Hyperspectral technology-based hydrangea leaf nitrogen and phosphorus content prediction method
Through hyperspectral technology and machine learning algorithms, a nitrogen-phosphorus prediction model for large-leaf hydrangea was constructed, which solved the problem of inaccurate prediction of nitrogen-phosphorus content in the existing technology, and achieved efficient monitoring of nitrogen-phosphorus content and optimized application of fertilizers. Reduced environmental pollution.
Patent Information
- Application Number
- CN202410782667.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-17
- Publication Date
- 2024-11-22
AI Technical Summary
The prior art is difficult to accurately predict the nitrogen and phosphorus content of large-leaf hydrangea, resulting in improper use of fertilizers, affecting plant growth and environmental pollution.
Hyperspectral technology is used to combine continuous wavelet transformation and partial least squares regression modeling, and through different spectral conversion methods and machine learning algorithms, a prediction model for nitrogen and phosphorus content of large-leaf hydrangea leaves is constructed, and band screening is optimized to improve prediction accuracy.
It significantly improves the accuracy of nitrogen and phosphorus prediction, provides real-time monitoring of the nutritional status of large leaf hydrangea and key technical support for efficient fertilizer application, and reduces environmental pollution.
Smart Images

Figure CN119000559A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for predicting nitrogen and phosphorus contents in hydrangea leaves. Background Art
[0002] Hydrangea is a popular horticultural ornamental plant worldwide. Hydrangea can be used for landscape construction in gardening and can also be sold as potted flowers and cut flowers in the flower market to generate high economic benefits. It has rich flower shapes and varied flower colors. Currently, there are more than 1,000 cultivated varieties. Hydrangea has large leaves and full inflorescences, and has a high demand for nutrients. Nitrogen and phosphorus are essential elements for plants. They are directly involved in the physiological and biochemical processes of plants and play a vital role in the structure, yield and quality of plants. Nitrogen deficiency can cause leaves to become smaller, turn yellow, and reduce chlorophyll content, resulting in reduced biomass. Phosphorus deficiency interferes with the production of chlorophyll, causing anthocyanins to accumulate and causing the leaves to turn purple. In production practice, in order to ensure the quality of seedlings, nursery producers usually over-apply fertilizers based on experience. From an environmental perspective, nitrogen and phosphorus nutrient management is crucial. Nitrogen and phosphorus runoff generated by irrigation in nursery production is one of the main causes of eutrophication of surface water bodies. In addition, compared with field crops, the soilless substrate (pine scales, peat, vermiculite or sand) used in the production process of ornamental seedlings in containers has higher nutrient leaching after irrigation and leaching. In recent years, with the enhancement of ecological awareness, higher requirements have been put forward for the application of chemical fertilizers, and achieving zero growth or negative growth of chemical fertilizers has become a general trend. Therefore, the rational application of nitrogen and phosphorus is of great significance to the seedling production of large-leaf hydrangea and the reduction of environmental pollution. Research in recent years has focused on applying the least amount of fertilizer to achieve the growth quality required by the plant. Studies have shown that the minimum phosphorus fertilizer concentration for hydrangea to achieve maximum aboveground biomass (SDW) is 4.7 mg·L -1 ; The plant growth index and biomass of hydrangea were higher when 1.5gN / plant was applied, but excessive nutrient stress would occur at a concentration of 2.0gN / plant.
[0003] In recent years, more efficient and accurate technologies have been gradually applied to plant nutrition diagnosis. Traditional nutrition diagnosis methods (such as Kjeldahl nitrogen determination) are time-consuming, labor-intensive and costly. Hyperspectral technology has the advantages of accuracy, speed, non-destructiveness and environmental friendliness. It can obtain high-resolution, high-dimensional continuous, narrow-band image data, with rich spectral information, and can be applied to plant nitrogen, phosphorus, potassium diagnosis, water use efficiency evaluation and other scenarios. Hyperspectral technology has good application prospects in plant nutrition monitoring. It is currently widely used in single nutrient element monitoring of rice, wheat, cotton and other crops. Hyperspectral data contains information on multiple nutrient elements, and single nutrient element inversion modeling will result in hyperspectral data not being fully mined. Due to differences in nutrient element content, leaf tissue structure and spectral characteristics of each plant, the prediction accuracy of models established with different plants and different collected data is also significantly different. It is necessary to carry out spectral conversion, band screening and model comparison in a targeted manner to improve the prediction accuracy of the model. Summary of the invention
[0004] The purpose of the present invention is to solve the defects in the prior art and provide a method for predicting nitrogen and phosphorus in Hydrangea macrophylla that can effectively improve the prediction accuracy.
