SP-URNet-based marine non-uniform troposphere atmospheric waveguide propagation loss space-time prediction algorithm

Abstract

The invention discloses an SP-URNet-based marine heterogeneous troposphere atmospheric waveguide propagation loss space-time prediction algorithm, which comprises the following steps: step 1, parameterization calculation of an atmospheric waveguide atmospheric correction refractive index M and waveguide propagation loss image data calculation; step 2, spatial feature extraction modeling of a three-dimensional convolution automatic encoder; a U-TCN module formed by a time convolution network TCN performs modeling on beyond visual range propagation loss time features; a fourth step of integrating time and space features of a three-dimensional convolution automatic decoder; and a fifth step of establishing an SP-URNet prediction network model. According to the algorithm disclosed by the invention, the defect that space-time prediction and nonlinear prediction cannot be carried out on offshore non-uniform troposphere atmospheric waveguide propagation loss data in the prior art is overcome.

Description

Technical Field

[0001] The present invention belongs to the field of ocean atmosphere remote sensing, and in particular relates to a spatiotemporal prediction algorithm for sea clutter non-uniform atmospheric waveguide propagation loss based on a deep convolutional network in the field. Background Art

[0002] The marine tropospheric atmospheric waveguide is a special radio wave propagation environment. When electromagnetic waves propagate in the tropospheric waveguide, there is an important concept - propagation loss, which refers to the reduction of signal strength when electromagnetic waves pass through the medium, and experience attenuation or weakening along the way. It can reflect the trapping characteristics of the tropospheric waveguide structure on electromagnetic waves and plays a vital role in long-distance communications. Predicting the time and space propagation loss of marine atmospheric waveguides has an important impact on various fields such as communication, navigation, scientific research and emergency response.

[0003] For a long time, the main methods for calculating electromagnetic wave propagation loss in the ocean atmospheric duct environment are: theoretical propagation modeling, numerical simulation, and parabolic equation (PE) calculation. Theoretical propagation models are all based on the assumption that the atmospheric environment is uniform, while the actual atmospheric environment shows discontinuous changes over time. Numerical simulation methods usually consume a lot of computing resources. The PE method can easily solve long-distance problems with complex atmospheric refraction and irregular terrain. However, in the actual atmospheric duct environment, the quantitative calculation of the propagation loss of the non-uniform atmospheric duct has a large error. Summary of the invention

[0004] In order to solve the technical problem that the spatiotemporal characteristics modeling capability of offshore non-uniform atmospheric duct propagation is insufficient, which restricts the development of spatiotemporal prediction of atmospheric duct propagation loss, the present invention provides a spatiotemporal prediction algorithm for offshore non-uniform tropospheric atmospheric duct propagation loss based on SP-URNet, which realizes the spatiotemporal prediction of non-uniform atmospheric duct propagation loss based on sea clutter data, and supports the research on offshore atmospheric remote sensing related technologies.

[0005] The present invention adopts the following technical solution:

[0006] A spatiotemporal prediction algorithm for atmospheric waveguide propagation loss in marine non-uniform troposphere based on SP-URNet, the improvement of which is that it includes the following steps:

[0007] Step 1: Parameterization calculation of atmospheric waveguide atmospheric correction refractive index M and waveguide propagation loss image data calculation:

[0008] Firstly, the atmospheric temperature, humidity, wind direction, pressure and atmospheric height in the non-uniform atmospheric waveguide environment are obtained by reanalyzing data in the ERA5 model, and the atmospheric corrected refractive index M is calculated. Based on the different atmospheric refractive indices corresponding to different altitudes above the sea surface, the spatiotemporal variation diagram of the propagation loss is obtained.

[0009] Step 2: 3D convolutional autoencoder spatial feature extraction modeling:

[0010] A spatial information autoencoder composed of three-dimensional convolution, normalization and activation function is used to realize the feature extraction of nonlinear mapping relationship of inhomogeneous atmospheric waveguide. The input layer inputs data with a time series length of T. The data includes images of the input frame sequence length and data of the frame sequence length to be predicted. The shape is (B×T)×C×H×W. Each frame is regarded as a single sample and focuses on the spatial features of a single frame level without considering the time change. The output data is input to the time information learning module, and the output shape is reshaped to B×(T×C)×H×W.

