Method for losslessly predicting chlorophyll of basil leaves based on hyperspectral close-range imaging

Abstract

The invention provides a method for losslessly predicting chlorophyll of basil leaves based on hyperspectral close-range imaging, which comprises the following steps: 1) extracting spectral data of different positions of leaves in a single basil canopy based on an original spectral image of a basil plant; (2) preprocessing the extracted spectral data by adopting a normalized variable (VSN); and (3) constructing a chlorophyll regression model for detecting the chlorophyll content of the basil leaves by adopting a PLS (Partial Least Squares) deep supervised mechanical learning method according to the spectral data after VSN preprocessing. The method can rapidly, accurately and nondestructively detect the chlorophyll content of the leaves in situ, and has important significance for monitoring the chlorophyll condition in time and improving the planting management strategy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological component detection, and relates to a method for nondestructive prediction of chlorophyll in Ocimum basilicum leaves based on hyperspectral close-range imaging. BACKGROUND

[0002] Ocimum is an important genus of Lamiaceae, containing a variety of annual and perennial herbs and shrubs, and most plants in the genus have aromatic flavor (Abirami et al., 2023). This plant is widely cultivated, from Europe to Asia, and different varieties have different morphologies, leaf shapes and aromas (Bajomo et al., 2022; -Stanko et al., 2010). Healthy Ocimum basilicum is widely concerned due to its rich secondary metabolites with biological activity, and aromatic essential oils and polyphenols are used in the perfume, cosmetic and pharmaceutical industries (Li & Chang, 2016; Pandey et al., 2014). However, abiotic stress, especially drought, can significantly inhibit the growth of Ocimum basilicum, which is manifested as a decrease in plant height, diameter and leaf fresh weight, and this growth impairment further hinders the synthesis and accumulation of secondary metabolites with important economic value (Radácsi et al., 2025). Chlorophyll is a key indicator for evaluating photosynthetic capacity, nutritional status and growth stage, and can directly reflect the health level of plants (FILELLA & PENUELAS, 1994). Therefore, timely detection of the chlorophyll content of Ocimum basilicum at a certain growth stage or environmental condition is crucial for timely grasping its physiological state.

[0003] Conventional methods for measuring chlorophyll content include tissue disruption, chemical extraction and spectrophotometer detection (Wellburn & Lichtenthaler, 1984). These methods in the past are invasive and have complicated procedures and time-consuming. A portable chlorophyll meter (SPAD-502) can avoid direct damage to plant tissues, and can quickly and conveniently establish a correlation curve to detect leaf chlorophyll content (Netto et al., 2005). However, the SPAD point measurement method can only collect the chlorophyll content of a single leaf at a time, and when the sample size is large, it cannot meet the rapid capture of the chlorophyll content in the entire canopy. SUMMARY

[0004] In view of the above technical problems, the present application aims to provide a method for nondestructive prediction of chlorophyll in Ocimum basilicum leaves based on hyperspectral close-range imaging, which can quickly realize the rapid nondestructive detection of the chlorophyll distribution of Ocimum basilicum leaves.

[0005] The technical scheme adopted by the present application to achieve the technical purpose is as follows:

[0006] A method for non-destructive prediction of chlorophyll in basil leaves based on hyperspectral close-range imaging, comprising:

[0007] 1) Extracting spectral data of different positions of leaves in the canopy of a single basil plant based on the original spectral image of the basil plant;

[0008] 2) Preprocessing the extracted spectral data using variable normalization (VSN);

[0009] 3) Constructing a chlorophyll regression model using the PLS deep supervision machine learning method on the VSN-processed spectral data for detection of chlorophyll content in basil leaves.

[0010] Preferably, in step 1), the original spectral image of the basil plant is collected by a hyperspectral close-range imaging system, which includes an FS-IQ-VISNIR camera, a 25mm lens, a white reference plate, and two LED lights for supplementary lighting.

[0011] More preferably, before image collection, the lens is covered to perform black reference on the camera to reduce noise by deducting dark current.

[0012] More preferably, during the collection process, a white reference plate is placed beside the basil plant, and the exposure time of the camera and the intensity of the lighting are adjusted according to the reflectance of the white plate to prevent overexposure.

[0013] More preferably, the exposure time is set to 50000ms, the spectral merging is 4, the spatial merging is 4, and the spectral range is 400-1000nm.

[0014] Preferably, in step 1), the first to third leaves of a single basil plant are collected, and each leaf is collected on both sides, with six rectangles collected per plant, and the corresponding single reflectance spectral data of each sampling area is extracted.