[0005] In order to achieve the above object, the technical solution provided by the present invention is as follows:
[0006] A method for predicting nitrogen and phosphorus content in hydrangea leaves based on hyperspectral technology, the method comprising the following steps:
[0007] (1) Different nitrogen or phosphorus treatments were applied to large-leaf hydrangea to obtain leaf-scale hyperspectral images and corresponding leaf nitrogen or phosphorus contents at different growth stages;
[0008] (2) Calculate the average spectral curve of all pixel points of the leaves in each hyperspectral image, obtain the spectral reflectance data, and then obtain the original spectral reflectance data through standard normal transformation and filter smoothing;
[0009] (3) After the original spectrum was processed by continuous wavelet transform (CWT), the correlation coefficients between the spectral reflectance at each scale of 1-10 and the corresponding leaf nitrogen content or phosphorus content were calculated, the correlation coefficient values were sorted in descending order, and the top 5% were retained as the spectral characteristic bands; then partial least squares regression modeling (PLSR) was used
[0010] Obtain the PLSR prediction model based on CWT;
[0011] (4) Take leaves of hydrangea plants to be tested for nitrogen and phosphorus content, obtain the spectral reflectance data of the spectral characteristic band described in step (3) and input it into the model, and use the CWT-based PLSR prediction model described in step (3) to estimate the corresponding nitrogen content or phosphorus content.
[0012] The present invention uses different spectral conversion methods and machine learning algorithms to compare the prediction accuracy of nitrogen and phosphorus content in leaves of different models, and determines the optimal model of the present invention for estimating nitrogen and phosphorus content in hydrangea leaves.
[0013] Furthermore, in step (1), the large-leaf hydrangea is planted as potted cuttings, and different nitrogen treatments or phosphorus treatments are performed after the plants grow to 4-5 pairs of leaves; the different nitrogen treatments use the following four nitrogen application rates: N 0 :N=0g·pot -1 , N 1 :N=0.4g·pot -1 , N 2 :N=0.8g·pot -1 , N 3 :N=1.2g·pot -1 The nitrogen application rate of each treatment was divided into 7 parts and one part was applied every 7 days; the phosphorus treatment used the following 4 treatment methods: P 0 :P 2 O 5 =0g·pot -1 , P 1 :P 2 O 5 =0.15g·pot -1 , P 2 :P 2 O 5 =0.3g·pot -1 , P 3 :P 2 O 5 =0.45g·pot -1 The phosphorus application rate for each treatment was divided into three parts and one part was applied every 15 days; and potassium fertilizer was applied routinely to each treatment.
[0014] Urea (containing 46.6% N), superphosphate (containing 46.6% P), and potassium fertilizers were selected respectively. 2 O 5 14.5%) and potassium sulfate (containing K 2 O 54.1%).
[0015] Furthermore, in step (1), the different growth stages are selected at 0, 20, 40 and 60 days after the first fertilization of each treatment.
[0016] Furthermore, when constructing the CWT-based PLSR prediction model in step (3), 75% of the total samples are randomly selected as the training set, and the remaining 25% are used as the test set.
[0017] Compared with the prior art, the present invention has the following advantages:
[0018] Based on the technology and theory of hyperspectral, the present invention reveals the relationship and rules between the leaf nitrogen and phosphorus contents and leaf characteristics and spectral reflectivity that reflect the growth status of bigleaf hydrangea. Through machine learning modeling and other means, a nitrogen and phosphorus prediction model based on leaf spectral reflectivity is constructed, which provides key technical support and theoretical basis for real-time monitoring of the nutritional status of bigleaf hydrangea and optimization of the efficient application of nitrogen and phosphorus fertilizers in bigleaf hydrangea.
