the combined effects of environmental background interference, plant epigenetic features, and intrinsic physiological metabolism. In practical applications, effectively eliminating interference from non-target components under complex outdoor lighting conditions and physiological dynamics, precisely decoupling the deep characteristics characterizing the nutritional status of honeysuckle, and constructing a robust detection system that balances local precision and global generalization has become a serious challenge in the field of precision cultivation of Chinese medicinal herbs.

[0008] Therefore, developing a rapid, non-invasive method for detecting the nutritional status of honeysuckle plants, which is highly sensitive, specific, and tailored to their unique physiological and biochemical characteristics, has become a key technical challenge that urgently needs to be addressed by those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide a method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy, aiming to solve the technical problems of low detection accuracy and lack of model robustness caused by severe diffuse reflection induced by the microstructure of the cuticle and villi on the leaf surface, signal shielding under high water content background, and overlapping of multi-component spectral features during in situ detection of honeysuckle in vivo.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy includes the following steps: S1: Controlled polarized light covering the 400nm to 2500nm wavelength band is emitted towards the leaves of the honeysuckle. The specular reflection signal on the surface of the honeysuckle leaves is suppressed by an orthogonal extinction system, and the deep diffuse reflectance spectrum data is collected by a near-infrared spectral sensor.

[0011] An integrated diffuse reflection acquisition unit is employed, comprising a self-cooling quartz halogen tungsten lamp light source assembly. The light source assembly has a rated power of 20W, a spectral coverage range of 400nm to 2500nm, and a pair of polarizers with mutually perpendicular polarization directions are configured in its output optical path. The first polarizer is positioned at the emission end of the light source, and the second polarizer is positioned in front of the incident window of the near-infrared spectral sensor. By adjusting the polarization axis angles of the first and second polarizers to ensure they are orthogonal, specular reflection light from the cuticle layer of the honeysuckle leaf surface is suppressed, ensuring that the signals entering the sensor are all deep diffuse reflection light carrying nutrient information from within the plant.

[0012] The acquisition unit incorporates a hemispherical integrating sphere with an inner wall coated with a barium sulfate diffuse reflectance coating, achieving a reflectivity exceeding 98% within the 900nm to 2500nm wavelength range. A sampling hole with a diameter of 10mm is located at the bottom of the integrating sphere. An annular pressure sensor with a range of 0 to 5N is positioned at the edge of the sampling hole. This pressure sensor is connected to the actuator via a data feedback circuit to ensure that the contact pressure between the probe and the honeysuckle leaf remains constant at 1.5N during spectral acquisition, thus eliminating variations in the optical path distribution within the leaf tissue caused by different contact stresses.

[0013] S2: The near-infrared spectral sensor automatically acquires the reference spectrum of the polytetrafluoroethylene standard reflector before and after sample acquisition, and performs real-time normalization calculations based on the sensor's dark current background to obtain the real-time spectral reflectance R(λ).

[0014] A real-time calibration strategy based on dynamic background difference was implemented during the spectral acquisition process. This strategy is achieved by introducing a rotating reference calibration wheel into the acquisition optical path. A polytetrafluoroethylene (PTFE) standard reflector and a neutral density filter are evenly spaced on the calibration wheel. During each sampling cycle, the controller drives a stepper motor to rotate the calibration wheel at high speed, ensuring that the sensor automatically acquires the reference spectrum of the standard reflector before and after acquiring the spectrum of the honeysuckle sample. The real-time reflectivity R(λ) is calculated according to the formula: R (λ)=( I s (λ)− I d (λ)) / ( I r (λ)− I d (λ)), I r (λ) represents the reference light intensity of the standard reflector; I r (λ) represents the background light intensity of the sensor's dark current; I s(λ) represents the original reflected light intensity of the sample. This real-time ratio calculation can eliminate the interference of water source fluctuations and ambient stray light on the spectral baseline.

[0015] S3: An algorithm combining orthogonal signal correction and external parameter orthogonalization is adopted to extract the water absorption feature vector and construct the orthogonal complement space of the water-sensitive subspace. The original spectrum is then projected and transformed to achieve physical-level decoupling of the water background signal.

[0016] To address the submergence effect of the high water content of honeysuckle leaves on the characteristic absorption peaks of nitrogen, phosphorus, and potassium, this invention employs a spectral preprocessing algorithm combining orthogonal signal correction (OSC) and external parameter orthogonalization (EPO). The algorithm first identifies feature vectors related to water absorption in the spectral data through principal component analysis (PCA). Specifically, a water-sensitive subspace is established for the strong water absorption bands at 1450 nm and 1940 nm. Then, a projection operator is used to project the original spectral matrix onto the orthogonal complement of this subspace, thereby achieving physical-level decoupling of the water background signal while preserving nutrient component characteristic information.

[0017] S4: The preprocessed signal is decomposed into time-frequency domains at different scales using a multi-scale continuous wavelet transform algorithm, and the characteristic wavelength vectors reflecting the content of total nitrogen, total phosphorus, total potassium and key secondary metabolites of honeysuckle are selected by combining the correlation coefficient method.

[0018] In the feature extraction stage, a multi-scale continuous wavelet transform (CWT) algorithm was introduced. The algorithm used the Daubechies4 (db4) wavelet basis function to decompose the preprocessed spectral signal into scales 1 to 6. By calculating the energy distribution of the wavelet coefficients at each scale, characteristic spectral bands representing the chlorogenic acid, total nitrogen, and total potassium content of *Lonicera japonica* were identified. The correlation coefficient method was used to screen for characteristic wavelengths with an absolute correlation value greater than 0.85. The characteristic wavelengths after multi-scale decomposition were distributed in the regions of 1050nm-1220nm (first-order overtone absorption belonging to NH bonds), 1460nm-1680nm (second-order overtone absorption belonging to CH bonds), and 2100nm-2350nm (combined overtone absorption belonging to OH and CN bonds).

[0019] S5: Input the feature wavelength vector into a pre-trained nonlinear model based on the fusion of deep residual network and self-attention mechanism, and introduce a phenological period adaptive compensation algorithm to output the nutrient quality fraction of the plant.

[0020] To establish a robust nutritional status prediction model, this invention constructs a nonlinear inversion model based on the fusion of a deep residual network (ResNet) and an attention mechanism. The input layer of the model receives feature wavelength vectors filtered by wavelet transform. The model internally comprises four residual modules, each containing two one-dimensional convolutional layers, two batch normalization layers, and one ReLU activation function layer. The convolutional kernel size is set to 3, and the stride is set to 1. Following the residual modules, a self-attention layer is introduced. This self-attention layer assigns differentiated weight coefficients to feature information at different wavelength positions by calculating a weight matrix; the calculation formula is as follows: ; Q, K, V: These are the query, key, and value matrices obtained by linear transformation of feature F, respectively; d k : The dimension of matrix K, used to scale the dot product and prevent gradient vanishing; SoftMax(·): SoftMax activation function, used to normalize attention weights. This mechanism can significantly enhance the model's ability to capture weak nutrient feature signals while suppressing noise interference in redundant frequency bands.

