Remote sensing high-precision inversion method for chlorophyll content of vegetation leaves
By combining wavelet transform and machine learning, the universality and robustness issues of remote sensing inversion of leaf chlorophyll content were solved, achieving high-precision and non-destructive monitoring of leaf chlorophyll content, applicable to different vegetation types and environments.
Patent Information
- Application Number
- CN202511400000.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-28
- Publication Date
- 2026-02-13
Smart Images
Figure CN121528342A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a high-precision remote sensing inversion method for vegetation leaf chlorophyll content, which is mainly suitable for correcting and constructing a general inversion model for leaf chlorophyll content of different vegetation types and evaluating and verifying, and can be widely applied to quantitative inversion of vegetation biochemical parameters in the fields of agricultural remote sensing monitoring and forest ecosystem research. BACKGROUND
[0002] Leaf chlorophyll content (C ab or LCC) Leaf chlorophyll is a key factor for plants to carry out photosynthesis, and its main function is to absorb light energy and convert it into chemical energy, and then synthesize organic matter. Leaf chlorophyll content directly affects the photosynthetic efficiency of plants, and then determines the growth rate of vegetation, and is closely related to primary productivity, therefore, leaf chlorophyll content is often used as an important indicator to evaluate the health status of vegetation. In addition, the change of leaf chlorophyll content can also reflect the adaptability of vegetation to environmental conditions (such as light, temperature and nutrient supply, etc.). As can be seen from the leaf reflectance spectrum, in the visible light spectrum range of 380nm-740nm, the main factor affecting the leaf reflectance spectrum is the various pigments contained in the leaf, especially chlorophyll, and due to the strong absorption of chlorophyll to blue light band and red light band and the insensitivity to green light band, normal leaves appear green. Therefore, according to such characteristics, there is a possibility to invert the leaf chlorophyll from remote sensing data.
[0003] Traditional measurement methods mostly rely on ground sampling and laboratory analysis, which has high accuracy, but is time-consuming and laborious, and is difficult to popularize on a large scale, and is destructive to leaves. There are disadvantages such as time-consuming, high cost and personnel error, which limit their application in large-scale and long-term plant monitoring. In contrast, the emergence of remote sensing detection technology provides a fast, continuous and non-destructive monitoring method, which can respond to the changes of leaf biochemical parameters by analyzing the spectral information of plant leaves, and provides a possibility for fast, continuous and non-destructive monitoring of plant biochemical parameter conditions.
[0004] Currently, methods for estimating LCC using remote sensing technology can be divided into three categories: (i) Empirical model estimation based on spectral indices: Based on mathematical statistical analysis methods, this method analyzes the relationship between vegetation indices and vegetation biochemical parameters, and uses this relationship to predict vegetation biochemical parameters. However, it lacks universality due to limitations in the data used during model establishment, the time and location of data collection. (ii) Physical model estimation based on radiative transfer mechanisms: PROSPECT is the most widely used leaf-scale radiative transfer model. By inputting leaf biochemical and structural parameters, it simulates reflectance and transmittance in the 400-2500nm spectral range. When using an inverse model for inversion, both reflectance and transmittance are used as inputs to obtain high-precision inversion results. However, in reality, transmittance data is difficult to obtain, and only reflectance can be used as input, which can lead to ill-conditioned inversion. (iii) Machine learning algorithms can handle the nonlinear relationship between leaf biochemical parameters and reflectance spectra, simplifying complex inversion processes and improving the accuracy of leaf biochemical parameter inversion. However, the accuracy of the model is related to the quality of the training dataset, and high-quality training data is key to improving inversion accuracy.
[0005] In summary, although many studies have made good progress in the quantitative inversion of leaf biochemical parameters, these studies are usually based on specific study areas or specific observation conditions to build local optimal models. These models have not undergone comprehensive applicability and robustness evaluation and do not meet the requirements for general application. Summary of the Invention
[0006] The technical problem solved by this application is to overcome the above-mentioned deficiencies in the prior art and to provide a high-precision remote sensing inversion method for chlorophyll content of vegetation leaves based on wavelet transform.
