Inversion method of vegetation photosynthetically active radiation absorption rate based on multi-source data fusion

Abstract

The invention discloses a method for inverting photosynthetically active radiation absorption rate of vegetation based on multi-source data fusion, comprising: obtaining a modeling data set for inverting photosynthetically active radiation absorption rate; constructing a pre-trained long short-term memory network inversion model PreT LSTM , where the pre-trained long short-term memory network inversion model PreT LSTM The training set is obtained by training, and the training set is the modeling data generated based on the soil-leaf-canopy coupling model SLC; the pre-trained long short-term memory network inversion model PreT LSTM As the source model, a migration training set is constructed through satellite data and measured data, and the source model is trained using the migration training set to obtain the migration training inversion model TL LSTM ; Input the modeling data set into the migration training inversion model TL LSTM , and obtain the estimated result of vegetation photosynthetically active radiation absorption rate. The present invention can better capture the complex ecological and biophysical processes in remote sensing data.

Description

Technical Field

[0001] The invention belongs to the technical field of ecological remote sensing, and in particular relates to a method for inverting vegetation photosynthetically active radiation absorption rate based on multi-source data fusion. Background Art

[0002] The Fraction of absorbed Photosynthetically Active Radiation (FAPAR) is defined as the ratio of solar radiation energy absorbed by vegetation with a wavelength between 400 nm and 700 nm. It is the core parameter for measuring the photosynthetic capacity of vegetation and significantly affects the physiological and ecological processes of vegetation photosynthesis, respiration and transpiration in terrestrial ecosystems. Since photosynthetically active radiation (PAR) is composed of direct and diffuse lighting conditions, FAPAR can be divided into three types according to different sky conditions: black sky, white sky and blue sky FAPAR. Black sky FAPAR (FAPAR bs ) refers to the fraction of radiation absorbed by plants under a clear, cloudless sky, representing the maximum potential absorption capacity under ideal conditions. ws ) takes into account the scattered radiation under cloudy or overcast conditions, providing an estimate of plant radiation absorption in a non-ideal, shadow-affected environment, reflecting a more realistic plant absorption situation. tol ) combines direct radiation and diffuse radiation, providing a comprehensive measurement that can reflect the radiation absorption of plants under typical sunlight conditions. These three types of FAPAR provide researchers with multi-level information. Therefore, in-depth analysis of FAPAR will help explore the carbon absorption and conversion capacity of vegetation under different environments and reveal the dynamic changes of vegetation in the ecosystem carbon cycle. At present, remote sensing estimation methods of FAPAR can be roughly divided into three categories: data-driven empirical methods, model-driven physical methods, and hybrid methods that integrate the two.

[0003] The empirical method has the advantages of simple operation and high computational efficiency by establishing a relationship model between measured FAPAR data and remote sensing observation parameters (such as reflectivity, vegetation index, etc.). However, due to its failure to consider physical laws, the universality of this type of method is limited and it is difficult to be widely used in diverse environments. In contrast, the physical method inverts FAPAR by constructing complex models, has a solid theoretical basis, and can accurately reflect the physiological and ecological processes of vegetation, but its calculation is complex and the inversion accuracy is easily affected by input data and parameterization methods. The hybrid method combines the advantages of empirical methods and physical methods, taking into account the theoretical support of physical models and the advantages of artificial intelligence algorithms in terms of simplicity of operation, efficiency, accuracy and practicality, and gradually gains more attention in practical applications. However, current related studies usually rely on machine learning algorithms such as support vector machines, artificial neural networks and random forests to achieve FAPAR inversion. The inversion relationships constructed by these methods mainly rely on data sets simulated by radiation transfer models and lack physical constraints from the remote sensing field. Compared with traditional machine learning algorithms, deep learning algorithms have stronger capabilities in constructing complex nonlinear regression relationships, and a variety of auxiliary technologies in the field of deep learning provide more possibilities for introducing remote sensing field knowledge.

