High spatial resolution vegetation leaf area index estimation method based on fusion of multi-source data

Abstract

The present invention belongs to the field of remote sensing technology and relates to a method for estimating high-spatial-resolution vegetation leaf area index by fusing multi-source data. The method includes generating a modeling data set based on a physical radiation transfer model, constructing an LSTM-PG eff model, constructing a migrated-trained LSTM-TL eff model and an LSTM-TL true model, and LAI encapsulation, conversion, splicing and saving of the remote sensing estimation model. By using the radiation transfer model to generate simulation data and applying the knowledge of the migrated measured data set to deep learning training to increase physical constraints, the present invention improves LAI the fitting ability of the remote sensing estimation model for complex non-linear regression relationships, effectively improving the model accuracy; the generated LAI products have high spatial resolution, reducing the uncertainty caused by the mixed pixel problem in the coarse spatial resolution algorithm, and greatly ensuring LAI the accuracy of remote sensing estimation.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing technology, and in particular relates to a high spatial resolution vegetation leaf area index estimation method integrating multi-source data. Background Art

[0002] Leaf area index (abbreviated as ) is defined as half of the total plant leaf area per unit surface area. It is not only an important parameter for describing the structure and function of vegetation canopy, but also an important input parameter for many ecological, hydrological and climate models. It has also been listed as an important climate variable in the field of global climate change research. The data has important practical significance. Compared with ground measurement methods, remote sensing technology has the ability to observe synchronously and continuously over a large range. Remote sensing estimation methods are mainly divided into data-driven empirical methods, model-driven physical methods and hybrid methods that combine the two.

[0003] The empirical method is simple to calculate and can establish the reflectivity or vegetation index and The relationship between Fast inversion. However, since physical rules are not taken into account, the universality is insufficient. Physical methods can be applied to any location in theory. First, a simulated data set is generated based on a remote sensing physical model, and then a simulated data set is constructed from the simulated data set by iterative optimization or lookup table. Estimation model. These two inversion methods are not only time-consuming but also have great limitations in algorithm accuracy. The hybrid method combines the advantages of empirical methods and physical methods. It not only takes into account the theoretical basis of the physical model, but also utilizes the characteristics of artificial intelligence algorithms, which are simple, efficient, accurate and practical.

[0004] Affected by bad weather, inversion algorithms and sensor equipment, different global scales There are obvious differences between remote sensing products, and there are still problems such as low quantitative verification accuracy and large product uncertainty. The spatial resolution of existing products is low, and the existing algorithms generate When the product is applied to heterogeneous surfaces, there will be serious mixed pixel problems, which makes it 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 images with high spatial resolution. Products are available for use.

[0005] In addition, the prior art The estimation model only relies on the modeling data set and does not take into account the constraints of remote sensing physical knowledge. The estimation model is oriented towards crop modeling and cannot be applied to other vegetation types; The estimation model is based on Sentinel-2 data and it is difficult to produce global long-term series products. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a high spatial resolution vegetation leaf area index estimation method integrating multi-source data, including generating a modeling data set based on a physical radiation transfer model, constructing an LSTM-PG eff Model, build transfer-trained LSTM-TL eff Model and build transfer-trained LSTM-TL true Model, and Encapsulation, conversion, splicing and preservation of remote sensing estimation models;

[0007] Generate a modeling data set based on the physical radiation transfer model, including: obtain spectral simulation data of leaves, canopy and soil based on the physical radiation transfer model, and generate pure vegetation spectral simulation data and pure soil spectral simulation data; fit the generated pure vegetation spectral simulation data and pure soil spectral simulation data to the Landsat satellite band to generate mixed pixel simulation reflectance; screen the data to obtain the modeling data set;

[0008] Building LSTM-PG eff Model, including: Building LSTM-PG eff Model training data set, set LSTM-PG eff The network structure of the model, defines the loss function of the network parameter fitting process, and uses LSTM-PG according to the loss function eff The training data set of the model is LSTM-PG eff The model performs parameter fitting of network node weights and biases to obtain the trained LSTM-PG eff Model;

[0009] Building a transfer-trained LSTM-TL eff Model, including: collating Landsat satellite image reflectivity and measured data to generate effective Migration training set, defined with trained LSTM-PG eff The network structure of the model is the same as the source model, and the node defines the loss function of the network parameter fitting process. The migration training set is used to fine-tune the parameters of the network node weights and biases of the source model to obtain the migration-trained LSTM-TL eff Model;

[0010] Building a transfer-trained LSTM-TL true Model, including: collating Landsat satellite image reflectivity and measured data to generate real Migration training set, defined with trained LSTM-TL eff The network structure of the model is the same as the source model. The node defines the loss function of the network parameter fitting process. The LSTM-TL true The model fine-tunes the parameters of network node weights and biases to obtain the LSTM-TL model after migration training. true Model;

[0011] For the transfer-trained LSTM-TL true The model is packaged and saved and uses the LSTM-TL trained after migration true Model Estimate.

