A SAR sea clutter amplitude distribution prediction method based on dual-stream multi-task learning network

Abstract

The present invention discloses a SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network, which relates to the technical field of synthetic aperture radar image interpretation. The method comprises the following steps: collecting GF-3 satellite wave mode SAR data to obtain a plurality of SAR amplitude images; constructing a SAR sea clutter amplitude distribution prediction data set using a label corresponding to each SAR amplitude image; constructing a SAR sea clutter amplitude distribution prediction model based on the dual-stream multi-task learning network; training the SAR sea clutter amplitude distribution prediction model using a training set, and optimizing and adjusting the parameters of the SAR sea clutter amplitude distribution prediction model using a validation set; testing and evaluating the trained SAR sea clutter amplitude distribution prediction model; and inputting a to-be-predicted SAR image into the trained SAR sea clutter amplitude distribution prediction model to obtain prediction results of the sea clutter amplitude distribution type and distribution parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of synthetic aperture radar (SAR) image interpretation, and in particular to a SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network. Background Art

[0002] Sea clutter, radar echoes reflected from the sea surface, is the primary source of interference to target echoes in marine environments. Synthetic aperture radar (SAR) is an active microwave imaging radar that enables all-day, all-weather observations and has been widely used in marine remote sensing. The amplitude distribution characteristics of sea clutter directly impact the accuracy of SAR target detection. Research on the amplitude characteristics of SAR sea clutter has become an integral part of marine target detection. Influenced by marine environmental parameters and radar parameters, the amplitude distribution characteristics of sea clutter vary over time. Consequently, the amplitude distribution of sea clutter in different SAR images varies significantly. Accurately predicting the type and parameters of sea clutter amplitude distribution in SAR images is crucial for improving SAR target detection performance.

[0003] The study of sea clutter amplitude distribution has undergone a long development process. In the early days, researchers mainly used the Rayleigh distribution to characterize the amplitude of sea clutter. This model was suitable for low sea conditions and medium and low resolution radars. However, with the improvement of radar resolution, the statistical distribution of sea clutter deviated from the Rayleigh model and showed enhanced peaks and tails. Subsequently, models such as the lognormal distribution and the Weibull distribution were proposed to improve the fitting accuracy of the sea clutter amplitude distribution. In recent years, researchers have developed more complex amplitude distribution models. For example, the K distribution and the generalized gamma distribution. These distributions can better characterize the non-Gaussian and peak characteristics of sea clutter. Although many sea clutter amplitude distribution models are available, there is no universal model that can be applied to SAR images in all ocean environments and radar imaging conditions. Therefore, how to predict the optimal amplitude distribution of sea clutter in SAR images is a key issue.

[0004] Deep learning technology provides a new paradigm for predicting sea clutter amplitude distribution, bypassing various tedious preprocessing steps and achieving end-to-end prediction of sea clutter amplitude distribution. Many researchers have adopted convolutional neural networks (CNNs) to predict sea clutter amplitude distribution and have achieved positive results. However, most existing deep learning-based sea clutter amplitude distribution prediction methods use one-dimensional sea clutter sequence data and are not suitable for two-dimensional SAR images. The existing paper "SCA-Net: A Network Based on Multi-task Learning for Sea Clutter Amplitude Distribution Prediction of SAR Images" does not consider the impact of ocean environmental parameters (wind speed, wind direction, wave height, and wave direction) on sea clutter amplitude distribution characteristics, and cannot achieve dynamic prediction of sea clutter amplitude distribution under different sea conditions.

[0005] A joint prediction method for sea clutter amplitude distribution type and parameters in a prior art patent document is based on measured X-band sea clutter data. Long-sequence features and statistical features are extracted from the sea clutter data through feature engineering, and a data set is constructed. The MT1dCNN model is then used to predict the sea clutter amplitude distribution type and parameters. This technology has the following technical defects: (1) Long-sequence features and statistical features need to be extracted from the sea clutter data before inputting into the neural network model, making it impossible to achieve end-to-end sea clutter amplitude distribution prediction; (2) The one-dimensional convolution layer of the model limits the input data to one-dimensional data, making it impossible to predict the sea clutter amplitude distribution of two-dimensional SAR images. (3) The ocean environment parameters are not considered. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network, comprising:

[0007] Step 1: collect GF-3 satellite wave mode SAR data and obtain multiple SAR amplitude images through data processing;

[0008] Step 2: Based on the multiple SAR amplitude images obtained in step 1, a SAR sea clutter amplitude distribution prediction dataset is constructed using the labels corresponding to each SAR amplitude image.