[0015] Preferably, in step 2), the VSN data processing comprehensive formula is represented as:

[0016]

[0017] x ij represents the original spectral data, u j and σ j represent the mean and standard deviation of the jth spectrum, and Xij represents the normalized spectral data.

[0018] Preferably, in step 3), the PLS deep supervision machine learning method includes selecting 80% of the data samples as the training set and 20% as the test set.

[0019] More preferably, based on the training set, the model training and prediction are carried out by using leave-one-out cross-validation (LOO-CV), and the root mean square error value (RMSEP) of prediction is recorded, and when the RMSEP value reaches the minimum and its change curve starts to tend to be stable, the corresponding component number is the optimal state of the model.

[0020] More preferably, the first 10 main components are obtained for cross-validation.

[0021] The beneficial effects of the present application are:

[0022] The present application is aimed at cultivating basil under different drought stress conditions, analyzes the influence of drought on the chlorophyll of basil leaves, and proposes a method for quickly realizing the rapid and non-destructive detection of the chlorophyll distribution of basil leaves by reducing noise and light influence. The method is applied to predict the distribution of chlorophyll content of basil under different drought stress degrees, and good results are obtained, and the prediction R 2 value reaches 0.858. The overall results show that, based on the analysis of the influence of light deviation, PLS combined with VSN close-range HSI pipeline significantly reduces these influences, so as to quickly, accurately and non-destructively detect the chlorophyll content of leaves in situ. The method has important significance for timely monitoring of chlorophyll condition and improving planting management strategy. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 Four kinds of drought stress treatments of basil are grown in a climate chamber, a: basil with two true leaves after planting for 20 days; b: basil after 20 days of stress after transplanting; c: basil after 30 days of stress after transplanting.

[0024] Figure 2 Hyperspectral close-range imaging system.

[0025] Figure 3 Framework diagram of data pipeline system.

[0026] Figure 4 Basil plants after 30 days of drought stress, a: 100%, b: 75%, c: 50%, d: 25%.

[0027] Figure 5 Planting height and fresh weight of basil under different drought stress intensities, 100%, 75%, 50%, 25% represent 4 groups of different drought stress intensities, respectively, and the data is expressed by mean value ± standard deviation. Different letters represent significant difference (p<0.05), the same below.

[0028] Figure 6 Chlorophyll content of basil planted under different drought stress intensities.

[0029] Figure 7 Six sample areas (A1-C2) extracted on the hyperspectral image of the canopy of a single basil plant.

[0030] Figure 8 Figure 7 Spectral curves of six sample areas of a single L. camara plant, (a) raw reflectance spectra, (b) SG smoothed spectra, (c) VSN smoothed spectra.

[0031] Figure 9 Spectral curves of multiple pixels of L. camara under four different stress levels, (a) raw reflectance spectra, (b) SG smoothed spectra, (c) VSN smoothed spectra.

[0032] Figure 10 Model performance of two spectral preprocessing training sets, (a) SG preprocessing cross-validation prediction, (b) SG preprocessing score plot, (c) VSN preprocessing cross-validation prediction, (d) VSN preprocessing score plot.

[0033] Figure 11 Correlation of two preprocessing model predictions, (a) SG preprocessing, (b) VSN preprocessing. DETAILED DESCRIPTION

[0034] The present application will be described in detail below with specific reference cases, which are helpful for those skilled in the art to further understand the present application, but do not limit the present application in any form.

[0035] EXAMPLE

[0036] I. Materials and methods

[0037] 1.1 Experimental plant cultivation

[0038] In this experiment, L. camara was used as the research object. L. camara was cultivated in a light box in the laboratory of the Forestry College of the Tropical Agriculture and Forestry College of Hainan University in Danzhou, Hainan Province on March 1, 2025. The soil used for cultivation was peat soil (Table 1). The climate box was set to 14 / 10 h (light / dark cycle) for 14 / 10 h (light / dark cycle), day / night temperature of 28℃ / 18℃, and relative humidity of 70%.