[0012] Furthermore, the spectrum simulation data of leaves, canopy and soil are obtained based on the physical radiation transfer model, including: defining the input parameters of the radiation transfer model and constraining the consistency between the parameters to generate the spectrum simulation data of leaves, canopy and soil; setting represents randomly generated values ​​of the input parameters of the radiative transfer model, represents the input parameter value after the consistency constraint between parameters, represents the leaf area index, Indicates when =0 when the maximum value of the input parameter is Indicates when =0 is the minimum value of the input parameter. Indicates when The minimum value of the input parameter when the maximum value is obtained. Indicates when The maximum value of the input parameter when the maximum value is obtained. Indicates when The minimum value of the input parameter when obtaining a fixed value, Indicates when The maximum value of the input parameter when obtaining a fixed value, then:

[0013] ;

[0014] ;

[0015] .

[0016] Furthermore, the leaf reflectance and transmittance were generated based on the PROSPECT model; the GSV model was used to generate pure soil spectrum simulation data; the pure vegetation spectrum simulation data was generated based on the SAIL model by combining the simulation results of the PROSPECT model and the GSV model; It means that the simulated blade has a wavelength of The reflectivity at It means that the simulated blade has a wavelength of The transmittance at , N represents the blade structure parameter, Indicates the chlorophyll content, Indicates the carotenoid content, Indicates the anthocyanin content, Indicates the brown pigment content. represents the equivalent water thickness, Indicates the dry matter content, represents the solar zenith angle, represents the observation zenith angle, represents the relative azimuth angle between the sun and the sensor, represents the canopy reflectance under specific observation conditions, represents the average leaf inclination angle, represents the hotspot parameter, represents the leaf area index, represents the simulated soil reflectivity, , and Represent the three spectral vectors of dry soil, represents the spectrum vector of wet soil, represents the forward model used to simulate the reflectivity and transmittance of the leaves, represents the radiative transfer model used to simulate the reflectivity of vegetation canopies, represents the spectral vector model used to simulate soil reflectance, then:

[0017] ;

[0018] ;

[0019]

[0020] Furthermore, the generated pure vegetation spectral simulation data and pure soil spectral simulation data are fitted to the Landsat satellite band to generate mixed pixel simulation reflectance, including:

[0021] set up Indicates that the mixed pixel has a wavelength of The reflectivity, represents the leaf area index, Represents mixed pixel scale ; Indicates the proportion of pure vegetation coverage in mixed pixels, with a value between 0 and 1. represents the pure soil coverage ratio in the mixed pixel, then the mixed pixel has a wavelength of The reflectivity is:

[0022] ;

[0023] .

[0024] Further, the data is screened to obtain a modeling data set, including: The positive correlation with the normalized difference vegetation index is 5. ° The fixed step size is the interval and The simulated data set was screened at a fixed step size of 0.5 to ensure that the data were filtered at different solar zenith angles and The amount of simulated data within the interval combination is the same.

[0025] Furthermore, the reflectance in the modeling dataset is fitted to the Landsat satellite band in combination with the spectral response function of the Landsat sensor, and Gaussian truncated noise is added according to the uncertainty of the sensor signal in each band; It represents the first The reflectance value of the band, Indicates that the sensor receives The starting wavelength of the band, Indicates that the sensor receives The end wavelength of the band, The wavelength is The reflectivity at The sensor is The received signal strength, represents the probability density function of Gaussian noise, is the value of noise, is the mean of the noise, is the standard deviation of the noise Gaussian distribution, then:

[0026] ;

[0027] .