[0011] Step 3: The U-TCN module composed of the time convolutional network TCN models the temporal characteristics of over-the-horizon propagation loss:

[0012] A temporal information learning module parser composed of a U-net type connected temporal convolutional network is proposed for feature extraction. The input is a tensor of shape B×(T×C)×H×W reshaped by the spatial information learning module. The features are extracted by stacking along the time axis. The output data with spatial and temporal information is input into the spatiotemporal information integration learning module and reshaped into (B×T)×C×H×W.

[0013] Step 4, 3D convolutional automatic decoder time and space feature integration:

[0014] A decoder that automatically integrates learned temporal and spatial information is constructed using three-dimensional convolution, normalization, and activation functions. The input is a tensor with the shape of B×(T×C)×H×W with learned temporal and spatial information, and the output shape is T'×C×H×W.

[0015] Step 5: SP-URNet prediction network model construction:

[0016] The model structure includes: input data, model parameter setting, spatial encoder, temporal parser, spatiotemporal integration decoder construction and network output;

[0017] The input data comes from the atmospheric refractive index data calculated by ERA5 model reanalysis data, and the atmospheric waveguide propagation loss image data is obtained, and the data is divided into input frame data and predicted frame data;

[0018] The parameters used are Adam optimizer, leakyReLU activation function, weight loss, dilated convolution, and weight normalization;

[0019] Based on the training of network model parameters, the most accurate spatiotemporal prediction results of propagation loss are output and the model is saved.

[0020] Furthermore, in step 2, after N cycles of convolutional layers, normalization, and activation functions, the input shape of [T, C, H, W] data is converted to [T×C, H / 2N, W / 2N], the data is converted from four-dimensional to three-dimensional data, and the image size is reduced. The formula is as follows:

[0021] z i = σ(LayerNorm(Conv2d(z i-1 ))),1 ≤ i ≤ Ns, (1)

[0022] The input z i-1 and output z i The shapes are (T, C, H, W) and

[0023] Furthermore, in step 3, the parser is composed of 6 temporal convolutional network modules TCN connected in the form of a U-Net architecture. The temporal convolutional network TCN consists of two branches. The first branch performs 1*1Conv2d convolution, and the other branch consists of four modules: dilated convolution, standardization, activation function, and regularization. After the two branches complete the operation, residual connection is performed. The formula is as follows:

[0024] z j = TCN(z j -1),Ns < j ≤ Ns + Nt (2)

[0025] The input z j-1 and output z j The shape is (T×C, H, W) and

[0026] Furthermore, in step 4, the decoder reconstructs the spatiotemporal live frame, which requires the reconstruction and decoding of the spatial information of the encoder and the temporal information parameters of the parser. In the decoding stage, after N cycles of convolutional layers, normalization and activation functions, the input shape is The data is converted into [T′, C, H, W], the data is converted from three-dimensional to four-dimensional data, and the image size is enlarged to finally obtain data with the same dimension and size as the original data. The formula is as follows:

[0027] z k = σ(GroupNorm(unConv2d(z k-1 ))) , Ns + Nt < k ≤ 2Ns + Nt (3)

[0028] The input z k-1 and output z k The shapes are and (T′, C, H, W).

[0029] The beneficial effects of the present invention are:

[0030] The algorithm disclosed in the present invention overcomes the lack of existing technologies in meeting the requirements for spatiotemporal prediction and nonlinear prediction of atmospheric waveguide propagation loss data in the non-uniform troposphere at sea. The image data of atmospheric waveguide propagation loss at different times and altitudes is used as input, and the image data of atmospheric waveguide propagation loss after prediction is used as output. The nonlinear mapping relationship between the two is obtained through the constructed SP-URNet network model, which realizes the effective prediction of non-uniform atmospheric waveguide propagation loss of sea clutter, the effective prediction of atmospheric waveguide propagation loss in the offshore troposphere, and the extraction and accurate prediction of spatiotemporal characteristics of long-distance propagation loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the flow chart of the algorithm of the present invention;

[0032] Figure 2 It is a schematic diagram of the SP-URNet model;

[0033] Figure 3 This is a schematic diagram of the temporal convolutional network TCN module;

[0034] Figure 4 This is the process details diagram of the SP-URNet model;

[0035] Figure 5 It is a graph showing the spatiotemporal prediction results of atmospheric waveguide propagation loss data by various models;

[0036] Figure 6 This is a comparison chart of the prediction results of different models at different heights. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] Example 1. This example discloses a spatiotemporal prediction algorithm for marine non-uniform tropospheric atmospheric waveguide propagation loss based on SP-URNet, which realizes the spatiotemporal feature extraction and modeling of marine non-uniform atmospheric waveguide propagation loss. On this basis, the atmospheric waveguide propagation loss image data at different times and different heights is used as input, and the future atmospheric waveguide propagation loss image data is used as output. The nonlinear mapping relationship between the two is obtained through the constructed SP-URNet network model, so as to realize the effective prediction of the sea clutter non-uniform atmospheric waveguide propagation loss in time and space, such as Figure 1As shown, the specific steps include:

