A remote sensing image albedo inversion method and system based on LSTM network

Abstract

The present invention discloses a remote sensing image albedo inversion method and system based on an LSTM network, comprising: (1) obtaining the solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth and satellite observation zenith angle at different observation times in a study area; (2) obtaining remote sensing image albedo product data of a continuous time series in the study area, sampling and obtaining albedo values ​​corresponding to different observation times; (3) taking the four angles and albedo values ​​corresponding to the same observation time as a set of sequences, obtaining a data set consisting of all observation times and dividing it into a training set and a test set; (4) constructing an optimized LSTM model, inputting the training set into the constructed model for iterative training, terminating the training when a preset number of training times is reached and obtaining a trained model; (5) inputting the test set into the trained model for testing, and obtaining the albedo values ​​predicted at different observation times. The present invention can invert the remote sensing image albedo and obtain accurate albedo.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image inversion, and in particular to a remote sensing image albedo inversion method and system based on an LSTM network. Background Art

[0002] Currently, there are relatively few global-scale and long-term albedo products produced based on remote sensing satellite data. Furthermore, these satellite-based albedo products suffer from low temporal and spatial resolution. However, with the increasing number of remote sensing satellites launched, the massive amounts of data captured by these satellites are providing a wealth of information for albedo inversion, making it possible to perform albedo inversion using remote sensing as a technical tool. Remote sensing technology can meet the needs of albedo inversion for continuous time series in both space and time. Commonly used remote sensing albedo inversion methods include statistical inversion, albedo inversion that considers underlying surface anisotropy, surface binomial reflectance models based on multiple satellite observations, and semi-empirical kernel-driven models.

[0003] Currently, there are few studies on albedo inversion based on satellite data. Most of them focus on CCD data and high-resolution satellite data. These data have high resolution but limited coverage, making it difficult to generate global-scale albedo products. Therefore, it is necessary to provide a study on albedo inversion based on satellite data such as Fengyun 3D data. Summary of the Invention

[0004] Purpose of the Invention: To address the above shortcomings, the present invention discloses a remote sensing image albedo inversion method based on an LSTM network. This method uses an optimized LSTM network model to invert the albedo of remote sensing images, obtaining relatively accurate inversion results. Furthermore, the present invention provides a remote sensing image albedo inversion system based on an LSTM network, which can implement the above method to obtain accurate albedo.

[0005] Technical solution: To solve the above problems, the present invention provides a remote sensing image albedo inversion method based on LSTM network, which specifically includes the following steps:

[0006] The specific steps include:

[0007] (1) Obtain the solar incidence azimuth θ at different observation times in the study area i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 ;

[0008] (2) Obtain the albedo product data of the continuous time series of remote sensing images of the study area, sample the obtained albedo product data, and obtain the albedo value R corresponding to different observation times;

[0009] (3) The solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth, satellite observation azimuth, and albedo values ​​corresponding to the same observation time are taken as a set of sequences D i ={θ i1 i ,θ i2 i ,θ v1 i ,θ v2 i ,R i}, get the data set consisting of all observation moments i is the observation time, n is a positive integer; the obtained data set D is divided into a training set and a test set;

[0010] (4) Construct an optimized LSTM model. The input layer of the LSTM model is the solar incident azimuth angle θ i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 , the output layer is the albedo value R; the formula of the optimized LSTM model is:

[0011] f t 2 =σ(W f 2 ·[h t 1 ,x t ]+b f 2 )

[0012] i t 2 =σ(W i 2 ·[h t 1 ,x t ]+b i 2 )

[0013] o t 2 =σ(W o 2 ·[h t 1 ,x t ]+b o 2 )

[0014]

[0015]

[0016] h t 2 =o t 2 ×tanh(c t 2 )+h t 3

[0017] Where, f t 2 For the second layer of the forget gate; i t 2 is the input gate of the second layer; o t 2 is the output gate of the second layer; x t is the input; h t 2 is the output value of the second layer, that is, the output at the current moment, h t 1 is the output value of the first layer; h t 3 is the output value of the third layer; W f 2 、W i 2 、W o 2 、W c 2 are all learning weights; b f 2 、b i 2 、b o 2 、b c 2 are bias; σ represents the sigmoid function; c t 2 represents the impact factor of the second layer output, c t 3 represents the third layer output impact factor, c t 1 is the first-level impact factor; Indicates the value of the gate when the second layer is updated;

[0018] The training set is input into the optimized LSTM model for iterative training. When the preset number of training times is reached, the training ends and the optimized LSTM model is obtained.