[0021] Considering the physiological differences in honeysuckle at different growth stages (seedling stage, budding stage, flowering stage, etc.), a phenological stage adaptive compensation algorithm based on multi-task learning is introduced. This algorithm adds an auxiliary classification branch to the deep residual network to identify the growth stage of the current sample. By sharing the underlying feature extraction parameters during training, the model can learn general nutritional features across growth stages and perform nonlinear bias correction on the output of the regression branch based on the phenological stage category output by the classification branch.

[0022] During model training and validation, this invention employs Leave-One-out Cross-Validation (LOOCV) for hyperparameter optimization. A learning rate of 0.001 is set, and the Adam optimizer is used for gradient descent with 500 iterations and a batch size of 32. The optimization aims to minimize the root mean square error (RMSE) and maximize the coefficient of determination (R²). 2 This ensures that the model has extremely high prediction accuracy.

[0023] S6: Perform nutritional diagnosis, compare the output nutrient content data with the pre-stored database of nutrient surplus and deficit thresholds for the entire growth period of honeysuckle, calculate the nutrient surplus and deficit index, and obtain the nutritional status test results.

[0024] In this step, a database of nutrient surplus / deficit thresholds for honeysuckle throughout its entire growth cycle is pre-stored. This database is built based on extensive field trial data and covers standard nutrient values ​​for honeysuckle leaves under different soil types and altitudes. After the detection system outputs real-time nutrient content data, the measured values ​​are compared with the standard values ​​to calculate the nutrient surplus / deficit index. I i Used to quantify the profitability of nutrient provision: ; in, I i :nutrient i The profit and loss index, where a negative value indicates a deficit and a positive value indicates an excess; y i The S5 output includes the predicted mass fraction values ​​of each nutrient, including... y N ,y P ,y K ,y C .

[0025] Preferably, the specific process of step S1 is as follows: S11: Emits a continuous spectrum in the 400nm to 2500nm band, which is modulated into linearly polarized light by the first polarizer. The intensity of the emitted light is denoted as... I 0; S12: Separation of specular reflection from the blade surface and deep transmission: After the incident light reaches the blade, it splits into two parts: Intensity of light reflected from the mirror surface of the blade I s It follows Fresnel's law of reflection: I s = I 0· R ( θ,n 1 ,n 2 ,l ); in, I s : The intensity of light reflected from the specular surface of the blade; I 0: Intensity of the incident ray polarized light after passing through the first polarizer; R ( θ,n 1 , n 2 ,l Fresnel reflectance; i Angle of incidence of light; n 1: Air refractive index; n 2: Refractive index of the cuticle of honeysuckle leaves; lNear-infrared light wavelength; Depolarized diffuse reflection light, intensity denoted as I d ; S13: Orthogonal polarization extinction and specular reflection signal filtering: The transmittance of incident light through the polarizer follows Malus's law: T = cos 2 (Δ i ); in, T Δ: The transmittance of the polarizer to incident light; i The angle between the polarization direction of the incident light and the polarization axis of the second polarizer; Since the reflected light from the mirror retains its original polarization state and is orthogonal to the second polarizer, i.e., Δ i =90°, therefore the intensity of the outgoing light after the specular reflected light passes through the second polarizer is: I s_out =I s · cos 2 (90°) = 0; S14: Effective Deep Diffuse Reflection Signal Acquisition: Due to scattering and depolarization, some of the deep diffuse reflection light passes through the second polarizer. The intensity of the outgoing light after passing through the second polarizer is denoted as... I d_out ; The total light intensity ultimately received by the near-infrared spectral sensor is: I total = I d_out + I s_out ≈ I d_out .

[0026] Preferably, the specific process of step S2 is as follows: S21: Acquire reference light intensity from the standard reflector I r (λ), sensor dark current background light intensity I d (λ), original reflected light intensity of the sample I s (λ); S22: Real-time normalization calculation of spectral reflectance R (λ): Using the reference light intensity of the standard reflector as the benchmark and the dark current background light intensity as the zero-point compensation, the original reflected light intensity of the sample is subjected to real-time background subtraction and normalization processing to obtain the real-time spectral reflectance. The calculation formula is as follows: R (λ)=( I s (λ)− I d (λ)) / ( I r (λ)− I d (λ)).

[0027] Preferably, the specific process of step S3 is as follows: S31: Constructing a moisture-sensitive subspace: Selecting honeysuckle leaf samples with different moisture content gradients, and obtaining the original calibrated spectral matrix. X ; Principal component analysis was used to extract eigenvectors highly correlated with moisture at 1450 nm and 1940 nm in the near-infrared band, forming an eigenvector matrix for the moisture-sensitive subspace. M : M =[ m 1, m 2,··· m k ]; in, m 1, m 2,··· m k :forward k A water-dominant feature vector; S32: Construct the orthogonal projection operator P: Based on the EPO algorithm, a projection operator is constructed to project onto the orthogonal complement space of water molecules: P=I−M(M T M) −1 M T ; Where P: orthogonal projection operator, used to remove moisture interference; I: identity matrix, with dimensions consistent with the spectral matrix; M T : The transpose of the moisture eigenvector matrix M; (M T M) −1 Matrix M T The inverse matrix of M; S33: Orthogonal projection transformation of the original spectrum: The original calibration spectral matrix X By projecting the signal onto the orthogonal complement space of the moisture-sensitive subspace, physical-level stripping of the moisture signal is achieved. X clean =X · P ; in, XOriginal calibration spectral matrix; X clean : Pure spectral matrix after hydrolysis and coupling; S34: Further purification through joint quadrature signal correction: [The following text appears to be incomplete and requires further context:] X clean Iterative removal of orthogonal noise unrelated to nutrient composition is performed to ultimately output the target spectrum of completely decoupled water. X final : X final =OSC( X clean ); Where OSC(·) is the orthogonal signal correction function.

[0028] Preferably, the original calibration spectral matrix in step S31 X The acquisition process is as follows: Calculate the real-time reflectivity for each wavelength λ R (λ), arranged in order of wavelength from 400nm to 2500nm, representing all wavelength points. R (λ) forms a single-row spectral vector: x =[ R (λ1), R (λ2),···, R (λ m )]; When there is n When dealing with a sample, the spectral vectors of all samples are stacked row-wise to form the original calibration spectral matrix. X : ; Where n: number of samples; m: number of spectral wavelength points; R i ( l j ): No. i The sample at the th j Real-time spectral reflectance at each wavelength; l j : No. j Near-infrared wavelength.

