Method for constructing radiation inversion model considering correlation of plant traits

Abstract

The application discloses a radiation inversion model construction method considering plant trait correlation. By considering the plant trait correlation in the construction task of the radiation inversion model, the obtained leaf mixed inversion model and canopy mixed inversion model can realize high-precision inversion of key leaf functional traits based on hyperspectral remote sensing.

Description

Technical Field

[0001] This invention relates to the field of vegetation remote sensing technology, and in particular to a method for constructing a radiometric inversion model that considers the correlation of plant traits. Background Technology

[0002] Plant functional traits refer to the core attributes that significantly affect plant establishment, survival, growth, and death, reflecting the plant's response and adaptation to changes in the external environment. Leaf functional traits include parameters related to leaf physiology, biochemistry, and morphology, such as chlorophyll content, leaf nitrogen content, specific leaf area, specific leaf weight, and water content. These are related to numerous physiological and ecological processes such as photosynthesis, respiration, and transpiration, and are closely associated with primary productivity, litter decomposition, and nutrient cycling. They influence ecosystem processes, functions, and services, and are key parameters for assessing biodiversity.

[0003] Traditional methods for measuring plant functional traits mainly involve field sampling and indoor chemical analysis, which are time-consuming, labor-intensive, and costly. Hyperspectral remote sensing can extract and analyze fine, narrow-band information about plants to achieve rapid and efficient estimation of plant functional traits.

[0004] Currently, there are two main types of hyperspectral remote sensing inversion methods. The first is the statistical analysis method, which directly establishes the statistical relationship between plant functional traits and spectra. This method offers high inversion accuracy but has low generalizability. The second is the physical model method, which uses optical laws to describe the reflection, absorption, and scattering of solar radiation in leaves and canopies. This method has high generalizability, but often suffers from "ill-conditioned inversion," resulting in lower accuracy. Most past studies have relied solely on these two inversion methods, leading to limitations in predictive models, such as low portability or low inversion accuracy.

[0005] The limitation of hyperspectral remote sensing inversion methods that use physical models is that they invert the functional traits of a single leaf, ignoring the intrinsic relationship between different leaf functional traits. This causes the model input parameters to change independently, resulting in unrealistic combinations of model parameters. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method for constructing a radiometric inversion model that considers the correlation of plant traits. This method aims to take into account the intrinsic correlation between leaf functional traits in hyperspectral remote sensing inversion, thereby improving inversion accuracy and ensuring that the inversion results conform to ecological rules.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0008] A method for constructing a radiation inversion model that considers the correlation of plant traits includes the following steps:

[0009] Step 1: Randomly obtain several sets of leaf trait parameters, and select parameter sets with variable correlation from the set of leaf trait parameters based on the leaf trait correlation data. Define these parameter sets as the first parameter set, and calculate the contribution value of leaf traits to leaf reflectance based on the first parameter set.

[0010] Step 2: Randomly obtain several sets of leaf trait parameters and canopy input parameters. Based on the correlation data of leaf traits, select parameter sets with variable correlation from the canopy input parameter sets and define them as the second parameter set. Calculate the contribution value of leaf traits to canopy reflectivity based on the second parameter set.

[0011] Step 3: Generate a preset number of simulated spectra and corresponding leaf input parameters through forward simulation. Add a preset proportion of Gaussian noise to the simulated spectra and the leaf parameters respectively. Using the first parameter set as input and the simulated spectra as output, optimize and restrict the simulated spectral library according to the correlation of leaf traits. Use the Kolesky matrix decomposition method to associate the correlation of leaf traits between the first parameter set and the simulated spectral library to train and generate a leaf hybrid inversion model.

[0012] Step 4: Set the background soil and observation geometry to fixed values, change the second parameter group, and generate a preset number of model spectra and corresponding model input parameters in a forward simulation. Add a preset proportion of Gaussian noise to each model and train to generate a canopy hybrid inversion model.