[0028] Further, set LSTM-PG eff The network structure of the model consists of the input gate, output gate and forget gate of LSTM. represents the weight matrix of the forget gate, represents the bias matrix of the forget gate, represents the weight matrix of the input gate, represents the bias matrix of the input gate, represents the weight matrix of the output gate, represents the bias matrix of the output gate, represents the weight matrix of candidate memory cells, represents the bias matrix of candidate memory cells, represents the characteristic value of the current input, represents the forget gate of the LSTM network, represents the input gate of the LSTM network, represents the output gate of the LSTM network, represents the previous cell state, Indicates the current cell state. represents the previous hidden neuron state, represents the current hidden neuron state, represents the hyperbolic tangent function as the activation function, Represents the sigmoid activation function, then:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] Define the loss function of the network parameter fitting process, Indicates the total amount of data. Indicates The measured value of the data, Indicates The model prediction value of the data, express The probability density function of the predicted value, express The probability density function of the target value, Represents the calculated value of the loss function under the conventional training method, is the loss value calculated by KL divergence, It is a nonlinear function in deep learning. It is the penalty loss calculated by limiting the boundary of the model prediction value less than 0. It is the penalty loss calculated by the boundary limit when the model prediction value is greater than 8. Used to calculate LSTM-PG eff The loss function value of the model, yes The proportion of loss items, yes The proportion of loss items, yes The coefficient of additional penalty for the predicted value outside the range of less than 0 or greater than 8 is:

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] .

[0040] Further, construct a transfer-trained LSTM-TL eff Model, node defines the loss function of the network parameter fitting process, set Indicates The measured value of the data, Indicates The model prediction value of the data, Indicates The weight of the measured data, Indicates the number of data in the current interval in the statistical measured values. and Represents the coefficient set to strengthen the quantitative characteristics, Indicates the total amount of data. Indicates the loss value affected by the weight when calculating the MSE loss. Indicates that the final use is effective Migrate the loss value of the training data set, It represents the proportion of weighted mean square error in quantifying the loss value. Indicates the proportion of KL divergence error, The coefficient representing the additional penalty for out-of-bounds cases where the model predicts values ​​less than 0 and greater than 8, then:

[0041] ;

[0042] ;

[0043] .

[0044] Further, construct a transfer-trained LSTM-TL true Model, node defines the loss function of the network parameter fitting process, set Indicates The model prediction value of the data, It is a nonlinear function in deep learning. It is the penalty loss calculated by the boundary limit when the model prediction value is greater than 10. Indicates that it is ultimately used for real Migrate the loss value of the training data set, It represents the proportion of weighted mean square error in quantifying the loss value. Indicates the proportion of KL divergence error, The coefficient representing the additional penalty for outliers where the model predicts values ​​less than 0 and greater than 10, then:

[0045] ;

[0046] .

[0047] Furthermore, the LSTM-TL true The model is packaged and saved and uses the LSTM-TL trained after migration true Model Estimation, including: LSTM-TL eff Model and LSTM-TL true The model is packaged and set to call LSTM-TL eff Model and LSTM-TL true Model interface; using LSTM-TL true The model predicts the Landsat satellite image data and obtains the first real ; For the first real For areas beyond the set range, LSTM-PG eff The model predicts Landsat satellite image data and obtains effective ; Effective construction based on measured data With the truth The empirical relationship between Convert to second reality ; Splicing the first reality The Second Reality Finally get the truth Estimate the results and save them.

[0048] The beneficial effects of the present invention are as follows: the present invention generates simulated data by using a physical radiation transfer model, and uses the knowledge of the migrated measured data set for deep learning training to increase physical constraints. Compared with the traditional method of modeling based only on data, the present invention can introduce remote sensing field knowledge to assist the modeling process; at the same time, the application of deep learning algorithms improves the model's ability to fit complex nonlinear regression relationships, and effectively improves the model accuracy; with satellite image data with high spatial resolution as the application target, the generated The product also has high spatial resolution. Compared with the coarse spatial resolution algorithm, it reduces the uncertainty caused by mixed pixel problems in the algorithm, greatly ensuring The accuracy of remote sensing estimates can also produce global long-term series The potential of the product. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A schematic diagram of a high spatial resolution leaf area index remote sensing estimation method provided in Example 1 of the present invention;

[0050] Figure 2 To build a transfer-trained LSTM-TL true Flowchart of the specific implementation of the model;

[0051] Figure 3 This is a flow chart of remote sensing estimation of vegetation leaf area index according to the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0053] Example 1

[0054] As an example, Figure 1 As shown, in order to solve the above technical problems, this embodiment provides a high spatial resolution vegetation leaf area index estimation method that integrates multi-source data, including generating a modeling data set based on a physical radiation transfer model, constructing an LSTM-PG eff Model, build transfer-trained LSTM-TL eff Model and build transfer-trained LSTM-TL true Models, and remote sensing Encapsulation, conversion, splicing and preservation of estimation models; as shown in the attached Figure 2 The following figure shows how to construct a transfer-trained LSTM-TL true Flowchart of the specific implementation of the model.