[0029] Preferably, the specific process of step S4 is as follows: S41: Obtain the pure spectrum of water decoupling based on step S3 X final : X final =[ X final (λ1), X final(λ2),···, X final (λ m [ ], for the purification spectrum after dehydration X final Perform continuous wavelet decomposition at scales 1 to 6 to obtain wavelet coefficients at each scale: ; in, W ( a,b ):scale a Translation amount b The corresponding wavelet coefficients; a Wavelet transform scale; b Wavelength shift factor; :db4 wavelet basis functions; : Conjugate of wavelet basis functions; S42: Single-scale reconstruction of wavelet coefficients at each scale yields spectral feature components at different scales: ; in, X recon ( a ,λ): Scale a Reconstructed spectral characteristics at the next wavelength λ; C: wavelet admissibility constant; S43: Reconstruction features based on each wavelength X recon (λ) and measured nutrient content y Calculate the Pearson correlation coefficient r (λ): ; in, r (λ): Pearson correlation coefficient between the spectrum and nutrient content at wavelength λ; n : Sample size; X recon,i (λ): the first i Reconstructed features of each sample at wavelength λ; : The mean value of the wavelength λ reconstruction feature; y i : No. i Measured nutrient values ​​for each sample; : Average measured nutrient values; S44: Feature Wavelength Selection and Feature Vector Construction: Set the threshold to 0.85 and retain | r Highly correlated wavelengths with (λ)|≥0.85 form a characteristic wavelength set. : ; Sort by wavelength from smallest to largest to obtain the characteristic wavelength vector. W : W =[ λw 1, λw 2,···, l wp ]; in, p Number of characteristic wavelengths; λw 1, λw 2,···, l wp : The characteristic wavelength after screening.

[0030] Preferably, the specific process of step S5 is as follows: S51: Transfer the characteristic wavelength vector W The corresponding spectral eigenvalues ​​are standardized to obtain the standardized eigenvectors. ; S52: The standardized feature vectors Inputting a deep residual network, the residual blocks extract deep nonlinear features. The output formula of the residual blocks is: ; in, F l : No. l Input features of layer residual blocks; F l+1 : No. l Output features of the residual block; Conv(·): Convolutional layer operation; W l : No. l The weight matrix of each convolutional layer; S53: Deep features of the residual network output F Apply a self-attention mechanism, calculate feature weights, and weight them accordingly; S54: Input the weighted attention features into the fully connected layer to obtain a preliminary prediction of the nutrient quality score: ; : Vector of preliminary predicted values ​​of uncompensated nutrient quality fraction; F attn : The feature vector after weighting by the self-attention mechanism; FC(·): Fully connected layer operation; S55: Revise the preliminary forecast: ; Among them, phenological period t The corresponding compensation term vector ΔY( tLet t be a function of phenological period t, and represent it using a piecewise linear model: ; in, Y : The vector of predicted nutrient quality fractions after final compensation; : Vector of preliminary predicted values ​​of uncompensated nutrient mass fraction; Δ Y ( t ): Phenological period t The corresponding compensation term vector; t: phenological period, t 1 is the seedling stage, t 2 is the budding stage, t 3 is the peak flowering period, t 4 is the end of the flowering period; α i ,β i : Linear compensation coefficients for each phenological stage.

[0031] Preferably, the specific process of step S6 is as follows: S61: Obtain the nutrient quality fraction prediction vector based on step S5. Y , Y =[ y N ,y P ,y K ,y C ], y N ,y P ,y K ,y C They are respectively N,P,K,C The predicted quality fraction value is obtained by matching the nutrient critical value parameters corresponding to the current phenological stage t of the plant from the pre-stored database of nutrient surplus and deficit critical values ​​for the entire growth period of honeysuckle, including: deficiency critical value. L i (t), suitable critical value U i (t) and excess critical value E i ( t ); S62: Calculate the nutrient break-even index: based on the predicted value. y N ,y P ,y K ,y CThe relationship with the critical value is calculated using a piecewise linear model to determine the nutrient surplus / deficit index. I i Used to quantify the profitability of nutrient provision: ; in, I i :nutrient i Profit and loss index; y i Predicted mass fraction values ​​for each nutrient; S63: Determine nutrient status classification based on nutrient surplus / deficit index. I i The nutritional status of a single nutrient is classified. S64: Based on the grading results of all nutrients, a weighted summation method is used to calculate the overall nutritional status index of the plant. I total : ; w i :nutrient i The weight coefficients are determined by the analytic hierarchy process (AHP) or the entropy weight method, and .

[0032] Preferably, after step S6, the process further includes generating corresponding fertilization decision recommendations, as follows: Based on the profit and loss index of each nutrient I i The recommended fertilizer application rate is calculated using the following formula: F i =k i ·( U i (t)− y i ); in, F i :nutrient i Recommended fertilizer application rate per unit area; k i :nutrient i Correction coefficients; according to F i Based on the sign and magnitude of the value, corresponding fertilization decision recommendations are generated: F i >0: Recommended to supplement with nutrients i Dosage is F i ; F i≤0: No need to supplement, if I i A value >0.5 indicates a need to control fertilization or increase the application of organic fertilizer to improve the soil.

[0033] The beneficial effects of this invention include: 1. Suppressing surface scattering interference from the physical source and significantly improving the original spectral signal-to-noise ratio: An orthogonal extinction system composed of a first polarizer and a second polarizer is used. Based on Fresnel's law of reflection and Malus's law, the specular reflection signal generated by the cuticle and villi of honeysuckle leaves is selectively blocked, and only the deep diffuse reflection light carrying internal nutrient information is retained, thereby improving the effective spectral signal-to-noise ratio and solving the baseline drift and nonlinear noise problems caused by the microstructure of the leaves to a certain extent.

[0034] 2. Real-time background differential calibration significantly improves the stability of field environment testing: By synchronously acquiring and real-time normalizing the sample light intensity, standard reflector light intensity, and dark current background, the effects of light source fluctuations, environmental temperature drift, stray light, and circuit noise can be eliminated. Baseline drift can be controlled within the temperature fluctuation range of 0℃ to 40℃, enabling the equipment to maintain laboratory-level testing accuracy in complex field environments.

[0035] 3. Physical-level decoupling of moisture background for accurate restoration of nutrient component characteristic signals: The joint algorithm of orthogonal signal correction (OSC) and external parameter orthogonalization (EPO) is adopted to construct a moisture-sensitive subspace and remove moisture interference through orthogonal projection. Under the condition of large fluctuations in leaf moisture content, it can still completely preserve key nutrient characteristics such as total nitrogen (NH), total phosphorus, total potassium, and chlorogenic acid (C–H), solving the problem of signal shielding of weak absorption peaks by high moisture background.