[0009] Step 3: Build a SAR sea clutter amplitude distribution prediction model based on a dual-stream multi-task learning network;

[0010] Step 4: using the training set of the SAR sea clutter amplitude distribution prediction data set constructed in step 2 to train the SAR sea clutter amplitude distribution prediction model, and using the validation set to optimize and adjust the parameters of the SAR sea clutter amplitude distribution prediction model;

[0011] Step 5: test and evaluate the trained SAR sea clutter amplitude distribution prediction model;

[0012] Step 6: Perform the data processing of step 1 on the SAR image to be predicted to obtain the SAR amplitude image to be predicted, and input the SAR amplitude image to be predicted into the trained SAR sea clutter amplitude distribution prediction model to obtain the prediction results of the sea clutter amplitude distribution type and distribution parameters.

[0013] In a preferred embodiment, in step 2, the probability distribution histogram of the SAR amplitude image is statistically analyzed, and the logarithmic cumulant algorithm is used to perform parameter estimation and calculate the KL divergence value:

[0014] ;

[0015] Among them, g(x) and f(x) represent two different probability density distributions, x Represents the amplitude value of each pixel in the SAR image; the distribution model with the smallest KL divergence value is taken as the sea clutter amplitude distribution type of the SAR amplitude image.

[0016] In a preferred embodiment, step 3 includes:

[0017] Step 3.1: Construct an image feature encoder and input the SAR amplitude image into the image feature encoder to obtain the high-dimensional image feature Feature_image.

[0018] Step 3.2: Construct an environmental feature encoder, input the ocean environmental parameters into the environmental feature encoder, and obtain high-dimensional parameter features Feature_param;

[0019] Step 3.3, performing cross-modal feature fusion on the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param to obtain a fused feature map Feature_fusion;

[0020] Step 3.4: Construct a distribution type decoder and input the fused feature map Feature_fusion into the distribution type decoder to obtain the prediction result of the sea clutter amplitude distribution type;

[0021] Step 3.5: Construct a distribution parameter decoder and input the fused feature map Feature_fusion into the distribution parameter decoder to obtain the prediction results of the sea clutter amplitude distribution parameters.

[0022] In a preferred embodiment, in step 3.2, the ocean environment parameters are input into the environmental feature encoder for channel combination to obtain the parameter matrix Input_P; the parameter matrix Input_P is passed to the first convolution layer L1, and the convolution kernel is used to extract the spatial gradient features of different parameters to generate the feature map F1:

[0023] ;

[0024] Among them, Conv 3×3 Indicates a convolution calculation of size 3×3;

[0025] Perform batch normalization on the feature map F1, convert the distribution of the feature map F1 into a normal distribution, and obtain the normalized result BN1:

[0026] ;

[0027] in, represents mean calculation, represents variance calculation, To prevent the denominator from being a decimal of 0, the ReLU function is used to activate the normalized result BN1 to obtain the convolutional layer feature map Re1:

[0028] ;

[0029] in, Indicates the maximum value of the calculated variable;

[0030] The convolutional layer feature map Re1 of the first convolutional layer L1 is input into the first pooling layer P1 to obtain the pooling layer feature map Pool1:

[0031] ;

[0032] in, Represents the maximum pooling operation;

[0033] Perform multiple convolutional layers and pooling layers in sequence to obtain a high-dimensional feature map Feature_param.

[0034] In a preferred embodiment, in step 3.3, the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param are spatially aligned, and the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param are spliced ​​in the channel dimension to obtain a fused feature map Feature_fusion:

[0035] ;

[0036] in, Indicates concatenate calculation.

