Hyperspectral retrieval method for phytoplankton absorption coefficient based on multispectral satellite

Abstract

This invention relates to a hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites. The method includes fitting a Gaussian decomposition model to the measured phytoplankton absorption coefficient spectrum, selecting several predicted center wavelengths from the center wavelengths of multiple Gaussian curves in the model; establishing a fitting model between the remaining center wavelengths and the predicted center wavelengths; acquiring multispectral remote sensing images, using the remote sensing reflectance and spectral index of the sensitive bands of phytoplankton absorption coefficients as input to train a machine learning model, and obtaining a Gaussian parameter inversion model; using the inversion model and the fitting model to obtain Gaussian parameter values ​​from the multispectral remote sensing images of the lake under test, and substituting them into the Gaussian decomposition model to obtain the phytoplankton absorption coefficient spectrum of the lake under test. This method can solve the problem of the limited applicability of remote sensing inversion algorithms for inherent optical quantities of inland water bodies, and realize the monitoring of phytoplankton absorption coefficients in inland water bodies under different optical conditions.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, specifically to a hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites. Background Technology

[0002] Phytoplankton, through the absorption of light, provide a fundamental energy source for biogeochemical processes in aquatic bodies, playing a crucial role in the dynamic changes of ecosystems. The phytoplankton absorption coefficient (ACO) is a key indicator of this process. ph The a (light absorption capacity) index reflects the absorption characteristics of phytoplankton to different spectra in water bodies. It varies depending on the species and quantity of phytoplankton in the water body and is an important indicator for measuring the light absorption efficiency of phytoplankton. In inland lakes, especially eutrophic water bodies, remote sensing technology can be used to invert a (light absorption capacity) index. ph This value can help us to delve into the biogeochemical mechanisms and aquatic ecological conditions of water bodies, and has important academic value and practical significance for accurately assessing the content and types of phytoplankton pigments in water bodies. ph A spectrum is a comprehensive reflection of the absorption coefficients of different pigments; existing a ph Most products are based on a single or a few characteristic bands, making it difficult to obtain a ph Continuous spectrum. In addition, due to the complex optical properties of inland water bodies and the significant regional differences in surrounding aerosol conditions, as well as their susceptibility to the influence of neighboring land effects, remote sensing inversion algorithms developed for specific regions are often difficult to generalize and have limited application effects. Summary of the Invention

[0003] The purpose of this invention is to provide a hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites, the method comprising:

[0006] The center wavelengths of multiple Gaussian curves in the model were obtained by fitting the Gaussian decomposition model of the phytoplankton absorption coefficient spectrum with the measured phytoplankton absorption coefficient spectrum.

[0007] Several center wavelengths are selected from the center wavelengths of the multiple Gaussian curves as predicted center wavelengths; for the remaining center wavelengths, fitting models are established with at least one of the predicted center wavelengths respectively.

[0008] The remote sensing reflectance of multispectral remote sensing images is obtained, and sensitive bands for phytoplankton absorption coefficients and spectral indices constructed based on the sensitive bands are screened. The remote sensing reflectance of the sensitive bands and the spectral indices are used as inputs, and the Gaussian parameter values ​​of the Gaussian decomposition model corresponding to the predicted center wavelength are used as outputs to train a machine learning model and obtain a Gaussian parameter inversion model.

[0009] A multispectral remote sensing image of the lake to be tested is acquired. Based on the Gaussian parameter inversion model, the Gaussian parameter value of the predicted center wavelength is obtained, and the Gaussian parameter value of the remaining center wavelength is obtained using the fitting model. Based on the predicted and fitted Gaussian parameter values, the Gaussian decomposition model is substituted to obtain the phytoplankton absorption coefficient spectrum of the lake to be tested.

[0010] In some embodiments of the present invention, the predicted center wavelength is selected based on the following method:

[0011] Perform pairwise correlation analysis on the center wavelengths of the multiple Gaussian curves;

[0012] Select several center wavelengths such that the correlation between the remaining center wavelengths and at least one selected center wavelength meets the preset requirements.

[0013] In some embodiments of the present invention, the fitting model is in the form of a power function.

[0014] In some embodiments of the present invention, the sensitive bands include the blue light band to the near-infrared band of multispectral remote sensing images.

[0015] In some embodiments of the present invention, the spectral indices include the phytoplankton index FAI or AFAI, the near-infrared / green band ratio N / G, the green / blue band ratio G / B, the near-infrared band normalized data n1, and the blue band normalized data b1.