[0039] Step 1: Parameterization calculation of atmospheric waveguide atmospheric correction refractive index M and waveguide propagation loss image data calculation:

[0040] In order to obtain the atmospheric duct propagation loss data, the atmospheric temperature, humidity, wind direction, pressure and atmospheric height in the non-uniform atmospheric duct environment are firstly obtained through the ERA5 model reanalysis data, and the atmospheric correction refractive index M is calculated. According to the different atmospheric refractive indices corresponding to different altitudes above the sea surface, the spatiotemporal variation diagram of the propagation loss is obtained by combining ray optics and PE with SSF methods.

[0041] Step 2: 3D convolutional autoencoder spatial feature extraction modeling:

[0042] A spatial information autoencoder composed of three-dimensional convolution, normalization and activation function is used to realize the feature extraction of nonlinear mapping relationship of inhomogeneous atmospheric waveguide. The input layer inputs data with a time series length of T. The data includes images of the input frame sequence length and data of the frame sequence length to be predicted. The shape is (B×T)×C×H×W. Each frame is regarded as a single sample and focuses on the spatial features of a single frame level without considering the time change. The output data is input to the time information learning module, and the output shape is reshaped to B×(T×C)×H×W.

[0043] Step 3: The U-TCN module composed of the time convolutional network TCN models the temporal characteristics of over-the-horizon propagation loss:

[0044] In order to perform nonlinear mapping of the time series of over-the-horizon propagation loss in non-uniform atmospheric waveguides, a temporal information learning module parser (U-TCN) composed of a U-net type connected temporal convolutional network (TCN) is proposed for feature extraction. The input is a tensor of shape B×(T×C)×H×W reshaped by the spatial information learning module, which is stacked along the time axis to extract features. The output data with spatial and temporal information is input into the spatiotemporal information integration learning module, and the shape is reshaped to (B×T)×C×H×W.

[0045] Step 4, 3D convolutional automatic decoder time and space feature integration:

[0046] A decoder that automatically integrates learned temporal and spatial information is constructed using three-dimensional convolution, normalization, and activation functions. The input is a tensor with the shape of B×(T×C)×H×W with learned temporal and spatial information, and the output shape is T'×C×H×W.

[0047] Step 5: SP-URNet prediction network model construction:

[0048] like Figure 4 As shown, in order to capture the nonlinear spatiotemporal sequence relationship of the over-the-horizon propagation loss output, the present invention proposes the SP-URNet prediction model. The model structure includes: input data, model parameter setting, spatial encoder (Encoder), time resolver (U-TCN) and spatiotemporal integrated decoder (Decoder) construction and network output;

[0049] The input data comes from the atmospheric refractivity data calculated by the ERA5 model reanalysis data, and is input into the atmospheric waveguide propagation loss image data obtained by the AREPS software. The data is divided into input frame data and predicted frame data.

[0050] The hyperparameters of the model play an important role in the evaluation process of the entire network. The main parameters used in this paper are Adam optimizer, leakyReLU activation function, weight loss, dilated convolution (dilation factor) and weight normalization.

[0051] According to step 2, the nonlinear relationship of the spatial sequence of atmospheric waveguide propagation loss data can be learned, and the data with spatial information is reshaped and input into step 3. The time series learning module in step 3 is composed of TCN modules connected in a U-shaped manner to obtain the nonlinear relationship of the time series. Finally, the data with spatial and temporal information is input into step 4 to complete the information integration. Based on the training of the network model parameters, the most accurate spatiotemporal prediction results of propagation loss are output and the model is saved.

[0052] 1) Input data: The input data is obtained through software after calculation of ERA5 model reanalysis data, which is the over-the-horizon propagation loss image data in the Yellow Sea of ​​China from January 1, 2015 to December 31, 2016.

[0053] 2) Model parameter setting: The model parameter setting plays an important role in the entire network evaluation process. The main parameters in the SP-URNet network model are as follows:

[0054] Optimizer: Adam (Adaptive Moment Estimation) optimizer is a commonly used optimization algorithm for training neural networks. It combines the first-order moment estimation (mean) and second-order moment estimation (uncentered variance) of the gradient to adaptively adjust the learning rate of each parameter, so it is very suitable for solving the correlation problem of over-the-horizon propagation loss parameters.

