Remote Sensing-Based Vegetation Leaf Parameter Inversion Method and Device

Abstract

The present invention provides a remote sensing-based vegetation leaf parameter inversion device, which is obtained based on the inverse process of expressing the PROSAIL model using a deep neural network, and has not only a strong physical mechanism, but also high accuracy. The remote sensing-based vegetation leaf parameter inversion method provided in the present application uses the remote sensing-based vegetation leaf parameter inversion device provided in the present application to invert the remote sensing data of the vegetation leaves to obtain the parameters of the vegetation leaves, which has high credibility and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing, and more specifically, to a remote sensing-based vegetation leaf parameter inversion method and device. Background Art

[0002] Leaf parameters are important indicators for characterizing the growth status of vegetation. Quantitative acquisition of leaf parameters is an important area of ​​precision agriculture research and is of great significance for supporting research such as carbon cycle.

[0003] At present, remote sensing applications based on deep learning are becoming more and more extensive, especially in the problem of nonlinear parameter inversion. Deep neural network models can find the intrinsic connection between model input parameters and output parameters through continuous learning and training, and realize parameter inversion.

[0004] However, there is still room for improvement in the accuracy of traditional blade parameter inversion methods based on deep learning. Summary of the invention

[0005] The present invention proposes a vegetation leaf parameter inversion method and device based on remote sensing, aiming to improve the accuracy of the leaf parameter inversion method based on deep learning.

[0006] In order to achieve the above objectives, the proposed solution is as follows:

[0007] A vegetation leaf parameter inversion device based on remote sensing, comprising:

[0008] Input module, SAIL-Net subnetwork, PROSPECT-Net subnetwork, and output module;

[0009] The input module is used to input the remote sensing data of vegetation leaves into the SAIL-Net sub-network;

[0010] The SAIL-Net subnetwork is used to acquire and output the reflectivity and transmittance of the vegetation leaves based on the remote sensing data;

[0011] The PROSPECT-Net sub-network is used to obtain parameter values ​​of the vegetation leaves based on the reflectivity and transmittance;

[0012] The output module is used to output the parameter values ​​of the vegetation leaves.

[0013] Optionally, the SAIL-Net sub-network includes:

[0014] The first module is used to implement the inverse process of solving the reflectivity by the SAIL model based on the remote sensing data;

[0015] The second module is used to implement the SAIL model to solve the bidirectional reflection parameter rso , and the directional reflection parameter r of the diffuse reflection do The reverse process of

[0016] The third module is used to implement the inverse process of solving the first coefficient of the SAIL model;

[0017] The fourth module is used to implement the inverse process of solving the extinction and scattering coefficients of the SAIL model;

[0018] The fifth module is used to implement the inverse process of solving the singularity point of the SAIL model;

[0019] The sixth module is used to implement the inverse process of solving hot spots by the SAIL model;

[0020] The seventh module is used to implement the inverse process of solving the reflectivity and transmittance of the blade using the SAIL model.

[0021] Optionally, the first module consists of a convolutional layer and a ReLU layer;

[0022] The second module is composed of a transposed convolutional layer and a ReLU layer;

[0023] The third module is composed of a convolutional layer and a ReLU layer;

[0024] The fourth module is composed of a convolutional layer and a ReLU layer;

[0025] The fifth module is composed of a convolutional layer and a ReLU layer;

[0026] The sixth module is composed of a convolutional layer and a ReLU layer;

[0027] The seventh module is composed of a maximum pooling layer, a convolutional layer, and a ReLU layer.

[0028] Optionally, the output data of the first module is used as the input data of the second module;

[0029] The output data of the second module is used as the input data of the third module and the fifth module;

[0030] The output data of the third module is used as the input data of the fourth module and the sixth module.

[0031] Optionally, the SAIL-Net sub-network further includes:

[0032] The splicing module is used to splice the output data of the fourth module, the fifth module and the sixth module, and input the spliced ​​data into the seventh module.

[0033] Optionally, the PROSPECT-Net sub-network includes:

[0034] The eighth module is used to implement the inverse process of the PROSPECT model to solve the leaf transmittance and refractive index when N≠1;

[0035] The ninth module is used to implement the PROSPECT model to solve the leaf transmittance ρ when N=1 α and the refractive index τ α The reverse process of

[0036] The tenth module is used to implement the inverse process of solving the transmission coefficient θ by the PROSPECT model;

[0037] The eleventh module is used to implement the PROSPECT model to solve the blade parameters N, C m , C w and C ab The reverse process of .