[0019] (5) The test set is input into the trained optimized LSTM model for testing to obtain the predicted albedo values ​​at different observation times.

[0020] Furthermore, the method further includes step (6) of verifying the albedo value obtained in step (5). The specific verification steps include:

[0021] (6.1) The solar incident azimuth angle θ in the test set is i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 Input into the BRDF model to calculate the albedo value; the specific formula of the BRDF model is:

[0022] R(θ i1 ,θ V1 ,θ i2 ,θ v2 )=f iso +f geo k geo (θ i1 ,θ V1 ,θ i2 ,θ v2 )+f vol k vol (θ i1 ,θ V1 ,θ i2 ,θ v2 )

[0023] Where, f iso represents the isotropic scattering kernel coefficient; f geo represents the geometric optics scattering kernel coefficient; f vol represents the scale factor of the volume scattering kernel; k geo represents the geometric optics kernel function; k vol is the volume scattering kernel function;

[0024] (6.2) Compare the calculated albedo value with the predicted albedo value obtained in step (5). If the difference between the two is within the preset value range, the predicted albedo value is considered correct; otherwise, update the learning weights in the LSTM model, and retrain and test the LSTM model.

[0025] Furthermore, the step (2) of obtaining the remote sensing image albedo product data of the continuous time series of the study area specifically includes the following steps:

[0026] (2.1) Obtaining a continuous time series of remote sensing images of the study area, extracting image data from the remote sensing images, and dividing the obtained images into small blocks;

[0027] (2.2) Use the cloud mask to filter out the cloud-free image blocks to create a dataset, and put the cloud-free image blocks into the cloud-free folder;

[0028] (2.3) Read the albedo data of the area corresponding to the image in the cloud-free folder.

[0029] Furthermore, the optimized LSTM model has three hidden layers.

[0030] In addition, the present invention provides a remote sensing image albedo inversion system based on LSTM network, which includes the following modules:

[0031] The acquisition module is used to obtain the solar incidence azimuth θ at different observation times in the study area i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 ; Obtain the albedo product data of the continuous time series of remote sensing images of the study area, sample the obtained albedo product data, and obtain the albedo value R corresponding to different observation times;

[0032] The data preprocessing module is used to process the solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth, satellite observation azimuth, and albedo values ​​corresponding to the same observation time as a set of sequence D i ={θ i1 i ,θ i2 i ,θ v1 i ,θ v2 i ,R i}, get the data set consisting of all observation moments i is the observation time, n is a positive integer; the obtained data set D is divided into a training set and a test set;

[0033] Optimize the LSTM model building module; used to build an optimized LSTM model. The input layer of the LSTM model is the solar incidence angle θ i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 , the output layer is the albedo value R; the formula of the optimized LSTM model is:

[0034] f t 2 =σ(W f 2 ·[h t 1 ,x t ]+b f 2 )

[0035] it 2 =σ(W i 2 ·[h t 1 ,x t ]+b i 2 )

[0036] o t 2 =σ(W o 2 ·[h t 1 ,x t ]+b o 2 )

[0037]

[0038]

[0039] h t 2 =o t 2 ×tanh(c t 2 )+h t 3

[0040] Where, f t 2 For the second layer of the forget gate; i t 2 is the input gate of the second layer; o t 2 is the output gate of the second layer; x t is the input; h t 2 is the output value of the second layer, that is, the output at the current moment, h t 1 is the output value of the first layer; h t 3 is the output value of the third layer; W f 2 、W i 2 、W o 2 、W c 2 are all learning weights; b f 2 、b i 2 、b o 2 、b c2 are bias; σ represents the sigmoid function; c t 2 represents the impact factor of the second layer output, c t 3 represents the third layer output impact factor, c t 1 is the first-level impact factor; Indicates the value of the gate when the second layer is updated;

[0041] Optimize the LSTM model training module; it is used to input the training set into the optimized LSTM model for iterative training. When the preset number of training times is reached, the training ends and the optimized LSTM model is obtained.