[0055] Generate a modeling data set based on the physical radiation transfer model, including: obtain spectral simulation data of leaves, canopy and soil based on the physical radiation transfer model, and generate pure vegetation spectral simulation data and pure soil spectral simulation data; fit the generated pure vegetation spectral simulation data and pure soil spectral simulation data to the Landsat satellite band to generate mixed pixel simulation reflectance; screen the data to obtain the modeling data set;

[0056] Building LSTM-PG eff Model, including: Building LSTM-PG eff Model training data set, set LSTM-PG eff The network structure of the model, defines the loss function of the network parameter fitting process, and uses LSTM-PG according to the loss function eff The training data set of the model is LSTM-PG eff The model performs parameter fitting of network node weights and biases to obtain the trained LSTM-PG eff Model;

[0057] Specifically, the reflectance of the green, red, near-infrared, short-wave infrared-1, and short-wave infrared-2 bands simulated by the radiation transfer model and the cosine value of the solar zenith angle are used as input. As output, construct LSTM-PG eff Model training data set; setting LSTM-PG eff The network structure of the model determines the number of network layers and the number of neurons in each hidden layer, defines the loss function of the network parameter fitting process, and quantifies the LSTM-PG according to the value of the loss function. eff The difference between the model's predictions and the reference values; Adjusting LSTM-PG eff Model training parameters, using LSTM-PG eff The training data set of the model is LSTM-PG eff The model performs parameter fitting of network node weights and biases, and the trained LSTM-PG eff The model is saved as a file.

[0058] Building a transfer-trained LSTM-TL eff Model, including: collating Landsat satellite image reflectivity and measured data to generate effective Migration training set, defined with trained LSTM-PG eff The network structure of the model is the same as the source model, and the node defines the loss function of the network parameter fitting process. The migration training set is used to fine-tune the parameters of the network node weights and biases of the source model to obtain the migration-trained LSTM-TL eff Model;

[0059] Specifically, Landsat satellite image reflectivity and measured data are collated to generate effective Migration training set; definition and trained LSTM-PG eff Model the same network structure and load the saved LSTM-PG eff Model as LSTM-TL eff The source model of the model; the node defines the loss function of the network parameter fitting process, setting the effective Migrate the sample imbalance coefficient in the training set and weight the loss function; adjust the LSTM-TL eff The training parameters of the model are effectively The migration training set of LSTM-PG eff The model fine-tunes the parameters of network node weights and biases, and the fine-tuned LSTM-TL eff The model is saved as a file;

[0060] Building a transfer-trained LSTM-TL true Model, including: collating Landsat satellite image reflectivity and measured data to generate real Migration training set, defined with trained LSTM-TL eff The network structure of the model is the same as the source model. The node defines the loss function of the network parameter fitting process. The LSTM-TL true The model fine-tunes the parameters of network node weights and biases to obtain the LSTM-TL model after migration training. true Model;

[0061] Specifically, the reflectance of the green, red, near-infrared, short-wave infrared-1, and short-wave infrared-2 bands of the Landsat satellite sensor and the cosine value of the solar zenith angle are used as feature values, and the measured real As the target value, the reflectivity of Landsat satellite images and measured data are collated to generate the real Migration training set; definition and trained LSTM-TL eff Model the same network structure and load the saved LSTM-TL eff Model as LSTM-TL true Source model; node defines the loss function of the network parameter fitting process, setting the real The sample imbalance coefficient in the migration training set weights the loss function; adjust the LSTM-TL true The training parameters of the model are used to train the LSTM-TL true The model fine-tunes the parameters of network node weights and biases, and uses the fine-tuned LSTM-TL trueThe model is saved as a file;

[0062] For the transfer-trained LSTM-TL true The model is packaged and saved and uses the LSTM-TL trained after migration true Model Estimate.

[0063] Optionally, the spectrum simulation data of leaves, canopy and soil are obtained based on the physical radiation transfer model, including: defining the input parameters of the radiation transfer model and constraining the consistency between the parameters to generate the spectrum simulation data of leaves, canopy and soil; setting represents randomly generated values ​​of the input parameters of the radiative transfer model, represents the input parameter value after the consistency constraint between parameters, represents the leaf area index, Indicates when =0 when the maximum value of the input parameter is Indicates when =0 is the minimum value of the input parameter. Indicates when The minimum value of the input parameter when the maximum value is obtained. Indicates when The maximum value of the input parameter when the maximum value is obtained. Indicates when The minimum value of the input parameter when obtaining a fixed value, Indicates when The maximum value of the input parameter when obtaining a fixed value, then:

[0064] ;

[0065] ;

[0066] .