[0013] In some embodiments, the set of leaf trait parameters includes at least leaf structure, chlorophyll, carotenoids, water content, specific leaf weight, and leaf protein content.

[0014] In some embodiments, the leaf trait correlation data includes measured correlation data between chlorophyll and carotenoids, correlation data between specific leaf weight and leaf water content, correlation data between leaf water content and leaf nitrogen content, and correlation data between chlorophyll and leaf nitrogen content.

[0015] In some embodiments, the leaf trait correlation data are expressed using the Pearson correlation coefficient.

[0016] In some implementations, in step one, based on the leaf trait correlation data, the leaf trait parameter set is decomposed using the Kolesky matrix and transformed into the first parameter set.

[0017] The beneficial effects of this invention are as follows: by considering the correlation of plant traits in the construction task of the radiometric inversion model, the resulting leaf hybrid inversion model and canopy hybrid inversion model can achieve high-precision inversion of key leaf functional traits based on hyperspectral remote sensing. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of the method for constructing a radiation inversion model that considers the correlation of plant traits, as disclosed in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram showing the sensitivity analysis results of leaf radiative transfer models that do not consider trait correlation and those that do consider trait correlation.

[0020] Figure 3 This is a schematic diagram showing the sensitivity analysis results of canopy radiative transfer models that do not consider trait correlations and those that do consider trait correlations.

[0021] Figure 4 A schematic diagram of the inversion results of the PROSEPCT-D / -PRO leaf radiative transfer model, considering trait correlations, for the output of the simulated dataset.

[0022] Figure 5 A schematic diagram of the inversion results of the PROSAIL canopy radiative transfer model considering trait correlations, assuming the output of the simulated dataset.

[0023] Figure 6 A schematic diagram of the hybrid inversion results of the leaf radiative transfer model considering trait correlation, given the output of the measured dataset.

[0024] Figure 7 This is a schematic diagram of the hybrid inversion results of the canopy radiative transfer model considering trait correlations, given the output of the measured dataset. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.

[0026] Example:

[0027] This embodiment proposes a method for constructing a radiation inversion model that considers the correlation of plant traits, such as... Figure 1 As shown, it includes the following steps:

[0028] Step 1: Randomly obtain several sets of leaf trait parameters. Based on the correlation data of leaf trait parameters, select the parameter set with variable correlation from the set of leaf trait parameters and define it as the first parameter set. Calculate the contribution value of leaf traits to leaf reflectance based on the first parameter set.

[0029] In one example, step one first randomly samples 1000 sets of leaf trait parameters, which include at least leaf structure, chlorophyll, carotenoids, water content, specific leaf weight, and leaf protein content. To consider the correlations between functional traits of plant leaves, measured leaf trait correlation data is also obtained, expressed using the Pearson correlation coefficient. This leaf trait correlation data includes the correlation between chlorophyll and carotenoids (Pearson correlation coefficient r = 0.95), the correlation between specific leaf weight and leaf water (r = 0.85), the correlation between leaf water and leaf nitrogen (r = 0.66), and the correlation between chlorophyll and leaf nitrogen (r = 0.42).

[0030] Based on the correlation data of leaf traits, the leaf trait parameter set was decomposed using the Kolesky matrix and transformed into the first parameter set.

[0031] The aforementioned leaf radiative transfer models are important tools for describing the process and mechanism of light radiative transfer within leaves. These models, by simulating the absorption, reflection, and transmission of light by leaves, help to understand the optical properties of leaves and the spectral response of vegetation. The input parameters of leaf radiative transfer models typically include leaf structural parameters (such as the number of layers and thickness) and biochemical components (such as chlorophyll, water, and dry matter content). These parameters have a significant impact on the model's output. The output parameters of leaf radiative transfer models mainly include leaf reflectivity, transmittance, and absorptivity, which reflect the leaf's response characteristics to different wavelengths of light.