[0042] Optimize the LSTM model testing module; input the test set into the trained optimized LSTM model for testing to obtain the predicted albedo values ​​at different observation times.

[0043] Furthermore, a verification module is included: used to verify the albedo value obtained by the optimized LSTM model test module; the verification module specifically includes:

[0044] BRDF model calculation module, used to convert the solar incident angle θ in the test set i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 Input into the BRDF model to calculate the albedo value; the specific formula of the BRDF model is:

[0045] R(θ i1 ,θ V1 ,θ i2 ,θ v2 )=f iso +f geo k geo (θ i1 ,θ V1 ,θ i2 ,θ v2 )+f vol k vol (θ i1 ,θ V1 ,θ i2 ,θ v2 )

[0046] Where, f iso represents the isotropic scattering kernel coefficient; f geo represents the geometric optics scattering kernel coefficient; f vol represents the scale factor of the volume scattering kernel; k geo represents the geometric optics kernel function; kvol is the volume scattering kernel function;

[0047] The comparison module is used to compare the calculated albedo value with the predicted albedo value obtained by the optimized LSTM model test module. If the difference between the two is within the preset value range, the predicted albedo value is considered correct; otherwise, the learning weights in the LSTM model are updated, and the LSTM model is retrained and tested.

[0048] The present invention also provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the aforementioned methods when executing the computer program. A computer-readable storage medium having the computer program stored thereon is provided, wherein the computer program implements the steps of any of the aforementioned methods when executed by the processor.

[0049] Beneficial effects: The advantages of the remote sensing image albedo inversion method based on the LSTM network provided by the present invention over the prior art are: 1. By modifying the memory gate mechanism of LSTM, the outputs of the two moments before and after a certain moment are used as influencing factors for module training in the memory gate of the current moment, and a trained optimized LSTM model is obtained, thereby obtaining the final predicted inversion result; 2. By comparing the predicted inversion result with the calculation result of the BRDF model, the prediction model is further optimized to improve the albedo inversion accuracy. In addition, the remote sensing image albedo inversion system based on the LSTM network described in the present invention has the advantages over the prior art in that: the above-mentioned inversion method can be implemented by this system to obtain the final inversion result. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Shown is a schematic diagram of the method of the present invention;

[0051] Figure 2 Shown is a flow chart of inversion using an LSTM network module in the method of the present invention;

[0052] Figure 3 The figure shows the structure of the optimized LSTM network module of the present invention;

[0053] Figure 4 The figure shows the module structure diagram of the optimized LSTM network memory gate mechanism according to the present invention;

[0054] Figure 5 Shown is a schematic diagram of the BRDF model of the present invention;

[0055] Figure 6 Shown is an albedo inversion map obtained according to the method of the present invention. DETAILED DESCRIPTION

[0056] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0057] like Figure 1 and Figure 2 The present invention provides a remote sensing image albedo inversion method based on LSTM network, which specifically includes the following steps:

[0058] Step 1: Obtain the solar incidence azimuth θ at different observation times in the study area i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 .

[0059] Step 2: Obtain the albedo product data of the continuous time series of remote sensing images of the study area, sample the obtained albedo product data, and obtain the albedo value R corresponding to different observation times; the specific steps include:

[0060] (1) Extracting frame data from remote sensing images of a continuous time series of the study area; and dividing the corresponding images into small blocks;

[0061] (2) Using the cloud mask, filter out the cloud-free image blocks to make a dataset, and put the cloud-free image blocks into the cloud-free folder;

[0062] (3) Read the albedo data of the area corresponding to the image in the cloud-free folder.

[0063] Specifically, in actual operation, albedo data is generally read for each pixel of the remote sensing image, and the four angles of the study area reflected by each pixel in the remote sensing image at the same time are also obtained pixel by pixel. i ) are shown in Table 1 below. The four angle values ​​and albedo values ​​composed of all pixel points at the same moment can be represented in matrix form respectively.