[0067] Optionally, generate leaf reflectance and transmittance based on the PROSPECT model; use the GSV model to generate pure soil spectrum simulation data; combine the simulation results of the PROSPECT model and the GSV model, and generate pure vegetation spectrum simulation data based on the SAIL model; set It means that the simulated blade has a wavelength of The reflectivity at It means that the simulated blade has a wavelength of The transmittance at , N represents the blade structure parameter, Indicates the chlorophyll content, Indicates the carotenoid content, Indicates the anthocyanin content, Indicates the brown pigment content. represents the equivalent water thickness, Indicates the dry matter content, represents the solar zenith angle, represents the observation zenith angle, represents the relative azimuth angle between the sun and the sensor, represents the canopy reflectance under specific observation conditions, represents the average leaf inclination angle, represents the hotspot parameter, represents the leaf area index, represents the simulated soil reflectivity, , and Represent the three spectral vectors of dry soil, represents the spectrum vector of wet soil, represents the forward model used to simulate the reflectivity and transmittance of the leaves, represents the radiative transfer model used to simulate the reflectivity of vegetation canopies, represents the spectral vector model used to simulate soil reflectance, then:

[0068] ;

[0069] ;

[0070]

[0071] Optionally, the generated pure vegetation spectral simulation data and pure soil spectral simulation data are fitted to the Landsat satellite band to generate mixed pixel simulated reflectance, including:

[0072] set up Indicates that the mixed pixel has a wavelength of The reflectivity, represents the leaf area index, Represents mixed pixel scale ; Indicates the proportion of pure vegetation coverage in mixed pixels, with a value between 0 and 1. represents the pure soil coverage ratio in the mixed pixel, then the mixed pixel has a wavelength of The reflectivity is:

[0073] ;

[0074] .

[0075] Optionally, data screening to obtain a modeling data set includes: The positive correlation with the normalized difference vegetation index is 5. °The fixed step size is the interval and The simulated data set was screened at a fixed step size of 0.5 to ensure that the data were filtered at different solar zenith angles and The amount of simulated data within the interval combination is the same, ensuring that the characteristics of the modeled data set are relatively uniform.

[0076] Optionally, the reflectance in the modeling dataset is fitted to the Landsat satellite band in combination with the Landsat sensor spectral response function, and Gaussian truncated noise is added according to the uncertainty of the sensor signal in each band; It represents the first The reflectance value of the band, Indicates that the sensor receives The starting wavelength of the band, Indicates that the sensor receives The end wavelength of the band, The wavelength is The reflectivity at The sensor is The received signal strength, represents the probability density function of Gaussian noise, is the value of noise, is the mean of the noise, is the standard deviation of the noise Gaussian distribution, then:

[0077] ;

[0078] .

[0079] Optionally, set LSTM-PG eff The network structure of the model consists of the input gate, output gate and forget gate of LSTM. represents the weight matrix of the forget gate, represents the bias matrix of the forget gate, represents the weight matrix of the input gate, represents the bias matrix of the input gate, represents the weight matrix of the output gate, represents the bias matrix of the output gate, represents the weight matrix of candidate memory cells, represents the bias matrix of candidate memory cells, represents the characteristic value of the current input, represents the forget gate of the LSTM network, represents the input gate of the LSTM network, represents the output gate of the LSTM network, represents the previous cell state, Indicates the current cell state. represents the previous hidden neuron state, represents the current hidden neuron state, represents the hyperbolic tangent function as the activation function, Represents the sigmoid activation function, then:

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] Define the loss function of the network parameter fitting process, Indicates the total amount of data. Indicates The measured value of the data, Indicates The model prediction value of the data, express The probability density function of the predicted value, express The probability density function of the target value, Represents the calculated value of the loss function under the conventional training method, is the loss value calculated by KL divergence, It is a nonlinear function in deep learning. It is the penalty loss calculated by limiting the boundary of the model prediction value less than 0. It is the penalty loss calculated by the boundary limit when the model prediction value is greater than 8. Used to calculate LSTM-PG eff The loss function value of the model, yes The proportion of loss items, yes The proportion of loss items, yes The coefficient of additional penalty for the predicted value outside the range of less than 0 or greater than 8 is:

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] .

[0091] Optionally, build a transfer-trained LSTM-TL eff Model, node defines the loss function of the network parameter fitting process, set Indicates The measured value of the data, Indicates The model prediction value of the data, Indicates The weight of the measured data, Indicates the number of data in the current interval in the statistical measured values. and Represents the coefficient set to strengthen the quantitative characteristics, Indicates the total amount of data. Indicates the loss value affected by the weight when calculating the MSE loss. Indicates that the final use is effective Migrate the loss value of the training data set, It represents the proportion of weighted mean square error in quantifying the loss value. Indicates the proportion of KL divergence error, The coefficient representing the additional penalty for out-of-bounds cases where the model predicts values ​​less than 0 and greater than 8, then:

[0092] ;

[0093] ;

[0094] .