[0004] At present, some global-scale FAPAR remote sensing products have been produced and publicly released, such as GLASS FAPAR, GEOV2 FAPAR, and MODIS FAPAR. However, due to the influence of bad weather, inversion algorithms, sensor equipment, etc., there are obvious differences between different FAPAR products, and there are still problems such as low quantitative verification accuracy and large product uncertainty. The most critical thing is that the spatial resolution of existing products is relatively coarse. For example, GEOV2 FAPAR has a spatial resolution of 1000m, MODIS FAPAR has a spatial resolution of 500m, and GLASS FAPAR provides two spatial resolutions of 500m and 250m. This shows that the FAPAR products generated by the existing algorithm will have serious mixed pixel problems when applied to heterogeneous surfaces, and it is difficult to meet the requirements of fine information extraction at a small regional scale. Therefore, it is urgent to propose new algorithms and produce global high-precision FAPAR products with high spatial resolution (such as 30m) for application. Summary of the invention

[0005] To solve the above technical problems, the present invention proposes a vegetation photosynthetically active radiation absorption rate inversion method based on multi-source data fusion, which is of great significance in improving the accuracy of FAPAR remote sensing estimation and the level of practical application.

[0006] The present invention provides a method for inverting vegetation photosynthetically active radiation absorption rate based on multi-source data fusion, comprising:

[0007] Obtain a modeling data set for inverting photosynthetically active radiation absorption rate;

[0008] Constructing a pre-trained long short-term memory network inversion model PreT LSTM , wherein the pre-trained long short-term memory network inversion model PreT LSTM The training set is obtained by training, wherein the training set is modeling data generated based on a soil-leaf-canopy coupling model SLC;

[0009] The pre-trained long short-term memory network inversion model PreT LSTM As the source model, a migration training set is constructed through satellite data and measured data, and the source model is trained using the migration training set to obtain a migration training inversion model TL LSTM ;

[0010] The modeling data set is input into the migration training inversion model TL LSTM , obtain the estimated results of vegetation photosynthetically active radiation absorption rate.

[0011] Optionally, obtaining the training set includes:

[0012] Using the soil-leaf-canopy coupling model SLC, pure vegetation simulation spectra and pure soil spectra are obtained:

[0013]

[0014] Among them, R l (λ) and T l (λ) represents the reflectivity and transmittance of the simulated blade at wavelength λ; N, C ab , C ar , C an , C bp , C wrel and C m represents leaf structure parameters, chlorophyll content, carotenoid content, anthocyanin content, brown pigment content, equivalent water thickness and dry matter content, ALA, hot and LAI represent the average leaf inclination angle, hot spot parameter and leaf area index, respectively. s ,θ v and φ are the solar zenith angle, observation zenith angle and relative azimuth angle between the sun and the sensor, respectively. c (λ, θ s ,θ v , φ) represents the canopy reflectance under specific observation conditions; R s (λ) is the simulated soil reflectivity, c1, c2, c3 and c sm They are three spectral vectors representing dry soil and one spectral vector representing wet soil;

[0015] The photosynthetically active radiation absorption rate is simulated using the soil-leaf-canopy coupling model SLC:

[0016]

[0017] Among them, FAPAR bs , FAPAR ws and FAPAR tol are the photosynthetically active radiation absorption rates of black sky, white sky and blue sky respectively. represents the canopy black sky spectral albedo at a specific wavelength and solar zenith angle, represents the black sky spectral albedo of soil; It indicates the transmittance of solar radiation directly penetrating into the soil, skyl is the proportion of scattered PAR energy, is a specific solar zenith angle The black sky photosynthetically active radiation absorption rate value under is the solar zenith angle The cosine value of is the solar zenith angle The sine value of ;

[0018] Acquiring the simulation data set according to the pure vegetation simulation spectrum, the pure soil spectrum and the photosynthetically active radiation absorption rate value;

[0019] The simulated data set is screened to obtain the training set.

[0020] Optionally, obtaining a simulation data set according to the pure vegetation simulation spectrum, the pure soil simulation spectrum and the photosynthetically active radiation absorption rate value includes:

[0021] Combined with the sensor spectral response function, the pure vegetation simulation spectrum and the pure soil simulation spectrum are simulated to the corresponding band reflectance to obtain pure vegetation reflectance data and pure soil reflectance data;

[0022] The pure vegetation reflectance data and the pure soil reflectance data are linearly combined to obtain mixed pixel simulated reflectance;

[0023] Using the photosynthetically active radiation absorption rate value and the pure vegetation coverage ratio, obtaining the photosynthetically active radiation absorption rate of the mixed pixel;

[0024] Based on the simulated reflectivity of the mixed pixel, Gaussian noise is added to obtain a simulated data set.