[0064] Table 1

[0065]

[0066] Step 3: Preprocess the acquired data and divide it into training data set and test data set.

[0067] Specifically, the solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth, satellite observation azimuth, and albedo values ​​corresponding to the same observation time are taken as a set of sequences D i ={θ i1 i ,θ i2i ,θ v1 i ,θ v2 i ,R i}; The data set consisting of all observation moments The obtained data set D is divided into a training set and a test set; where i is the observation time, and t1, t2, ..., t n , n is a positive integer.

[0068] Step 4: Build an optimized LSTM model and use the training dataset to train the optimized LSTM model.

[0069] (1) Set the input layer of the LSTM model to the solar incident azimuth angle θ i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 , the output of the output layer is the albedo value R; the LSTM model contains three hidden layers;

[0070] The optimized LSTM model is specifically improved by the memory gate mechanism of the LSTM model, and the influence factors and output values ​​of the two moments before and after a certain moment are input into the output obtained at the current moment. Figure 4 As shown, the model is simplified and part of it is improved to the influencing factors of the previous and next moments are input into the current moment to affect the output;

[0071] like Figure 3 As shown, the optimized LSTM model has a three-layer structure. The output of the second layer is the output value at the current moment. The specific formula is:

[0072] f t 2 =σ(W f 2 ·[h t 1 ,x t ]+b f 2 )

[0073] i t 2 =σ(W i 2 ·[h t 1 ,x t ]+b i 2 )

[0074] o t 2 =σ(Wo 2 ·[h t 1 ,x t ]+b o 2 )

[0075]

[0076]

[0077] h t 2 =o t 2 ×tanh(c t 2 )+h t 3

[0078] Where, f t 2 For the second layer of the forget gate; i t 2 is the input gate of the second layer; o t 2 is the output gate of the second layer; x t is the input; h t 2 is the output value of the second layer, h t 1 is the output value of the first layer; h t 3 is the output value of the third layer; W f 2 、W i 2 、W o 2 、W c 2 are all learning weights; b f 2 、b i 2 、b o 2 、b c 2 are bias; σ represents the sigmoid function; c t 2 represents the impact factor of the second layer output, c t 3 represents the third layer output impact factor, c t 1 is the first-level impact factor; Indicates the value of the gate when the second layer is updated;

[0079] Among them, for h t 1 、h t 3 The specific expression is as follows:

[0080] f t 1 =σ(W f 1 ·[h t-1 ,x t ]+b f 1 )

[0081] i t 1 =σ(W i 1 ·[h t-1 ,x t ]+b i 1 )

[0082] o t 1 =σ(W o 1 ·[h t-1 ,x t ]+b o 1 )

[0083]

[0084]

[0085] h t 1 =o t 1 ×tanh(c t 1 )

[0086] Where h t-1 is the output value at the previous moment; b f 1 、b i 1 、b o 1 、b c 1 are all biases of the first layer; W f 1 、W i 1 、W o 1 、W c 1 are the weights of the first layer; ,f t1 is the forget gate of the first layer; i t 1 is the input gate of the first layer; o t 1 is the output gate of the first layer; Indicates that when the first layer updates the door; c t-1 is the impact factor at the previous moment;

[0087] f t 3 =σ(W f 3 ·[h t+1 ,x t ]+b f 3 )

[0088] i t 3 =σ(W i 3 ·[h t+1 ,x t ]+b i 3 )

[0089] o t 3 =σ(W o 3 ·[h t+1 ,x t ]+b o 3 )

[0090]

[0091]

[0092] h t 3 =o t 1 ×tanh(c t 3 )

[0093] Where h t+1 is the output value at the next moment; b f 3 、b i 3 、b o 3 、b c 3 The bias of the third layer; W f 3 、W i 3 、Wo 3 、W c 3 is the weight of the third layer; f t 3 For the third layer of the forget gate; i t 3 is the input gate of the third layer; o t 3 is the output gate of the third layer; Indicates that when the third layer updates the door; c t+1 is the impact factor at the next moment;

[0094] (2) Input the training set into the optimized LSTM model for iterative training. When the preset number of training times N is reached, the training is terminated and the optimized LSTM model is obtained.

