CN116384237A - Thermal infrared atmospheric parameter inversion method and device and electronic equipment - Google Patents

Abstract

The invention belongs to the technical field of remote sensing image processing, and relates to a thermal infrared atmospheric parameter inversion method, device and electronic equipment. The method includes: extracting surface emissivity information and atmospheric information; determining atmospheric parameters and establishing a simulated data set; constructing a two-way long-term short-term memory Neural network model; model training to determine the structural parameters of the model; use the attention mechanism to weight the information of each channel of the output feature vector of the two-way long-term short-term memory neural network model; model training; update iteratively until the output of the model converges, and obtain the atmospheric parameter feedback Inversion model; obtain thermal infrared atmospheric parameter inversion results. The present invention effectively solves the problem that the deep neural network has weak data mining ability, proposes a two-way long-term and short-term memory neural network structure, uses the channel attention module to weight different channel information, and obtains the correlation characteristics of the radiance data of different channels , to obtain the inversion results of thermal infrared atmospheric parameters.

Description

Thermal infrared atmospheric parameter inversion method, device and electronic equipment

technical field

The invention belongs to the technical field of remote sensing image processing, and in particular relates to an attention mechanism-based long-short-term memory neural network thermal infrared atmospheric parameter inversion method, device and electronic equipment.

Background technique

Atmospheric parameter inversion has received extensive attention in recent years and has been widely used in many disciplines. The inversion of atmospheric parameters has also aroused the interest of many scholars, such as modeling the correlation of atmospheric transmittance at different wavelengths, inverting and calculating the atmospheric transmittance of radiation near wavelengths; by modeling water vapor and atmospheric transmittance in different bands The statistical relationship of the rate is used to obtain the transmittance. In the actual application of the existing atmospheric parameter inversion method, the observed brightness temperature is obtained from the actual observation of the microwave radiometer, and its spatial resolution is low, resulting in low accuracy of the atmospheric parameter inversion, and the deep neural network has the problem of weak data mining ability .

Contents of the invention

In order to solve the above technical problems, the present invention provides a thermal infrared atmospheric parameter inversion method, device and electronic equipment.

In the first aspect, the present invention provides a thermal infrared atmospheric parameter inversion method, including:

Extract the surface emissivity information and atmospheric information from the object spectral library and atmospheric profile library;

determining atmospheric parameters and establishing a simulation data set according to the surface emissivity information and the atmospheric information;

Construct a bidirectional long-short-term memory neural network model;

Using the hyperspectral data in the simulation data set to train the bidirectional long-short-term memory neural network model, and determine the structural parameters of the bidirectional long-term short-term memory neural network model;

Using the attention mechanism to weight each channel information of the output feature vector of the two-way long-short-term memory neural network model to obtain the target two-way long-short-term memory neural network model;

performing data training on the target bidirectional long-short-term memory neural network model through the atmospheric parameters in the simulated data set;

Using a backpropagation algorithm to continuously iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model until the output of the target bidirectional long-short-term memory neural network model converges to obtain an atmospheric parameter inversion model;

The thermal infrared hyperspectral radiance data is obtained and used as the input of the atmospheric parameter inversion model to obtain the thermal infrared atmospheric parameter inversion result.

In a second aspect, the present invention provides a thermal infrared atmospheric parameter inversion device, including an extraction unit, a simulated data set establishment unit, a first model construction unit, a first model training unit, a second model construction unit, a second model training unit, Iterative update unit, input unit and output unit;

The extraction unit is used to extract the surface emissivity information and atmospheric information of the surface object spectrum library and the atmospheric profile library;

The simulation data set establishment unit is used to determine atmospheric parameters and establish a simulation data set according to the surface emissivity information and the atmospheric information;

The first model construction unit is configured to construct a bidirectional long-short-term memory neural network model;

The first model training unit is configured to use the hyperspectral data in the simulated data set to train the bidirectional long-short-term memory neural network model, and determine structural parameters of the bidirectional long-short-term memory neural network model;

The second model construction unit is configured to use an attention mechanism to weight each channel information of the output feature vector of the bidirectional long-short-term memory neural network model to obtain a target bidirectional long-short-term memory neural network model;

The second model training unit is configured to perform data training on the target bidirectional long-short-term memory neural network model through the atmospheric parameters in the simulated data set;

The iterative update unit is used to continuously iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model by using a back-propagation algorithm until the output of the target bidirectional long-term short-term memory neural network model converges to obtain atmospheric parameter responses. play model;

The input unit is used to obtain thermal infrared hyperspectral radiance data and serve as the input of the atmospheric parameter inversion model;

The output unit is used to output the inversion results of thermal infrared atmospheric parameters.