[0095] Optionally, build a transfer-trained LSTM-TL true Model, node defines the loss function of the network parameter fitting process, set Indicates The model prediction value of the data, It is a nonlinear function in deep learning. It is the penalty loss calculated by the boundary limit when the model prediction value is greater than 10. Indicates that it is ultimately used for real Migrate the loss value of the training data set, It represents the proportion of weighted mean square error in quantifying the loss value. Indicates the proportion of KL divergence error, The coefficient representing the additional penalty for outliers where the model predicts values ​​less than 0 and greater than 10, then:

[0096] ;

[0097] .

[0098] Optionally, transfer trained LSTM-TL true The model is packaged and saved and uses the LSTM-TL trained after migration true Model Estimation, including: LSTM-TL eff Model and LSTM-TL true The model is packaged and set to call LSTM-TL eff Model and LSTM-TL true Model interface; using LSTM-TL true The model predicts the Landsat satellite image data and obtains the first real ; For the first real For areas beyond the set range, LSTM-PG eff The model predicts Landsat satellite image data and obtains effective ; Effective construction based on measured data With the truth The empirical relationship between Convert to second reality ; Splicing the first reality The Second Reality Finally get the truth Estimate the results and save them.

[0099] Specifically, see the saved LSTM-TL eff Model files and LSTM-TL true The network weights and biases in the model file are for LSTM-TL eff Model and LSTM-TL true The model is packaged and set to call LSTM-TL eff Model and LSTM-TL true The interface of the model;

[0100] With truth As the inversion target, select LSTM-TL true The model is used as the main algorithm to predict Landsat satellite image data and obtain the first real ; Filter the main algorithm prediction result area outside the set range value, using LSTM-PG effThe model is effective as a backup algorithm Predictions and effective With the truth The empirical relationship between Convert to second reality ; Combine the prediction results obtained using the main algorithm with the prediction results obtained using the backup algorithm to finally get the true The estimated results and the actual The estimation results are saved as GeoTIFF raster image data.

[0101] Optionally, when filtering the main algorithm prediction results of the set range value, filter the main algorithm prediction results that are less than 0 or greater than 10, as shown in the attached Figure 3 As shown, the LSTM-PGeff model is used as a backup algorithm for prediction, and the statistically effective and truth The relationship converts the backup algorithm prediction results into real , and get the 30m spatial resolution true Estimation results.

[0102] This paper proposes a global high spatial resolution leaf area index remote sensing estimation method based on radiation transfer model and deep learning with physical constraints, which can be used for regional or global scale land vegetation. Remote sensing estimation provides a reference, physical radiation transfer models are used to generate simulated data sets, and measured data are used to generate migration training sets to build more accurate Remote sensing estimation model.

[0103] The present invention selects a universal remote sensing radiation transmission model, determines the values ​​that can represent various vegetation types when defining the input parameters, and the samples of migration training cover a variety of vegetation types, ensuring that the model can be widely applied to different vegetation types.

[0104] The algorithm constructed by the present invention can be adapted to the Landsat series images and can generate a long-term global series with a spatial resolution of 30m. Products, solutions Mixed pixel problem in products to achieve high resolution accurate estimate of .

[0105] The present invention generates simulated data by using a physical radiation transfer model and uses the knowledge of the migrated measured data set for deep learning training to increase physical constraints. Compared with the traditional method of modeling based only on data, it can introduce remote sensing field knowledge to assist the modeling process; at the same time, the application of deep learning algorithms improves the model's ability to fit complex nonlinear regression relationships and effectively improves the model accuracy; targeting satellite image data with high spatial resolution as the application target, the generated The product also has high spatial resolution. Compared with the coarse spatial resolution algorithm, it reduces the uncertainty caused by mixed pixel problems in the algorithm, greatly ensuring The accuracy of the estimate is also capable of producing global long-term series The potential of the product.

