A sp-ur net-based marine non-uniform troposphere atmospheric waveguide propagation loss space-time prediction method

Abstract

The application discloses a kind of SP-URNet-based offshore non-uniform troposphere atmospheric waveguide propagation loss space-time prediction method, comprising the following steps: step 1, atmospheric waveguide atmospheric correction refractive index M parameterization calculation and waveguide propagation loss image data calculation: step 2, three-dimensional convolution automatic encoder space feature extraction modeling: step 3, the U-TCN module of time convolution network TCN composition models the time feature of over-the-horizon propagation loss: step 4, three-dimensional convolution automatic decoder time and space feature integration: step 5, SP-URNet prediction network model building.The algorithm disclosed in the application overcomes the lack that the prior art cannot satisfy the space-time prediction and nonlinear prediction of offshore non-uniform troposphere atmospheric waveguide propagation loss data.

Description

Technical Field

[0001] This invention belongs to the field of marine atmospheric remote sensing, and specifically relates to a spatiotemporal prediction method for propagation loss of sea clutter non-uniform atmospheric waveguides based on deep convolutional networks. Background Technology

[0002] Marine tropospheric waveguides present a unique electromagnetic wave propagation environment. A crucial concept in electromagnetic wave propagation within tropospheric waveguides is propagation loss, which refers to the reduction in signal strength as the electromagnetic wave passes through the medium, experiencing attenuation or weakening along its path. Propagation loss reflects the trapping characteristics of electromagnetic waves within the tropospheric waveguide structure and plays a vital role in long-distance communication. Predicting the temporal and spatial propagation losses of marine atmospheric waveguides has significant implications for various fields, including communication, navigation, scientific research, and emergency response.

[0003] For a long time, the main methods for calculating electromagnetic wave propagation loss in marine atmospheric waveguide environments have included theoretical propagation modeling, numerical simulation, and parabolic equation (PE) calculations. Theoretical propagation models are all based on the assumption of a homogeneous atmospheric environment, while the actual atmospheric environment varies discontinuously over time. Numerical simulation methods typically consume significant computational resources. The PE method can conveniently solve long-distance problems with complex atmospheric refraction and irregular terrain effects. However, in actual atmospheric waveguide environments, quantitative calculations of propagation loss in non-uniform atmospheric waveguides often result in large errors. Summary of the Invention

[0004] To address the technical problem of insufficient spatiotemporal characteristic modeling capability for non-uniform atmospheric waveguide propagation at sea, which restricts the development of spatiotemporal prediction of atmospheric waveguide propagation loss, this invention provides a spatiotemporal prediction method for non-uniform tropospheric atmospheric waveguide propagation loss at sea based on SP-URNet. This method enables spatiotemporal prediction of non-uniform atmospheric waveguide propagation loss based on sea clutter data, supporting research on related technologies for marine atmospheric remote sensing.

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

[0006] A spatiotemporal prediction method for propagation loss of non-uniform tropospheric atmospheric waveguides at sea based on SP-URNet is improved by including the following steps:

[0007] Step 1: Calculation of atmospheric waveguide atmospheric corrected refractive index M parameter and waveguide propagation loss image data:

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

[0009] Step 2, Spatial Feature Extraction and Modeling of a 3D Convolutional Autoencoder:

[0010] A spatial information autoencoder composed of three-dimensional convolution, normalization, and activation functions is used to extract features of nonlinear mapping relationships in non-uniform atmospheric waveguides. The input layer takes in data of time series length T, which 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 treated as a single sample, and attention is paid to the spatial features at the single frame level, without considering temporal changes. The output data is input to the temporal information learning module, and the shape of the output is reshaped to B×(T×C)×H×W.

[0011] Step 3: The U-TCN module, composed of a temporal convolutional network (TCN), models the temporal characteristics of beyond-line-of-sight propagation loss.

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

[0013] Step 4, Integration of temporal and spatial features in the 3D convolutional autodecoder:

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

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

[0016] The model structure includes: input data, model parameter settings, spatial encoder, temporal resolver, spatiotemporal integrated decoder construction, and network output;

[0017] The input data comes from atmospheric refractive index data calculated using ERA5 model reanalysis data, atmospheric waveguide propagation loss image data, and is divided into input frame data and prediction 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 the network model parameters, the most accurate spatiotemporal prediction result of propagation loss is output and the model is saved.