[0038] Optionally, the eighth module, the ninth module, the tenth module and the eleventh module are all composed of a fully connected layer and a LeakReLU layer.

[0039] Optionally, the SAIL-Net sub-network and the PROSPECT-Net sub-network constitute a PROSAlL-Net network;

[0040] The training process of the PROSAlL-Net network includes:

[0041] The PROSAlL-Net network is trained using forward propagation and back propagation.

[0042] A vegetation leaf parameter inversion method based on remote sensing, comprising:

[0043] Obtain remote sensing data of vegetation leaves;

[0044] Based on the PROSAIL-Net network and the remote sensing data, the inversion parameters of the vegetation leaves are obtained, and the PROSAIL-Net network is composed of the SAIL-Net subnetwork described in any one of claims 1 to 8 and the PROSPECT-Net subnetwork.

[0045] An electronic device comprising: a memory and a processor;

[0046] The memory is used to store programs;

[0047] The processor is used to execute the program to implement the above-mentioned vegetation leaf parameter inversion method based on remote sensing.

[0048] A readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the above-mentioned vegetation leaf parameter inversion method based on remote sensing is implemented.

[0049] The remote sensing-based vegetation leaf parameter inversion device provided in the present application is obtained based on the inverse process of expressing the PROSAIL model using a deep neural network, and not only has a strong physical mechanism, but also has high accuracy. The remote sensing-based vegetation leaf parameter inversion method provided in the present application uses the remote sensing-based vegetation leaf parameter inversion device provided in the present application to invert the remote sensing data of vegetation leaves to obtain the parameters of vegetation leaves, which has high credibility and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0051] Figure 1 A structural example diagram of a PROSAIL-Net network provided in an embodiment of the present invention;

[0052] Figure 2 This is an example diagram of the relationship between the loss function and the number of iterations based on the training set and validation set loss of the SAIL-Net sub-network in the PROSAIL-Net network;

[0053] Figure 3 This is an example diagram of the relationship between the loss function and the number of iterations of the PROSPECT-Net sub-network in the PROSAIL-Net network based on the training set and the validation set loss;

[0054] Figure 4 The flowchart of training the PROSAIL-Net network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] Deep neural networks can find the intrinsic connection between input and output through learning and training based on the sample data provided.

[0057] During the research, the inventors found that the use of deep neural networks in the blade parameter inversion problem has the following problems:

[0058] Deep neural networks (equivalent to "black boxes") lack physical mechanisms and have poor interpretability. Therefore, it is difficult to improve the accuracy of the inversion results of deep neural networks for blade parameters.

[0059] In order to solve the above problems, the remote sensing-based vegetation leaf parameter inversion method provided in the embodiment of the present application reconstructs the deep neural network based on the physical mechanism of leaf parameter inversion, so that the structure of the deep neural network conforms to the physical logic of leaf parameter inversion. Therefore, not only the inversion results of the deep neural network are interpretable, but also have higher accuracy.

[0060] The following is a detailed description of the process of constructing a deep neural network based on the physical mechanism of blade parameter inversion.

[0061] The role of the PROSAIL model is to map the parameters of vegetation leaves into the spectrum of vegetation leaves. The PROSAIL model is a coupling model of the leaf optical property model PROSPECT and the canopy bidirectional reflectance model SAIL.

[0062] The PROSPECT model simulates the reflectance and transmittance of plant leaves between 400nm and 2500nm. The simplified formula of the PROSPECT model is as follows:

[0063]

[0064] Where α is the maximum incident angle of the solid angle, n is the refractive index, and t αv (α,n) is the transmittance of the dielectric planar surface.

[0065] According to the physical and mathematical process of the PROSPECT model, the PROSPECT model can be divided into three modules. The first module solves the transmission coefficient θ, and the second module solves the leaf transmittance ρ when N = 1. α and the refractive index τ α The third module solves the leaf transmittance R when N≠1 N,α and refractive index T N,α .