[0016] In some embodiments of the present invention, the machine learning model is selected from XGBoost, RF, or DNN models.

[0017] In some embodiments of the present invention, the measured phytoplankton absorption coefficient spectrum includes measured data from lakes in multiple regions and with various water body optical types.

[0018] In some embodiments of the present invention, the center wavelength of the Gaussian curve is the center wavelength obtained by optimization using the least squares method.

[0019] In some embodiments of the present invention, the spectral remote sensing image is an OLCI image.

[0020] In some embodiments of the present invention, the remote sensing reflectance is the remote sensing reflectance corrected by the DSF method.

[0021] To address the challenges of complex optical properties, significant differences in spectral shapes, and limited applicability of remote sensing inversion algorithms for inherent optical quantities in inland water bodies, which hinder continuous spectral monitoring of phytoplankton absorption coefficients, this invention provides a method for acquiring continuous spectral data of phytoplankton absorption coefficients by combining multispectral satellite data sources and utilizing Gaussian decomposition models, machine learning, and fitting models. This method is universally applicable to different inland water bodies, providing methodological and data support for monitoring phytoplankton absorption coefficients in inland water bodies with various optical conditions, and has significant scientific value. Attached Figure Description

[0022] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein:

[0023] Figure 1 This is a flowchart of the method shown in the embodiment.

[0024] Figure 2 These are measured biological optical parameters and a ph Statistical histograms of the data; where (a) is the distribution of statistical parameters, and (b) is a ph (443), (c)a ph (510), (d)a ph (560), (e)a ph (620), (f)a ph (674).

[0025] Figure 3 This is a frequency distribution chart of derivatives, where (a) is the first derivative and (b) is the second derivative.

[0026] Figure 4 It is the result of Gaussian function decomposition curve.

[0027] Figure 5 a is the measured data ph Inversion accuracy verification.

[0028] Figure 6 It is the satellite-to-ground synchronization data a ph Spatial distribution verification results; where (a)a ph (412), (b)a ph (443), (c)a ph (490), (d)a ph (510), (e)a ph (560), (f)a ph (620), (g)a ph(665), (h)a ph (674), (i)a ph (682).

[0029] Figure 7 This represents the fitting accuracy of the DSF method in different bands; where (a)a ph (412), (b)a ph (443), (c)a ph (490), (d)a ph (510), (e)a ph (560), (f)a ph (620), (g)a ph (665), (h)a ph (674), (i)a ph (709).

[0030] Figure 8 Lake A was on March 15, 2022. ph (443), a ph (490), a ph (510), a ph (560), a ph (620), a ph (665), a ph (674) and a ph (682) Spatial distribution.

[0031] Figure 9 It is the average monthly a of a typical lake ph (665) Spatial distribution; where (a) monthly average a of Lake A ph (665) Spatial distribution, (b) Monthly average a of Lake B ph (665) Spatial distribution, (c) monthly average a of lake C ph (665) Spatial distribution.

[0032] In the aforementioned Figures 1-9, the coordinates, symbols, or other representations expressed in English are all well-known in the field and will not be elaborated upon in this example. Detailed Implementation

[0033] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0034] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. The embodiments of this disclosure are not necessarily defined to include all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0035] Example 1

[0036] This embodiment illustrates the hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites according to the present invention.

[0037] This embodiment uses a hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites, and the implementation is as follows:

[0038] Measured phytoplankton absorption coefficient data from 2008 to 2023 were obtained. A Gaussian decomposition model was used to determine the phytoplankton absorption coefficients in characteristic bands. Different machine learning methods were employed for fitting, and the fitting accuracy of different models and input data combinations was compared to select the most suitable machine learning algorithm and input data combination. This model was then applied to OLCI data corrected for DSF atmospheric conditions to obtain the a... ph Hyperspectral dataset.

[0039] As an example, the following description, in conjunction with Figure 1 As shown, the implementation of the aforementioned method will be explained in detail.

[0040] 1) Obtain measured phytoplankton absorption coefficient data for 25 lakes in different lake regions from 2008 to 2023. Figure 2 Using a Gaussian decomposition model, the phytoplankton absorption coefficients of characteristic bands were determined;

[0041] Based on the Gaussian function model (Formula (1)), the Gaussian model was fitted using N = 847 data points collected from field measurements between 2008 and 2023.