[0036] 4. Multi-scale wavelet fine feature extraction to achieve efficient dimensionality reduction of high-dimensional spectral data: Time-frequency domain decomposition is performed through continuous wavelet transform of 1 to 6 scales, and highly correlated feature wavelengths are selected by correlation coefficient method. While retaining effective information, the dimensionality of model input is greatly reduced, improving model training efficiency and prediction response speed, and avoiding modeling failure caused by the overlap of multi-component spectral features.

[0037] 5. A closed-loop system for nutrient diagnosis and fertilization decision-making supports the precise cultivation of honeysuckle: Based on a database of critical nutrient values ​​throughout the entire growth period, the system automatically calculates the nutrient surplus / deficient index, quickly determines the nutrient status, and generates quantitative fertilization recommendations. This achieves the integration of non-destructive testing, status assessment, and precise fertilization, without the need for chemical reagents or damage to the plants. Each test is quick, which can significantly reduce testing costs and improve the quality and yield of medicinal materials. Attached Figure Description

[0038] Figure 1 This is a schematic flowchart of the method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to the present invention.

[0039] Figure 2 This is a schematic diagram of the neural network architecture of the present invention. Detailed Implementation

[0040] The following is in conjunction with the appendix Figure 1~Figure 2 The present invention will be further described in detail below: The design logic of this invention is based on a deep decoupling of the unique physiological and morphological characteristics of *Lonicera japonica* and its near-infrared spectral response mechanism. The surface of *Lonicera japonica* leaves is covered with a dense cuticle and numerous non-glandular trichomes. This microstructure induces strong specular reflection and diffuse loss under traditional diffuse reflection acquisition modes, resulting in an extremely low proportion of characteristic photons carrying effective chemical bond information entering the mesophyll tissue. Based on this, the following quantitative analysis procedure is constructed.

[0041] Its core hardware architecture is built upon a precisely controlled diffuse reflection acquisition unit. The power source for this unit is a set of quartz halogen tungsten lamps, with a rated power precisely set to 20W. To ensure long-term stability of the spectral output, the light source integrates a self-cooling module based on the Peltier effect. A closed-loop PID control algorithm maintains the lamp's operating temperature at 45℃±0.2℃, thereby suppressing thermal drift modulation of the full-band spectral output characteristics from 400nm to 2500nm. In the light source's output path, the light first passes through a collimating lens group to form a parallel beam, and then passes through a first polarizer. This first polarizer uses a high extinction ratio polarizing material, and its polarization axis direction serves as the system's initial reference plane.

[0042] To eliminate specular reflection interference induced by the pubescence on the surface of *Lonicera japonica* leaves, a second polarizer is coaxially positioned in front of the incident window of the near-infrared spectral sensor. During assembly, a precision rotating mechanism adjusts the polarization axis of the second polarizer to ensure it is strictly orthogonal to the polarization axis of the first polarizer (i.e., at the extinction position). According to Fresnel's law of reflection, light directly reflected from the leaf cuticle maintains the polarization state of the incident light and is therefore blocked by the second polarizer. Only diffusely reflected light, which penetrates the cuticle and enters the deep mesophyll tissue, losing its polarization characteristics after multiple scattering, can pass through the second polarizer and be received by the near-infrared spectral sensor. This optical path design physically improves the signal-to-noise ratio by approximately 12 dB, ensuring that the acquired signal primarily originates from the absorption and modulation of photons by nutrients within the leaf.

[0043] In the engineering design of the sampling interface, a hemispherical integrating sphere structure is adopted. The inner wall of the integrating sphere is covered with a barium sulfate (BaSO4) diffuse reflection coating with a thickness of not less than 0.5 mm through a multi-layer spraying process. The sampling hole at the bottom of the integrating sphere has a diameter of 10 mm, which can cover a sufficient number of honeysuckle leaf mesophyll cells to ensure sample representativeness, while avoiding excessively large openings that would cause internal energy to dissipate. An annular pressure sensor is embedded in the outer periphery of the sampling hole. This sensor adopts a high-sensitivity strain gauge architecture, with a measurement range covering 0 to 5 N and a resolution as high as 0.01 N.

[0044] Example 1 See appendix Figure 1 As shown, a method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy includes the following steps: S1: Construct an integrated diffuse reflection acquisition unit. Use a quartz halogen tungsten lamp light source group with self-cooling function to emit controlled polarized light covering the 400nm to 2500nm wavelength band to the honeysuckle leaves. Through the orthogonal extinction system composed of the first polarizer and the second polarizer set in the optical path, the specular reflection signal of the honeysuckle leaf surface is suppressed. The deep diffuse reflection spectral data carrying the internal nutrient information of the plant is collected by the near-infrared spectral sensor.

[0045] S2: Perform real-time background differential calibration. Drive the rotating reference calibration wheel to rotate during the acquisition cycle, so that the near-infrared spectral sensor can automatically acquire the reference spectrum of the polytetrafluoroethylene standard reflector before and after sample acquisition. Combine the sensor dark current background to perform real-time normalization calculation and obtain the real-time spectral reflectance R(λ).

[0046] S3: Perform in-depth preprocessing of the spectral signal, using an algorithm that combines orthogonal signal correction and external parameter orthogonalization to extract the moisture absorption feature vector and construct the orthogonal complement space of the moisture-sensitive subspace. Perform projection transformation on the original spectrum to achieve physical-level decoupling of the moisture background signal.

[0047] S4: Conduct fine extraction of characteristic spectral bands, use multi-scale continuous wavelet transform algorithm to decompose the preprocessed signal into time-frequency domain at scales of 1 to 6, and combine the correlation coefficient method to screen out characteristic wavelength vectors that reflect the content of total nitrogen, total phosphorus, total potassium and key secondary metabolites of honeysuckle.

[0048] S5: Input the feature wavelength vector into a pre-trained nonlinear model based on the fusion of deep residual network and self-attention mechanism, and introduce a phenological adaptive compensation algorithm to output the nutrient quality fraction of the plant. S6: Perform nutritional diagnosis, compare the output nutrient content data with the pre-stored database of nutrient surplus and deficit thresholds for the entire growth period of honeysuckle, and calculate the nutrient surplus and deficit index.