[0066] The SAIL model is based on solving four radiation fluxes E + , E - , E s , E o The SAIL model simulates the bidirectional reflectance of the vegetation canopy in turbid media by the scattering and absorption of the turbid medium. The differential equation shown in formula (2) is used to calculate E +, E - , E s , E o To solve:

[0067]

[0068] Among them, E s is the direct radiation flux density transmitted from top to bottom, L is the leaf area index LAI, and k is the extinction coefficient in the direction of the sun. - is the downward diffused irradiance, s' is the scattering coefficient of the back-direct radiation, a is the extinction coefficient, σ is the backscattering coefficient, E + is the upward diffuse irradiance. s is the scattering coefficient of direct radiation in the same direction. E o is the flux density in the observation direction, w is the upward bidirectional scattering coefficient, v is the backscattering coefficient in the direction of the scattered incident, v` is the forward scattering coefficient in the direction of the scattered incident, K is the extinction coefficient in the observation direction, and x is the relative optical height, ranging from -1 at the bottom to 0 at the top.

[0069] The process of solving equation (2) can be divided into 6 steps:

[0070] The first step is to solve the extinction and scattering coefficients k, s', a, σ, w, v, v', K. Solving differential equations may lead to singularity problems, resulting in numerical errors, so the second step is to solve the singularity. In order to avoid hot spot effects, the third step is to solve the hot spots. The fourth step is to solve ρ so , τ sd The parameters are functions of LAI, extinction coefficient and scattering coefficient, ρ so is the bidirectional reflectivity of a layer, τ sd is the diffuse reflection transmittance of the mirror incident. Some parameters are affected by the singular point problem. The fifth step is to solve r so 、r do , where r so Affected by hot spots, so is the bidirectional reflection parameter, r do is the directional reflection parameter of diffuse reflection. The sixth step is to solve the reflectivity R.

[0071] Based on the function of the PROSAIL model, it can be known that vegetation leaf inversion is the inverse process of the PROSAIL model function. However, because the inverse processes of the PROSPECT model and the SAIL model are not analytic, in the embodiments of the present application, a deep neural network is used to replace the inverse processes of the two.

[0072] The idea of ​​PROSAIL-Net, a deep neural network based on the inverse process simulation of the PROSAIL model, is to pass a 1×244 spectrum into the network that replaces the inverse process of the SAIL model to obtain the reflectivity and transmittance of the blade, and then pass the reflectivity and transmittance of the blade into the network that replaces the inverse process of the PROSPECT model to obtain the inversion value of the blade parameters.

[0073] The following describes the network that replaces the inverse process of the SAIL model and the design that replaces the PROSPECT model.

[0074] The inverse module of the SAIL model, that is, the network corresponding to the inverse process of the SAIL model, implements the inverse process of the six-step functions of the above SAIL model:

[0075] 1. The first module implements the inverse process of the SAIL model to solve the reflectivity R.

[0076] The reflectivity formula is shown in formula (3):

[0077]

[0078] E o is the flux density in the observation direction, E sun is the solar irradiance on the ground, E sky is the sky irradiance on the ground, SKYL represents the proportion of diffuse sky light in the solar incidence (known value), r so is the bidirectional reflectivity, r do is the directional reflectivity when diffuse scattering occurs.

[0079] It can be understood that the first module implements the inverse process of formula (3): the remote sensing data of vegetation leaves include reflectivity, and when SKYL is known, the bidirectional reflectivity and the directional reflectivity under diffuse scattering incidence are obtained.

[0080] Based on the functions of the convolution layer and the ReLU layer and because the inverse process of the reflectance module is a nonlinear process, combined with formula (3), the convolution layer can be used to extract the characteristic image of the spectrum, and the ReLU layer expresses nonlinearity. Therefore, this step uses the convolution plus ReLU function to express. In order to fully express each spectrum, the convolution size should not be too large. In some implementations, the convolution size is set to 1×3, and the number is 8. The first module is Figure 1 The 11 in it indicates.

[0081] The first module mainly extracts the main features of the spectrum. The convolution has the characteristics of incomplete connection and "parameter sharing", which greatly reduces the network parameters, ensures the sparsity of the network, prevents overfitting, and obtains 8 1×242 spectral features.

[0082] 2. The second module implements the SAIL model to solve rso , and r do The reverse process of .