[0055] Loss function: LeakyReLU introduces a small non-zero slope for negative input values ​​to prevent neurons from becoming inactive. In SP-URNet model training, it helps to avoid saturation of certain units during training, that is, to prevent the output of a unit from remaining in a state with extremely small gradients, thereby slowing down or preventing learning.

[0056] Weight loss: By adjusting the weight W between hidden layers, the sequence model can be optimized and the effectiveness of model training can be improved. This method is also applied to the SP-URNet network.

[0057] Dilated convolution: Dilated convolution introduces a dilation factor strategy in the traditional convolution operation. Dilated convolution allows a larger receptive field to be obtained while keeping the number of parameters relatively small, thereby better capturing the global information of the input data. In addition, no additional parameters are introduced when increasing the receptive field, which helps to reduce the complexity of the model.

[0058] Weight Normalization: Decouples the magnitude of a weight from its direction. Normalization is applied independently to each weight in the network. The normalized weights are then scaled by a learnable parameter (usually a scalar). Introducing weight normalization in U-TCN convolution kernels enables the identification of effective regularization filters to improve the speed of model convergence.

[0059] 3) Spatial encoder (Encoder), temporal parser (U-TCN) and spatiotemporal integration decoder (Decoder): 17520 (24*365) hours of atmospheric waveguide propagation loss data are divided into 292 groups, each of which has an input frame length of 48 and a prediction frame length of 12. Each group consists of 60 frames of continuous image data. Spatiotemporal information is extracted sequentially through the spatial encoder, temporal parser and spatiotemporal integration decoder, and the network output is represented as follows:

[0060]

[0061] Where f and e represent the number of network layers of the U-TCN network and the Encoder_Decoder network, respectively. Represents the actual output of the Sp-URNet network.

[0062] 4) Network output: By training the model parameters, accurate prediction results of over-the-horizon propagation loss can be obtained. Figure 5 , as shown in Figure 6, if the model obtains the best prediction result for the test data set, the prediction result of the propagation loss is output and the trained model is saved. If the result is the opposite, the model parameter layer is returned to adjust the network parameters to optimize the network prediction result.

[0063] In step 2, at the encoder stage, after N cycles of convolutional layers (Conv2d), normalization (GroupNorm), and activation functions (ReLU), the data size is reshaped, and the input shape of [T, C, H, W] data is converted to [T×C, H / 2N, W / 2N], the data is converted from four-dimensional to three-dimensional data, and the image size is reduced, so that spatial information can be learned and extracted without considering the time series problem, such as Figure 2 As shown, the formula is as follows:

[0064] z i = σ(LayerNorm(Conv2d(z i-1 ))),1 ≤ i ≤ Ns, (1)

[0065] The input z i-1 and output z i The shapes are (T, C, H, W) and

[0066] In step 3, the parser is composed of 6 temporal convolutional network modules TCN connected in the form of U-Net architecture. The temporal convolutional network TCN consists of two branches. The first branch performs 1*1Conv2d convolution, and the other branch consists of four modules: dilated casual Conv, normalization (Weight Norm), activation function (Relu) and regularization (Dropout). After the two branches complete the operation, residual connection is performed, as shown in Figure 3 As shown, the formula is as follows:

[0067] z j = TCN(z j -1),Ns < j ≤ Ns + Nt (2)

[0068] The input z j-1 and output z j The shape is (T×C, H, W) and

[0069] In step 4, the decoder reconstructs the spatiotemporal live frame. It needs to reconstruct the decoder of the spatial information of the encoder and the temporal information parameters of the parser. In the decoding stage, after the convolution layer (Conv2d), normalization (GroupNorm) and activation function (ReLU) have been cycled N times, the data size is reshaped and the input shape is changed to The data is converted into [T′, C, H, W], the data is converted from three-dimensional to four-dimensional data, and the image size is enlarged, and finally the data with the same dimension and size as the original data is obtained, but the input time length may be inconsistent with the output time length. Figure 2 As shown, the formula is as follows:

[0070] z k = σ(GroupNorm(unConv2d(z k-1 ))) , Ns + Nt < k ≤ 2Ns + Nt (3)

[0071] The input z k-1 and output z k The shapes are and (T′, C, H, W).