[0049] The specific process of step S1 is as follows: S11: Generation and incidence of linearly polarized light: A quartz halogen tungsten lamp light source assembly with self-cooling function emits a continuous spectrum in the 400nm to 2500nm wavelength range. This spectrum is modulated into linearly polarized light by a first polarizer, and the emitted light intensity is denoted as [insert intensity here]. I 0. The linearly polarized light is incident on the surface of the honeysuckle leaf at a fixed incident angle, forming a stable incident light path; S12: Separation of specular reflection from the blade surface and deep transmission: After the incident light reaches the blade, it splits into two parts: One type is specular reflection produced by the stratum corneum and vellus hair. This reflected light maintains the polarization state of the incident light, and its intensity follows Fresnel's law of reflection. I s = I 0· R ( θ,n 1 ,n 2 ,l ); in, I s : Light intensity reflected by the specular surface of the blade, unit: W / cm²; I 0: Intensity of incident ray polarized light after passing through the first polarizer, unit: W / cm²; R ( θ,n 1 ,n 2 ,l Fresnel reflectance; i Angle of incidence of light, unit: °; n 1: Air refractive index, value: 1.0003; n 2: Refractive index of the cuticle of honeysuckle leaves, ranging from 1.42 to 1.48; l Near-infrared light wavelength, unit: nm; Second, the transmitted light that penetrates the surface and enters the mesophyll tissue is scattered multiple times to form depolarized diffuse reflected light, the intensity of which is denoted as [insert intensity here]. I d The light carries information about the plant's internal nutrients, and its polarization state is completely destroyed. S13: Orthogonal polarization extinction and specular reflection signal filtering: The polarization axis of the second polarizer is 90° orthogonal to that of the first polarizer, forming an orthogonal extinction system; The transmittance of incident light through the polarizer follows Malus's law: T = cos 2 (Δ i ); in, T Δ: The transmittance of the polarizer to incident light; i : The angle between the polarization direction of the incident light and the polarization axis of the second polarizer, in °; Since the reflected light from the mirror retains its original polarization state and is orthogonal to the second polarizer, i.e., Δ i =90°, therefore the intensity of the outgoing light after the specular reflected light passes through the second polarizer is: I s_out =I s · cos 2 (90°) = 0; This achieves physical suppression of specular reflection signals; S14: Effective Deep Diffuse Reflection Signal Acquisition: Due to scattering and depolarization, some of the deep diffuse reflection light passes through the second polarizer. The intensity of the outgoing light after passing through the second polarizer is denoted as... I d_out ; The total light intensity ultimately received by the near-infrared spectral sensor is: I total = I d_out + I s_out ≈ I d_out ; The sensor only receives deep diffuse reflection signals carrying nutrient information, thus eliminating specular reflection interference and providing high signal-to-noise ratio raw spectral data for subsequent constant pressure acquisition, background difference, and water decoupling.

[0050] Example 2 Based on Example 1, the specific process of step S2 is as follows: After completing the optical path preparation for S1 orthogonal polarization constant voltage spectral acquisition, the acquisition cycle begins; the controller drives the stepper motor to rotate the rotary reference calibration wheel, so that the near-infrared spectral sensor sequentially acquires the standard reflection reference signal, dark current background signal and sample signal, realizing full closed-loop automatic calibration.

[0051] S21: Acquire reference light intensity from the standard reflector I r (λ): When the calibration wheel rotates to the position of the PTFE standard reflector, the polarized light emitted by the light source is diffusely reflected by the standard reflector and then collected by the near-infrared spectral sensor to obtain the reference light intensity of the standard reflector across the entire wavelength range, denoted as: I r (λ).

[0052] Dark current background light intensity of the sensor I d(λ): When the calibration wheel rotates to the shaded area, the effective incident light path is closed. The sensor collects its own thermal noise and circuit dark current signals under no-light conditions, and obtains the dark current background light intensity at wavelength λ, denoted as: I d (λ).

[0053] Original reflected light intensity of the sample I s (λ): When the calibration wheel rotates to the sampling channel, polarized light is incident on the honeysuckle leaf. After deep diffuse reflection, it is collected by the sensor to obtain the original reflected light intensity of the sample, denoted as: I s (λ); S22: Real-time normalization calculation of spectral reflectance R (λ): Using the reference light intensity of the standard reflector as the benchmark and the dark current background light intensity as the zero-point compensation, the original reflected light intensity of the sample is subjected to real-time background subtraction and normalization processing to obtain the real-time spectral reflectance. The calculation formula is as follows: R (λ)=( I s (λ)− I d (λ)) / ( I r (λ)− I d (λ)).

[0054] The specific process of step S3 is as follows: Step S3 uses the real-time calibrated spectral reflectance output from S2. R (λ) is the input. The Orthogonal Signal Correction (OSC) + External Parameter Orthogonalization (EPO) joint algorithm is used to selectively remove moisture interference and retain nutrient feature signals, providing a pure spectrum for subsequent feature extraction.

[0055] S31: Constructing a moisture-sensitive subspace: Selecting honeysuckle leaf samples with different moisture content gradients, and obtaining the original calibrated spectral matrix. X ; For the two strong absorption peaks of moisture at 1450nm and 1940nm in the near-infrared band, principal component analysis (PCA) was used to extract eigenvectors that are highly correlated with moisture, forming the eigenvector matrix M of the moisture-sensitive subspace: M =[ m 1, m 2,··· m k ]; in, m 1, m 2,··· m k:forward k A water-dominant feature vector; S32: Construct the orthogonal projection operator P: Based on the EPO algorithm, a projection operator is constructed to project onto the orthogonal complement space of water molecules: P = I - M ( M T M ) −1 M T ; Where, P: orthogonal projection operator, used to remove moisture interference; I : The identity matrix, with the same dimensions as the spectral matrix; M T Moisture feature vector matrix M The transpose of; M T M ) −1 :matrix M T M The inverse matrix; S33: Orthogonal projection transformation of the original spectrum: The original calibration spectral matrix X By projecting the signal onto the orthogonal complement space of the moisture-sensitive subspace, physical-level stripping of the moisture signal is achieved. X clean =X · P ; in, X Original calibration spectral matrix; X clean The pure spectral matrix after hydrolysis decoupling is free from water background interference. S34: Further refinement of Joint Orthogonal Signal Correction (OSC): right X clean Iterative removal of orthogonal noise unrelated to nutrient composition is performed to ultimately output the target spectrum of completely decoupled water. X final : X final =OSC( X clean ); Where OSC(·) is the orthogonal signal correction function.

[0056] In this embodiment, the original calibration spectral matrix in step S31 X The acquisition process is as follows: Calculate the real-time reflectivity for each wavelength λ R (λ), arranged in order of wavelength from 400nm to 2500nm, representing all wavelength points. R (λ) forms a single-row spectral vector: x =[ R (λ1), R (λ2),···, R (λ m )]; When there is n When dealing with a sample, the spectral vectors of all samples are stacked row-wise to form the original calibration spectral matrix. X : ; Where n: number of samples; m: number of spectral wavelength points; R i ( l j ): No. i The sample at the th j Real-time spectral reflectance at each wavelength; l j : No. j Near-infrared wavelengths, unit: nm.