[0042] The measured data sampling range includes three lake areas: IMXL, YGPL, and EPL, in order to increase the optical types of water bodies and improve the applicability of the model.

[0043] (1)

[0044] in, and They represent the first The Gaussian parameter value (variation range) at the center wavelength of the Gaussian curve and the width of the Gaussian curve.

[0045] Each Gaussian curve consists of the center wavelength value λ, the width of the Gaussian curve σ, and the variation amplitude a. gaus Determined by three parameters.

[0046] Based on (Wang, G., Lee, Z., Mishra, DR, & Ma, R. (2016). Retrievingabsorption coefficients of multiple phytoplankton pigments from hyperspectralremote sensing reflectance measured over cyanobacteria bloom waters. LIMNOLOGY AND Oceanography-Methods, 14 The parameter optimization method (432–447) is used to calculate the measured a respectively. ph First derivative of data and second derivative And statistically analyze the frequency of occurrence at different wavelengths when the first and second derivatives are zero. Figure 3 The wavelength position where the first derivative is 0 indicates a. ph The peaks or troughs of the spectrum, while the wavelength position where the second derivative is 0 indicates a. ph Inflection point of the spectrum.

[0047] Using the least squares optimization method integrated in MATLAB, eight initial Gaussian function center wavelengths and curve widths were determined. The resulting Gaussian function curves are shown below. Figure 4 As shown.

[0048] Table 1. Optimized Gaussian parameters (center wavelength and Gaussian curve width) for the eight spectral components.

[0049]

[0050] 2) Parameter selection and model optimization in machine learning;

[0051] This application employs three machine learning algorithms as the core architecture of the inversion algorithm, including Random Forest (RF), Extreme Gradient Boosting (XGBoost) algorithm, and Deep Neural Network (DNN) model. These algorithms were selected as the basic model framework for constructing the inversion method.

[0052] The inversion model designed in this method aims to obtain the Gaussian function value corresponding to the center wavelength of the Gaussian curve. This is achieved by designing an input model based on measured R0. rsDifferent combinations of wavebands and water quality parameters (spectral indices) were used to fit the data using different machine learning methods. The fitting accuracy of different models and different combinations of input data was compared to select the most suitable machine learning algorithm and input data combination. A total of N = 847 available data points were used for model construction. 565 data points were randomly selected to build the training set, and 282 data points were used to build the test set.

[0053] The spectral index calculation method constructed in this embodiment is as follows:

[0054] (2)

[0055] (3)

[0056] (4)

[0057] (5)

[0058] (6)

[0059] This embodiment compared various combinations of input data types and different machine learning methods. It found that this embodiment involves eight Gaussian function center wavelengths, meaning the machine learning model needs to predict eight parameters. Using different combinations of inputs and machine learning models failed to achieve good prediction results. Therefore, this embodiment compared the pairwise correlations between the eight Gaussian function center wavelengths. Gaussian function values ​​corresponding to center wavelengths of 490.0 nm, 540.1 nm, and 622.4 nm were selected as predicted values. The Gaussian function values ​​corresponding to the remaining five center wavelengths all showed good correlations with the Gaussian function value corresponding to at least one selected center wavelength.

[0060] After selecting the predicted wavelengths, the prediction effects of various input data combinations and different machine learning methods were re-compared. The results showed that using the XGBoost machine learning model with measured reflectance data at wavelengths of 412, 443, 490, 510, 560, 665, 674, 709, 754, and 865 nm, and water quality parameters (using the improved phytoplankton index AFAI, band normalized data n1 and b1, and band ratio data N / G and G / B as input data), the model predicted Gaussian function values ​​corresponding to center wavelengths of 490.0 nm, 540.1 nm, and 622.4 nm. The obtained fitting accuracy was relatively ideal. The R-values ​​of the model training set were [not specified in the original text]. 2 The value was 0.90, and the RMSE was 0.22 m. -1 The UAPD was 31.7%, and the R of the test set was... 2The value was 0.64, and the RMSE was 0.48m. -1 UAPD was 46.3% ( Figure 5 ).

[0061] Based on the Gaussian function values ​​corresponding to the center wavelengths of 490.0 nm, 540.1 nm and 622.4 nm predicted by machine learning methods, the relationship between the predicted center wavelength Gaussian function values ​​and the corresponding Gaussian function values ​​of the other five center wavelengths was constructed by power function fitting (Table 2).