[0007] The principle of this invention is as follows:
[0008] The type of vegetation and differences in growth environment affect the cellular structure inside leaves. Different vegetation types share similar basic cellular structures, and the combined absorption of solar radiation by the content and structural parameters of these basic components results in corresponding characteristics in the leaf reflectance spectrum. Therefore, it can be considered that leaf reflectance is related to the absorption coefficients of the leaf components, and the absorption coefficient of each leaf biochemical parameter is only related to the wavelength and not to the plant species. From the leaf reflectance spectrum curves, it can be seen that in the visible light spectrum range of 380nm-740nm, the various pigments contained within the leaf, especially chlorophyll, play a dominant role in the leaf reflectance spectrum. After 800nm, chlorophyll no longer has a significant absorption effect on solar radiation; the dominant factors thereafter are mainly the leaf's structural parameters and water content. Wavelet analysis, as a powerful signal processing tool, has been widely used in multi-resolution analysis and feature extraction of complex spectral data. Wavelet analysis has shown great potential in the field of weak information extraction in hyperspectral remote sensing. The core idea of wavelet transform is to decompose complex signals into a series of simple sub-signal components of different scales and frequencies, thereby achieving multi-scale signal extraction. When applied to vegetation spectral data analysis, wavelet transform can decompose complex spectral signals into sub-signal components of different frequencies, effectively capturing the overall structural features of spectral information and extracting weak information hidden in the spectral signal. This provides the possibility of finding the optimal combination of sub-signal components to accurately estimate the chlorophyll content of vegetation leaves.
[0009] Wavelet transform effectively enhances the sensitivity of spectral data to chlorophyll content. It separates spectral information based on signal frequency variations. The characteristic absorption of vegetation chlorophyll is often embedded in spectral variations at specific frequencies, making it easily decomposed into different scales by wavelet transform, thus enabling separation and extraction. In practical applications, changes in background reflectance often significantly interfere with chlorophyll content retrieval results. Wavelet transform, by extracting spectral information at specific scales, can effectively reduce background noise interference to a certain extent, thereby improving the accuracy and universality of the retrieval model. Therefore, the wavelet transform method can be flexibly applied to chlorophyll content retrieval studies of different types of vegetation and water bodies.
[0010] This application, through research and analysis, extracts chlorophyll information features through continuous wavelet transformation, and then couples the chlorophyll information features extracted by wavelet transformation (referred to as wavelet coefficient features) with a machine learning model for inversion. The radiative transfer model is used to verify the general inversion effect, thereby achieving high-precision general inversion of chlorophyll in plant leaves.
[0011] The technical solution adopted in this application to solve the above-mentioned technical problems is: a high-precision remote sensing inversion method for chlorophyll content in vegetation leaves, which mainly includes the following steps:
[0012] S1: Constructing the leaf sample dataset
[0013] S11: Construct a real-world dataset of blade measurements;
[0014] S12: Generate a simulation dataset using the PROSPECT radiative transfer model;
[0015] S13: Using a stratified sampling strategy, the wavelet coefficient features of the measured dataset are divided into a measured training set and a measured validation set in a 6:4 ratio. The measured training set is used for universal model parameter calibration, and the measured validation set is used to evaluate the model's predictive performance in the same source data.
[0016] S2: Extracting chlorophyll information features using continuous wavelet transform
[0017] S21: Multi-scale analysis of chlorophyll information (LCC) is performed using the continuous wavelet transform (CWT) method. The expression for the continuous wavelet transform is as follows:
[0018]
[0019] In the formula, f(t) represents the input original spectral signal (reflectance spectrum curve in the range of 400nm-2500nm), and W(a,b) represents the calculated wavelet coefficients. 'a' is the scale factor, which controls the scaling of the wavelet function. A larger value of 'a' results in a wider wavelet function in the time domain and a narrower value in the frequency domain, reflecting the low-frequency, general characteristics of the spectral signal; a smaller value of 'a' expresses the opposite. 'b' is the translation factor, which controls the translation of the wavelet function on the x-axis, used to locate the characteristics of the signal at different time points. 'ψ(t)' is the wavelet basis function, also called the 'mother wavelet,' which is a short-time signal with oscillatory characteristics and a zero mean.
[0020] S22: Perform a 128-scale continuous wavelet transform analysis on the original reflectance spectrum of each leaf sample using Morlet basis functions. The expression is as follows:
[0021]
[0022] In the formula, t is the wavelength of the reflection spectrum, and w o is the center frequency of the wave, used to control the oscillation frequency of the Morlet wavelet, and i is the imaginary unit.