[0057] Example 3 Based on the above description in Example 1 or Example 2, the specific process of step S4 is as follows: Step S4 is to perform water decoupling purification spectrum on the output of S3. X final Fine extraction of characteristic spectral bands was carried out; continuous wavelet transform at multiple scales from 1 to 6 was performed using the db4 wavelet basis to obtain the wavelet coefficients W( a,b And reconstruct the spectral characteristics at each scale. X recon ( a The correlation coefficients between each wavelength characteristic and nutrient content were calculated using the Pearson correlation coefficient method. r (λ), filter | r The wavelength (λ)|≥0.85 was used as the characteristic wavelength, and finally the characteristic wavelength vector W reflecting the content of total nitrogen, total phosphorus, total potassium and key secondary metabolites of honeysuckle was constructed.

[0058] S41: Obtain the pure spectrum of water decoupling based on step S3 X final : X final =[ X final (λ1), X final(λ2),···, X final (λ m [ ], for the purification spectrum after dehydration X final Perform continuous wavelet decomposition at scales 1 to 6 to obtain wavelet coefficients at each scale: ; Among them, W( a,b ):scale a Translation amount b The corresponding wavelet coefficients; a Wavelet transform scale, with values ​​of 1, 2, 3, 4, 5, and 6; b : Wavelength shift factor, corresponding to the wavelength position; :db4 wavelet basis functions; : Conjugate of wavelet basis functions; S42: Wavelet coefficient reconstruction and feature information extraction: Single-scale reconstruction of wavelet coefficients at each scale yields spectral feature components at different scales: ; in, X recon ( a ,λ): Scale a Reconstructed spectral characteristics at the next wavelength λ; C: wavelet allowance constant, a fixed constant for db4; S43: Calculate the correlation coefficient between spectral characteristics and nutrient content: Based on reconstruction features at each wavelength X recon (λ) and measured nutrient content y Calculate the Pearson correlation coefficient for (total nitrogen, total phosphorus, total potassium, chlorogenic acid, etc.) r (λ): ; in, r (λ): Pearson correlation coefficient between the spectrum and nutrient content at wavelength λ; n : Sample size; X recon,i (λ): the first i Reconstructed features of each sample at wavelength λ; : The mean value of the wavelength λ reconstruction feature; y i : No. i Measured nutrient values ​​for each sample; : Average measured nutrient values; S44: Feature Wavelength Selection and Feature Vector Construction: Set the threshold to 0.85 and retain |r Highly correlated wavelengths with (λ)|≥0.85 form a characteristic wavelength set. : ; Sort by wavelength from smallest to largest to obtain the characteristic wavelength vector. W : W =[ λw 1, λw 2,···, l wp ]; in, p Number of characteristic wavelengths p ≪ m ; λw 1, λw 2,···, l wp : The characteristic wavelength after screening.

[0059] Example 4 Based on Example 1, Example 2, or Example 3, the specific process of step S5 is as follows: This step is based on a neural network architecture that integrates a deep residual network (ResNet) and a self-attention mechanism. See [link to documentation]. Figure 2 As shown, the model's input layer receives a high-dimensional vector filtered by wavelet processing, and its core consists of four cascaded residual modules. Each residual module contains two layers of one-dimensional convolutional operations with a kernel size of 3 and a stride of 1, and is coupled with a batch normalization layer to accelerate convergence. The introduction of residual connections effectively solves the gradient vanishing problem in deep networks, enabling the model to learn extremely weak feature evolutions in the spectrum. After the residual modules, a crucial self-attention layer is introduced. This layer calculates the dependencies between wavelengths by constructing a query matrix Q, a key matrix K, and a value matrix V. This mechanism can automatically identify and amplify the spectral frequency band weights that contribute the most to nutrient prediction, while suppressing redundant features caused by environmental fluctuations.

[0060] The feature wavelength vector output by S4 is input into a pre-trained nonlinear model based on the fusion of a deep residual network and a self-attention mechanism. First, the feature vector is standardized. Then, deep features are extracted through the residual network, and the features are weighted using the self-attention mechanism. Finally, a fully connected layer outputs a preliminary prediction of the nutrient quality fraction. A phenological adaptive compensation algorithm is introduced to correct the preliminary prediction based on the current phenological stage, ultimately outputting the nutrient quality fraction of the plant. The specific process is as follows: S51: Transfer the characteristic wavelength vector WThe corresponding spectral characteristic values ​​are standardized to eliminate dimensional differences: ; in, Standardized feature vectors; The mean of the feature vectors in the training set; : Standard deviation of the feature vectors in the training set.

[0061] S52: The standardized feature vectors Inputting a deep residual network, the residual blocks extract deep nonlinear features. The output formula of the residual blocks is: ; in, F l : No. l Input features of layer residual blocks; F l+1 : No. l Output features of the residual block; Conv(·): Convolutional layer operation; W l : No. l The weight matrix of the convolutional layer.

[0062] S53: Self-attention mechanism feature weighting: deep features of residual network output F A self-attention mechanism is applied, feature weights are calculated and weighted, as shown in the formula: ; Where Q, K, and V are the query, key, and value matrices obtained by linear transformation of feature F, respectively; d k : The dimension of matrix K, used to scale the dot product and prevent gradient vanishing; SoftMax(·): SoftMax activation function, used to normalize attention weights.

[0063] S54: Preliminary prediction of nutrient quality score: The weighted attention features are input into the fully connected layer to obtain a preliminary prediction of the nutrient quality score. ; : Vector of preliminary predicted values ​​of uncompensated nutrient quality fraction; F attn : The feature vector after weighting by the self-attention mechanism; FC(·): Fully connected layer operation.

[0064] S55: Phenological Stage Adaptive Compensation: An adaptive compensation algorithm for phenological stages is introduced to correct the initial prediction values ​​and eliminate systematic biases at different growth stages. ; Among them, phenological period t The corresponding compensation term vector Δ Y ( t Let t be a function of phenological period t, and represent it using a piecewise linear model: ; in, Y : The vector of predicted nutrient quality fractions after final compensation; : Vector of preliminary predicted values ​​of uncompensated nutrient mass fraction; Δ Y ( t ): Phenological period t The corresponding compensation term vector; t: phenological period, t 1 is the seedling stage, t 2 is the budding stage, t 3 is the peak flowering period, t 4 is the end of the flowering period; α i ,β i : Linear compensation coefficients for each phenological stage, obtained by fitting the training set.