In a third aspect, the present invention discloses an electronic device, comprising:

processor and memory;

The memory is used to store computer operation instructions;

The processor is configured to execute the thermal infrared atmospheric parameter inversion method by invoking the computer operation instruction.

The beneficial effects of the present invention are: the present invention not only effectively solves the problem of weak data mining ability of the deep neural network, but also strengthens the network's emphasis on the correlation information between hyperspectral adjacent channel data during the inversion process. The present invention proposes a two-way long-term and short-term memory neural network structure, which is used to extract the correlation features of different channels; then use the channel attention module to weight the information of different channels, so that the proportion of channels with higher weights in the features is increased; finally The fully connected structure is used to map the features to the sample label space, and the features are converted into the estimation results of the atmospheric uplink radiation, downlink radiation and transmittance of each channel; The cyclic neural network module of the network adds an attention mechanism to obtain the correlation between radiance and atmospheric parameters, and inverts the upward radiation, downward radiation and transmittance of the atmosphere.

On the basis of the above technical solutions, the present invention can also be improved as follows.

Further, the bidirectional long-short-term memory neural network model includes two LSTM neural networks whose parameters are independent of each other, and the hyperspectral data in the simulated data set are respectively input into the LSTM neural network in forward and reverse order for feature extraction , two feature extraction vectors are obtained, and the feature extraction vectors are concatenated to form an output vector.

Further, using the attention mechanism to weight each channel information of the output feature vector of the bidirectional long-short-term memory neural network model, including:

Through the convolution operation, the input data with the given number of feature channels as the first channel number is transformed to obtain a feature vector with the number of feature channels as the second channel number;

Using a globalization method, compressing the feature vector in the spatial dimension, compressing the two-dimensional feature vector of each channel into a real number, as the feature value of each channel;

Forming a Bottleneck structure through several fully connected layers generates weight values for the eigenvalues of each channel, and normalizes the weight values;

The normalization weight is weighted to the feature value of each channel, and the weight value and the feature vector are weighted and summed.

Further, the fully-connected layer uses three-layer fully-connected modules for mapping; the fully-connected modules of adjacent layers are activated using activation functions.

Further, using the backpropagation algorithm to iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model until the output of the target bidirectional long-short-term memory neural network model converges to obtain an atmospheric parameter inversion model, including:

Obtain hyperspectral atmospheric parameter inversion data from the simulated data set as training data;

Read the atmospheric parameters corresponding to the radiance under several different atmospheric parameters from the tag data, and pair the atmospheric parameters with the radiance information under the same atmospheric parameter state;

Using the backpropagation algorithm to update the parameters of the target bidirectional long-term short-term memory neural network model, input the radiance under different states, so that the root mean square error value of the target bidirectional long-term short-term memory neural network model gradually decays to a set value ;

The mean square error is used as the loss function, and the network module with the smallest final loss function is selected as the atmospheric parameter inversion model.

Further, by setting the channel attention module and setting the number of down-sampling channels at equal intervals in the channel attention module, the channel attention module is determined according to the convergence speed of the target bidirectional long-short-term memory neural network model The number of downsampling channels.

Description of drawings

Fig. 1 is the flowchart of the thermal infrared atmospheric parameter inversion method provided by embodiment 1 of the present invention;

Fig. 2 is a schematic diagram of the thermal infrared atmospheric parameter inversion device provided by Embodiment 1 of the present invention;

FIG. 3 is a schematic diagram of an electronic device provided by Embodiment 3 of the present invention.

Icons: 30-electronic device; 310-processor; 320-bus; 330-memory; 340-transceiver.

Detailed ways

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 It is a part of embodiments of the present invention, but not all embodiments. The components of the embodiments of the invention generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations.