[0025] Optionally, the method of simulating the pure vegetation simulation spectrum and the pure soil simulation spectrum to the corresponding band reflectance is:

[0026]

[0027] in, Represents a specific spectral wave The vegetation canopy reflectance under the band; W min and W max They are the minimum wavelength and maximum wavelength within the spectral band respectively; Represents the difference between adjacent discrete wavelength points; Indicates wavelength The reflectivity at Indicates the sensor at wavelength The spectral response intensity at ;

[0028] Optionally, the method for obtaining the reflectance of the mixed pixel is:

[0029]

[0030] The method for obtaining the photosynthetically active radiation absorption rate of the mixed pixel is:

[0031]

[0032] in, Indicates that the wavelength is The reflectivity of the mixed pixel when It is the proportion of pure vegetation coverage in mixed pixels, and its value range is between 0 and 1; is the photosynthetically active radiation absorption rate value of the mixed pixel.

[0033] Optionally, construct the PreT LSTM The inversion model includes:

[0034] Determine the number of network layers and the number of hidden neurons in each layer, and obtain the initial PreT LSTM Inversion model;

[0035] Based on the initial PreT LSTM Inversion model, introduce boundary penalty term, make the preliminary inversion result conform to the corresponding physical law, and optimize the initial PreT LSTM The training parameters of the inversion model are obtained by LSTM Inversion model.

[0036] Optionally, build TL LSTM The inversion model includes:

[0037] The pre-trained long short-term memory network inversion model PreT LSTM As the source model, the reflectivity of Landsat satellite images and measured data were collated to generate the FAPAR migration training set;

[0038] By partially freezing the source model, retaining its learned features, and training only the remaining layers, re-adjusting the weights and biases of the model network nodes to obtain the final TL LSTM Inversion model.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] First, by combining the physical radiation transfer model to simulate the data and introducing the boundary penalty term in the loss function, the knowledge of the measured data set is transferred to the deep learning training process by combining the transfer learning method, thereby adding physical constraints. Compared with the traditional method of relying solely on data-driven modeling, this method can effectively introduce professional knowledge in the field of remote sensing and improve the scientificity and accuracy of the model. Secondly, the application of deep learning algorithms in the model greatly enhances the ability to fit complex nonlinear regression relationships and significantly improves the accuracy of FAPAR estimation. This method can better capture the complex ecological and biophysical processes in remote sensing data and improve the limitations of traditional methods in nonlinear regression. In addition, the high-resolution products generated based on this method have the potential to produce global long-term series FAPAR data, providing an important tool for monitoring and assessment of large-scale ecosystems. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0042] Figure 1 It is a flow chart of a method for inverting vegetation photosynthetically active radiation absorption rate based on multi-source data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] A new solution is proposed to address the problem that existing technologies lack sufficient physical constraints and have a coarse spatial resolution. By performing consistency correction on the input parameters of the physical radiation transfer model, combined with deep learning algorithms and transfer learning strategies, the algorithm proposed in this implementation case effectively meets the actual needs of FAPAR remote sensing estimation. This method combines the advantages of physical models in providing interpretability and physical feature extraction with the automatic feature learning capabilities of deep learning. With the help of transfer learning, it overcomes the heterogeneity of multi-source data under different environments and conditions and significantly improves the generalization ability of the model. In particular, the algorithm is optimized for Landsat satellite image data and can generate global long-term series FAPAR products with a spatial resolution of 30 meters, meeting the requirements of high-precision remote sensing estimation. Overall, this technology fills the gap in existing research and provides a more accurate FAPAR data inversion method, which can be used in research in the fields of terrestrial ecosystem dynamic monitoring and carbon cycle. It is of great significance in improving the accuracy of FAPAR remote sensing estimation and the level of practical application.

[0046] This embodiment proposes a method for inverting vegetation photosynthetically active radiation absorption rate based on multi-source data fusion, such as Figure 1 As shown, the specific steps include:

[0047] Obtain a modeling data set for inverting photosynthetically active radiation absorption rate;

[0048] Constructing a pre-trained long short-term memory network inversion model PreT LSTM , where the pre-trained long short-term memory network inversion model PreT LSTM It is obtained through training with a training set, which is modeling data generated based on the soil-leaf-canopy coupling model SLC;

[0049] The pre-trained long short-term memory network inversion model PreT LSTM As the source model, a migration training set is constructed through satellite data and measured data, and the source model is trained using the migration training set to obtain the migration training inversion model TL LSTM ;

[0050] Input the modeling dataset to the transfer training inversion model TL LSTM , obtain the estimated results of vegetation photosynthetically active radiation absorption rate.

[0051] Furthermore, obtaining a training set includes:

[0052] The soil-leaf-canopy coupling model (SLC) was used to obtain pure vegetation simulation spectra and pure soil spectra.