[0083] SAIL model to solve the bidirectional reflection parameter r so , and the directional reflection parameter r of the diffuse reflection do The formula based on (i.e. the positive process) is:

[0084]

[0085] ρ so is the bidirectional reflectivity of a layer, ρ sd is the diffuse reflection of a layer of specular incident light, ρ do is the diffuse directional reflectivity of a layer, ρ dd is the diffuse reflectivity of a layer, τ sd is the diffuse reflection transmittance of specular incident light, τ ssoo is the bidirectional direct transmittance, τ ss is the direct transmittance from the sun, τ oo is the direct transmittance in the viewing direction, τ do is the directional transmittance r for diffuse incidence so is bidirectional reflection, r do Directional reflection for diffuse reflection.

[0086] According to the bidirectional reflection parameter r obtained in the first module so , and the directional reflection parameter r of the diffuse reflection do , the second module obtains the parameters on the right side of the equal sign in formula (4). For the convenience of description, the parameters on the right side of the equal sign in formula (4) are called the first coefficients.

[0087] The second module uses deconvolution (i.e. transposed convolution) and ReLU functions to express. The second module consists of transposed convolution layers and ReLU layers. The transposed convolution layer selects an even convolution size. In order to avoid the checkerboard phenomenon, the convolution size is set to 4×4, the step size is 2, and the number is 16. The second module expands the spectral feature size, especially the width, to facilitate subsequent operations and processing, and obtains 16 4×486 spectral features. The second module is based on Figure 1 The 12 in it means.

[0088] 3. The third module implements the SAIL model to solve ρ so , sd , do , dd , τ sd , τ ssoo , τ ss , τ oo , τ doThe inverse process of the equation coefficients. The equation coefficients are obtained by the second module, and the third module obtains the second coefficients based on the first coefficients.

[0089] Because the third module needs to solve many formulas and they are highly nonlinear, we choose convolution plus ReLU function to express it. That is, the third module consists of convolution layer and ReLU layer. The convolution size is 3×3, the padding is 1, and the number is 16. The third module mainly extracts spectral features. The purpose of padding is to make the input and output have the same height and width. It also increases the use of edge feature information and can fully increase the receptive field. 16 4×486 spectral features are obtained. Figure 1 The 13 in it indicates.

[0090] 4. The fourth module implements the inverse process of the SAIL model to solve the coefficients k, s', a, σ, w, v, v', K. The fourth module is also composed of convolutional layers and ReLU layers. The convolution kernel size is 3×3, the padding is 1, and the step size is (1, 3). The fourth module further extracts spectral features and reduces the size. In preparation for subsequent splicing, 4 2×162 spectral features are obtained. The fourth module is based on Figure 1 The 14 in it indicates.

[0091] 5. The fifth module implements the inverse process of the SAIL model to solve the singular point problem. The fifth module consists of a convolution layer and a ReLU layer. The convolution size is 3×3, the step size is (1, 3), and the number is 4. The purpose is to extract spectral features. Four 2×162 spectral features are obtained. Figure 1 The 15 in the figure indicates that the input of 15 is the output of 12, and the output of 12 is also the input of 13.

[0092] 6. The sixth module implements the inverse process of the SAIL model to solve hot spots. The sixth module is also composed of convolutional layers and RELU layers. The convolution size is 5×5, the step size is (1, 3), and the number is 4. The purpose is to extract spectral features. The 5×5 convolution increases the receptive field and obtains 4 2×162 spectral features. Figure 1 16 in it represents. The input of 16 is the output of 13, and the output of 13 is also the input of 14. 7. The seventh module implements the inverse process of the SAIL model to solve the reflectance and transmittance of the leaves. The seventh module selects the maximum pooling layer, convolution layer, and ReLU function to express the inverse process. The maximum pooling size is 2×2, and the padding is 1. This step mainly extracts the most important features and reduces the dimension of the spectrum; the convolution size is 1×3, the padding is 1, and the number is 1. The purpose is to keep the number of channels and width unchanged, and a 2×244 spectral feature is obtained, that is, the reflectance and transmittance of the leaves. The network constructed in this step is based on Figure 1 17 in represents that the input of 17 is the concatenated features of the features output by 14, 15 and 16.

[0093] It should be noted that, because the SAIL model actually has more than two input parameters, namely, blade reflectivity and transmittance, a selection is required, i.e., executing the seventh step.