Example 1

As an embodiment, as shown in Figure 1, in order to solve the above technical problems, this embodiment provides a method for inversion of thermal infrared atmospheric parameters, including:

Extract the surface emissivity information and atmospheric information from the object spectral library and atmospheric profile library;

According to the surface emissivity information and atmospheric information, determine the atmospheric parameters and establish a simulation data set;

Construct a bidirectional long-short-term memory neural network model;

Using the hyperspectral data in the simulated data set to train the bidirectional long-term short-term memory neural network model, and determine the structural parameters of the bidirectional long-term short-term memory neural network model;

Using the attention mechanism to weight the information of each channel of the output feature vector of the bidirectional long-term short-term memory neural network model to obtain the target bidirectional long-term short-term memory neural network model;

Perform data training on the target bidirectional long-short-term memory neural network model by simulating the atmospheric parameters in the data set;

Using the backpropagation algorithm to iteratively update the parameters of the trained target bidirectional long-term short-term memory neural network model until the output of the target bidirectional long-term short-term memory neural network model converges to obtain an atmospheric parameter inversion model;

Obtain the thermal infrared hyperspectral radiance data and use it as the input of the atmospheric parameter inversion model to obtain the thermal infrared atmospheric parameter inversion result.

In the actual application process, the surface emissivity information and atmospheric information of the surface object spectrum library and the thermodynamic initial guess retrieval (TIGR) library (thermodynamic initial guess retrieval, TIGR) are extracted, and the atmospheric parameters are calculated using MODTRAN to establish a simulation data set. The hyperspectral data is obtained by simulating the data set, the radiance under different emissivity and the atmospheric parameters corresponding to the radiance, and the atmospheric parameters are paired with the radiance information under the same state. The original database contains 2,311 atmospheric profiles in different regions. After cloud removal and water vapor and bottom temperature screening, 945 atmospheric profiles were selected for the experiment, and different surface emissivity information was extracted using the ground object emissivity library. 66 different ground objects emissivity, including bare soil, vegetation, water and other object types, using MODTRAN to simulate the radiance, by setting different surface objects and different atmospheric profiles, a total of 74844000 pieces of simulated radiance data were generated, using the arithmetic mean (The average value) 10 adjacent wave numbers are processed into one band, and a total of 24 bands are obtained.

Optionally, the bidirectional long-short-term memory neural network model includes two LSTM neural networks whose parameters are independent of each other. The hyperspectral data in the simulated data set are input to the LSTM neural network in forward and reverse order for feature extraction, and two feature extractions are obtained. Vector, the feature extraction vector is concatenated to form an output vector.

In the actual application process, a bidirectional long-term short-term memory neural network model is constructed, and the feature extraction of the simulated data set is carried out by using a bidirectional long-term short-term memory neural network. The long-term short-term neural network operation maps the input data set X of size H×W×S to H The output data set Y of the size '×W'×T, the formula is as follows:

为输入数据；C_m,n,s,t为当前某一位置的权重；b_m,n,s,t为当前某一位置的偏置；H为数据的批次；W为通道数；S为数据长度；m为当前批次；n为当前通道数；s为当前数据长度；t为时刻。Y _{h', w', t} are the output data;

is the input data; C _{m, n, s, t} is the weight of a certain current position; b _{m, n, s, t} is the bias of a certain current position; H is the batch of data; W is the number of channels; S is the data length; m is the current batch; n is the current channel number; s is the current data length; t is the time.

The specific calculation process is:

f(t)=σ(Wfh _t-1 +Ufx _t +bf);

i(t)=σ(W _i h _t-1 + U _i x _t + b _i );

a(t)=tanh(W _a h _t-1 +U _a x _t +b _a );

o(t)=σ(W _o h _t-1 +U _o x _t +b _o );

Among them, x _t represents the input at time t, h _t-1 represents the state value of the hidden layer at time t-1; W _f , W _i , W _o and W _a represent the forgetting gate, input gate, output gate and feature extraction process respectively The weight coefficient of h _t-1 in ; U _f , U _i , U _o and U _a represent the weight coefficient of x _t in the process of forgetting gate, input gate, output gate and feature extraction respectively; b _f , b _i , b _o and b _{and a} respectively represent the bias values in the process of forgetting gate, input gate, output gate and feature extraction, tanh(·) represents the tangent hyperbolic function, σ(·) represents the activation function Sigmoid, expressed as:

The results calculated by the forget gate and the input gate act on c(t-1) to form the variable state c(t) at time t, which is expressed as:

c(t)=c(t-1)⊙f(t)+i(t)⊙a(t);

Among them, ⊙ is the Hadamard product, f(t) is the information forgotten by the forgetting gate at time t; i(t) is the updated information of the input gate at time t; a(t) is the transition value to the variable state at time t. Finally, the state h(t) of the hidden layer at time t is obtained from the output gate o(t) and the variable state c(t) at the current moment, expressed as:

h(t)=o(t)⊙tanh(c(t));

The bidirectional long-term and short-term neural network module is composed of two LSTMs whose parameters are independent of each other. The data set is input to the two LSTM neural networks in forward and reverse order respectively for feature extraction, and two output vectors (namely, feature extraction vectors) are spliced to form The new feature vector of , the new vector is expressed as the final feature of the data. Under this concept, the bidirectional LSTM module enables the feature data obtained at time t to have both past and future correlation information.

Optionally, use the attention mechanism to weight the information of each channel of the output feature vector of the bidirectional long-short-term memory neural network model, including:

Through the convolution operation, the input data with the given number of feature channels as the first channel number is transformed to obtain a feature vector with the number of feature channels as the second channel number;

Use the globalization method to compress the feature vector in the spatial dimension, compress the two-dimensional feature vector of each channel into a real number, and use it as the feature value of each channel;

The Bottleneck structure is composed of several fully connected layers to generate weight values for the feature values of each channel, and normalize the weight values;

The normalized weights are weighted to the eigenvalues of each channel, and the weighted sums are weighted with the eigenvectors.

In the actual application process, through the convolution operation, for a given input transformation with the number of feature channels T, a feature with the number of feature channels T ¹ is obtained; the global pooling method is used to compress the feature in the spatial dimension, and each The two-dimensional features of channels are compressed into a real number, and the image area corresponding to the feature value is obtained; among them, the C-th element Z _C value of the calculation feature Z is expressed as:

Among them, H is the height of the image; W is the width of the image; H×W is the two-dimensional feature of the channel; i is the counter accumulated on the height; j is the counter accumulated on the width; u _c is the height at the Cth channel is the pixel value at i and width j;

The Bottleneck structure is composed of two fully connected layers to evaluate the importance of each channel, and a weight value is generated for each feature channel of _ZC through Bottleneck, and the weight is normalized; finally, the normalized weight is weighted to each On the features of each channel, the weights and feature vectors are weighted and summed.

Optionally, the fully connected layer uses three layers of fully connected modules for mapping; the fully connected modules of adjacent layers are activated using activation functions.

In the actual application process, the channel weighted features are output to the fully connected layer, and the extracted channel features are mapped to the sample label space. After testing, when the network uses a three-layer fully connected layer structure, the regression atmospheric parameters are stable and accurate. higher degree.

Optionally, the backpropagation algorithm is used to iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model until the output of the target bidirectional long-term short-term memory neural network model converges to obtain an atmospheric parameter inversion model, including:

Obtain hyperspectral atmospheric parameter inversion data from the simulated data set as training data;

Read the atmospheric parameters corresponding to the radiance under several different atmospheric parameters from the tag data, and pair the atmospheric parameters with the radiance information under the same atmospheric parameter state;

Use the backpropagation algorithm to update the parameters of the target bidirectional long-term short-term memory neural network model, and input the radiance in different states, so that the root mean square error value of the target bidirectional long-term short-term memory neural network model gradually decays to the set value;

The mean square error is used as the loss function, and the network module with the smallest final loss function is selected as the atmospheric parameter inversion model.