Claims

1. A spatiotemporal prediction algorithm for atmospheric waveguide propagation loss in the marine non-uniform troposphere based on SP-URNet, characterized in that: The steps include: Step 1: Parameterization calculation of atmospheric waveguide atmospheric correction refractive index M and waveguide propagation loss image data calculation: Firstly, the atmospheric temperature, humidity, wind direction, pressure and atmospheric height in the non-uniform atmospheric waveguide environment are obtained by reanalyzing data in the ERA5 model, and the atmospheric corrected refractive index M is calculated. Based on the different atmospheric refractive indices corresponding to different altitudes above the sea surface, the spatiotemporal variation diagram of the propagation loss is obtained. Step 2: 3D convolutional autoencoder spatial feature extraction modeling: A spatial information autoencoder composed of three-dimensional convolution, normalization and activation function is used to realize the feature extraction of nonlinear mapping relationship of inhomogeneous atmospheric waveguide. The input layer inputs data with a time series length of T. The data includes images of the input frame sequence length and data of the frame sequence length to be predicted. The shape is (B×T)×C×H×W. Each frame is regarded as a single sample and focuses on the spatial features of a single frame level without considering the time change. The output data is input to the time information learning module, and the output shape is reshaped to B×(T×C)×H×W. Step 3: The U-TCN module composed of the temporal convolutional network (TCN) models the temporal characteristics of over-the-horizon propagation loss: A temporal information learning module parser composed of a U-net type connected temporal convolutional network is proposed for feature extraction. The input is a tensor of shape B×(T×C)×H×W reshaped by the spatial information learning module. The features are extracted by stacking along the time axis. The output data with spatial and temporal information is input into the spatiotemporal information integration learning module and reshaped into (B×T)×C×H×W. Step 4, 3D convolutional automatic decoder time and space feature integration: A decoder that automatically integrates learned temporal and spatial information is constructed using three-dimensional convolution, normalization, and activation functions. The input is a tensor with the shape of B×(T×C)×H×W with learned temporal and spatial information, and the output shape is T'×C×H×W. Step 5: SP-URNet prediction network model construction: The model structure includes: input data, model parameter setting, spatial encoder, temporal parser, spatiotemporal integration decoder construction and network output; The input data comes from the atmospheric refractive index data calculated by ERA5 model reanalysis data, and the atmospheric waveguide propagation loss image data is obtained, and the data is divided into input frame data and predicted frame data; The parameters used are Adam optimizer, leakyReLU activation function, weight loss, dilated convolution, and weight normalization; Based on the training of network model parameters, the most accurate spatiotemporal prediction results of propagation loss are output and the model is saved.

2. The spatiotemporal prediction algorithm for atmospheric waveguide propagation loss in the marine non-uniform troposphere based on SP-URNet according to claim 1 is characterized by: In step 2, after N cycles of convolutional layers, normalization, and activation functions, the input shape of [T, C, H, W] data is converted to [T×C, H / 2N, W / 2N], the data is converted from four-dimensional to three-dimensional data, and the image size is reduced. The formula is as follows: With i = σ(LayerNorm(Conv2d(z i-1 ))),1 ≤ i ≤ Ns, (1) The input z i-1 and output z i The shapes are (T, C, H, W) and 3. The spatiotemporal prediction algorithm for atmospheric waveguide propagation loss in the marine non-uniform troposphere based on SP-URNet according to claim 1 is characterized by: In step 3, the parser is composed of 6 temporal convolutional network modules TCN connected in the form of U-Net architecture. The temporal convolutional network TCN consists of two branches. The first branch performs 1*1Conv2d convolution, and the other branch consists of four modules: dilated convolution, standardization, activation function and regularization. After the two branches complete the operation, residual connection is performed. The formula is as follows: z j = TCN(z j -1),Ns < j ≤ Ns + Nt (2) The input z j-1 and output z j The shape is (T×C, H, W) and 4. The spatiotemporal prediction algorithm for atmospheric waveguide propagation loss in the marine non-uniform troposphere based on SP-URNet according to claim 1 is characterized by: In step 4, the decoder reconstructs the spatiotemporal live frame, which requires the spatial information of the encoder and the temporal information parameters of the parser to be reconstructed and decoded. In the decoding stage, after N cycles of convolutional layers, normalization and activation functions, the input shape is The data is converted into [T′, C, H, W], the data is converted from three-dimensional to four-dimensional data, and the image size is enlarged to finally obtain data with the same dimension and size as the original data. The formula is as follows: z k = σ(GroupNorm(unConv2d(z k-1 ))) ,Ns + Nt < k ≤ 2Ns + Nt (3) The input z k-1 and output z k The shapes are and (T′, C, H, W).

Images