[0020] Furthermore, in step 2, after N iterations of the convolutional layer, normalization, and activation function, the input data of shape [T,C,H,W] is transformed into [T×C,H / 2N,W / 2N], converting the data from four-dimensional to three-dimensional, and the image size is reduced, as shown in the following formula:

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

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

[0023] Furthermore, in step 3, the parser consists of six Temporal Convolutional Network (TCN) modules connected in a U-Net-like architecture. Each TCN comprises two branches: the first branch performs 1x1 Conv2d convolutions, and the second branch consists of four modules: dilated convolution, normalization, activation function, and regularization. After these operations, residual connections are performed between the two branches, as shown in the following formula:

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

[0025] Where 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 real-time frame, requiring reconstruction of the decoding based on the spatial information from the encoder and the temporal information from the decoder. During the decoding phase, the convolutional layer, normalization, and activation function undergo N iterations to convert the input into a shape... The data is transformed into [T′, C, H, W], converting the three-dimensional data back into four-dimensional data, and the image size is enlarged to finally obtain data with the same dimensions and size as the original data. The formula is shown below:

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

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

[0029] The beneficial effects of this invention are:

[0030] The method disclosed in this invention overcomes the shortcomings of existing technologies in providing spatiotemporal and nonlinear prediction of propagation loss data of non-uniform tropospheric atmospheric waveguides at sea. Using atmospheric waveguide propagation loss image data at different times and altitudes as input, and the predicted atmospheric waveguide propagation loss image data as output, a nonlinear mapping relationship between the two is obtained through a constructed SP-URNet network model. This achieves effective prediction of propagation loss of non-uniform atmospheric waveguides in sea clutter, effective prediction of propagation loss of tropospheric atmospheric waveguides at sea, and extraction and accurate prediction of spatiotemporal features of long-distance propagation loss. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the method of the present invention;

[0032] Figure 2 This 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 a flowchart detailing the SP-URNet model;

[0035] Figure 5 These are spatiotemporal prediction results of various models for atmospheric waveguide propagation loss data;

[0036] Figure 6 This is a comparison chart of prediction results from different models at different altitudes. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] Example 1 discloses a spatiotemporal prediction method for propagation loss of non-uniform tropospheric atmospheric waves at sea based on SP-URNet, realizing the extraction and modeling of spatiotemporal features of propagation loss of non-uniform atmospheric waves at sea. Based on this, using atmospheric waveguide propagation loss image data at different times and altitudes as input, and future atmospheric waveguide propagation loss image data as output, the nonlinear mapping relationship between the two is obtained through the constructed SP-URNet network model, achieving effective prediction of the propagation loss of non-uniform atmospheric waves in the sea clutter in both time and space. Figure 1 As shown, the specific steps include the following:

[0039] Step 1: Calculation of atmospheric waveguide atmospheric corrected refractive index M parameter and waveguide propagation loss image data:

[0040] To obtain atmospheric waveguide propagation loss data, we first obtained parameters such as atmospheric temperature, humidity, wind direction, pressure, and atmospheric height in a non-uniform atmospheric waveguide environment using ERA5 mode reanalysis data. We then calculated the atmospheric corrected refractive index M. Based on the calculated atmospheric refractive indices at different altitudes above the sea surface, we used a combination of ray optics and PE and SSF methods to obtain the spatiotemporal variation map of propagation loss.

[0041] Step 2, Spatial Feature Extraction and Modeling of a 3D Convolutional Autoencoder:

[0042] A spatial information autoencoder composed of three-dimensional convolution, normalization, and activation functions is used to extract features of nonlinear mapping relationships in non-uniform atmospheric waveguides. The input layer takes in data of time series length T, which 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 treated as a single sample, and attention is paid to the spatial features at the single frame level, without considering temporal changes. The output data is input to the temporal information learning module, and the shape of the output is reshaped to B×(T×C)×H×W.

[0043] Step 3: The U-TCN module, composed of a temporal convolutional network (TCN), models the temporal characteristics of beyond-line-of-sight propagation loss.

[0044] To establish a nonlinear time-series mapping relationship for the time series of propagation loss beyond line of sight in non-uniform atmospheric waveguides, a time 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. Features are extracted by stacking along the time axis. The output data, which contains spatial and temporal information, is input into the spatiotemporal information integration learning module, and the shape is reshaped into (B×T)×C×H×W.