In the actual application process, the present invention uses the channel attention mechanism and the combined structure of the fully connected layer to form a channel attention module, and upgrades the spectral features extracted by the long-term and short-term networks. The condition of whether the channel attention module converges is used as a judgment indicator, and the number of downsampling channels of the attention module is finally determined by setting the number of downsampling channels of multiple channel attention modules at equal intervals, and the extracted features are mapped to the fully connected layer (Linear), and the features are normalized by the normalization function, and the Sigmoid activation function is used for feature activation. Use a three-layer fully connected module (Linear1, Linear2, Linear3) for feature mapping. After testing, use the Sigmoid activation function between Linear1 and Linear2 for feature activation, and use the Softmax activation function between Linear2 and Linear3 for feature activation.

The features extracted by the channel attention module are mapped to the sample space through a multi-layer fully connected module, and the output of the module is the required atmospheric parameters.

Optionally, by setting the channel attention module and setting the number of downsampling channels at equal intervals in the channel attention module, determine the downsampling channel of the channel attention module according to the convergence speed of the target bidirectional long-term short-term memory neural network model number.

The invention not only effectively solves the problem that the deep neural network has weak data mining ability, but also strengthens the network's emphasis on the correlation information between hyperspectral adjacent channel data during the inversion process. The present invention proposes a two-way long-term and short-term memory neural network structure, which is used to extract the correlation features of different channels; then use the channel attention module to weight the information of different channels, so that the proportion of channels with higher weights in the features is increased; finally The fully connected structure is used to map the features to the sample label space, and the features are transformed into the estimation results of the atmospheric uplink radiation, downlink radiation and transmittance of each channel.

Example 2

Based on the same principle as the method shown in Embodiment 1 of the present invention, as shown in Figure 2, a thermal infrared atmospheric parameter inversion device is also provided in the embodiment of the present invention, including an extraction unit and an analog data set establishment unit , a first model construction unit, a first model training unit, a second model construction unit, a second model training unit, an iterative update unit, an input unit and an output unit;

The extraction unit is used to extract the surface emissivity information and atmospheric information of the object spectrum library and the atmospheric profile library;

The simulation data set establishment unit is used to determine the atmospheric parameters and establish the simulation data set according to the surface emissivity information and atmospheric information;

The first model construction unit is used to construct a bidirectional long-short-term memory neural network model;

The first model training unit is used to use the hyperspectral data in the simulated data set to train the bidirectional long-term short-term memory neural network model, and determine the structural parameters of the bidirectional long-term short-term memory neural network model;

The second model construction unit is used to weight each channel information of the output feature vector of the two-way long-short-term memory neural network model by using the attention mechanism to obtain the target two-way long-short-term memory neural network model;

The second model training unit is used to perform data training on the target bidirectional long-short-term memory neural network model by simulating the atmospheric parameters in the data set;

The iterative update unit is used to continuously iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model by using the back propagation algorithm until the output of the target bidirectional long-term short-term memory neural network model converges to obtain an atmospheric parameter inversion model;

The input unit is used to obtain thermal infrared hyperspectral radiance data and serve as the input of the atmospheric parameter inversion model;

The output unit is used to output the retrieval results of thermal infrared atmospheric parameters.

Optionally, the bidirectional long-short-term memory neural network model includes two LSTM neural networks whose parameters are independent of each other. The hyperspectral data in the simulated data set are input to the LSTM neural network in forward and reverse order for feature extraction, and two feature extractions are obtained. Vector, the feature extraction vector is concatenated to form an output vector.

Optionally, use the attention mechanism to weight the information of each channel of the output feature vector of the bidirectional long-short-term memory neural network model, including:

Through the convolution operation, the input data with the given number of feature channels as the first channel number is transformed to obtain a feature vector with the number of feature channels as the second channel number;

Use the globalization method to compress the feature vector in the spatial dimension, compress the two-dimensional feature vector of each channel into a real number, and use it as the feature value of each channel;

The Bottleneck structure is composed of several fully connected layers to generate weight values for the feature values of each channel, and normalize the weight values;

The normalized weights are weighted to the eigenvalues of each channel, and the weighted sums are weighted with the eigenvectors.

Optionally, the fully connected layer uses three layers of fully connected modules for mapping; the fully connected modules of adjacent layers are activated using activation functions.