[0023] S23: Perform correlation analysis on the wavelet coefficient features and LCC values, using the Pearson correlation coefficient to quantify the linear correlation between the wavelet coefficient features and LCC values. The calculation formula is as follows:
[0024]
[0025] In the formula, x i It is the i-th value of the wavelet coefficients; y i It is the i-th value of LCC; It is the average value of the wavelet coefficients; R is the average value of LCC; the value of R ranges from [-1, 1], where r = 1 indicates perfect positive correlation; r = -1 indicates perfect negative correlation; and r = 0 indicates no linear correlation.
[0026] S24: Set a threshold to filter out wavelet coefficient features that are sensitive to LCC spectral information;
[0027] S3: Constructing a chlorophyll inversion model based on wavelet coefficient feature coupling machine learning
[0028] S31: Construct a chlorophyll model based on spectral indices, that is, calculate the chlorophyll spectral index using the original reflectance. The formula for the chlorophyll spectral index is:
[0029] Chlorophyll spectral index CIred = R 750 / R 720 -1
[0030] In the formula, R 720 and R 750 These are the leaf reflectances at 720nm and 750nm;
[0031] S32: Constructing an LCC inversion model by coupling machine learning methods with the most relevant wavelet coefficient features.
[0032] The chlorophyll spectral index is calculated using the most relevant raw reflectance. The result is then normalized and used as a spectral index feature. This feature, along with the wavelet coefficient features obtained from continuous wavelet transform, is then used to construct the machine learning LCC inversion model.
[0033] This application can also include an S4 step.
[0034] S4: Add noise to enhance model robustness
[0035] S41: Perform correlation analysis on wavelet coefficient features of the PROSPECT simulation dataset with added relative and absolute noise according to step S2, and evaluate the stability of wavelet coefficient features;
[0036] Step S32 described in this application may include:
[0037] S321: The Support Vector Machine (SVM) method is used to optimize the parameters of the LCC inversion model. The Radial Basis Function (RBF) is used as the kernel function to handle the nonlinear characteristics of the LCC inversion model data. At the same time, the regularization parameter and tolerance parameter values (C=1.0, epsilon=0.01) are adjusted to balance the complexity and training error of the LCC inversion model and optimize the generalization ability of the model.
[0038] The S32 step described in this application also includes the following steps:
[0039] S322: Construct a universal LCC inversion model using wavelet coefficient feature coupling machine learning method, use the obtained measured training set to participate in model training, and use the measured validation set to participate in model validation;
[0040] S323: Construct a general LCC inversion model using wavelet coefficient features coupled with machine learning methods, use the obtained simulated dataset wavelet coefficient features for model training, and use the measured dataset wavelet coefficient features for model validation.
[0041] This application can also include an S5 step.
[0042] S5: Comparison and Verification of Inversion Results
[0043] S51: Comparison of inversion results between the leaf chlorophyll inversion model constructed based on the original reflectance and the leaf chlorophyll inversion model constructed by the wavelet coefficient feature coupling machine learning method;
[0044] S52: Comparison of inversion results of LCC inversion model constructed by wavelet coefficient feature coupling machine learning method before and after adding spectral index as a feature to wavelet coefficient;
[0045] S53: Comparison of inversion results of LCC inversion model constructed by wavelet coefficient feature coupling machine learning method before and after adding noise.
[0046] The PROSPECT radiative transfer model adopts the PROSPECT-D radiative transfer model, and the simulation dataset includes the chlorophyll content, dry matter content, equivalent water thickness, carotenoid content, brown pigment content, structural parameters, and anthocyanin content of plant leaves.
[0047] This application uses normalized root mean square error (nRMSE) and coefficient of determination (R²). 2 These two evaluation metrics are used to verify the accuracy of the model, and the formulas are as follows:
[0048]
[0049] In the formula, n is the total sample size, and y i These are the measured values of the sample. These are the model's predicted values. Let y be the sample mean. max y min These are the maximum and minimum values of the sample measured values, respectively.