[0094] The first step of the inverse process of the PROSPECT model is to solve the leaf transmittance R when N≠1 N,α and refractive index T N,α The reverse process of , its forward process formula is:

[0095]

[0096]

[0097] That is, the first module of the PROSPECT-Net subnetwork (called the eighth module in the above order) consists of a fully connected layer and a LeakReLU layer, with 128 LeakReLU nodes. The LeakReLU activation function is added to make it a nonlinear structure. Since the network structure is very deep, the LeakReLU activation function is used to avoid gradient disappearance.

[0098] The second step of the inverse process of the PROSPECT model is to solve the leaf transmittance ρ when N = 1 α and the refractive index τ α The reverse process of is:

[0099]

[0100]

[0101] That is, the second module of the PROSPECT-Net subnetwork (called the ninth module in the above order) is also composed of a fully connected layer and a LeakReLU layer, with 64 LeakReLU nodes.

[0102] The third step of the inverse process of the PROSPECT model is to solve the transmission coefficient θ. The forward process formula is:

[0103]

[0104] k(λ)=∑K i (λ)C i (10)

[0105] Where λ is the wavelength, K i (λ) is the spectral specific absorption coefficient relative to the leaf biochemical parameter i, C i is the content of leaf biochemical parameter i per unit leaf area.

[0106] That is, the third module of the PROSPECT-Net sub-network (called the tenth module in the above order) is also composed of a fully connected layer and a LeakReLU layer, with 16 nodes.

[0107] The fourth step of the inverse process of the PROSPECT model is to solve the blade parameters N, C m , C w and C ab The inverse process of the PROSPECT-Net sub-network, that is, the fourth module (the eleventh module in the above order) is composed of a fully connected layer and a LeakReLU layer.

[0108] The modules of PROSPECT-Net sub-network are Figure 1 Among them, 18, 19, 20 and 21. Figure 1 The output module output in outputs 4 values ​​of leaf parameters.

[0109] In summary, after using deep neural networks to express the inverse process of the PROSAIL model, we get Figure 1 The deep neural network PROSAlL-Net shown.

[0110] According to the function, Figure 1 The network deep neural network PROSAlL-Net shown is divided into two sub-networks: SAIL-Net sub-network and PROSPECT-Net sub-network.

[0111] The inventor discovered during the research that Figure 1 The SAIL-Net sub-network shown can invert the blade reflectivity and transmittance with high accuracy, and the PROSPECT-Net sub-network can invert the blade parameters with high accuracy.

[0112] The above conclusions will be proved as follows:

[0113] First, the PROSAIL model was used to simulate 50,000 data points, and the spectral reflectance, blade reflectance, blade transmittance, and blade parameters N, C were saved respectively. ab , C w and C m Then divide it into training set, validation set and test set according to the ratio of 60%, 20% and 20%. Then pass the corresponding training set and validation set samples into SAIL-Net network and PROSPECT-Net network to train the two networks, and then pass the corresponding test samples into the trained network. Finally, get the corresponding network inversion value and calculate the R with its true value 2 and root mean square error RMSE.

[0114] During the training of the SAIL-Net subnetwork, it was found that the accuracy of the loss function of the training set and the validation set reached 10 -3 And gradually decreases and converges, such as Figure 2 As shown, it can be seen that the SAIL-Net sub-network is stable. By training the SAIL-Net sub-network, the reflectivity and transmittance of the blade are inverted. The R 2 is 0.9489, RMSE is 0.0380, and the leaf transmittance R 2 The value is 0.9473, and the RMSE is 0.0387. It can be seen that the SAIL-Net network can invert the reflectivity and transmittance of the leaves with high accuracy, so the design of the SAIL-Net subnetwork is reasonable.

[0115] In the process of training the PROSPECT-Net sub-network, it is also found that the loss function accuracy of the training set and the validation set reaches 10 -7 is gradually reduced and converged, such as Figure 3 As shown, by training the PROSPECT-Net sub-network, the blade parameter C is inverted ab , N, C m and C w , its R 2 Both reached 0.99, and the RMSE were 0.053 and 9.3×10 -4 , 3×10 -5 , 2×10 -5 ,It can be seen that the PROSPECT-Net network can invert the values ​​of blade parameters 4 with high precision, so the design of the PROSPECT-Net subnetwork is reasonable.