[0019] The present invention combines different spectral conversion methods (first-order derivative and continuous wavelet transform) and machine learning algorithms, compares the prediction accuracy of leaf nitrogen and phosphorus content of different models, and determines the optimal model for estimating the nitrogen and phosphorus content of hydrangea leaves; and optimizes the band screening method of continuous wavelet transform, using the first 5% of the correlation coefficients between spectral reflectance and leaf nitrogen and phosphorus content at each scale in descending order as spectral feature bands, further improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 Here are photos of bigleaf hydrangea at different growth stages;
[0021] Figure 2 is the spectral reflectance of hydrangea leaves. In the figure, a is the original spectrum of the leaves, b is the spectrum after first-order differential processing, c is the average reflectance spectrum of nitrogen and phosphorus, and d is the average reflectance spectrum of nitrogen and phosphorus after first-order differential processing;
[0022] Figure 3 is the average spectral reflectance of hydrangea leaves under different fertilization treatments. In the figure, a is the original spectrum of leaves at different nitrogen levels, b is the spectral reflectance of leaves at nitrogen levels after first-order differential treatment, c is the original spectrum of leaves at different phosphorus levels, and d is the spectral reflectance of leaves at phosphorus levels after first-order differential treatment;
[0023] Figure 4 The continuous wavelet transform reflectance changes of leaf spectra under different nutritional conditions. In the figure, a is the nitrogen treatment, b is the phosphorus treatment, and c is the control group;
[0024] Figure 5 The correlation between different spectral data and nitrogen and phosphorus content. In the figure, a is the correlation analysis between the original spectral reflectance of leaves and nitrogen content, b is the correlation analysis between the spectral reflectance of leaves and nitrogen content after first-order differential processing, c is the correlation analysis between the original spectral reflectance of leaves and phosphorus content, and d is the correlation analysis between the spectral reflectance of leaves and phosphorus content after first-order differential processing;
[0025] Figure 6 is the absolute value of the correlation coefficient between the wavelet coefficient and the leaf nitrogen and phosphorus at different scales. In the figure, a is the nitrogen content and b is the phosphorus content;
[0026] Figure 7 This is a comparison of three prediction models for nitrogen and phosphorus content in Hydrangea macrophylla leaves based on first-order differential transformation;
[0027] Figure 8 It is a PLSR prediction model for nitrogen and phosphorus content in Hydrangea macrophylla leaves based on continuous wavelet transform.
[0028] Fig. 9 This is a graph showing the importance ranking of features in the PLSR model after continuous wavelet transform. DETAILED DESCRIPTION
[0029] The present invention is described in detail below in conjunction with specific embodiments and drawings.
[0030] Example 1
[0031] Determination of prediction model for nitrogen and phosphorus content
[0032] 1. Test materials and processing
[0033] The pot experiment was conducted in Shanghai Chenshan Botanical Garden. The test material was the cuttings of Hydrangea macrophylla 'Hanatemari', planted in a 13 cm flower pot. The substrate consisted of mudstone: coconut coir: perlite (volume ratio 3:1:1), and no fertilizer was used. The chemical properties of the substrate were as follows: pH 6.45, EC 1.27mS / cm, total nitrogen 7.96mg / g, total phosphorus 1.09mg / g, total potassium 3.33mg / g, and organic matter 865.18g / kg. The experiment was pre-cultivated in a climate chamber for 8 weeks, with a photoperiod of 16 hours, and the photosynthetic photon flux density PPFD provided by the LED light was 101.93μmol m -2s-1 During the pre-cultivation period, the plants were irrigated with unfertilized tap water.
[0034] Fertilization treatment was started when the plants grew to 4-5 pairs of leaves. The experiment adopted a completely randomized block design, with 4 nitrogen treatments and 4 phosphorus treatments, and no nitrogen and phosphorus fertilizers were used as controls (Table 1). The four nitrogen application rates were N 0 (N=0g·pot -1 ), N 1 (N = 0.4 g·pot -1 ), N 2 (N=0.8g·pot -1 ),N 3 (N=1.2g·pot -1 );P 0 (P 2O 5 =0g·pot -1 ),P 1 (P 2 O 5 =0.15g·pot -1 ), P 2 (P 2 O 5 =0.3g·pot -1 ),P 3 (P 2 O 5 =0.45g·pot -1 ). The amount of potassium applied per pot in each treatment (K 2 SO 4 =1.48g·pot -1 ) is the same. 2 , P 2 and K 2 Based on the conventional application rate. Nitrogen, phosphorus and potassium fertilizers are urea (containing N46.6%), superphosphate (containing P 2 O 5 14.5%) and potassium sulfate (containing K 2 O 54.1%). Each treatment was divided into seven equal portions according to the application amount of urea and potassium sulfate, and applied once every 7 days. Superphosphate was divided into three portions and applied once every 15 days. Hyperspectral images and plant tissues were collected at different growth stages: G0, G1, G2, and G3 corresponded to 0, 20, 40, and 60 days after the first fertilization, respectively (e.g. Figure 1 shown).