[0032] In this embodiment, the PROSPECT-D / -PRO model of the blade radiative transfer is used. The first set of parameters is used as the input of the blade radiative transfer model to generate the blade reflectance in a forward simulation. Then, the Sobel global sensitivity analysis method is used to calculate the contribution of the blade parameters to the blade reflectance under the condition of blade trait correlation.

[0033] Furthermore, due to the four types of correlations mentioned above among functional traits of plant leaves, more correlations will arise between various parameter combinations. Therefore, it is necessary to calculate the direct and indirect correlations between all parameter combinations. Among them, the direct correlations are mainly the four groups of parameters mentioned above, while the indirect correlations include carotenoids and water (r = 0.26), carotenoids and leaf proteins (r = 0.40), and chlorophyll and leaf water (r = 0.28).

[0034] In summary, the contribution of leaf traits to leaf reflectivity can remove redundant variables, eliminate trait parameters that contribute little to reflectivity, and reduce the input noise of subsequent leaf hybrid inversion models.

[0035] Strengthen correlation constraints: Only retain trait parameters with significant physical correlations to avoid inversion errors caused by biases in the parameter independence assumption in the leaf mixture inversion model. The subsequent generation of the simulated spectral library (step three) is based on the first parameter group, whose trait correlations have been screened through contribution values ​​to ensure the physical rationality of the spectral simulation.

[0036] Step 2: Randomly obtain several sets of leaf trait parameters and canopy input parameters. Based on the correlation data of leaf traits, select parameter sets with variable correlation from the canopy input parameter sets and define them as the second parameter set. Calculate the contribution value of leaf traits to canopy reflectivity based on the second parameter set.

[0037] In one example, by coupling the leaf radiative transfer model PROSPECT-D / -PRO and the canopy radiative transfer model SAIL, 2000 sets of random parameters were obtained by randomly sampling the leaf trait parameter set and the canopy input parameter set. The canopy input parameters included leaf area index and mean leaf tilt angle. Similarly, considering the correlation of leaf traits, parameter combinations with variable correlations were obtained (refer to step one). A second set of parameters was selected and used as the input parameters for both the leaf and canopy radiative transfer models. Forward simulation was used to obtain the canopy reflectance. Then, the Sobel global sensitivity analysis method was used to calculate the contribution of leaf parameters to canopy reflectance considering the correlation of leaf traits.

[0038] In summary, the contribution of leaf traits to canopy reflectivity can be modeled across scales, incorporating the correlation between leaf traits and canopy parameters into the model input, thus avoiding the decrease in inversion accuracy caused by ignoring cross-scale interactions.

[0039] Canopy model simplification: By setting the background soil and observation geometry to fixed values ​​(step four), the model spectrum is generated by changing only the second parameter set, reducing interference from non-critical variables.

[0040] Step 3: Generate a preset number of simulated spectra and corresponding leaf input parameters through forward simulation. Add a preset proportion of Gaussian noise to the simulated spectra and leaf parameters respectively. Using the first parameter group as input and the simulated spectra as output, optimize and restrict the simulated spectral library according to the correlation of leaf traits. Then, use the Kolesky matrix decomposition method to associate the correlation of leaf traits between the first parameter group and the simulated spectral library to train and generate a leaf hybrid inversion model.

[0041] In one example, step three uses a forward simulation of the blade radiative transfer model to generate 100,000 simulated spectra and blade input parameters, adding 2% Gaussian noise to the reflectance range of each band in the simulated spectra and blade input parameters. For the simulated spectra and blade parameters, the data are randomly divided into training and validation samples, with 2 / 3 for training and 1 / 3 for validation. Based on the training samples, a Gaussian process regression machine learning method is used for training to obtain a hybrid inversion model for each blade trait.

[0042] A leaf fusion inversion model considering plant trait correlations is proposed. For the aforementioned noisy simulated spectra and leaf input parameters, the simulated dataset is optimized and constrained based on leaf trait correlations. The Kolesky matrix factorization method is used to consider leaf trait correlations, resulting in a simulated database with trait correlations that better reflects real-world scenarios. The training process of the fusion model is repeated, and the model's performance is validated for accuracy.