[0095] Step 5: Input the test set into the trained optimized LSTM model for testing to obtain the predicted albedo values ​​at different observation times.

[0096] Step 6: The predicted albedo value obtained by the optimized LSTM model can be verified through the BRDF model. The specific verification steps include:

[0097] (1) The test is focused on the solar incident azimuth angle θ i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 The data value is input into the BRDF model to calculate the albedo value; the specific formula of the BRDF model is:

[0098] R(θ i1 ,θ V1 ,θ i2 ,θ v2 )=f iso +f geo k geo (θ i1 ,θ V1 ,θ i2 ,θ v2 )+f vol k vol (θ i1 ,θ V1 ,θ i2 ,θ v2 )

[0099] Where, f iso represents the isotropic scattering kernel coefficient; f geo represents the geometric optics scattering kernel coefficient; f vol represents the scale factor of the volume scattering kernel; kgeo represents the geometric optics kernel function; k vol is the volume scattering kernel function;

[0100] like Figure 5 As shown in FIG, the calculation of the BRDF coefficient in the BRDF model can be obtained by substituting the data in the training set into the BRDF coefficient for calculation.

[0101] (2) Compare the calculated albedo value with the predicted albedo value obtained in step 5. If the difference between the two is within the preset value range, the predicted albedo value is considered correct; otherwise, update the learning weights in the LSTM model and retrain and test the LSTM model.

[0102] like Figure 6 As shown, the albedo inversion is performed for remote sensing effects through this method, and the albedo result value can be accurately read from the final inversion result map.

[0103] In addition, the present invention also provides a remote sensing image albedo inversion system based on LSTM network, which includes the following modules:

[0104] The acquisition module is used to obtain the solar incidence azimuth θ at different observation times in the study area i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 ; Obtain the albedo product data of the continuous time series of remote sensing images of the study area, sample the obtained albedo product data, and obtain the albedo value R corresponding to different observation times;

[0105] The data preprocessing module is used to process the solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth, satellite observation azimuth, and albedo values ​​corresponding to the same observation time as a set of sequence D i ={θ i1 i ,θ i2 i ,θ v1 i ,θ v2 i ,R i}, get the data set consisting of all observation moments i is the observation time, n is a positive integer; the obtained data set D is divided into a training set and a test set;

[0106] Optimize the LSTM model building module; used to build an optimized LSTM model. The input layer of the LSTM model is the solar incidence angle θ i1 、Zenith angle of solar incidence θ i2、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 , the output layer is the albedo value R; the formula of the optimized LSTM model is:

[0107] f t 2 =σ(W f 2 ·[h t 1 ,x t ]+b f 2 )

[0108] i t 2 =σ(W i 2 ·[h t 1 ,x t ]+b i 2 )

[0109] o t 2 =σ(W o 2 ·[h t 1 ,x t ]+b o 2 )

[0110]

[0111]

[0112] h t 2 =o t 2 ×tanh(c t 2 )+h t 3

[0113] Where, f t 2 For the second layer of the forget gate; i t 2 is the input gate of the second layer; o t 2 is the output gate of the second layer; x t is the input; h t 2 is the output value of the second layer, that is, the output at the current moment, h t 1 is the output value of the first layer; ht 3 is the output value of the third layer; W f 2 、W i 2 、W o 2 、W c 2 are all learning weights; b f 2 、b i 2 、b o 2 、b c 2 are bias; σ represents the sigmoid function; c t 2 represents the impact factor of the second layer output, c t 3 represents the third layer output impact factor, c t 1 is the first-level impact factor; Indicates the value of the gate when the second layer is updated;

[0114] Optimize the LSTM model training module; it is used to input the training set into the optimized LSTM model for iterative training. When the preset number of training times is reached, the training ends and the optimized LSTM model is obtained.

[0115] Optimize the LSTM model testing module; input the test set into the trained optimized LSTM model for testing to obtain the predicted albedo values ​​at different observation times.