[0035] Table 1 Different fertilization treatments of hydrangea 'Hua Temari'
[0036]
[0037] 2. Image Acquisition and Processing
[0038] Hyperspectral images of leaf scale were acquired using a hyperspectral imaging system (GaiaSky-mini2-VN, Dualix Spectral Imaging Technology Co., Ltd, Sichuan, China). The sensor has a spectral range of 400-1000 nm, a spectral resolution of 3 ± 0.5 nm and 176 spectral channels. The hyperspectral imager was fixed on a tripod about 1 meter high and photographed under a fixed light source. The acquisition system was set in a black box to eliminate ambient light. The leaf sample was placed in the center of the black background plate so that the leaf was as flat as possible without obvious curling and folding. The data acquisition software Specview was used to record the hyperspectral images obtained during the measurement, and white plate correction was performed according to the method of Baranowski et al. The calibrated leaf hyperspectral images were imported into the R-4.2.1 software, and the Raster Package was used to identify the leaf edge, and the average spectral curve of all pixel points of the leaf in each image was calculated. The second leaf from top to bottom was selected to acquire hyperspectral images from the G0 to G3 growth period. Each treatment was repeated 10-20 times, and a total of 405 leaf samples were collected in the experiment.
[0039] 3. Determination of biomass, nitrogen and phosphorus content
[0040] All leaf samples were subjected to nitrogen and phosphorus content determination after hyperspectral imaging. Drying was performed at 70°C for at least 48 h to constant dry mass. Leaf samples were ground and leaf nitrogen and phosphorus content was determined by Kjeldahl method and molybdenum antimony colorimetry. Three plants per treatment were randomly selected for aboveground biomass determination at G1, G2 and G3 stages. Similar to leaf samples, aboveground biomass was determined by drying at 105°C for 30 min and then drying at 70°C until constant weight.
[0041] Tables 2 and 3 show the nitrogen and phosphorus content characteristics of the leaves of 'Hua Teju' under different nitrogen application conditions. It can be seen that different nitrogen application treatments had a significant effect on the aboveground biomass (P<0.05). The aboveground biomass of 'Hua Teju' was the highest under the N1 treatment (1.14 g·plant -1 ), while the lowest value was found under N3 treatment (0.18 g·plant -1 ), showing different growth states of nitrogen deficiency (N0), suitable (N1-N2) and excess (N3). From G1 to G3, the nitrogen content (LNC) of leaves gradually decreased over time. This trend was more obvious under lower nitrogen application rates (N0). Similarly, different phosphorus application treatments had a significant effect on the aboveground biomass of 'Hana Temari' (P<0.05), reaching a maximum under the P2 treatment (0.72 g·plant -1 ), and the lowest under P3 (0.19 g·plant -1). Over time, the leaf phosphorus content (LPC) did not change significantly with the addition of phosphorus. The differences in leaf nitrogen and phosphorus content under different fertilization treatments provide a reliable physiological basis for the application of hyperspectral technology to estimate leaf nitrogen and phosphorus content.
[0042] Table 2 Aboveground biomass and leaf nitrogen content of hydrangea 'Hua Temari'
[0043]
[0044] Different lowercase letters indicate significant differences (P<0.05), the same below.
[0045] Table 3 Aboveground biomass and leaf phosphorus content of 'Hua Temari' hydrangea
[0046]
[0047] 4. Spectral Data Preprocessing and Spectral Transformation
[0048] Before model building, the spectral reflectance data were preprocessed by standard normal transformation (SNV) and Savitzky-Golay (SG) filter smoothing to obtain the original spectrum. In order to reduce noise and accurately reveal the spectral characteristics, two transformation methods were used: first-order derivative transformation (FD) and continuous wavelet transformation (CWT). The sensitive bands were identified by correlation analysis between the transformed spectrum and the nutrient content.