[0053] The photosynthetically active radiation absorption rate values ​​were simulated using the soil-leaf-canopy coupling model (SLC);

[0054] Linearly combine pure vegetation reflectance data and pure soil reflectance data to obtain mixed pixel simulated reflectance;

[0055] The photosynthetically active radiation absorption rate of mixed pixels is obtained by using the photosynthetically active radiation absorption rate value and the pure vegetation coverage ratio;

[0056] Filter the simulated data set to obtain the training set.

[0057] Specifically include: (1) Generate modeling data set based on soil-leaf-canopy coupling model (SLC):

[0058] 1) Define the Soil-Leaf-Canopy Model (SLC), which combines the leaf-scale PROSPECT model, the canopy-scale SAIL model, and the spectral vector model (GSV). The coupled model simulates the pure vegetation spectrum, pure soil spectrum, and photosynthetically active radiation absorption rate.

[0059]

[0060] Among them, R l (λ) and T l (λ) represents the reflectivity and transmittance of the simulated blade at wavelength λ; N, C ab , C ar , C an , C bp , C wrel and C m represents leaf structure parameters, chlorophyll content, carotenoid content, anthocyanin content, brown pigment content, equivalent water thickness and dry matter content, ALA, hot and LAI represent the average leaf inclination angle, hot spot parameter and leaf area index, respectively. s ,θ v and φ are the solar zenith angle, observation zenith angle and relative azimuth angle between the sun and the sensor, respectively. c (λ, θ s ,θ v , φ) represents the canopy reflectance under specific observation conditions; R s (λ) is the simulated soil reflectivity, c1, c2, c3 and c sm There are three spectral vectors representing dry soil and one spectral vector representing wet soil. bs , FAPAR ws and FAPAR tol are the photosynthetically active radiation absorption rates of black sky, white sky and blue sky respectively. represents the canopy black sky spectral albedo at a specific wavelength and solar zenith angle, represents the black sky spectral albedo of soil; It represents the transmittance of solar radiation directly penetrating into the soil, and skyl is the proportion of scattered PAR energy.

[0061] 2) Combined with the spectral response function of the Landsat sensor, the generated pure vegetation and pure soil simulated spectral data are fitted to the corresponding band reflectance.

[0062]

[0063] in, Represents a specific spectral wave The vegetation canopy reflectance under the band; W min and W max They are the minimum wavelength and maximum wavelength within the spectral band respectively; Represents the difference between adjacent discrete wavelength points; Indicates wavelength The reflectivity at Indicates the sensor at wavelength The spectral response intensity at ;

[0064] 3) Linearly combine vegetation and soil reflectance to obtain mixed pixel reflectance; linearly combine the photosynthetically active radiation absorption rate value obtained by coupling model simulation and the pure vegetation coverage ratio to obtain the photosynthetically active radiation absorption rate of the mixed pixel.

[0065]

[0066] in, Indicates that the wavelength is , the reflectivity of the mixed pixel; It refers to the FAPAR at the mixed pixel scale (due to the FAPAR at the mixed pixel scale bs , FAPAR ws and FAPAR tol The calculation method is the same, and the subsequent modeling process is consistent, so FAPAR is used uniformly for description).

[0067] Gaussian noise is added to the calculated band reflectivity to simulate the interference or error that the sensor may encounter in actual work, making the simulated remote sensing reflectivity data closer to the actual observation data.

[0068] 4) According to the positive correlation between FAPAR and normalized difference vegetation index in the modeling dataset, the simulation dataset was screened with the solar zenith angle at a fixed step size of 5° and the FAPAR at a fixed step size of 0.1 to ensure that the amount of simulated data in different combinations of solar zenith angle and FAPAR interval was the same, so as to make the feature distribution of the modeling dataset more uniform.

[0069] Furthermore, constructing PreT LSTM The inversion model includes:

[0070] Using the modeling dataset to build a pre-trained long short-term memory network inversion model PreT LSTM The model is used as the source model to organize the reflectivity of Landsat satellite images and measured data to generate the FAPAR training set for transfer learning;

[0071] By partially freezing some layers in the source model, retaining its learned features, training and optimizing only the remaining parts, and re-adjusting the weights and biases of the model network nodes to obtain the final TL LSTM Inversion model.

[0072] Specifically, (2) construct a pre-trained long short-term memory network inversion model PreT LSTM :

[0073] 1) The reflectance of the green, red, near-infrared, shortwave infrared-1, shortwave infrared-2 bands and the cosine value of the solar zenith angle are used as input features, and the FAPAR value of the simulated scene is used as the target variable. The data filtered in step (1) are used to generate a training set for model construction.