[0039] Table 1 Characteristics of soil medium mixture

[0040]

[0041] 1.2 Experimental design and treatment

[0042] The experiment set up one control group with 100% soil water holding capacity (SWC), and three experimental groups with 75%, 50% and 25% SWC, corresponding to mild drought, light drought and severe drought, respectively. Five plants were grown in each drought level treatment with five replicates. The SWC of soil was measured by the gravimetric method modified by (Reynolds, 1970) and (Kulak et al., 2021). First, the peat soil was completely saturated with distilled water. Then, the saturated soil was left to stand at room temperature for 24 h to remove excess water. The soil was weighed (Ww) and placed in an oven set at 105°C for 24 h until the mass of the soil changed by no more than ±0.1 g. At this time, the dry weight of the soil (Wd) was measured. The formula for calculating the SWC (%) of the soil is as follows:

[0043]

[0044] SWC: soil water holding capacity, Ww: saturated soil weight after standing for 24 h, Wd: dry weight of soil.

[0045] Basil seeds were sown on March 1st in 50-hole plug trays. After 15-20 days, the basil seedlings with two true leaves were transplanted into single-hole hard plastic nursery cups with a diameter of 10 cm. Figure 1 a) After 5-7 days of stable growth, the plants were subjected to drought stress for 30 days. Figure 1 b and c) During the entire drought process, a layer of plastic film was placed at the bottom to prevent soil loss, and 4-5 small holes were made for ventilation. The soil water content of each treatment was weighed and recorded daily. To avoid errors in water content, the weight was measured at a fixed time each day to determine the water loss the previous day, and then additional water was added according to the weight loss to maintain the soil water content.

[0046] 1.3 Hyperspectral image acquisition

[0047] After 30 days of drought stress, the basil was removed from the phytotron for data collection. The first to third leaves of each plant were collected, with six rectangles collected from each plant. As shown in Figure 2 The hyperspectral close-range imaging system included a FS-IQ-VISNIR camera (Analytical Spectral Devices, Hangzhou, China), a 25 mm lens, a white reference plate, and two LED lights (140 W) for supplementary lighting. Before image acquisition, the camera was covered with a lens cap to perform dark calibration, which subtracted the dark current to reduce noise. During the acquisition process, a white reference plate was placed beside the basil plant. The camera exposure time and lighting intensity were adjusted according to the reflectance of the white plate to prevent overexposure. After several experiments, the exposure time was set to 50,000 ms, the spectral merging was set to 4, the spatial merging was set to 4, and the spectral range was set to 400-1000 nm.

[0048] The raw spectral data obtained after the instrument calibration were pre-processed by MATLAB 2023a.

[0049] 1.4 Measurement of chlorophyll content and growth indicators

[0050] After the hyperspectral imaging of basil plants, one leaf was immediately picked from the first, second, and third nodes of each basil plant. Each treatment group contained 5 plants, and there were 4 treatment groups (a total of 20 plants). Three leaves were collected from each plant, and a total of 60 samples were obtained for destructive chlorophyll content determination. The leaf samples were washed with distilled water and cut in half along the midrib for grinding the plant tissue. Then, the crushed samples were placed in an acetone extraction solution and soaked in the dark for 2-4 h until the plant residue turned into a white precipitate. Finally, 1 ml was taken for absorbance measurement at 645 and 663 nm using a spectrophotometer, and the method of (Wellburn & Lichtenthaler, 1984) and (Johan et al., 2014) was used for total chlorophyll content calculation. The growth indicators of basil were measured, including plant height and leaf fresh weight. The plant height was measured from the base to the top of the main stem using a vernier caliper. The leaf fresh weight was measured using an electronic balance.

[0051] 1.5 Data analysis pipeline

[0052] The data pipeline system framework used in this application is shown in Figure 3 After sample preparation and acquisition of hyperspectral images, chlorophyll content and other growth indicators were measured and statistically analyzed. Then, the effects of noise and light on canopy close-range imaging HSI were analyzed, and the spectral pretreatment method was selected accordingly and applied to the extracted spectrum. Next, different spectral pretreatment methods were combined with supervised machine learning methods to establish a chlorophyll regression model, and the effectiveness of the model was proved by cross-validation. Finally, the best close-range HSI analysis pipeline to reduce the effects of noise and light was determined and applied to the detection of basil leaf chlorophyll content.

[0053] 1.5.1 Preprocessing using SG

[0054] First, Savitzky-Golay (SG) smoothing is applied to the extracted spectral matrix. This method, based on the principles proposed by Savitzky and Golay, utilizes least-squares polynomial fitting within a local window to effectively suppress noise while maximizing the preservation of spectral peak characteristics (Savitzky & Golay, 1964). In the system acquiring the spectral data, the window width is set to 2M+1 (i.e., containing 2M+1 adjacent data points), with each point corresponding to the original absorbance value of one wavelength. SG smoothing achieves signal smoothing by fitting an Nth-order polynomial within this window (Schafer, 2011).