[0065] The specific process of step S6 is as follows: S61: Obtain the nutrient quality fraction prediction vector based on step S5. Y , Y =[ y N ,y P ,y K ,y C ], y N ,y P ,y K ,y C They are respectively N,P,K,C The predicted quality fraction value is determined by matching the nutrient critical value parameters corresponding to the current phenological stage t of the plant from the pre-stored database of nutrient surplus and deficit critical values ​​for the entire growth period of honeysuckle. These parameters include: Lack of critical value L i (t): The lower limit of nutrient deficiency; Suitable critical value U i (t): The upper limit of suitable nutrient levels; Excess threshold E i ( t ): The threshold for nutrient excess; S62: Calculate the nutrient break-even index: based on the predicted value. y N ,y P ,y K ,y C The relationship with the critical value is calculated using a piecewise linear model to determine the nutrient surplus / deficit index. I i Used to quantify the profitability of nutrient provision: ; in, I i :nutrient i The profit and loss index, where a negative value indicates a deficit and a positive value indicates an excess; y i The S5 output includes the predicted mass fraction values ​​of each nutrient, including... y N ,y P ,y K ,y C ; S63: Determine nutrient status classification based on nutrient surplus / deficit index. I i The nutritional status of individual nutrients is classified: I i <−0.5: Severe deficiency; -0.5≤ I i <0: Mild deficiency; 0≤ I i ≤0.5: Suitable; I i >0.5: Excessive; S64: Based on the grading results of all nutrients, a weighted summation method is used to calculate the overall nutritional status index of the plant. I total : ; w i :nutrient i The weight coefficients are determined by the analytic hierarchy process (AHP) or the entropy weight method, and .

[0066] Step S6 is followed by generating corresponding fertilization decision recommendations, the specific process of which is as follows: Based on the profit and loss index of each nutrient I iThe recommended fertilizer application rate is calculated using the following formula: F i =k i ·( U i (t)− y i ); in, F i :nutrient i Recommended fertilizer application rate per unit area; k i :nutrient i The correction coefficient is determined by factors such as soil fertility capacity and fertilizer utilization rate; according to F i Based on the sign and magnitude of the value, corresponding fertilization decision recommendations are generated: F i >0: Recommended to supplement with nutrients i Dosage is F i ; F i ≤0: No need to supplement, if I i A value >0.5 indicates a need to control fertilization or increase the application of organic fertilizer to improve the soil.

Claims

1. A method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy, characterized in that, Includes the following steps: S1: Emit controlled polarized light covering a specified wavelength band to the leaves of honeysuckle. Suppress the specular reflection signal on the surface of the honeysuckle leaves through an orthogonal extinction system. Collect deep diffuse reflectance spectral data by a near-infrared spectral sensor. S2: The near-infrared spectral sensor automatically acquires the reference spectrum of the polytetrafluoroethylene standard reflector before and after sample acquisition, and performs real-time normalization calculation in combination with the dark current background of the sensor to obtain the real-time spectral reflectance R(λ). S3: An algorithm combining orthogonal signal correction and external parameter orthogonalization is adopted to extract the water absorption feature vector and construct the orthogonal complement space of the water-sensitive subspace. The original spectrum is then projected and transformed to achieve physical-level decoupling of the water background signal. S4: The preprocessed signal is decomposed into time-frequency domains at different scales using a multi-scale continuous wavelet transform algorithm, and the characteristic wavelength vectors reflecting the content of total nitrogen, total phosphorus, total potassium and key secondary metabolites of honeysuckle are selected by combining the correlation coefficient method. S5: Input the feature wavelength vector into a pre-trained nonlinear model based on the fusion of deep residual network and self-attention mechanism, and introduce a phenological adaptive compensation algorithm to output the nutrient quality fraction of the plant. S6: Perform nutritional diagnosis, compare the output nutrient content data with the pre-stored database of nutrient surplus and deficit thresholds for the entire growth period of honeysuckle, calculate the nutrient surplus and deficit index, and obtain the nutritional status test results.

2. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 1, characterized in that, The specific process of step S1 is as follows: S11: Emits a continuous spectrum in the 400nm to 2500nm band, which is modulated into linearly polarized light by the first polarizer. The intensity of the emitted light is denoted as... I 0; S12: Separation of specular reflection from the blade surface and deep transmission: After the incident light reaches the blade, it splits into two parts: Intensity of light reflected from the mirror surface of the blade I s It follows Fresnel's law of reflection: I s = I 0· R ( θ,n 1 ,n 2 ,λ )； in, I s : The intensity of light reflected from the specular surface of the blade; I 0: Intensity of the incident ray polarized light after passing through the first polarizer; R ( θ,n 1 ,n 2 , λ Fresnel reflectance; θ Angle of incidence of light; n 1: Air refractive index; n 2: Refractive index of the cuticle of honeysuckle leaves; λ Near-infrared light wavelength; Depolarized diffuse reflection light, intensity denoted as I d ; S13: Orthogonal polarization extinction and specular reflection signal filtering: The transmittance of incident light through the polarizer follows Malus's law: T = cos 2 (D θ ); in, T Δ: The transmittance of the polarizer to incident light; θ The angle between the polarization direction of the incident light and the polarization axis of the second polarizer; Since the reflected light from the mirror retains its original polarization state and is orthogonal to the second polarizer, i.e., Δ θ =90°, therefore the intensity of the outgoing light after the specular reflected light passes through the second polarizer is: I s_out =I s · cos 2 (90°)=0； S14: Effective Deep Diffuse Reflection Signal Acquisition: Due to scattering and depolarization, some of the deep diffuse reflection light passes through the second polarizer. The intensity of the outgoing light after passing through the second polarizer is denoted as... I d_out ; The total light intensity ultimately received by the near-infrared spectral sensor is: I total = I d_out + I s_out ≈ I d_out 。 3. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 2, characterized in that, The specific process of step S2 is as follows: S21: Acquire reference light intensity from the standard reflector I r (λ), sensor dark current background light intensity I d (λ), original reflected light intensity of the sample I s (λ); S22: Real-time normalization calculation of spectral reflectance R (λ): Using the reference light intensity of the standard reflector as the benchmark and the dark current background light intensity as the zero-point compensation, the original reflected light intensity of the sample is subjected to real-time background subtraction and normalization processing to obtain the real-time spectral reflectance. The calculation formula is as follows: R (λ)=( I s (λ)− I d (λ)) / ( I r (λ)− I d (l)).

4. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 1, characterized in that, The specific process of step S3 is as follows: S31: Constructing a moisture-sensitive subspace: Selecting honeysuckle leaf samples with different moisture content gradients, and obtaining the original calibrated spectral matrix. X ; Principal component analysis was used to extract eigenvectors highly correlated with moisture at 1450 nm and 1940 nm in the near-infrared band, forming the eigenvector matrix M of the moisture-sensitive subspace. M =[ m 1, m 2,··· m k ]； in, m 1, m 2,··· m k :forward k One water-dominant feature vector; S32: Construct the orthogonal projection operator P: Based on the EPO algorithm, a projection operator is constructed to project onto the orthogonal complement space of water molecules: P=I−M(M T M) −1 M T ; Where P: orthogonal projection operator, used to remove moisture interference; I: identity matrix, with dimensions consistent with the spectral matrix; M T : The transpose of the moisture eigenvector matrix M; (M T M) −1 Matrix M T The inverse matrix of M; S33: Orthogonal projection transformation of the original spectrum: The original calibration spectral matrix X is projected onto the orthogonal complement space of the moisture-sensitive subspace to achieve physical-level stripping of the moisture signal: X clean =X · P ； in, X Original calibration spectral matrix; X clean : Pure spectral matrix after hydrolysis and coupling; S34: Further purification through joint quadrature signal correction: [The following text appears to be incomplete and requires further context:] X clean Iterative removal of orthogonal noise unrelated to nutrient composition is performed to ultimately output the target spectrum of completely decoupled water. X final : X final =OSC( X clean ); Where OSC(·) is the orthogonal signal correction function.

5. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 4, characterized in that, The original calibration spectral matrix in step S31 X The acquisition process is as follows: Calculate the real-time reflectivity for each wavelength λ R (λ), arranged in order of wavelength from 400nm to 2500nm, representing all wavelength points. R (λ) forms a single-row spectral vector: x =[ R (λ1), R (λ2),···, R (l m )]; When there is n When dealing with a sample, the spectral vectors of all samples are stacked row-wise to form the original calibration spectral matrix. X : ； Where n: number of samples; m: number of spectral wavelength points; R i ( λ j ): No. i The sample at the th j Real-time spectral reflectance at each wavelength; λ j : No. j Near-infrared wavelength.

6. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 4, characterized in that, The specific process of step S4 is as follows: S41: Obtain the pure spectrum of water decoupling based on step S3 X final : X final =[ X final (λ1), X final (λ2),···, X final (λ m [ ], for the purification spectrum after dehydration X final Perform continuous wavelet decomposition at scales 1 to 6 to obtain wavelet coefficients at each scale: ； Among them, W( a,b ):scale a Translation amount b The corresponding wavelet coefficients; a Wavelet transform scale; b Wavelength shift factor; :db4 wavelet basis functions; : Conjugate of wavelet basis functions; S42: Single-scale reconstruction of wavelet coefficients at each scale yields spectral feature components at different scales: ； in, X recon ( a ,λ): Scale a Reconstructed spectral characteristics at the next wavelength λ; C: wavelet admissibility constant; S43: Reconstruction features based on each wavelength X recon (λ) and measured nutrient content y Calculate the Pearson correlation coefficient r (λ): ； in, r (λ): Pearson correlation coefficient between the spectrum and nutrient content at wavelength λ; n : Sample size; X recon,i (λ): the first i Reconstructed features of each sample at wavelength λ; : The mean value of the wavelength λ reconstruction feature; y i : No. i Measured nutrient values ​​for each sample; : Average measured nutrient values; S44: Feature Wavelength Selection and Feature Vector Construction: Set the threshold to 0.85 and retain | r Highly correlated wavelengths with (λ)|≥0.85 form a characteristic wavelength set. : ； Sort by wavelength from smallest to largest to obtain the characteristic wavelength vector. W : W =[ λw 1, λw 2,···, λ wp ]； in, p Number of characteristic wavelengths; λw 1, λw 2,···, λ wp : The characteristic wavelength after screening.

7. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 6, characterized in that, The specific process of step S5 is as follows: S51: Transfer the characteristic wavelength vector W The corresponding spectral eigenvalues ​​are standardized to obtain the standardized eigenvectors. ; S52: The standardized feature vectors Inputting a deep residual network, the residual blocks extract deep nonlinear features. The output formula of the residual blocks is: ； in, F l : No. l Input features of layer residual blocks; F l+1 : No. l Output features of the residual block; Conv(·): Convolutional layer operation; W l : No. l The weight matrix of each convolutional layer; S53: Deep features of the residual network output F Apply a self-attention mechanism, calculate feature weights, and weight them accordingly; S54: Input the weighted attention features into the fully connected layer to obtain a preliminary prediction of the nutrient quality score: ； : Vector of preliminary predicted values ​​of uncompensated nutrient mass fraction; F attn : The feature vector weighted by the self-attention mechanism; FC(·): Fully connected layer operation; S55: Revise the preliminary forecast: ； Among them, phenological period t The corresponding compensation term vector ΔY( t Let t be a function of phenological period t, and represent it using a piecewise linear model: ； Where Y: the vector of predicted nutrient quality fractions after final compensation; : Preliminary predicted value vector of uncompensated nutrient mass fraction; ΔY( t ): Phenological period t The corresponding compensation term vector; t: phenological period, t 1 is the seedling stage, t 2 is the budding stage, t 3 is the peak flowering period, t 4 is the end of the flowering period; α i ,β i : Linear compensation coefficients for each phenological stage.

8. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 7, characterized in that, The specific process of step S6 is as follows: S61: Based on step S5, obtain the nutrient quality fraction prediction vector Y, Y=[ y N ,y P ,y K ,y C ], y N ,y P ,y K ,y C They are respectively N,P,K,C The predicted quality fraction value is obtained by matching the nutrient critical value parameters corresponding to the current phenological stage t of the plant from the pre-stored database of nutrient surplus and deficit critical values ​​for the entire growth period of honeysuckle, including: deficiency critical value. L i (t), suitable critical value U i (t) and excess critical value E i ( t ); S62: Calculate the nutrient break-even index: based on the predicted value. y N ,y P ,y K ,y C The relationship with the critical value is calculated using a piecewise linear model to determine the nutrient surplus / deficit index. I i Used to quantify the profitability of nutrients: ； in, I i :nutrient i Profit and loss index; y i Predicted mass fraction values ​​of each nutrient; S63: Determine nutrient status classification based on nutrient surplus / deficit index. I i The nutritional status of a single nutrient is classified. S64: Based on the grading results of all nutrients, a weighted summation method is used to calculate the overall nutritional status index of the plant. I total : ; w i :nutrient i The weight coefficients are determined by the analytic hierarchy process (AHP) or the entropy weight method, and .

9. The method for detecting the nutritional status of honeysuckle plants based on near-infrared spectroscopy according to claim 8, characterized in that, Step S6 is followed by generating corresponding fertilization decision recommendations, the specific process of which is as follows: Based on the profit and loss index of each nutrient I i The recommended fertilizer application rate is calculated using the following formula: F i =k i ·( U i (t)− y i )； in, F i :nutrient i Recommended fertilizer application rate per unit area; k i :nutrient i Correction coefficients; according to F i Based on the sign and magnitude of the value, corresponding fertilization decision recommendations are generated: F i >0: Recommended to supplement with nutrients i Dosage is F i ; F i ≤0: No need to supplement, if I i A value >0.5 indicates a need to control fertilization or increase the application of organic fertilizer to improve the soil.

Images