[0043] Step 4: Set the background soil and observation geometry to fixed values, change the second parameter group, and generate a preset number of model spectra and corresponding model input parameters in a forward simulation. Add a preset proportion of Gaussian noise to each model and train to generate a canopy hybrid inversion model.

[0044] In one example, step four uses the canopy radiative transfer model INFORM as the basis for spectral simulation. INFORM couples the forest light model FLIM, the canopy model SAILH, and the leaf radiative transfer model PROSPECT-D / -PRO, simulating forest bidirectional-directional reflectance. This model strikes a good balance between finely characterizing canopy structure and model invertibility. The input parameters of INFORM include leaf attributes, canopy structure, background soil, and observation geometry. To reduce the number of input parameters and thus decrease model uncertainty, the background soil and observation geometry are set to fixed values. The second set of parameters, namely leaf parameters and canopy structure parameters such as leaf area index, average leaf tilt angle, tree height, and tree density, are mainly changed. 100,000 model spectra and input parameters are generated using forward simulation with INFORM, and 2% Gaussian noise is added. Subsequent model training follows the same process as the simulated spectra and leaf parameters described in step three, resulting in a canopy hybrid inversion model that considers the correlation of plant traits.

[0045] Comparative example:

[0046] Referring to steps one through four above, the difference is that parameter groups with variable correlation are not screened, resulting in leaf mixed inversion models and canopy mixed inversion models that do not consider the correlation of plant functional traits.

[0047] The accuracy verification process of the blade hybrid inversion model includes:

[0048] The leaf hybrid inversion model was applied to the validation samples of the leaf horizontal simulation dataset to obtain predicted leaf trait values. The coefficient of determination (R²), root mean square error (RMSE), relative error (RMSE% = RMSE / mean), and bias (BIAS) were calculated to verify the model's inversion accuracy. Simultaneously, the leaf hybrid inversion model was applied to the leaf horizontal field measurement dataset, including leaf reflectance and leaf traits measured by ground object spectroscopy, to calculate the aforementioned four statistical indicators and verify the model's accuracy.

[0049] The accuracy verification process of the canopy hybrid inversion model includes:

[0050] Canopy spectra of different geographical regions, plant functional types and species were acquired using an airborne imaging spectrometer. Leaf samples were collected to measure plant leaf functional traits, including chlorophyll, carotenoids, leaf water and dry matter. The coefficient of determination, root mean square error, relative error and bias were calculated to verify the accuracy of the prediction results of the canopy mixing inversion model.

[0051] The inversion accuracy of the model is shown in Tables 1 to 4 below.

[0052] Table 1. Inversion accuracy of leaf hybrid inversion models without considering trait correlation and with considering trait correlation (leaf level simulation dataset).

[0053]

[0054]

[0055] Table 2. Inversion accuracy of canopy hybrid inversion models without considering trait correlation and with considering trait correlation (canopy level simulation dataset).

[0056]

[0057] Table 3. Inversion accuracy of leaf hybrid inversion models without considering trait correlation and with considering trait correlation (actual measured data set of leaf level).

[0058]

[0059] Table 4. Inversion accuracy of canopy hybrid inversion models without considering trait correlation and with considering trait correlation (canopy level measured dataset).

[0060]

[0061] Figure 2The results show that in the PROSPECT-D leaf model, chlorophyll plays a significant role primarily in the red to red-edge wavelength range, while carotenoids contribute mainly in the green wavelength range, exhibiting a narrow but strong absorption effect. After considering the correlation between chlorophyll and carotenoids, chlorophyll content incorporates the contributions of carotenoids, showing a high contribution of over 80% across the entire green-red-red-edge wavelength range. However, the contribution of carotenoids to leaf reflectance is masked, showing a very low contribution of approximately 10%. For specific leaf weight and leaf water content, after considering their correlation, similar results are observed as with leaf pigments: the contribution of leaf water is amplified to 80%, while the contribution of specific leaf weight is significantly reduced to only about 5%. For PROSPECT-PRO, without considering the correlation of leaf traits, leaf proteins show a small contribution at 1600-1800 nm and 2100-2300 nm, but this contribution almost disappears after considering the correlation. Meanwhile, chlorophyll shows a small contribution in the short-wave infrared band.