[0045] Step 4, Integration of temporal and spatial features in the 3D convolutional autodecoder:

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

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

[0048] like Figure 4As shown, to capture the nonlinear spatiotemporal sequence relationship of beyond-line-of-sight propagation loss output, this invention proposes the SP-URNet prediction model. The model structure includes: input data, model parameter settings, construction of a spatial encoder (Encoder), a temporal resolver (U-TCN), and a spatiotemporal integrated decoder (Decoder), and network output;

[0049] The input data comes from atmospheric refractive index data calculated from ERA5 model reanalysis data and is input into atmospheric waveguide propagation loss image data obtained in AREPS software. The data is then divided into input frame data and prediction frame data.

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

[0051] Based on the nonlinear relationship of the spatial sequence of atmospheric waveguide propagation loss data learned in step 2, the data with spatial information is reshaped and input into step 3. The time series learning module in step 3 is composed of U-shaped connections of TCN modules 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 result of propagation loss is output and the model is saved.

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

[0053] 2) Model Parameter Settings: Model parameter settings play a crucial role in the entire network evaluation process. The main parameters in the SP-URNet network model are as follows:

[0054] Optimizer: The Adam (Adaptive Moment Estimation) optimizer is a commonly used optimization method for training neural networks. It combines the first-moment estimate (mean) and second-moment estimate (uncentered variance) of the gradient to adaptively adjust the learning rate of each parameter, making it well-suited for solving the correlation problem of propagation loss parameters beyond line-of-sight.

[0055] Loss function: LeakyReLU introduces a small non-zero slope for negative input values, preventing neurons from becoming inactive. In SP-URNet model training, this helps avoid saturation of certain units during training, i.e., preventing the output of a unit from remaining in a state of minimal gradient, thus slowing down or halting learning.

[0056] Weight loss: By adjusting the weights 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 into traditional convolution operations. It allows for a larger receptive field while maintaining a relatively small number of parameters, thus better capturing global information from the input data. Furthermore, increasing the receptive field does not introduce additional parameters, helping to reduce model complexity.

[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 using a learnable parameter (usually a scalar). Incorporating weight normalization into U-TCN convolutional kernels can identify effective regularization filters to improve model convergence speed.

[0059] 3) Spatial Encoder, Temporal Resolution Unit (U-TCN), and Spatiotemporal Integrated Decoder: 17520 (24*365) hours of atmospheric waveguide propagation loss data are divided into 292 groups. Each group contains 48 input frames and 12 prediction frames, with each group consisting of 60 consecutive image frames. The spatiotemporal information is extracted sequentially through the spatial encoder, temporal resolution unit, and spatiotemporal integrated decoder. The network output is represented as follows:

[0060]

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

[0062] 4) Network Output: By training the model parameters, accurate prediction results of beyond-line-of-sight propagation loss can be obtained. For example... Figure 5 As shown in Figure 6, if the model achieves the best prediction result on the test dataset, the prediction result of the propagation loss is output and the trained model is saved. If the result is the opposite, the process returns to the model parameter layer to adjust the network parameters to optimize the network prediction result.

[0063] In step 2, during the encoder stage, after N iterations of the convolutional layer (Conv2d), normalization (GroupNorm), and activation function (ReLU), the data size is reduced and the data shape is reshaped, transforming the input data of shape [T,C,H,W] into [T×C,H / 2N,W / 2N]. This converts the data from four-dimensional to three-dimensional data and reduces the image size. This allows for the learning and extraction of spatial information without considering time series issues. Figure 2 As shown, the formula is as follows:

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

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

[0066] In step 3, the parser consists of six Temporal Convolutional Network (TCN) modules connected in a U-Net-like architecture. Each TCN comprises two branches: the first performs 1x1 Conv2d convolutions, and the second branch consists of four modules: dilated casual convolution, weight normalization, ReLU activation, and dropout regularization. After the operations on both branches are completed, residual connections are made, such as... Figure 3 As shown, the formula is as follows:

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

[0068] Where 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 real-time frame. This requires reconstructing the decoding based on the spatial information from the encoder and the temporal information from the decoder. During the decoding stage, the convolutional layer (Conv2d), normalization (GroupNorm), and activation function (ReLU) are processed N times. The data size is adjusted, and the data shape is reshaped, resulting in an input shape of... The data is transformed into [T′, C, H, W], converting the three-dimensional data back into four-dimensional data, and the image size is enlarged to obtain data with the same dimensions and size as the original data, except that the input time length may differ from the output time length. For example... 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] Where the input z k-1 and output z k The shapes are respectively And (T′, C, H, W).