[0055]

[0056] P(n) represents the value of the polynomial smoothed curve at a certain center point, a i Represents the polynomial smoothing coefficient.

[0057] To solve for the polynomial coefficients more efficiently, the smoothing value P( ) at each sample point can be simplified by matrix transformation. n First, design a matrix A, where the rows correspond to the sample indices within the window, and the columns correspond to the polynomial order. Let the light wave range sample vector be X = [a0, a1, ..., a...]. n ] T The corresponding original absorbance response vector is Y = [y0, y1, ..., a n ] T Therefore, the following normal equation after transpose can be obtained:

[0058] A T A X =A T Y

[0059] X[a0,a1,…,a n ]=((A T A)) -1 A T Y

[0060] This application uses the SG filter primarily because it effectively preserves key spectral features while smoothing the data. During spectral acquisition, raw spectral data typically contains noise from various sources, such as instrument and environmental noise. The SG filter selects a local cell to construct a moving window and performs polynomial fitting within the window, significantly suppressing baseline drift caused by noise, thus obtaining a more stable and reliable spectrum. This application uses a window size of 5, and all preprocessing steps are performed on the MATLAB platform.

[0061] 1.5.2 Preprocessing using VSN

[0062] Variable standardization (VSN) is a technique developed in recent years to correct for multivariate scattering. This technique is only affected by additive and multiplicative effects of signal variables, and ignores some variables that are not important to the target (Rabatele et al., 2020). VSN achieves uniform adjustment of data scale and minimizes chemical variables by assigning weights to each spectral variable, thereby enhancing the comparability and analysis effect of spectral data, and significantly improving the properties of signals and model interpretation. The VSN data processing comprehensive formula can be expressed as

[0063]

[0064] x ij represents the original spectral data, u j and σ j represent the mean and standard deviation of the jth spectrum, and Xij represents the normalized spectral data.

[0065] 1.5.3 Modeling method

[0066] Partial least squares regression (PLS) is a commonly used modeling toolkit in R language. When building a PLS model, the optimal number of latent variables needs to be determined first to avoid model underfitting or overfitting (Mevik & Wehrens, 2007). In this embodiment, 80% of the data samples are selected as the training set and 20% as the test set. Based on the training set, leave-one-out cross-validation (LOO-CV) is used for model training and prediction, and the root mean square error of prediction (RMSEP) is recorded. According to the description of Bjorn-Helge Mevik and Ron Wehrens, by obtaining the first 10 main components, through program verification, when the RMSEP value reaches the minimum and its change curve starts to stabilize, the corresponding component number is the optimal state of the model.

[0067] II. Results and discussion

[0068] 2.1 Growth and chlorophyll content of basil under different drought stress intensities

[0069] Figure 4 The figure shows the growth of basil plants under different stress intensities of SWC 100%, 75%, 50%, and 25% at the end of transplanting drought cultivation (30d). The growth vigor of basil crops under 50% and 25% soil water content is obviously weak, and the first node leaf yellowing is serious, and the lateral branches do not grow new buds. The growth of basil under 100% and 75% is stronger, the crown is more lush, the leaves are larger, and the first node grows new leaves.

[0070] Figure 5 and Figure 6The statistical data of plant height, fresh weight, and chlorophyll content of basil plants after stress are shown in four groups. The vertical distance from the base of the plant to the meristem of the main stem was measured. Each group of plants was measured five times, and the average value was taken as the plant height for that group. Fresh weight and chlorophyll content were measured at the first, second, and third nodes of the plants, and the average values ​​were calculated. It is evident that there are significant differences in plant height, fresh weight, and chlorophyll content among the four groups. Clearly, with increasing drought stress intensity, the growth of basil plants is significantly inhibited, specifically, plant height and fresh weight decrease significantly with increasing stress intensity. Under severe drought (25%), basil plant height and fresh weight decreased sharply, reaching the lowest level among the treatments. Figure 5 Chlorophyll content peaked under mild drought stress (75%). However, when drought stress intensity exceeded this range (≤50%), chlorophyll content decreased significantly with increasing stress intensity, showing a significant negative correlation between the two. Figure 6 ).