[0050] Compared with existing technologies, this application has the following advantages and effects: 1. Continuous wavelet transform can capture the transformation of chlorophyll information better; 2. By enhancing the expression of spectral information through wavelet coefficient features, the inversion accuracy surpasses that of traditional vegetation indices; 3. The inversion accuracy of the leaf chlorophyll inversion model constructed by coupling wavelet coefficient features with machine learning SVM method is significantly better than the inversion results of the leaf chlorophyll inversion model constructed based on the original reflectance; 4. Continuous wavelet transform has strong robustness to noise, and the wavelet coefficient features in the high-scale range of 750-800nm are less affected by noise, and the model still maintains high accuracy after being transferred between different datasets. Attached Figure Description
[0051] Figure 1 This is a flowchart illustrating the high-precision inversion method for chlorophyll content in vegetation leaves used in this application.
[0052] Figure 2 This is a scale plot showing the correlation between the continuous wavelet coefficients of the Morlet basis functions and the reflectance spectrum of the leaf. (On a composite measured dataset...) Figure 2 A), as shown in the figure, the region with the most significant correlation has a wavelength range of approximately 700 nm and 750 nm. (PROSPECT simulation dataset) Figure 2 B) The wavelet coefficients related to leaf LCC are mainly continuously distributed in the 700-800 nm band. Similar to the composite measured dataset, its morphology is relatively discrete, with several clustered "island"-shaped highly correlated patches appearing in the core region of each "fan" structure. From the above regions, the highly correlated areas are all located in the red edge and near-infrared spectral regions, corresponding to a larger analytical scale. This phenomenon is consistent with plant physiological theory, as the spectral characteristics of the red edge and near-infrared regions are known to be closely related to biophysical parameters such as vegetation chlorophyll content and cell structure. The top 1% of the results with the highest co-correlation between the composite measured dataset and the simulated dataset were extracted. Figure 2 As can be seen from C, the bright areas of the heatmap are significantly reduced and more concentrated, and highly converged in the red-edge to near-infrared band (700-850nm) and the mid-to-high scale range (80-100), presenting multiple more compact "island"-shaped highly correlated patches. Therefore, the values of these common highly correlated regions are also the optimal feature regions selected in the end.
[0053] Figure 3This diagram illustrates the principle of LCC inversion using the Support Vector Machine (SVM) method. During parameter tuning, the Radial Basis Function (RBF) is used as the kernel function to handle the nonlinear characteristics of the LCC inversion model data. At the same time, the regularization parameter and tolerance parameter values (C = 1.0, epsilon = 0.01) are adjusted to balance the complexity and training error of the LCC inversion model and optimize the model's generalization ability.
[0054] Figure 4 The inversion results are shown for two CIred index models that construct LCC spectral indices based on the original reflectance. The results show that, regardless of whether it is the general or universal model, the CIred index model exhibits high clustering characteristics and low error, with stable accuracy across different validation subsets.
[0055] Figure 5 The results show the inversion of two LCC machine learning (SVM) models built based on the original reflectance. The results indicate that the general model built based on the original reflectance also performs well, but the universal model is superior in both accuracy and nRMSE.
[0056] Figure 6 This shows the inversion results of two LCC machine learning models constructed using wavelet coefficients. Figure 5 The comparison shows that in the universality validation, the two feature construction methods perform similarly, with wavelet coefficients slightly outperforming in terms of error metrics, but the original reflectance method is slightly better at explaining variance. In the generality validation, the model constructed using wavelet coefficients significantly outperforms the original reflectance model, especially in… The better performance indicates that wavelet coefficient features are more robust in cross-dataset (simulation → measurement) generalization.
[0057] Figure 7 The validation results of the LCC inversion model obtained by incorporating spectral indices as features into wavelet coefficients and coupling them with two machine learning models are presented. The results show that, compared to... Figure 6 The machine learning model constructed by adding spectral indices as features to wavelet coefficients showed a comprehensive performance improvement in universality validation after incorporating the CIred index. higher The errors were smaller (nRMSE = 9.15%; MAE = 7.58%). This indicates that the CIred index provides additional effective information, enhancing the model's fitting ability. In the generalization validation, the addition of the CIred index significantly improved the model's generalization ability. The improvement was approximately 6 percentage points, while the error metrics (nRMSE and MAE) both decreased by approximately 2 percentage points. This indicates that the CIred index helps the model better adapt to the distribution differences between simulated and measured data. This further demonstrates that the introduction of the spectral index provides more interpretable features and enhances the model's utilization of chlorophyll-sensitive bands.