[0116] Verification module: used to verify the albedo value obtained by the optimized LSTM model test module; the verification module specifically includes:

[0117] BRDF model calculation module, used to convert the solar incident angle θ in the test set i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 Input into the BRDF model to calculate the albedo value; the specific formula of the BRDF model is:

[0118] R(θ i1 ,θ V1 ,θ i2 ,θ v2 )=f iso +f geo k geo (θ i1 ,θ V1 ,θi2 ,θ v2 )+f vol k vol (θ i1 ,θ V1 ,θ i2 ,θ v2 )

[0119] Where, f iso represents the isotropic scattering kernel coefficient; f geo represents the geometric optics scattering kernel coefficient; f vol represents the scale factor of the volume scattering kernel; k geo represents the geometric optics kernel function; k vol is the volume scattering kernel function;

[0120] The comparison module is used to compare the calculated albedo value with the predicted albedo value obtained by the optimized LSTM model test module. If the difference between the two is within the preset value range, the predicted albedo value is considered correct; otherwise, the learning weights in the LSTM model are updated, and the LSTM model is retrained and tested.

Claims

1. A remote sensing image albedo inversion method based on LSTM network, characterized in that: The specific steps include: (1) Obtain the solar incidence azimuth θ at different observation times in the study area i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 ; (2) Obtain the albedo product data of the continuous time series of remote sensing images of the study area, sample the obtained albedo product data, and obtain the albedo value R corresponding to different observation times; (3) The solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth, satellite observation azimuth, and albedo values ​​corresponding to the same observation time are taken as a set of sequences D i ={θ i1 i ,θ i2 i ,θ v1 i ,θ v2 i ,R i }, get the data set consisting of all observation times t i is the observation time, n is a positive integer; the obtained data set D is divided into a training set and a test set; (4) Construct an optimized LSTM model. The input layer of the LSTM model is the solar incident azimuth angle θ i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 , the output of the output layer is the albedo value R; the formula of the optimized LSTM model is: f t 2 =σ(W f 2 ·[h t 1 ,x t ]+b f 2 ) i t 2 =σ(W i 2 ·[h t 1 ,x t ]+b i 2 ) the t 2 =σ(W o 2 ·[h t 1 ,x t ]+b o 2 ) h t 2 =o t 2 ×tanh(c t 2 )+h t 3 Where, f t 2 For the second layer of the forget gate; i t 2 is the input gate of the second layer; o t 2 is the second layer output gate; x t is the input; h t 2 is the output value of the second layer, that is, the output at the current moment, h t 1 is the output value of the first layer; h t 3 is the output value of the third layer; W f 2 、W i 2 、W o 2 、W c 2 are all learning weights; b f 2 、b i 2 、b o 2 、b c 2 are bias; σ represents the sigmoid function; c t 2 represents the impact factor of the second layer output, c t 3 represents the third layer output impact factor, c t 1 is the first-level impact factor; Indicates the value of the gate when the second layer is updated; The training set is input into the optimized LSTM model for iterative training. When the preset number of training times is reached, the training ends and the optimized LSTM model is obtained. (5) Input the test set into the trained optimized LSTM model for testing to obtain the predicted albedo values ​​at different observation times; (6) Verify the albedo value obtained in step (5). The specific verification steps include: (6.1) The solar incident azimuth angle θ in the test set is i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 Input into the BRDF model to calculate the albedo value; the specific formula of the BRDF model is: R(θ i1 ,i V1 ,i i2 ,i v2 )=f iso +f geo k geo (i i1 ,i V1 ,i i2 ,i v2 )+f vol k vol (i i1 ,i V1 ,i i2 ,i v2 ) Where, f iso represents the isotropic scattering kernel coefficient; f geo represents the geometric optics scattering kernel coefficient; f vol represents the scale factor of the volume scattering kernel; k geo represents the geometric optics kernel function; k vol is the volume scattering kernel function; (6.2) Compare the calculated albedo value with the predicted albedo value obtained in step (5). If the difference between the two is within the preset value range, the predicted albedo value is considered correct. Otherwise, the learning weights in the LSTM model are updated, and the LSTM model is retrained and tested.