[0049] The original reflectance data with wavelengths ranging from 400 to 1000 nm were obtained from the leaf samples of 'Hana Temari', and a total of 405 valid sample data were obtained. Figure 2 -a shows the original spectrum of each leaf used for preprocessing. The original spectral data is complex due to its multicollinearity, nonlinearity of reflectance, baseline variation, additional scattering variation, and superposition of random noise. It is necessary to perform certain transformation processing on the original spectral data to eliminate noise and improve the estimation accuracy. In order to eliminate the influence of baseline drift and background interference, Figure 2 -b shows the spectrum after the first-order differential transformation. It can be seen that some wavelength information with small differences but important is easier to identify from the original spectrum. Overall, the spectral reflectance of the leaves of 'Hana Temari' showed a similar trend of change under different fertilization treatments. The collected leaf hyperspectral data were divided into three hyperspectral data sets: nitrogen level (N0, N1, N2, N3), phosphorus level (P0, P1, P2, P3) and control group (no nitrogen and phosphorus added), which can more intuitively show the changes in the average spectral reflectance of leaves under different nitrogen and phosphorus treatments. The wavelength corresponding to the maximum value of the first-order differential (the maximum slope of the original spectral curve) is also called the "red edge" position. Figure 2It can be seen from the cd that in the visible light and red edge range of 400-760nm, the average spectral reflectance of different nitrogen and phosphorus levels showed obvious differences. The average spectral reflectance of the nitrogen level was higher than that of the phosphorus level, and the reflectance of the control group was the highest. After the first-order differential transformation, it can be found that the average spectral reflectance of the control group showed an obvious "red shift" trend in the wavelength range of 648nm to 797nm.
[0050] Figure 3 The spectral reflectance response characteristics of 'Hana Temari' leaves under different nitrogen and phosphorus fertilization treatments were further analyzed. Overall, the spectral reflectance of 'Hana Temari' leaves under different nitrogen and phosphorus treatments showed basically the same trend. The results showed that the leaves in the N0 treatment showed higher reflectance in the visible light region. Unlike the nitrogen treatment, the reflectance of the P3 treatment in the visible light region was higher than that of the other treatments. In the far-infrared region, the reflectance increased to varying degrees with the increase in phosphorus application. The two spectral curves of the P1 and P2 treatments overlapped almost completely, with no obvious increase or decrease trend. In addition, it is worth noting that in the "red edge" region between visible light and near-infrared wavelengths, at the peak position near 710nm, the first-order differential values of different nitrogen treatments are N0>N3>N2>N1, and the red edge positions corresponding to N0, N1, N2, and N3 are 705, 713, 709, and 709nm, respectively; the first-order differential values of different phosphorus treatments are P3>P2>P1>P0, and the red edge positions corresponding to P0, P1, P2, and P3 are 728, 728, 728, and 713nm, respectively. In addition, in the near-infrared wavelength range, the spectral reflectance of leaves with nitrogen deficiency and phosphorus excess showed an obvious "red shift" phenomenon. In general, the difference in spectral reflectance between all bands for phosphorus treatment is smaller than that for nitrogen treatment.
[0051] Figure 4 The hydrangea leaf spectrum after the original leaf spectrum was processed by continuous wavelet transform (CWT) is shown. At scales [1, 2, 3], the spectral curve is relatively flat and lacks identifiable features. Starting from scales [4, 5, 6], the spectrum shows a wavy shape, turning into a parabola at scales [7, 8], and approximating a straight line at scales [9, 10]. It is worth noting that compared with the control, scales 6 and 7 show more obvious differences in distinguishing nitrogen and phosphorus spectral features.
[0052] 5. Correlation Analysis
[0053] 5.1 Correlation analysis between spectral reflectance and leaf nitrogen and phosphorus contents after first-order differential transformation
[0054] Figure 5It is a correlation diagram of the spectral reflectance of 'Flower Handball' leaves and their nitrogen and phosphorus contents. It can be seen that after the first-order differential processing of the original spectrum, more bands show extremely significant correlations (P<0.01). At the nitrogen level, there are many bands with negative correlation coefficients in the original spectrum. After the first-order differential transformation, the position with the highest correlation coefficient appears near 782nm, and the maximum correlation coefficient is 0.71; in the range of 599 - 660nm, the correlation coefficient is also at a relatively high level (higher than 0.65). At the phosphorus level, the correlation coefficient of the original spectrum is the highest near 652nm, showing a negative correlation, and the maximum correlation coefficient is 0.30; after the first-order differential transformation, the highest value appears near 458nm, showing a negative correlation, and the maximum correlation coefficient is 0.35; in the range of 652 - 674nm, the correlation coefficient values remain at a relatively high level (close to 0.3).