Optionally, the backpropagation algorithm is used to iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model until the output of the target bidirectional long-term short-term memory neural network model converges to obtain an atmospheric parameter inversion model, including:

Obtain hyperspectral atmospheric parameter inversion data from the simulated data set as training data;

Read the atmospheric parameters corresponding to the radiance under several different atmospheric parameters from the tag data, and pair the atmospheric parameters with the radiance information under the same atmospheric parameter state;

Use the backpropagation algorithm to update the parameters of the target bidirectional long-term short-term memory neural network model, and input the radiance in different states, so that the root mean square error value of the target bidirectional long-term short-term memory neural network model gradually decays to the set value;

The mean square error is used as the loss function, and the network module with the smallest final loss function is selected as the atmospheric parameter inversion model.

Optionally, by setting the channel attention module and setting the number of downsampling channels at equal intervals in the channel attention module, determine the downsampling channel of the channel attention module according to the convergence speed of the target bidirectional long-term short-term memory neural network model number.

Example 3

Based on the same principle as the method shown in the embodiment of the present invention, an electronic device is also provided in the embodiment of the present invention, as shown in Figure 3, the electronic device may include but not limited to: a processor and a memory The memory is used to store the computer program; the processor is used to execute the method shown in any embodiment of the present invention by calling the computer program.

An electronic device is provided in an optional embodiment, and the electronic device 30 shown in FIG. 3 includes: a processor 310 and a memory 350 . Wherein, the processor 310 is connected to the memory 350 , such as through a bus 320 .

Optionally, the electronic device 30 may further include a transceiver 340, and the transceiver 340 may be used for data interaction between the electronic device and other electronic devices, such as sending and/or receiving data. It should be noted that in practical applications, the transceiver 340 is not limited to one, and the structure of the electronic device 30 does not limit the embodiment of the present invention.

The processor 310 may be a CPU central processing unit, a general purpose processor, a DSP data signal processor, an ASIC application specific integrated circuit, an FPGA field programmable gate array or other programmable logic devices, hardware components or any combination thereof. The processor 310 may also be a combination that implements computing functions, for example, a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

Bus 320 may include a path for communicating information between the components described above. The bus 320 may be a PCI peripheral component interconnect standard bus or an EISA extended industry standard architecture bus or the like. The bus 320 can be divided into a control bus, a data bus, an address bus, and the like. For ease of representation, only one thick line is used in FIG. 3 , but it does not mean that there is only one bus or one type of bus.

The memory 350 can be a ROM read-only memory or other types of static storage devices that can store static information and instructions, a RAM random access memory or other types of dynamic storage devices that can store information and instructions, or an EEPROM electrically erasable programmable memory device. Read-only memory, CD-ROM or other optical disc storage, optical disc storage (including optical discs, laser discs, compact discs, digital versatile discs, etc.), magnetic disk storage media, or capable of carrying or storing information in the form of instructions or data structures desired program code and any other medium that can be accessed by a computer, but not limited thereto.

The memory 350 is used to store application program codes (computer programs) for implementing the solution of the present invention, and the execution is controlled by the processor 310 . The processor 310 is configured to execute the application program codes stored in the memory 350, so as to implement the contents shown in the foregoing method embodiments.

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 changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (8)