Claims

1. A spatiotemporal prediction method for propagation loss of marine non-uniform tropospheric atmospheric waveguides based on SP-URNet, characterized in that, Includes the following steps: Step 1: Calculation of atmospheric waveguide atmospheric corrected refractive index M parameter and waveguide propagation loss image data: First, atmospheric temperature, humidity, wind direction, pressure, and atmospheric height in a non-uniform atmospheric waveguide environment are obtained through ERA5 mode reanalysis data. The atmospheric corrected refractive index M is calculated. Based on the calculated atmospheric refractive indices corresponding to different altitudes on the sea surface, the spatiotemporal variation diagram of propagation loss is obtained. Step 2, Spatial Feature Extraction and Modeling of a 3D Convolutional Autoencoder: A spatial information autoencoder composed of three-dimensional convolution, normalization, and activation functions is used to extract features of nonlinear mapping relationships in non-uniform atmospheric waveguides. The input layer takes in data of time series length T, which 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 treated as a single sample, and attention is paid to the spatial features at the single frame level, without considering temporal changes. The output data is input to the temporal information learning module, and the shape of the output is reshaped to B×(T×C)×H×W. Step 3: The U-TCN module, composed of a temporal convolutional network (TCN), models the temporal characteristics of beyond-line-of-sight propagation loss. A temporal information learning module parser composed of a U-net-type connected temporal convolutional network is proposed to extract features. The input is a tensor of shape B×(T×C)×H×W reshaped by the spatial information learning module. Features are extracted by stacking along the time axis. The output data is input into the spatiotemporal information integration learning module with spatial and temporal information, and the shape is reshaped into (B×T)×C×H×W. Step 4, Integration of temporal and spatial features in the 3D convolutional autodecoder: A decoder that automatically integrates learned temporal and spatial information is constructed using three-dimensional convolution, normalization, and activation functions. The input is data with learned temporal and spatial information, a tensor of shape B×(T×C)×H×W, and the output is of shape T'×C×H×W. Step 5, SP-URNet prediction network model construction: The model structure includes: input data, model parameter settings, spatial encoder, temporal resolver, spatiotemporal integrated decoder construction, and network output; The input data comes from atmospheric refractive index data calculated using ERA5 model reanalysis data, atmospheric waveguide propagation loss image data, and is divided into input frame data and prediction frame data. The parameters used are Adam optimizer, leakyReLU activation function, weight loss, dilated convolution and weight normalization; Based on the training of the network model parameters, the most accurate spatiotemporal prediction result of propagation loss is output and the model is saved.

2. The spatiotemporal prediction method for propagation loss of marine non-uniform tropospheric atmospheric waveguides based on SP-URNet according to claim 1, characterized in that: In step 2, after N iterations of the convolutional layer, normalization, and activation function, the input data of shape [T,C,H,W] is transformed into [T×C,H / 2N,W / 2N], converting the data from four-dimensional to three-dimensional, and reducing the image size. The formula is as follows: With i = σ(LayerNorm(Conv2d(z i-1 ))),1 ≤ i ≤ Ns, (1) Where the input z i-1 and output z i The shapes are (T, C, H, W) and 3. The spatiotemporal prediction method for propagation loss of marine non-uniform tropospheric atmospheric waveguides based on SP-URNet according to claim 1, characterized in that: In step 3, the parser consists of six Temporal Convolutional Network (TCN) modules connected in a U-Net-like architecture. Each TCN comprises two branches: the first performs 1x1 Conv2d convolutions, and the second branch consists of four modules: dilated convolution, normalization, activation function, and regularization. After these operations, residual connections are performed between the two branches, as shown in the following formula: z j = TCN(z j -1),Ns < j ≤ Ns + Nt (2) Where the input z j-1 and output z j The shape is (T×C, H, W) and 4. The spatiotemporal prediction method for propagation loss of marine non-uniform tropospheric atmospheric waveguides based on SP-URNet according to claim 1, characterized in that: In step 4, the decoder reconstructs the spatiotemporal real-time frame, requiring reconstruction of the decoding based on the spatial information from the encoder and the temporal information from the decoder. During the decoding stage, the convolutional layer, normalization, and activation function are processed N times, resulting in an input with the shape of... The data is transformed into [T′, C, H, W], converting the three-dimensional data back into four-dimensional data, and the image size is enlarged to finally obtain data with the same dimensions and size as the original data. The formula is shown below: z k = σ(GroupNorm(unConv2d(z k-1 ))) ,Ns + Nt < k ≤ 2Ns + Nt (3) Where the input z k-1 and output z k The shapes are respectively And (T′, C, H, W).

Images