[0062] Figure 3 The results show that in the PROSAIL-D canopy model, after considering the correlation of leaf traits, the contribution of carotenoids almost disappears, the contribution of chlorophyll decreases by about 10%, the contribution of leaf water increases somewhat but decreases significantly compared to the contribution of leaf weight. The contribution of leaf area index increases by about 10%, while the contribution of leaf tilt angle distribution decreases accordingly. For the coupled PROSAIL-PRO model, without considering the correlation of leaf traits, proteins contribute to canopy reflectance twice, at 1600-1800 nm and 2100-2300 nm, but when the correlation is considered, their contribution decreases to almost nothing. Meanwhile, chlorophyll shows about 5% contribution in the shortwave infrared band at 1300-2500 nm. Without considering the correlation of leaf traits, leaf area index contributes about 40% in the near-infrared band, and average leaf tilt angle contributes significantly in visible, near-infrared, and shortwave infrared. However, after considering the correlation, the contribution of leaf area index to canopy reflectance decreases slightly, while the contribution of average leaf tilt angle decreases significantly.

[0063] Figure 4 As shown in Table 1, for the leaf radiative transfer models PROSEPCT-D and PROSEPCT-PRO, without considering leaf trait correlations, all leaf parameters in the simulation dataset can be accurately inverted. Among them, the inversion accuracy of chlorophyll, carotenoids, leaf water, and specific leaf weight is higher than that of leaf protein, and the relative errors of all five parameters are below 5%. After considering leaf trait correlations, the inversion accuracy of all five leaf parameters is improved to some extent, and the relative error of the model is reduced, especially the leaf protein content.

[0064] Figure 5As shown in Table 2, at the canopy level, without considering leaf trait correlations, the leaf parameters of the simulated dataset can still be accurately inverted, although the inversion accuracy is slightly lower than at the leaf level, with a relative error of 2.94-6.21%. After considering leaf trait correlations, the inversion accuracy of the leaf parameters of the simulated dataset is improved, and the relative error decreases to 2.37-4.65%.

[0065] Figure 6 As shown in Table 3, at the leaf level, chlorophyll in the measured dataset yielded accurate inversion results. Without considering leaf trait correlations, the R² and relative error were 0.56 and 14.38%, respectively. Considering leaf trait correlations significantly improved the chlorophyll inversion accuracy, with R² and relative error of 0.66 and 12.73%, respectively. Carotenoid inversion accuracy was very low, with R² and relative error of 0.06 and 55.68%, respectively. Considering leaf trait correlations, R² increased to 0.30, with a relative error of 60.05%. Using the leaf radiative transfer model, without considering leaf trait correlations, leaf water and specific leaf weight yielded relatively accurate inversion results, with R² of 0.60 and 0.75, and relative errors of 7.46% and 14.45%, respectively. Considering leaf trait correlations improved the inversion accuracy for both, with R² of 0.65 and 0.76, and relative errors of 6.83% and 12.15%, respectively. For the leaf radiative transfer model PROSEPCT-PRO, the accuracy of leaf protein inversion is very low when leaf trait correlation is not considered, with R² and relative error of 0.37 and 108.07%, respectively. When leaf trait correlation is considered, the inversion accuracy is greatly improved, and the relative error is reduced to 24.53%.