[0116] The principle of training is to minimize the loss function through forward and backward propagation. Since the blade parameter inversion problem is a regression problem, the mean square error loss function is used to estimate the error, see formula (11). The method used is the gradient descent method, which continuously updates the parameters through the forward and reverse processes to make the loss function reach the global minimum.

[0117]

[0118] where y i is the label value of the i-th leaf parameter, y i ' is the predicted value of the network, n is the number of blade parameters to be solved, and in this example it is equal to 4.

[0119] Before executing the training step, the data is preprocessed first. A certain amount of samples are randomly selected from the total sample as training samples, verification samples, and test samples. The three types of sample data are not repeated, and the proportions are 60%, 20%, and 20% respectively. The samples specifically include the spectral reflectance as the input value and the corresponding label value: N, C m , C w and C ab Data. The validation sample is used to adjust the network's hyperparameters, prevent or fit, and provide the ability to perform a preliminary evaluation of the model. In order to facilitate convergence, the leaf parameter 4 values ​​were normalized.

[0120] like Figure 4 As shown, the training process includes the following steps:

[0121] S11, after initializing the weights of the PROSAIL-Net network, the spectral reflectances of the training samples and the verification samples are transmitted to the PROSAIL-Net network in batches. That is, the training samples and the verification samples are input to the PROSAIL-Net network.

[0122] S12. The blade parameters are calculated by the PROSAIL-Net network. Here, the values ​​calculated by the PROSAIL-Net network are called the target values ​​of the blade parameters. Forward propagation is based on formula (12):

[0123] y=σ(xW+B) (12)

[0124] y is the output leaf parameter, which is a four-dimensional vector, x is the input spectrum number, where W and b are the weights and bias terms of the convolutional layer and the fully connected layer, and σ(x) is the activation function.

[0125] S13. Use the mean loss function to solve the error MSE between the target value and the label value.

[0126] S14: Determine whether the MSE is within the allowable range. If yes, execute S15; if no, execute S16.

[0127] S15. End the training and fix the values ​​of parameters such as weights and bias terms.

[0128] At this point, the forward propagation process is completed.

[0129] The back propagation process includes the following steps:

[0130] S16. Calculate the error of each layer in the network.

[0131] S17. Calculate the error gradient based on the errors of each layer.

[0132] In this embodiment, the Aadm algorithm is used to calculate the gradient, and its parameters are the default values ​​under the Pytorch framework. The network hyperparameters are set to batch size batchsize = 100, iteration number epoch = 10000, initial learning rate lr = 0.001. Since a large learning rate will cause the model to miss the optimal solution, the learning rate decay rate is set. In the first 1000 iterations, lr = lr × 0.6 for every 100 iterations, and after 1000, lr = lr × 0.6 for every 500 iterations. If lr < 10 -6 , it will no longer decrease. Since the data is input in batches, the mean of the mean square error will be used, as shown in formula (11).

[0133] S18. Update weights based on gradients.

[0134] The weight update formulas are as follows (13) and (14):

[0135]

[0136]

[0137] in,

[0138] After the above steps, one forward and one backward propagation are completed. Through such forward propagation and backward propagation, the network parameters, weight values, and bias terms are updated. Each iteration is updated once, and one iteration is completed. Through continuous iteration, the loss function is continuously reduced and finally convergence is achieved.

[0139] The trained PROSAIL-Net network can be used to invert leaf parameters of vegetation:

[0140] The remote sensing data of vegetation leaves are input into the PROSAIL-Net network, and the parameter values ​​of vegetation leaves output by the PROSAIL-Net network are obtained.

[0141] The embodiment of the present application further provides an electronic device, the electronic device comprising: a memory and a processor. The memory is used to store a program. The processor is used to execute the program to implement a vegetation leaf parameter inversion method based on remote sensing.

[0142] An embodiment of the present application also provides a readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, a vegetation leaf parameter inversion method based on remote sensing is implemented.

[0143] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0144] In this article, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0145] Each embodiment in this specification focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referenced to each other, and the features recorded in each embodiment in this specification can be replaced or combined with each other.