[0058] Figure 8 This is a scaled correlation plot of wavelet coefficients and the simulated dataset after adding noise. Low-scale wavelet coefficient features are highly sensitive to noise variations; noise masks weakly absorbing signals, and correlation features disappear due to noise. However, in the comprehensive plot (… Figure 8 As shown in C), the wavelet coefficient features at high scales remain stable, and adding noise does not affect the expression of LCC spectral information of this part of the wavelet coefficients, indicating that this part of the features has strong robustness.
[0059] Figure 9 A comparison of the inversion accuracy of the Cired spectral index model (based on raw reflectance) and the machine learning SVM model on the LCC general model after adding noise is presented. The results show that the Cired spectral index model maintains high accuracy even after adding noise. nRMSE = 16.42%; MAE = 15.58%, regardless of The error metrics nRMSE and MAE both outperform the machine learning model SVM, while the SVM model's generalization ability on real-world data declines. nRMSE=17.01%; MAE=15.66%).
[0060] Figure 10 This is the inversion result of a generalized LCC machine learning model constructed based on the original reflectance and wavelet coefficients, respectively, using spectral indices as features. From the results, we can see that the wavelet coefficients + spectral indices method (i.e....) Figure 10 B) is optimal in all metrics, compared to the inversion results of the SVM model constructed using pure wavelet coefficients. Figure 9 B), the performance is significantly improved after adding the spectral index, and compared with the simple spectral index method ( Figure 9 A) The combination of wavelet coefficients and spectral indices is also more advantageous. nRMSE=13.31%, MAE=11.96%). Detailed Implementation
[0061] Key terms explained in this application:
[0062] Continuous wavelet transform: Perform continuous wavelet transform analysis on the original reflectance spectrum of each leaf sample across 128 scales.
[0063] Wavelet coefficient characteristics: The wavelet coefficients obtained after extracting the continuous wavelet transform using Pearson coefficients exhibit a high correlation with LCC.
[0064] LCC universality model: The model is built using a training set of measured data and validated using a validation set of measured data.
[0065] LCC general model: The model is trained using a simulated dataset and validated using a real-world dataset.
[0066] A high-precision remote sensing inversion method for chlorophyll content in vegetation leaves mainly includes the following steps:
[0067] S1: Constructing the leaf sample dataset
[0068] S11: Construct a real-world dataset of blade measurements;
[0069] S12: Generate a simulation dataset using the PROSPECT radiative transfer model;
[0070] S13: Using a stratified sampling strategy, the wavelet coefficient features of the measured dataset are divided into a measured training set and a measured validation set in a 6:4 ratio. The measured training set is used for universal model parameter calibration, and the measured validation set is used to evaluate the model's predictive performance in the same source data.
[0071] S2: Extracting chlorophyll information features using continuous wavelet transform
[0072] S21: Multi-scale analysis of chlorophyll information (LCC) is performed using the continuous wavelet transform (CWT) method. The expression for the continuous wavelet transform is as follows:
[0073]
[0074] In the formula, f(t) represents the input original spectral signal (reflectance spectrum curve in the range of 400nm-2500nm), and W(a,b) represents the calculated wavelet coefficients. 'a' is the scale factor, which controls the scaling of the wavelet function. A larger value of 'a' results in a wider wavelet function in the time domain and a narrower value in the frequency domain, reflecting the low-frequency, general characteristics of the spectral signal; a smaller value of 'a' expresses the opposite. 'b' is the translation factor, which controls the translation of the wavelet function along the x-axis, used to locate the signal's characteristics at different time points. 'ψ(t)' is the wavelet basis function, also called the "mother wavelet," which is a short-time signal with oscillatory characteristics and a zero mean.
[0075] S22: Perform a 128-scale continuous wavelet transform analysis on the original reflectance spectrum of each leaf sample using Morlet basis functions. The expression is as follows:
[0076]
[0077] In the formula, t is the wavelength of the reflection spectrum, and w o is the center frequency of the wave, used to control the oscillation frequency of the Morlet wavelet; i is the imaginary unit.
[0078] S23: Perform correlation analysis on the wavelet coefficient characteristics and LCC values to quantify the degree of linear correlation;
[0079] Since there are differences between the measured data and the simulated data, correlation analysis was performed on the measured dataset and the simulated dataset separately.