[0106] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A high spatial resolution vegetation leaf area index estimation method integrating multi-source data, characterized in that: Including generating modeling data sets based on physical radiation transfer models, building LSTM-PG eff Model, build transfer-trained LSTM-TL eff Model and build transfer-trained LSTM-TL true Model, and Encapsulation, conversion, splicing and preservation of remote sensing estimation models; Generate a modeling data set based on the physical radiation transfer model, including: obtain spectral simulation data of leaves, canopy and soil based on the physical radiation transfer model, and generate pure vegetation spectral simulation data and pure soil spectral simulation data; fit the generated pure vegetation spectral simulation data and pure soil spectral simulation data to the Landsat satellite band to generate mixed pixel simulation reflectance; screen the data to obtain the modeling data set; Building LSTM-PG eff Model, including: Building LSTM-PG eff Model training data set, set LSTM-PG eff The network structure of the model, defines the loss function of the network parameter fitting process, and uses LSTM-PG according to the loss function eff Model training data set for LSTM-PG eff The model performs parameter fitting of network node weights and biases to obtain the trained LSTM-PG eff Model; Building a transfer-trained LSTM-TL eff Model, including: collating Landsat satellite image reflectivity and measured data to generate effective Migration training set, defined with trained LSTM-PG eff The network structure of the model is the same as the source model, and the node defines the loss function of the network parameter fitting process. The migration training set is used to fine-tune the parameters of the network node weights and biases of the source model to obtain the migration-trained LSTM-TL eff Model; Building a transfer-trained LSTM-TL true Model, including: collating Landsat satellite image reflectivity and measured data to generate real Migration training set, defined with trained LSTM-TL eff The network structure of the model is the same as the source model. The node defines the loss function of the network parameter fitting process. The LSTM-TL true The model fine-tunes the parameters of network node weights and biases to obtain the LSTM-TL model after migration training. true Model; For the transfer-trained LSTM-TL true The model is packaged and saved and uses the LSTM-TL trained after migration true Model Estimation, including: LSTM-TL eff Model and LSTM-TL true The model is packaged and set to call LSTM-TL eff Model and LSTM-TL true Model interface; using LSTM-TL true The model predicts the Landsat satellite image data and obtains the first real ; For the first real For areas beyond the set range, LSTM-PG eff The model predicts Landsat satellite image data and obtains effective ; Effective construction based on measured data With the truth The empirical relationship between Convert to second reality ; Splicing the first reality The Second Reality Finally get the truth Estimate the results and save them.

2. The high spatial resolution vegetation leaf area index estimation method based on multi-source data fusion according to claim 1 is characterized in that: The spectrum simulation data of leaves, canopy and soil are obtained based on the physical radiation transfer model, including: defining the input parameters of the radiation transfer model and constraining the consistency between the parameters to generate the spectrum simulation data of leaves, canopy and soil; setting represents randomly generated values ​​of the input parameters of the radiative transfer model, represents the input parameter value after the consistency constraint between parameters, represents the leaf area index, Indicates when =0 when the maximum value of the input parameter is Indicates when =0 is the minimum value of the input parameter. Indicates when The minimum value of the input parameter when the maximum value is obtained. Indicates when The maximum value of the input parameter when the maximum value is obtained. Indicates when The minimum value of the input parameter when obtaining a fixed value, Indicates when The maximum value of the input parameter when obtaining a fixed value, then: ； ； 。 3. The high spatial resolution vegetation leaf area index estimation method based on multi-source data fusion according to claim 2 is characterized in that: Generate leaf reflectance and transmittance based on PROSPECT model; use GSV model to generate pure soil spectrum simulation data; combine the simulation results of PROSPECT model and GSV model, and generate pure vegetation spectrum simulation data based on SAIL model; set up It means that the simulated blade has a wavelength of The reflectivity at It means that the simulated blade has a wavelength of The transmittance at , N represents the blade structure parameter, Indicates the chlorophyll content, Indicates the carotenoid content, Indicates the anthocyanin content, Indicates the brown pigment content. represents the equivalent water thickness, Indicates the dry matter content, represents the solar zenith angle, represents the observation zenith angle, represents the relative azimuth angle between the sun and the sensor, represents the canopy reflectance under specific observation conditions, represents the average leaf inclination angle, represents the hotspot parameter, represents the leaf area index, represents the simulated soil reflectivity, , and Represent the three spectral vectors of dry soil, represents the spectrum vector of wet soil, represents the forward model used to simulate the reflectivity and transmittance of the leaves, represents the radiative transfer model used to simulate the reflectivity of vegetation canopies, represents the spectral vector model used to simulate soil reflectance, then: ； ； 4. The high spatial resolution vegetation leaf area index estimation method based on fusion of multi-source data according to claim 1 is characterized in that: The generated pure vegetation spectral simulation data and pure soil spectral simulation data are fitted to the Landsat satellite band to generate mixed pixel simulation reflectance, including: set up Indicates that the mixed pixel has a wavelength of The reflectivity, represents the leaf area index, Represents mixed pixel scale , represents the canopy reflectance under specific observation conditions, is the simulated soil reflectivity, Indicates the proportion of pure vegetation coverage in mixed pixels, with a value between 0 and 1. represents the pure soil coverage ratio in the mixed pixel, then the mixed pixel has a wavelength of The reflectivity is: ； 。 5. The high spatial resolution vegetation leaf area index estimation method based on multi-source data fusion according to claim 1, characterized in that: Data screening to obtain a modeling data set includes: The positive correlation with the normalized difference vegetation index is 5. ° The fixed step size is the interval and The simulated data set was screened at a fixed step size of 0.5 to ensure that the The amount of simulated data within the interval combination is the same.