[0055] 5.2 Correlation Analysis of Spectral Reflectance Processed by Continuous Wavelet Transform and Leaf Nitrogen and Phosphorus Contents
[0056] Figure 6 It shows the correlation between the hydrangea leaf spectrum and nutrient contents after continuous wavelet transform (CWT) processing. The correlation coefficient for leaf nitrogen content (LNC) is significantly higher than that for leaf phosphorus content (LPC). The sensitive bands (P<0.05) for LNC are mainly concentrated in scales [5, 6, 7], while the sensitive bands (P<0.05) for LPC are concentrated in scales [6, 7]. At the nitrogen level, the highest coefficient appears at 462nm in scale 6 (scale 6) (r = -0.90). At the phosphorus level, the highest coefficient appears at 809nm in scale 1 (r = -0.87). The research shows that compared with FD, CWT significantly improves the correlation between leaf nitrogen and phosphorus contents and spectral reflectance.
[0057] VI. Construction of Nitrogen and Phosphorus Prediction Models
[0058] The pls, randomForest packages and nnet package in R language are used for model construction. A total of 405 valid data are obtained in the experiment. 75% of the total samples are randomly selected as the training set, and the remaining 25% are used as the test set. To prevent model overfitting, a ten-fold cross-validation method is used to evaluate the prediction ability of the model. The determination coefficient (R 2 ) and relative root mean square error (RRMSE) are selected to evaluate the model performance. When the RRMSE value is smaller, it indicates that the prediction ability of the model is stronger. When RRMSE < 10%, the model performance is excellent; when 10% < RRMSE < 20%, the model performance is good; if 20% < RRMSE < 30%, the model performance is average; if RRMSE ≥ 30%, the model prediction ability is poor.
[0059] Figure 7The results showed that the accuracy of three machine learning models, namely partial least squares regression (PLSR), random forest (RF) and one-dimensional convolutional neural network (CNN), was evaluated. The full-band spectral data after first-order differential processing was input into the model to predict the nitrogen and phosphorus content of leaves. The coefficient of determination (R 2 ) and relative root mean square error (RRMSE) were used to evaluate the performance of the model. At the same time, the fitting curve of the true value and the predicted value was compared with the 1:1 line to indicate the consistency and reliability of the estimation model for leaf nitrogen content (LNC) and leaf phosphorus content (LPC) of 'Hana Temari'. The goodness of fit was evaluated by R 2 Quantification, RRMSE is used to select the model with the best prediction performance. The lower the RRMSE value, the higher the model accuracy. The maximum R 2 The values reached 0.928 and 0.760, respectively. The RRMSE values of the three models for LNC and LPC were between 18.3% and 19.7% and 25.4% and 64.1%, respectively. For the prediction of leaf nitrogen content, the determination coefficient was PLSR>RF>CNN, among which the largest determination coefficient was 0.928 for the PLSR model. For the prediction of leaf phosphorus content, the prediction accuracy of the three models was lower than that of leaf nitrogen content. The RRMSE values of the PLSR, RF and CNN models were 25.4%, 64.1% and 26.8%, respectively. Overall, the prediction performance of random forest (RF) was the worst, while PLSR showed the best prediction performance.
[0060] In order to evaluate whether wavelet transform can further improve the prediction accuracy of the model, a PLSR model combined with continuous wavelet transform (CWT) was developed. The correlation coefficient (r) between spectral reflectance and leaf nitrogen and phosphorus content at each scale was first calculated, and then the correlation coefficient values were sorted in descending order, and only the first 5% were retained as spectral feature bands (88). The selected spectral bands were significantly correlated with LNC and LPC (P<0.05). The maximum threshold of the absolute value of the top 5% correlation coefficient of LNC was 0.823, while that of LPC was 0.684. Figure 8 It shows that the PLSR model based on continuous wavelet transform significantly improves the prediction accuracy of nitrogen and phosphorus content in hydrangea leaves. The prediction accuracy of LNC is R 2 is 0.947, RRMSE is 9.6%, and LPC’s R 2 is 0.827 and the RRMSE is 10.6%.