[0080] S24: Set a threshold to filter out wavelet coefficient features that are sensitive to LCC spectral information. Perform a weighted average filter on the wavelet coefficient features of the measured dataset and the simulated dataset. Extract the band scales that show high correlation in both the measured dataset and the simulated dataset, and mark them in red in the correlation scale map. The remaining features are marked with gradient colors.
[0081] S3: Constructing a chlorophyll inversion model based on wavelet coefficient coupled machine learning method
[0082] S31: Construct a chlorophyll model based on spectral indices, that is, calculate the chlorophyll spectral index using the original reflectance. The specific formula for the chlorophyll spectral index is as follows:
[0083] CIred = R 750 / R 720 -1
[0084] In the formula, R 720 and R 750 These are the leaf reflectances at 720nm and 750nm, respectively.
[0085] S32: Construct an LCC inversion model using wavelet coefficients coupled with machine learning methods that have the highest relevance;
[0086] S321: The Support Vector Machine (SVM) method is used to optimize the parameters of the LCC inversion model. The Radial Basis Function (RBF) is used as the kernel function to handle the nonlinear characteristics of the LCC inversion model data. At the same time, the regularization parameter and tolerance parameter values (C=1.0, epsilon=0.01) are adjusted to balance the complexity and training error of the LCC inversion model and optimize the generalization ability of the model.
[0087] S322: Construct a universal LCC inversion model using wavelet coefficient coupled machine learning method, use the obtained measured training set for model training, and use the measured validation set for validation;
[0088] S323: Construct a general LCC inversion model using wavelet coefficient coupled machine learning method, use the wavelet coefficient features of the obtained simulated dataset to participate in model training, and use the wavelet coefficient features of the measured data to participate in model verification;
[0089] S4: Add noise to enhance model robustness
[0090] S41: Add relative and absolute noise to the PROSPECT simulation dataset (this step can also be done in step S12), analyze the correlation of wavelet coefficient features according to the wavelet coefficient correlation process (referring to step S2), and evaluate the stability of wavelet coefficients.
[0091] S5: Comparison and Verification of Inversion Results
[0092] S51: Comparison of inversion results between the machine learning LCC inversion model constructed based on the original reflectivity and the LCC inversion model constructed using the wavelet coefficient coupled machine learning method;
[0093] S52: Comparison of inversion results of LCC inversion model constructed by wavelet coefficient coupled with machine learning method before and after adding spectral index as a feature to wavelet coefficient;
[0094] S53: Comparison of inversion results of LCC inversion model constructed by wavelet coefficient coupled machine learning method before and after adding noise.
[0095] This application proposes a high-precision remote sensing inversion method for chlorophyll content in plant leaves. It employs a multi-source data collaborative construction strategy, integrating a large-scale leaf spectral dataset from field measurements (covering different species, growth stages, and environmental conditions) and generating a simulated dataset covering various possible scenarios using the PROSPECT radiative transfer model. To enhance the model's robustness and generalization ability, a composite noise model, including systematic absolute and relative noise, is introduced into the simulated data, making the data distribution closer to real measurement conditions.
[0096] The leaf measurement dataset in step S11 of this application uses the reflectance spectrum of fresh leaves and the chlorophyll content of fresh leaves, and the specific details are shown in Table 1.
[0097] Table 1. Description of the measured sample dataset
[0098]
[0099] Table 2. PROSPECT-D Model Input Parameters and Value Range Settings Instructions
[0100]
[0101] The PROSPECT simulation dataset described in this application was generated using the PROSPECT-D radiative transfer model, which has seven input variables: leaf structure parameters (N, unitless), leaf chlorophyll content (μg / cm³), and leaf chlorophyll content (μg / cm³). 2 ), leaf carotenoid content (C ar μg / cm 2 ), brown pigment content (C bp (unitless), dry matter content (C) m g / cm 2 ), equivalent water thickness (C) w (cm) and anthocyanin content (Ant, μg / cm) 2 In order to make the simulation dataset representative, this application consulted publicly available literature to statistically analyze the extreme range of biochemical parameters of normal leaves, and based on this, constrained and set the input parameters (see Table 2). The generated simulation dataset sample includes leaf reflectance spectra and corresponding chlorophyll content.
[0102] This application uses normalized root mean square error (nRMSE) and coefficient of determination. To evaluate the model's performance on the test set and to verify the model's accuracy.