[0146] The above description of the embodiments disclosed in the present invention enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A vegetation leaf parameter inversion device based on remote sensing, characterized in that: include: Input module, SAIL-Net subnetwork, PROSPECT-Net subnetwork, and output module; The input module is used to input the remote sensing data of vegetation leaves into the SAIL-Net sub-network; The SAIL-Net subnetwork is used to acquire and output the reflectivity and transmittance of the vegetation leaves based on the remote sensing data; The PROSPECT-Net sub-network is used to obtain parameter values ​​of the vegetation leaves based on the reflectivity and transmittance; The output module is used to output the parameter value of the vegetation leaf; The SAIL-Net sub-network includes: The first module is used to implement the inverse process of solving the reflectivity by the SAIL model based on the remote sensing data; The second module is used to implement the SAIL model to solve the bidirectional reflection parameter r so , and the directional reflection parameter r of the diffuse reflection do The reverse process of The third module is used to implement the inverse process of solving the first coefficient of the SAIL model; The fourth module is used to implement the inverse process of solving the extinction and scattering coefficients of the SAIL model; The fifth module is used to implement the inverse process of solving the singularity point of the SAIL model; The sixth module is used to implement the inverse process of solving hot spots by the SAIL model; The seventh module is used to implement the inverse process of solving the reflectivity and transmittance of the blade by the SAIL model; The PROSPECT-Net sub-network includes: The eighth module is used to implement the PROSPECT model to solve the leaf transmittance R when N≠1 N,α and refractive index T Nα The reverse process of , its forward process formula is: The ninth module is used to implement the PROSPECT model to solve the leaf transmittance ρ when N=1 α and the refractive index τ α The reverse process of is: Where α is the maximum incident angle of the solid angle, n is the refractive index, and t αv (α,n) is the average transmittance within a solid angle of α; The tenth module is used to implement the inverse process of solving the transmission coefficient θ of the PROSPECT model, and its forward process formula is: k(λ)=∑K i (λ)C i ; where λ is the wavelength, K i (λ) is the spectral specific absorption coefficient relative to the leaf biochemical parameter i, C i is the content of leaf biochemical parameter i per unit leaf area, k is the extinction coefficient in the direction of the sun; The eleventh module is used to implement the PROSPECT model to solve the blade parameters N, C m , C w and C ab The reverse process of .

2. The device according to claim 1, characterized in that The first module consists of a convolutional layer and a ReLU layer; The second module is composed of a transposed convolutional layer and a ReLU layer; The third module is composed of a convolutional layer and a ReLU layer; The fourth module is composed of a convolutional layer and a ReLU layer; The fifth module is composed of a convolutional layer and a ReLU layer; The sixth module is composed of a convolutional layer and a ReLU layer; The seventh module is composed of a maximum pooling layer, a convolutional layer, and a ReLU layer.

3. The device according to claim 2, characterized in that The output data of the first module is used as the input data of the second module; The output data of the second module is used as the input data of the third module and the fifth module; The output data of the third module is used as the input data of the fourth module and the sixth module.

4. The device according to claim 3, characterized in that The SAIL-Net sub-network also includes: The splicing module is used to splice the output data of the fourth module, the fifth module and the sixth module, and input the spliced ​​data into the seventh module.

5. The device according to claim 1, characterized in that The eighth module, the ninth module, the tenth module and the eleventh module are all composed of a fully connected layer and a LeakReLU layer.

6. The device according to claim 1, characterized in that The SAIL-Net sub-network and the PROSPECT-Net sub-network constitute a PROSAlL-Net network; The training process of the PROSAlL-Net network includes: The PROSAlL-Net network is trained using forward propagation and back propagation.

7. A vegetation leaf parameter inversion method based on remote sensing, characterized in that: include: Obtain remote sensing data of vegetation leaves; Based on the PROSAlL-Net network and the remote sensing data, the inversion parameters of the vegetation leaves are obtained, and the PROSAlL-Net network is implemented by the remote sensing-based vegetation leaf parameter inversion device described in any one of claims 1-6.

8. An electronic device, characterized in that: include: Memory and processor; The memory is used to store programs; The processor is used to execute the program to implement the vegetation leaf parameter inversion method based on remote sensing as described in claim 7.

9. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the remote sensing-based vegetation leaf parameter inversion method as claimed in claim 7 is implemented.

Images