1. The thermal infrared atmospheric parameter inversion method is characterized in that, comprising: Extract the surface emissivity information and atmospheric information from the object spectral library and atmospheric profile library; determining atmospheric parameters and establishing a simulation data set according to the surface emissivity information and the atmospheric information; Construct a bidirectional long-short-term memory neural network model; Using the hyperspectral data in the simulation data set to train the bidirectional long-short-term memory neural network model, and determine the structural parameters of the bidirectional long-term short-term memory neural network model; Using the attention mechanism to weight each channel information of the output feature vector of the two-way long-short-term memory neural network model to obtain the target two-way long-short-term memory neural network model; performing data training on the target bidirectional long-short-term memory neural network model through the atmospheric parameters in the simulated data set; Using a backpropagation algorithm to continuously iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model until the output of the target bidirectional long-short-term memory neural network model converges to obtain an atmospheric parameter inversion model; The thermal infrared hyperspectral radiance data is obtained and used as the input of the atmospheric parameter inversion model to obtain the thermal infrared atmospheric parameter inversion result. 2. according to the described thermal infrared atmospheric parameter inversion method of claim 1, it is characterized in that, described two-way long-short-term memory neural network model comprises the LSTM neural network that two parameters are independent of each other, the described high-level in the described simulated data set Spectral data is input to the LSTM neural network in forward and reverse order respectively for feature extraction, and two feature extraction vectors are obtained, and the feature extraction vectors are concatenated to form an output vector. 3. according to the described thermal infrared atmospheric parameter inversion method of claim 1, it is characterized in that, utilize attention mechanism to carry out weighting to each channel information of the output feature vector of described two-way long-short-term memory neural network model, comprising: Through the convolution operation, the input data with the given number of feature channels as the first channel number is transformed to obtain a feature vector with the number of feature channels as the second channel number; Using a globalization method, compressing the feature vector in the spatial dimension, compressing the two-dimensional feature vector of each channel into a real number, as the feature value of each channel; Forming a Bottleneck structure through several fully connected layers generates weight values for the eigenvalues of each channel, and normalizes the weight values; The normalization weight is weighted to the feature value of each channel, and the weight value and the feature vector are weighted and summed. 4. The method for inversion of thermal infrared atmospheric parameters according to claim 3, wherein the fully-connected layer is mapped using three-layer fully-connected modules; activation functions are used between fully-connected modules of adjacent layers to activate. 5. according to the described thermal infrared atmospheric parameter inversion method of claim 1, it is characterized in that, adopt backpropagation algorithm to the described target two-way long-short-term memory neural network model parameter after training iteratively update until described target two-way long-short-term The output of the memory neural network model converges, and the atmospheric parameter inversion model is obtained, including: Obtain hyperspectral atmospheric parameter inversion data from the simulated data set as training data; Read the atmospheric parameters corresponding to the radiance under several different atmospheric parameters from the tag data, and pair the atmospheric parameters with the radiance information under the same atmospheric parameter state; Using the backpropagation algorithm to update the parameters of the target bidirectional long-term short-term memory neural network model, input the radiance under different states, so that the root mean square error value of the target bidirectional long-term short-term memory neural network model gradually decays to a set value ; The mean square error is used as the loss function, and the network module with the smallest final loss function is selected as the atmospheric parameter inversion model. 6. according to the described thermal infrared atmospheric parameter inversion method of claim 1, it is characterized in that, by arranging channel attention module, and the mode of the down-sampling channel number of equal intervals is set in described channel attention module, according to the target The convergence speed of the bidirectional long-short-term memory neural network model determines the number of down-sampling channels of the channel attention module. 7. The thermal infrared atmospheric parameter inversion device is characterized in that it includes an extraction unit, a simulated data set establishment unit, a first model construction unit, a first model training unit, a second model construction unit, a second model training unit, an iterative update unit, input unit and output unit; The extraction unit is used to extract the surface emissivity information and atmospheric information of the surface object spectrum library and the atmospheric profile library; The simulation data set establishment unit is used to determine atmospheric parameters and establish a simulation data set according to the surface emissivity information and the atmospheric information; The first model construction unit is configured to construct a bidirectional long-short-term memory neural network model; The first model training unit is configured to use the hyperspectral data in the simulated data set to train the bidirectional long-short-term memory neural network model, and determine structural parameters of the bidirectional long-short-term memory neural network model; The second model construction unit is configured to use an attention mechanism to weight each channel information of the output feature vector of the bidirectional long-short-term memory neural network model to obtain a target bidirectional long-short-term memory neural network model; The second model training unit is configured to perform data training on the target bidirectional long-short-term memory neural network model through the atmospheric parameters in the simulated data set; The iterative update unit is used to continuously iteratively update the parameters of the trained target bidirectional long-short-term memory neural network model by using a back-propagation algorithm until the output of the target bidirectional long-term short-term memory neural network model converges to obtain atmospheric parameter responses. play model; The input unit is used to obtain thermal infrared hyperspectral radiance data and serve as the input of the atmospheric parameter inversion model; The output unit is used to output the inversion results of thermal infrared atmospheric parameters. 8. An electronic device, characterized in that it comprises: processor and memory; The memory is used to store computer operation instructions; The processor is configured to execute the thermal infrared atmospheric parameter inversion method according to any one of claims 1 to 6 by invoking the computer operation instructions.

Images