[0103] The following analysis of the inversion results further describes this application:
[0104] The wavelet coefficients were correlated with the LCC values of both the measured dataset and the simulated dataset before adding noise. Figure 2 As can be seen, the wavelet coefficients related to LCC content are mainly distributed in the 700-800 nm band. The highly correlated regions are all located in the red-edge and near-infrared spectral regions, corresponding to larger analytical scales. This phenomenon is consistent with plant physiological theories, as the spectral characteristics of the red-edge and near-infrared regions are known to be closely related to biophysical parameters such as chlorophyll content and cell structure. However, after adding noise to the simulated dataset (0.005 relative noise and 0.003 absolute noise), the wavelet coefficient features at low scales are more sensitive to noise changes; noise masks weak absorption signals, and the correlated features disappear due to noise. But the wavelet coefficient features at mid- and high scales remain stable; adding noise does not affect the expression of LCC spectral information in these wavelet coefficients, indicating that these features have strong robustness.
[0105] In LCC inversion, spectral indices have a certain estimation capability, from Figure 4Comparative results show that, regardless of whether it's the general or universal model, the CIred index model exhibits high clustering characteristics and low error, with stable accuracy across different validation subsets. This is because the red-edge index CIred is highly sensitive to chlorophyll content and can effectively capture information about changes in chlorophyll content. Similarly, the SVM model also performs well in terms of raw reflectance. Figure 5 ).
[0106] Therefore, the spectral index model results are normalized and added to wavelet coefficients, then coupled again with the machine learning model to compare the inversion results of the LCC inversion model. Before adding the spectral index as a feature ( Figure 6 SVM machine learning models are comparable in their ability to explain variance in terms of generality and universality. (Basically unchanged), but the error increased significantly (nRMSE:
[0107] (9.15%→16.27%; MAE: 7.58%→16.29%), indicating a significant difference in dataset distribution. Despite using wavelet coefficients to extract features, performance degradation still occurred in cross-dataset validation, demonstrating that wavelet coefficients alone are insufficient to completely overcome the distributional differences between simulated and measured data. After adding spectral indices ( Figure 7 In the general model, it can be seen that the model performance is slightly improved after adding the CIred index, indicating that feature fusion is effective. In the general model, the accuracy is significantly improved after adding the CIred index, indicating that the spectral index has strong physical meaning and generalization ability, and can effectively make up for the difference in spectral response between simulation and measured data.
[0108] To enhance the robustness and generalization ability of the LCC inversion model, noise was added to the simulated dataset, and then the Cired model based on the original reflectance and the general SVM model based on machine learning were reconstructed. Figure 9 The results show that the CIred spectral index model has clear physical meaning and noise resistance, especially in chlorophyll estimation. Even after adding noise, the spectral index model still maintains high accuracy, indicating its insensitivity to noise and strong generalization ability. In contrast, the SVM model may have overfitted the training data (simulated dataset), leading to a decrease in generalization ability on the actual test data.
[0109] Therefore, the final comparison focuses on the performance of a general LCC model based on wavelet coefficient features with added spectral indices and coupled with machine learning SVM methods. Figure 10 It is evident that extracting the time-frequency features of spectral indices through wavelet transform can better capture the spectral response features of chlorophyll. Under the condition that noise is added to the training set, the wavelet coefficient method has a natural noise reduction capability, can extract more robust spectral features, and reduce noise interference.
[0110] The above results demonstrate that combining continuous wavelet transformation with machine learning provides a new approach for general LCC inversion and has great potential.