[0074] 2) Design a pre-trained long short-term memory network inversion model PreT LSTM The network structure of the model, including determining the number of network layers and the number of hidden neurons in each layer; defining the loss function to quantify the difference between the predicted value and the reference value, mainly using the mean square error (MSE) as the basic loss function, and introducing the boundary penalty term to ensure that the inversion result conforms to the corresponding physical laws; optimizing PreT LSTM Model training parameters, including number of iterations, batch size, optimizer type, and learning rate, to improve model performance and training efficiency.

[0075]

[0076] Where W is the weight matrix, including the weights of the forget gate, input gate, output gate, and candidate cell states, [h t−1 ,x t ] is the concatenation of the hidden state at the previous moment and the input at the current moment, b is the bias term, and The sigmoid and tanh functions generate the outputs of each gate and candidate cell states, respectively.

[0077] (11)

[0078] (12)

[0079] (13)

[0080] (14)

[0081] Where N represents the total number of data used for model validation; and Respectively represent the measured value and model predicted value of the i-th data; Represents the calculated value of the loss function under conventional training mode; Relu is a nonlinear function in deep learning; and is the penalty loss calculated using the boundary limits less than 0 and greater than 1 respectively; It is a combination of several loss expressions and is used as PreT LSTM The loss value of the model, The coefficient representing the additional penalty for outliers when the model predicts values ​​less than 0 and greater than 1.

[0082] 3) PreT based on the modeling dataset LSTM The model performs parameter fitting of network node weights and biases, and stores the trained model as a .pth file.

[0083] Furthermore, construct TL LSTM The inversion model includes:

[0084] The pre-trained long short-term memory network inversion model PreT LSTM As the source model, the reflectivity of Landsat satellite images and measured data were collated to generate the FAPAR migration training set;

[0085] By partially freezing the source model, retaining its learned features, and training only the remaining layers, re-adjusting the weights and biases of the model network nodes to obtain the final TL LSTM Inversion model.

[0086] Specifically, (3) construct a TL trained by transfer LSTM Model;

[0087] 1) The reflectance of green, red, near-infrared, shortwave infrared-1, shortwave infrared-2 and the cosine value of the solar zenith angle are used as input feature values, and the measured FAPAR from the global validation network is used as the target value. The reflectance of Landsat satellite images and the measured data are collated to generate the FAPAR migration training set.

[0088] 2) Define the same deep learning network structure as in step (2) and load the previously saved PreT LSTMThe model is used as the source model for transfer learning. During the transfer learning process, the source PreT LSTM Model consistent loss function, while adjusting TL LSTM Model training parameters, such as number of iterations, batch size, optimizer, and learning rate, to achieve better migration effects and model optimization.

[0089] 3) PreT based on transfer learning training set LSTM The model is partially frozen by freezing the lower layers in the source model, retaining the common features it has learned, and only training the remaining layers. LSTM The model will be saved in .pth file format.

[0090] (4) Packaging, integration and application of remote sensing FAPAR estimation model:

[0091] 1) Check the adjusted network weights and biases in the saved .pth file; manually define a network with the same structure and assign the obtained weights and biases to the network nodes; encapsulate and encrypt the manually defined network, and provide an interface for calling various inversion algorithms by attributes or methods.

[0092] 2) Taking FAPAR as the inversion target and considering the actual application effect of the model, TL LSTM The model is used as an inversion model to estimate vegetation FAPAR based on Landsat satellite optical image data. The FAPAR estimation results are saved as raster image data in GeoTIFF format for subsequent analysis.