[0061] Table 4 shows the model prediction results of different band screening methods. In order to optimize the band screening method, the influence of the scale selection of spectral characteristic bands in continuous wavelet transform on the model prediction accuracy was evaluated, and the model prediction performance of the single scale (scale 6 and scale 7) and multi-scale (scale 1-10) band screening methods with high correlation with leaf nitrogen and phosphorus content was compared. Only the top 5% of the selected scale with the highest correlation was retained as the spectral characteristic band. It can be found that compared with the single scale, the multi-scale band screening method significantly improved the model performance.
[0062] Table 4 Model prediction results of different band screening methods
[0063]
[0064] Fig. 9 This is a ranking diagram of the importance of spectral features in the PLSR model based on multi-scale CWT transformation (Top 10). It can be found that for leaf nitrogen content, the feature importance of 550nm at scale 5 is the highest (0.06), with a contribution rate of 6%; for leaf phosphorus content, the feature importance of 1030nm at scale 3 is the highest (0.08), with a contribution rate of 8%.
[0065] The present invention establishes a prediction model for nitrogen and phosphorus in hydrangea based on hyperspectral technology. After the original spectrum is processed by first-order derivative (FD) transformation and continuous wavelet transform (CWT), the effective information of the spectrum is further mined. By comparing the correlation and model performance, it is proved that CWT is a more effective spectral preprocessing method than FD. The PLSR model combined with CWT with multi-scale band selection significantly improves the accuracy of prediction of nitrogen and phosphorus content in hydrangea leaves, while significantly reducing the number of input variables (LNC, R 2 =0.947, RRMSE=9.6%; LPC, R 2 =0.827, RRMSE=10.6%), which can provide key technical support for the real-time monitoring of nitrogen and phosphorus nutrition status of hydrangea and the efficient application of fertilizers in nursery production.
Claims
1. A method for predicting nitrogen and phosphorus content in hydrangea leaves based on hyperspectral technology, characterized in that: The method for predicting nitrogen and phosphorus content in hydrangea leaves comprises the following steps: (1) Different nitrogen or phosphorus treatments were applied to large-leaf hydrangea to obtain leaf-scale hyperspectral images and corresponding leaf nitrogen or phosphorus contents at different growth stages; (2) Calculate the average spectral curve of all pixel points of the leaves in each hyperspectral image, obtain the spectral reflectance data, and then obtain the original spectral reflectance data through standard normal transformation and filter smoothing; (3) After the original spectrum is processed by continuous wavelet transform, the correlation coefficient between the spectral reflectance at each scale and the corresponding leaf nitrogen content or phosphorus content is calculated, the correlation coefficient values are sorted in descending order, and the top 5% are retained as the spectral characteristic bands; then partial least squares regression modeling is used to obtain the CWT-based PLSR prediction model; (4) Take leaves of the hydrangea plants to be tested for nitrogen and phosphorus content, obtain spectral reflectance data of the spectral characteristic band described in step (3), and use the CWT-based PLSR prediction model described in step (3) to estimate the corresponding nitrogen content or phosphorus content.
2. The method for predicting nitrogen and phosphorus contents in hydrangea leaves according to claim 1, characterized in that: In the step (1), the large-leaf hydrangea is planted as a potted cutting seedling, and different nitrogen treatments or phosphorus treatments are performed after the plant grows to 4-5 pairs of leaves; the different nitrogen treatments use the following four nitrogen application rates: N0: N = 0 g·pot -1 , N1: N = 0.4 g·pot -1 , N2: N = 0.8 g·pot -1 , N3: N = 1.2 g·pot -1 The nitrogen application rate of each treatment was divided into 7 parts and one part was applied every 7 days; the phosphorus treatment used the following 4 treatment methods: P0: P2O5 = 0g·pot -1 , P1:P2O5=0.15g·pot -1 , P2:P2O5=0.3g·pot -1 , P3:P2O5=0.45g·pot -1 The phosphorus application rate for each treatment was divided into three parts and one part was applied every 15 days; and potassium fertilizer was applied routinely to each treatment.
3. The method for predicting nitrogen and phosphorus contents in hydrangea leaves according to claim 2, characterized in that: In the step (1), the different growth stages are selected from 0, 20, 40 and 60 days after the first fertilization of each treatment.
4. The method for predicting nitrogen and phosphorus contents in hydrangea leaves according to claim 3, characterized in that: When constructing the CWT-based PLSR prediction model in step (3), 75% of the total samples are randomly selected as the training set, and the remaining 25% are used as the test set.
Citation Information
Cited By
Bagged fruit and vegetable sugar degree detection method based on spectrum correction
CN120275342A