[0066] Figure 7 As shown in Table 4, at the canopy level, without considering leaf correlation, the chlorophyll retrieval accuracy of the measured dataset was very low, with a relative error exceeding 200%. However, after considering leaf correlation, the relative error of chlorophyll decreased significantly to 61.10%. At the canopy level, the retrieval accuracy of carotenoids was even lower. Both without and with leaf correlation, the R² was 0.30, but the relative error decreased from 191.62% to 79.56%. At the canopy level, without considering leaf trait correlation, the retrieval results for leaf water and specific leaf weight were relatively accurate, with R² values ​​of 0.58 and 0.30, respectively, but the relative errors were large, at 84.25% and 128.08%, respectively. After considering leaf trait correlation, the retrieval R² for leaf water and specific leaf weight increased to 0.67 and 0.49, respectively, and the relative errors decreased to 81.65% and 57.71%, respectively. At the canopy scale, the accuracy of leaf protein inversion is very low when leaf trait correlation is not considered; after considering leaf trait correlation, the inversion accuracy is improved, with R² and relative error of 0.35 and 43.56%, respectively.

[0067] The above results indicate that the correlations among plant functional traits have a significant impact on the positive simulation of the model, especially the model sensitivity analysis. After considering trait correlations, the inversion accuracy of each leaf trait is improved at both the leaf and canopy scales. This demonstrates that considering leaf trait correlations can reduce parameter combinations that do not exist in reality, decrease pathological inversions, and improve the inversion accuracy of leaf parameters. The hybrid model method effectively combines the advantages of physical models and empirical models (machine learning), possessing both high inversion accuracy and good versatility. The results show that hybrid inversion can accurately invert each leaf trait, and the inversion accuracy is significantly improved when considering leaf correlations, enabling high-precision inversion of key plant leaf functional traits based on hyperspectral remote sensing. This research can enhance the understanding of ecosystem functions and processes, improve the accuracy of Earth system simulation and prediction, and enhance the monitoring capabilities of biological functional diversity.

[0068] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a radiation inversion model considering the correlation of plant traits, characterized in that, Includes the following steps: Step 1: Randomly obtain several sets of leaf trait parameters, and select parameter sets with variable correlation from the set of leaf trait parameters based on the leaf trait correlation data. Define these parameter sets as the first parameter set, and calculate the contribution value of leaf traits to leaf reflectance based on the first parameter set. Step 2: Randomly obtain several sets of leaf trait parameters and canopy input parameters. Based on the correlation data of leaf traits, select parameter sets with variable correlation from the canopy input parameter sets and define them as the second parameter set. Calculate the contribution value of leaf traits to canopy reflectivity based on the second parameter set. Step 3: Generate a preset number of simulated spectra and corresponding leaf input parameters through forward simulation. Add a preset proportion of Gaussian noise to the simulated spectra and the leaf parameters respectively. Using the first parameter set as input and the simulated spectra as output, optimize and restrict the simulated spectral library according to the correlation of leaf traits. Use the Kolesky matrix decomposition method to associate the correlation of leaf traits between the first parameter set and the simulated spectral library to train and generate a leaf hybrid inversion model. Step 4: Set the background soil and observation geometry to fixed values, change the second parameter group, and generate a preset number of model spectra and corresponding model input parameters in a forward simulation. Add a preset proportion of Gaussian noise to each model and train to generate a canopy hybrid inversion model.

2. The method for constructing a radiation inversion model considering the correlation of plant traits as described in claim 1, characterized in that, The set of leaf trait parameters includes at least leaf structure, chlorophyll, carotenoids, water content, specific leaf weight, and leaf protein content.

3. The method for constructing a radiation inversion model considering the correlation of plant traits as described in claim 1, characterized in that, The leaf trait correlation data includes the measured correlation data between chlorophyll and carotenoids, the correlation data between specific leaf weight and leaf water, the correlation data between leaf water and leaf nitrogen, and the correlation data between chlorophyll and leaf nitrogen.

4. The method for constructing a radiation inversion model considering the correlation of plant traits as described in claim 1, characterized in that, The correlation data of leaf traits are expressed using the Pearson correlation coefficient.

5. The method for constructing a radiation inversion model considering the correlation of plant traits as described in claim 1, characterized in that, In step one, based on the leaf trait correlation data, the leaf trait parameter set is decomposed using the Kolesky matrix and transformed into the first parameter set.

Images