[0071] The 75% group exhibited the best growth vigor, with the highest values ​​for all three indicators, followed by the 100% group. Plant height, fresh weight, and chlorophyll content decreased in the order of 75%, 100%, 50%, and 25%. Basil plants grown under severe drought stress (25%) showed significantly lower plant height, fresh weight, and chlorophyll content compared to other treatment groups. In contrast, plants under mild drought stress (75%) reached a peak chlorophyll content, but plant height remained at a moderate level, not the highest. This phenomenon also indicates that moderate drought stress can induce physiological responses in basil plants, promoting chlorophyll synthesis and increasing its content. However, severe drought stress not only significantly inhibited the accumulation of plant height and fresh weight but also severely reduced chlorophyll content (Jafari et al., 2019; Mulugeta et al., 2023).

[0072] 2.2 Analysis of local extraction spectra at different locations within the canopy of a single basil plant

[0073] To minimize interference from various factors such as location, light intensity, and noise, the chlorophyll content of leaves was assessed from multiple perspectives. Based on the original spectral images of basil plants, six small rectangles were manually identified. Figure 7 The data are denoted as A1 to C2, and the reflectance spectral data of each individual plant corresponding to each sampling area are extracted for further preprocessing and analysis. The subsequent training set data is taken from the first, second, and third nodes of each individual plant, with two rectangles identified on each leaf and two single points taken.

[0074] according to Figure 7 The raw spectral data obtained from individual plants were further preprocessed and analyzed, including SG smoothing and VSN normalization, to evaluate the effectiveness of the spectral preprocessing.

[0075] Due to the effects of noise and light,Figure 8 The original spectrum in (a) shows a significant difference in absorbance between 430–500 nm and 500–650 nm. However, in Figure 8 In (b), after SG smoothing, the difference from (a) is not significant. This is because the hyperspectral imager subtracts dark current during imaging, resulting in relatively low noise. Therefore, the spectral deviation mainly originates from the influence of illumination. Figure 8 (a) In the original spectrum, C1 has the highest absorbance, while A1 has the lowest. Besides the influence of the leaf's health status, the spatial heterogeneity of light conditions is the main factor causing this difference. In regions B and C, the leaf areas are nearly perpendicular to the incident light direction, resulting in concentrated light energy on a single surface and high absorbance in the original spectrum. In contrast, the leaves in region A are flatter, with a larger horizontal angle than in regions B and C, increasing light scattering under exposure conditions and thus reducing the absorbance of the original spectrum (Makdessiet et al., 2017). The absorbance of the original spectrum at site C2 in region C shows a sharp decline, and the peak-valley characteristics in the 500-650 nm band differ significantly from those at site C1. Besides the influence of light conditions, the shading effect caused by mutual shading of leaves is a key factor. The leaves in the upper right corner of region C and the leaves in front of it will more or less block the light reaching C2, so the main difference between C1 and C2 comes from the effects of blocking and shading (Zhu et al., 2024).

[0076] The six raw spectral curves were preprocessed by SG and VSN, respectively, as shown below. Figure 8 As shown in (b) and (c). Although Figure 8 The original spectrum in (a) was significantly shifted due to the working environment, but the influencing factors were greatly reduced after preprocessing. SG window smoothing mainly eliminates baseline shifts caused by noise. Because dark current is subtracted, the SG preprocessed spectrum is quite similar to the original spectrum, with only minor changes. VSN is a relatively new preprocessing method, which is essentially a weighted form of SNV preprocessing. Therefore, the curve difference is significantly reduced after VSN preprocessing (Rabatel et al., 2020). Figure 8In (a) and (b), the pre-processed spectral curves show typical reflectance of chlorophyll. In the visible light region of 400-500 nm, the reflectance is relatively low due to the strong absorption of leaf pigments in the blue light region. In the visible green light region of 500-570 nm, there is a clear peak-valley, which is caused by the strong reflection of chlorophyll molecules (Ollinger, 2011; Yendrek et al., 2017). At the infrared inflection point of 680-700 nm, the leaf reflectance slightly increases due to the reflection of infrared spectrum (FIELLLA and PENUELAS, 1994). To further test the effectiveness and practicality of spectral preprocessing, multiple pixel spectra were extracted from the four groups of basil under different drought stress intensities, as shown in Figure 9 (a) shows the original spectrum, and (b) and (c) show the SG and VSN pre-processed spectra, respectively. Overall, it can be observed that the original spectrum is mitigated, to some extent, verifying the effectiveness of the spectrum.

[0077] 2.3 Evaluation of modeling performance of PLS regression

[0078] The combination of the two spectral preprocessing methods of SG and VSN and the PLS deep supervision machine learning method obtained two modeling results, Figure 10 and Table 2. Compared with the previous results related to machine learning high-throughput prediction of plant chlorophyll, the performance of the model is generally improved to some extent, which benefits from the selection of the environment for obtaining the spectrum, the preprocessing method, the fitting of the model, and the removal of discrete values.