Claims
1. A high-precision remote sensing inversion method for chlorophyll content in vegetation leaves, comprising the following steps: S1: Constructing the leaf sample dataset S11: Construct a real-world dataset of blade measurements; S12: Generate a simulation dataset using the PROSPECT radiative transfer model; S13: A stratified sampling strategy is adopted to divide the wavelet coefficient features of the measured dataset into a measured training set and a measured validation set according to the proportion. The measured training set is used for universal model parameter calibration, and the measured validation set is used to evaluate the model's predictive performance in the same source data. S2: Extracting chlorophyll information features using continuous wavelet transform S21: Multi-scale analysis of leaf chlorophyll information is performed using the continuous wavelet transform method. The expression for the continuous wavelet transform is: In the formula, f(t) represents the input original spectral signal, W(a,b) represents the calculated wavelet coefficients; a is the scaling factor, which controls the scaling of the wavelet function; b is the translation factor, which controls the translation of the wavelet function on the x-axis; ψ(t) is the wavelet basis function. S22: Perform a 128-scale continuous wavelet transform analysis on the original reflectance spectrum of each leaf sample using Morlet basis functions. The expression is as follows: In the formula, t is the wavelength of the reflection spectrum, and w o is the center frequency of the wave, used to control the oscillation frequency of the Morlet wavelet; i is the imaginary unit. S23: Perform correlation analysis between the wavelet coefficient characteristics and the LCC value; S24: Set a threshold to filter out wavelet coefficient features that are sensitive to LCC spectral information; S3: Constructing a leaf chlorophyll inversion model based on wavelet coefficient feature coupling machine learning S31: Construct a chlorophyll model based on spectral indices, that is, calculate the chlorophyll spectral index using the original reflectance. The formula for the chlorophyll spectral index is: Chlorophyll spectral index CIred=R 750 / R 720 -1 In the formula, R 720 and R 750 These are the leaf reflectances at 720nm and 750nm; S32: Construct a leaf chlorophyll inversion model by coupling the most relevant wavelet coefficient features with machine learning methods, calculate the chlorophyll spectral index using the most relevant raw reflectance, normalize the obtained result and use it as the spectral index feature, and together with the wavelet coefficient features obtained by continuous wavelet transform, participate in the construction of the machine learning LCC inversion model.
2. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 1, characterized in that: Also set There is an S4 step: S4: Add noise to enhance model robustness S41: Extract wavelet coefficient features from the simulated dataset with added relative and absolute noise according to step S2 and perform correlation analysis to evaluate the stability of the wavelet coefficient features.
3. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 1, characterized in that: It also includes an S5 step: S5: Comparison and Verification of Inversion Results S51: Comparison of inversion results between the leaf chlorophyll inversion model constructed based on the original reflectance and the leaf chlorophyll inversion model constructed by the wavelet coefficient feature coupling machine learning method; S52: Comparison of inversion results of LCC inversion model constructed by wavelet coefficient feature coupling machine learning method before and after adding spectral index as a feature to wavelet coefficient; S53: Comparison of inversion results of LCC inversion model constructed by wavelet coefficient feature coupling machine learning method before and after adding noise.
4. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 1, characterized in that: Step S32 includes: S321: The LCC inversion model parameters are tuned using the support vector machine method. The radial basis function is used as the kernel function to handle the nonlinear characteristics of the LCC inversion model data. The regularization parameter C=1.0 and the tolerance parameter epsilon=0.
01.
5. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 3, characterized in that: Step S32 further includes: S322: Construct a universal LCC inversion model using wavelet coefficient feature coupling machine learning method, use the obtained measured training set to participate in model training, and use the measured validation set to participate in model validation.
6. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 1, characterized in that: Step S32 further includes: S323: Construct a general LCC inversion model using wavelet coefficient features coupled with machine learning methods, use the obtained simulated dataset wavelet coefficient features for model training, and use the measured dataset wavelet coefficient features for model validation.
7. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 1, characterized in that: The PROSPECT radiative transfer model adopts the PROSPECT-D radiative transfer model, and the simulation dataset includes the chlorophyll content, dry matter content, equivalent water thickness, carotenoid content, brown pigment content, structural parameters, and anthocyanin content of plant leaves.
8. The high-precision remote sensing inversion method for chlorophyll content in vegetation leaves according to claim 7, characterized in that: S2 In step 3, the Pearson correlation coefficient is used to quantify the linear correlation between wavelet coefficient features and LCC values. The calculation formula is as follows: In the formula, It is the first wavelet coefficient. i One value; It is the LCC's i One value; It is the average value of the wavelet coefficients; R is the average value of LCC; the value of R ranges from [-1, 1], where r = 1 indicates perfect positive correlation; r = -1 indicates perfect negative correlation; and r = 0 indicates no linear correlation.
9. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 2, characterized in that: The noise includes a relative noise of 0.005 and an absolute noise of 0.
003.
10. The method for high-precision remote sensing inversion of chlorophyll content in vegetation leaves according to claim 1, characterized in that: The measured dataset in step S11 includes the reflectance spectrum of fresh leaves and the chlorophyll content of fresh leaves.