[0062] Table 2 Gaussian function values ​​a at different center wavelengths gaus The power function relationship between them

[0063]

[0064] 3) a ph Inversion algorithm accuracy verification

[0065] Using the field measurements collected in step 1) from the eastern lake region in 2024, N = 101 a-type ... ph The data validates the constructed inversion algorithm. The obtained a from the inversion... ph Data and measured a ph Error results between data are as follows Figure 6 As shown in the figure. It can be seen that the constructed inversion algorithm is applicable to a. ph The data values ​​are somewhat overestimated; within the wavelength range of 400–700 nm, the average bias of the proposed algorithm is 0.07 m. -1 The average APD was 44.56%, and the average RMSE was 0.28 m. -1 .

[0066] The specific comparison algorithm inverts a at a specific wavelength in the OLCI band. ph Data and measured a ph Errors between data. At wavelengths less than 709 nm in the first 11 bands of OLCI, the inversion yields a... ph Data and measured a ph The average bias among the data is 0.07 m. -1 The average RMSE value is 0.28 m. -1 The average APD was 46.46%.

[0067] 4) Typical lake a ph Spatiotemporal distribution

[0068] The applicability of the DSF atmospheric correction method to OLCI imagery was compared. Figure 7Based on the aforementioned optimal model, a typical lake's a-value was constructed using OLCI imagery data from Sentinel-3A and Sentinel-3B from April 2016 to October 2024. ph Hyperspectral dataset, for typical lakes A, B, and C, a ph Data and typical wavelengths of a ph The results show the monthly spatiotemporal variations of the data. Figures 8-9 ).

Claims

1. A hyperspectral inversion method for phytoplankton absorption coefficients based on multispectral satellites, characterized in that, The method includes: The center wavelengths of multiple Gaussian curves in the model were obtained by fitting the Gaussian decomposition model of the phytoplankton absorption coefficient spectrum with the measured phytoplankton absorption coefficient spectrum. Several center wavelengths are selected from the center wavelengths of the multiple Gaussian curves as predicted center wavelengths; for the remaining center wavelengths, fitting models are established with at least one of the predicted center wavelengths respectively. The remote sensing reflectance of multispectral remote sensing images is obtained, and sensitive bands for phytoplankton absorption coefficients and spectral indices constructed based on the sensitive bands are screened. The remote sensing reflectance of the sensitive bands and the spectral indices are used as inputs, and the Gaussian parameter values ​​of the Gaussian decomposition model corresponding to the predicted center wavelength are used as outputs to train a machine learning model and obtain a Gaussian parameter inversion model. A multispectral remote sensing image of the lake to be tested is acquired. Based on the Gaussian parameter inversion model, the Gaussian parameter value of the predicted center wavelength is obtained, and the Gaussian parameter value of the remaining center wavelength is obtained using the fitting model. Based on the predicted and fitted Gaussian parameter values, the Gaussian decomposition model is substituted to obtain the phytoplankton absorption coefficient spectrum of the lake to be tested.

2. The method according to claim 1, characterized in that, The predicted center wavelength is selected based on the following method: Perform pairwise correlation analysis on the center wavelengths of the multiple Gaussian curves; Select several center wavelengths such that the correlation between the remaining center wavelengths and at least one selected center wavelength meets the preset requirements.

3. The method according to claim 1, characterized in that, The fitting model is in the form of a power function.

4. The method according to claim 1, characterized in that, The sensitive bands include the blue light band to the near-infrared band of multispectral remote sensing images.

5. The method according to claim 1, characterized in that, The spectral indices include the phytoplankton index FAI or AFAI, the near-infrared / green band ratio N / G, the green / blue band ratio G / B, the near-infrared band normalized data n1, and the blue band normalized data b1.

6. The method according to claim 1, characterized in that, The machine learning model selected is the XGBoost model, the RF model, or the DNN model.

7. The method according to claim 1, characterized in that, The measured phytoplankton absorption coefficient spectra include measured data from lakes in multiple regions and with various water body optical types.

8. The method according to claim 1, characterized in that, The center wavelength of the Gaussian curve is the center wavelength obtained through optimization using the least squares method.

9. The method according to claim 1, characterized in that, The spectral remote sensing image is an OLCI image.

10. The method according to claim 9, characterized in that, The remote sensing reflectance is the remote sensing reflectance corrected by the DSF method.

Images