[0079] PLS models can be quickly fitted by multivariate linear regression modeling of the spectral values of the chlorophyll samples obtained from the test set. For each data matrix, consider the lowest RMSEP value to select components to carry out the cross-validation process, while checking the variance of the relevant latent variables in x and y, in Figure 10 In (a) and (c), according to the lowest mean square error RMSEP 0.001143 and 0.001172, components 6 and 4 are obtained for cross-validation prediction. From the cross-validation process, it can be seen that the training values of 80% of the samples can be almost concentrated along the diagonal, and there is no bending or other abnormal signs, indicating that the model cross-validation is highly reliable. In Figure 10 In (b) and (d), the multivariate linear regression model is scored according to the three main components, and the score map can clearly find the anomalies of the data and evaluate the feasibility of the model. In these two models, there are no explicit outliers, which means that the distribution of the data on these 6 and 4 main components does not show obvious clustering or outliers, which may indicate that the variability of the data is evenly distributed on these components, or the data itself does not have a strong grouping structure.

[0080] Table 2 shows the performance of the regression model based on a combination of two spectral preprocessing methods and PLS deep supervised machine learning.

[0081]

[0082] Table 2 shows that after PLS modeling using the two preprocessing methods, SG smoothing performed the worst in terms of reflected spectral density, while VSN preprocessing significantly improved the results. This is because SG smoothing only eliminates noise and reduces baseline drift, but it does not eliminate the influence of illumination. VSN preprocessing, on the other hand, can mitigate these effects to some extent. As shown in Section 2.2, the deviation of the spectral curves under VSN preprocessing is very small, which is also reflected in the better model fitting effect after being combined with PLS. This model fits a 6-component model and includes predictions using leave-one-out (LOO) cross-validation. We can obtain an overview of the fitting and validation results through a summary method. Figure 9 a and b), ultimately R 2 It can be improved to 0.858, which is 4 percentage points higher than SG pretreatment, such as Figure 11 As shown, in (a), the predicted sample dispersion is significantly higher than that in (b).

[0083] Obviously, the above embodiments of the present invention are merely examples to illustrate the present invention more clearly, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all implementation methods here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A method for non-destructive prediction of chlorophyll in basil leaves based on hyperspectral close-range imaging, comprising: 1) extracting spectral data of leaves at different positions in the canopy of a single basil plant based on an original spectral image of the basil plant; 2) pre-processing the extracted spectral data using normalized variable VSN; 3) constructing a chlorophyll regression model using the VSN-processed spectral data by PLS deep supervision machine learning method for detection of chlorophyll content in basil leaves.

2. The method of claim 1, wherein, In step 1), the original spectral image of the basil plant is collected by a hyperspectral close-range imaging system, which comprises an FS-IQ-VISNIR camera, a 25mm lens, a white reference plate, and two LED lights for supplementary lighting.

3. The method of claim 2, wherein, Before image acquisition, the camera is black calibrated by covering the lens to reduce noise by deducting dark current.

4. The method of claim 2, wherein, During the acquisition process, a white reference plate is placed beside the basil plant, and the exposure time and illumination intensity of the camera are adjusted according to the reflectance of the white plate to prevent overexposure.

5. The method of claim 2, wherein, The exposure time is set to 50000ms, the spectral merging is 4, the spatial merging is 4, and the spectral range is 400-1000nm.

6. The method of claim 1, wherein, In step 1), the first to third leaves of a single basil plant are collected, and each leaf is collected on both sides. Six rectangles are collected from each plant, and the corresponding single reflectance spectral data of each sampling area are extracted.

7. The method of claim 1, wherein, In step 2), the VSN data processing comprehensive formula is represented as: x ij representing the original spectral data, u j and σ j representing the mean and standard deviation of the jth spectrum, Xij represents the normalized data of the spectrum.

8. The method of claim 1, wherein, In step 3), the PLS deep supervision machine learning method comprises: selecting 80% of the data samples as the training set and 20% as the test set.

9. The method of claim 8, wherein, Based on the training set, the model is trained and predicted by leave-one-out cross-validation (LOO-CV), and the root mean square error value (RMSEP) of prediction is recorded. When the RMSEP value reaches the minimum and its change curve starts to stabilize, the corresponding component number is the optimal state of the model.

10. The method of claim 9, wherein, The first 10 main components are obtained for cross-validation.

Images