6. The high spatial resolution vegetation leaf area index estimation method based on fusion of multi-source data according to claim 4 is characterized in that: Combined with the spectral response function of the Landsat sensor, the reflectance in the modeling dataset is fitted to the Landsat satellite band, and Gaussian truncation noise is added according to the uncertainty of the sensor signal in each band; It represents the first The reflectance value of the band, Indicates that the sensor receives The starting wavelength of the band, Indicates that the sensor receives The end wavelength of the band, The wavelength is The reflectivity at The sensor is The received signal strength, represents the probability density function of Gaussian noise, is the value of noise, is the mean of the noise, is the standard deviation of the noise Gaussian distribution, then: ； 。 7. The high spatial resolution vegetation leaf area index estimation method based on fusion of multi-source data according to claim 1, characterized in that: Setting up LSTM-PG eff The network structure of the model consists of the input gate, output gate and forget gate of LSTM. represents the weight matrix of the forget gate, represents the bias matrix of the forget gate, represents the weight matrix of the input gate, represents the bias matrix of the input gate, represents the weight matrix of the output gate, represents the bias matrix of the output gate, represents the weight matrix of candidate memory cells, represents the bias matrix of candidate memory cells, represents the characteristic value of the current input, represents the forget gate of the LSTM network, represents the input gate of the LSTM network, represents the output gate of the LSTM network, represents the previous cell state, Indicates the current cell state. represents the previous hidden neuron state, represents the current hidden neuron state, represents the hyperbolic tangent function as the activation function, Represents the sigmoid activation function, then: ； ； ； ； ； Define the loss function of the network parameter fitting process, Indicates the total amount of data. Indicates The measured value of the data, Indicates The model prediction value of the data, express The probability density function of the predicted value, express The probability density function of the target value, Represents the calculated value of the loss function under the conventional training method, is the loss value calculated by KL divergence, It is a nonlinear function in deep learning. It is the penalty loss calculated by limiting the boundary of the model prediction value less than 0. It is the penalty loss calculated by the boundary limit when the model prediction value is greater than 8. Used to calculate LSTM-PG eff The loss function value of the model, yes The proportion of loss items, yes The proportion of loss items, yes The coefficient of additional penalty for the predicted value outside the range of less than 0 or greater than 8 is: ； ； ； ； 。 8. The high spatial resolution vegetation leaf area index estimation method based on fusion of multi-source data according to claim 1, characterized in that: Building a transfer-trained LSTM-TL eff Model, node defines the loss function of the network parameter fitting process, set Indicates The measured value of the data, Indicates The model prediction value of the data, Indicates The weight of the measured data, Indicates the number of data in the current interval in the statistical measured values. and Represents the coefficient set to strengthen the quantitative characteristics, Indicates the total amount of data. Indicates the loss value affected by the weight when calculating the MSE loss. Indicates that the final use is effective Migrate the loss value of the training data set, It represents the proportion of weighted mean square error in quantifying the loss value. Indicates the proportion of KL divergence error, The coefficient representing the additional penalty for out-of-bounds cases where the model predicts values ​​less than 0 and greater than 8, then: ； ； 。 9. The high spatial resolution vegetation leaf area index estimation method based on fusion of multi-source data according to claim 1, characterized in that: Building a transfer-trained LSTM-TL true Model, node defines the loss function of the network parameter fitting process, set Indicates The model prediction value of the data, It is a nonlinear function in deep learning. It is the penalty loss calculated by the boundary limit when the model prediction value is greater than 10. Indicates that it is ultimately used for real Migrate the loss value of the training data set, It represents the proportion of weighted mean square error in quantifying the loss value. Indicates the proportion of KL divergence error, The coefficient representing the additional penalty for outliers where the model predicts values ​​less than 0 and greater than 10, then: ； 。

Images