[0093] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. The vegetation photosynthetically active radiation absorption rate inversion method based on multi-source data fusion is characterized by: include: Obtain a modeling data set for inverting photosynthetically active radiation absorption rate; Constructing a pre-trained long short-term memory network inversion model PreT LSTM , wherein the pre-trained long short-term memory network inversion model PreT LSTM The training set is obtained by training, wherein the training set is modeling data generated based on a soil-leaf-canopy coupling model SLC; Acquiring the training set includes: Using the soil-leaf-canopy coupling model SLC, pure vegetation simulation spectra and pure soil spectra are obtained: in, R l ( λ ) and T l (λ) represents the reflectivity and transmittance of the simulated blade at wavelength λ; N, C ab , C ar , C an , C bp , C wrel and C m represents leaf structure parameters, chlorophyll content, carotenoid content, anthocyanin content, brown pigment content, equivalent water thickness and dry matter content, ALA, hot and LAI represent the average leaf inclination angle, hot spot parameter and leaf area index, respectively. s ,θ v and φ are the solar zenith angle, observation zenith angle, and relative azimuth angle between the sun and the sensor, respectively. R c ( λ ,θ s ,θ v , φ) represents the canopy reflectance under target observation conditions; R s (λ) is the simulated soil reflectivity, c1, c2, c3 and c sm They are three spectral vectors representing dry soil and one spectral vector representing wet soil; The photosynthetically active radiation absorption rate is simulated using the soil-leaf-canopy coupling model SLC: Among them, FAPAR bs FAPAR ws and FAPAR tol are the photosynthetically active radiation absorption rates of black sky, white sky and blue sky respectively. represents the canopy dark sky spectral albedo at the target wavelength and solar zenith angle, represents the black sky spectral albedo of soil; It indicates the transmittance of solar radiation directly penetrating into the soil, skyl is the proportion of scattered PAR energy, is the target solar zenith angle The photosynthetically active radiation absorption rate of the black sky under is the solar zenith angle The cosine value of is the solar zenith angle The sine value of ; Acquire a simulation data set according to the pure vegetation simulation spectrum, the pure soil spectrum and the photosynthetically active radiation absorption rate value; Screening the simulated data set to obtain the training set; According to the pure vegetation simulation spectrum, the pure soil simulation spectrum and the photosynthetically active radiation absorption rate value, obtaining a simulation data set includes: Combined with the sensor spectral response function, the pure vegetation simulation spectrum and the pure soil simulation spectrum are simulated to the corresponding band reflectance to obtain pure vegetation reflectance data and pure soil reflectance data; The pure vegetation reflectance data and the pure soil reflectance data are linearly combined to obtain mixed pixel simulated reflectance; Using the photosynthetically active radiation absorption rate value and the pure vegetation coverage ratio, obtaining the photosynthetically active radiation absorption rate of the mixed pixel; Based on the simulated reflectivity of the mixed pixel, Gaussian noise is added to obtain a simulated data set; The method for simulating the pure vegetation simulation spectrum and the pure soil simulation spectrum to the corresponding band reflectance is: in, Represents the target spectrum wave The vegetation canopy reflectance under the band; W min and W max They are the minimum wavelength and maximum wavelength within the spectral band respectively; Represents the difference between adjacent discrete wavelength points; Indicates wavelength The reflectivity at Indicates the sensor at wavelength The spectral response intensity at ; The pre-trained long short-term memory network inversion model PreT LSTM As the source model, a migration training set is constructed through satellite data and measured data, and the source model is trained using the migration training set to obtain a migration training inversion model TL LSTM ; The modeling data set is input into the migration training inversion model TL LSTM , obtain the estimated results of vegetation photosynthetically active radiation absorption rate.

2. The method for inverting vegetation photosynthetically active radiation absorption rate based on multi-source data fusion according to claim 1, characterized in that: The method to obtain the reflectance of the mixed pixel is: The method for obtaining the photosynthetically active radiation absorption rate of the mixed pixel is: in, Indicates that the wavelength is The reflectivity of the mixed pixel when It is the proportion of pure vegetation coverage in mixed pixels, and its value range is between 0 and 1; is the photosynthetically active radiation absorption rate value of the mixed pixel.

3. The vegetation photosynthetically active radiation absorption rate inversion method based on multi-source data fusion according to claim 1 is characterized in that: Construct the pre-trained long short-term memory network inversion model PreT LSTM include: Determine the number of network layers and the number of hidden neurons in each layer, and obtain the initial PreT LSTM Inversion model; Based on the initial PreT LSTM Inversion model, introduce boundary penalty term, make the preliminary inversion result conform to the corresponding physical law, and optimize the initial PreT LSTM The training parameters of the inversion model are used to obtain the pre-trained long short-term memory network inversion model PreT LSTM .

4. The vegetation photosynthetically active radiation absorption rate inversion method based on multi-source data fusion according to claim 3 is characterized in that: Building TL LSTM The inversion model includes: The pre-trained long short-term memory network inversion model PreT LSTM As the source model, the reflectivity of Landsat satellite images and measured data were collated to generate the FAPAR migration training set; By partially freezing the source model, retaining its learned features, and training only the remaining layers, re-adjusting the weights and biases of the model network nodes to obtain the final TL LSTM Inversion model.

Images