2. The remote sensing image albedo inversion method based on LSTM network according to claim 1, characterized in that: The step (2) of obtaining the remote sensing image albedo product data of the continuous time series of the study area specifically includes the following steps: (2.1) Obtaining a continuous time series of remote sensing images of the study area, extracting image data from the remote sensing images; and dividing the obtained images into small blocks; (2.2) Use the cloud mask to filter out cloud-free image blocks to create a dataset, and put the cloud-free image blocks into a cloud-free folder; (2.3) Read the albedo data of the area corresponding to the image in the cloud-free folder.

3. The remote sensing image albedo inversion method based on LSTM network according to claim 1, characterized in that: The optimized LSTM model has three hidden layers.

4. A remote sensing image albedo inversion system based on LSTM network, characterized in that: Includes the following modules: The acquisition module is used to obtain the solar incidence azimuth θ at different observation times in the study area i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 ; Obtain the albedo product data of the continuous time series of remote sensing images of the study area, sample the obtained albedo product data, and obtain the albedo value R corresponding to different observation times; The data preprocessing module is used to process the solar incidence azimuth, solar incidence zenith angle, satellite observation azimuth, satellite observation azimuth, and albedo values ​​corresponding to the same observation time as a set of sequence D i ={θ i1 i ,θ i2 i ,θ v1 i ,θ v2 i ,R i }, get the data set consisting of all observation moments i is the observation time, n is a positive integer; the obtained data set D is divided into a training set and a test set; Optimize the LSTM model building module; used to build an optimized LSTM model. The input layer of the LSTM model is the solar incidence angle θ i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 , the output layer is the albedo value R; the formula of the optimized LSTM model is: f t 2 =σ(W f 2 ·[h t 1 ,x t ]+b f 2 ) i t 2 =σ(W i 2 ·[h t 1 ,x t ]+b i 2 ) the t 2 =σ(W o 2 ·[h t 1 ,x t ]+b o 2 ) h t 2 =o t 2 ×tanh(c t 2 )+h t 3 Where, f t 2 For the second layer of the forget gate; i t 2 is the input gate of the second layer; o t 2 is the second layer output gate; x t is the input; h t 2 is the output value of the second layer, that is, the output at the current moment, h t 1 is the output value of the first layer; h t 3 is the output value of the third layer; W f 2 、W i 2 、W o 2 、W c 2 are all learning weights; b f 2 、b i 2 、b o 2 、b c 2 are bias; σ represents the sigmoid function; c t 2 represents the impact factor of the second layer output, c t 3 represents the third layer output impact factor, c t 1 is the first-level impact factor; Indicates the value of the gate when the second layer is updated; Optimize LSTM model training module; It is used to input the training set into the optimized LSTM model for iterative training. When the preset number of training times is reached, the training ends and the optimized LSTM model is obtained. Optimize the LSTM model testing module; input the test set into the trained optimized LSTM model for testing to obtain the predicted albedo values ​​at different observation times; Verification module: used to verify the albedo value obtained by the optimized LSTM model test module; the verification module specifically includes: BRDF model calculation module, used to convert the solar incident angle θ in the test set i1 、Zenith angle of solar incidence θ i2 、Satellite observation azimuth θ v1 and the satellite observation zenith angle θ v2 Input into the BRDF model to calculate the albedo value; the specific formula of the BRDF model is: R(θ i1 ,i V1 ,i i2 ,i v2 )=f iso +f geo k geo (i i1 ,i V1 ,i i2 ,i v2 )+f vol k vol (i i1 ,i V1 ,i i2 ,i v2 ) Where, f iso represents the isotropic scattering kernel coefficient; f geo represents the geometric optics scattering kernel coefficient; f vol represents the scale factor of the volume scattering kernel; k geo represents the geometric optics kernel function; k vol is the volume scattering kernel function; The comparison module is used to compare the calculated albedo value with the predicted albedo value obtained by the optimized LSTM model test module. If the difference between the two is within the preset value range, the predicted albedo value is considered correct; otherwise, the learning weights in the LSTM model are updated, and the LSTM model is retrained and tested.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 3 when executing the computer program.

6. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.

Images