[0037] In a preferred embodiment, the image feature encoder is composed of a plurality of dimension-reduced residual blocks and a plurality of non-dimensionality-reduced residual blocks connected in series alternately, and each residual block includes a main branch, a shortcut connection branch, an addition operation, and an activation function;

[0038] In a preferred embodiment, in the dimensionality reduction residual block, the main branch reduces the input size of the SAR amplitude image through the first depthwise separable convolution layer, expands the number of channels, and then performs batch normalization and activation function calculation; then it is subjected to feature extraction and batch normalization through the second depthwise separable convolution layer, and then the spatial features are enhanced through the effective channel attention layer. At the same time, the shortcut connection branch synchronously adjusts the channel and size through convolution and batch normalization. The processing results of the main branch and the shortcut connection branch are added to output the target size feature map.

[0039] In a preferred embodiment, in the non-dimensionality reduction residual block, local features are enhanced by stacking two depth-wise separable convolutional layers of the main branch, and the residual connection of the identity mapping is combined to alleviate the gradient disappearance, and a high-dimensional feature map of the target size is output.

[0040] In a preferred embodiment, the environmental feature encoder is composed of multiple convolutional layers and multiple pooling layers alternately connected in series; the convolutional layers include shallow convolutional layers and deep convolutional layers, the shallow convolutional layers are used to extract the local spatial gradient features of each parameter, the deep convolutional layers are used to capture the spatial correlation between wind and waves, and the pooling layers are used to reduce dimensionality and enhance translation invariance.

[0041] In a preferred embodiment, the distribution type decoder is composed of a global average pooling layer, a flattening layer and a fully connected layer connected in series.

[0042] In a preferred embodiment, the distributed parameter decoder is composed of three fully connected layers in series.

[0043] Compared with the prior art, the present invention has the following beneficial technical effects:

[0044] Abstract: In order to improve the accuracy of sea clutter amplitude distribution prediction and the generalization ability of the model under different sea conditions, a dual-stream multi-task learning model SAR2SCAD was used to predict the amplitude distribution type and parameters of sea clutter using SAR image data and ERA5 meteorological grid data. Compared with the existing methods, the proposed method introduced marine environmental parameters such as wind speed and wind direction through the dual-stream multi-task learning model, realized the cross-modal feature fusion of marine environmental meteorological data and SAR images, and effectively improved the accuracy of sea clutter amplitude distribution prediction and the generalization ability of the model under different sea conditions. The F1 score of sea clutter amplitude distribution type prediction was 86.6%, and the RMSE of parameter prediction was 0.51. The proposed method achieved accurate prediction of SAR sea clutter amplitude distribution type and parameters, providing prior knowledge and important reference for SAR ship target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The SAR image and ERA5 data processing flow of the present invention;

[0046] Figure 2 The model structure of SAR2SCAD of the present invention;

[0047] Figure 3 is the basic structure of R2, R4, R6, and R8 of the present invention;

[0048] Figure 4 is the basic structure of R1, R3, R5, and R7 of the present invention;

[0049] Figure 5 is the sea clutter amplitude distribution type prediction result of the present invention;

[0050] Figure 6 This is the SAR sea clutter amplitude distribution prediction result of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] Step 1: Collect the GF-3 satellite wave mode SAR data to obtain multiple SAR amplitude images.

[0053] The Gaofen-3 satellite wave mode SAR data includes a total of 3597 HH polarization single-look complex images. First, the SAR data is cleaned to delete the poor quality images and obtain a SAR image containing pure sea clutter.

[0054] like Figure 1As shown, ERA5 data is collected, and then the filtered SAR image and ERA5 data are temporally and spatially matched to obtain the matched SAR data and the corresponding ERA5 data. The ERA5 data is further temporally and spatially interpolated to achieve pixel-level alignment with the SAR image. The wind speed, wind direction, wave height, and wave direction corresponding to the SAR image can be obtained from the ERA5 data, resulting in gridded data of wind speed, wind direction, wave height, and wave direction.

[0055] The SAR images after data cleaning were cropped to a size of 1000 pixels × 1000 pixels. The amplitude of each SAR image was then calculated and converted into TIFF image format, resulting in 3597 SAR amplitude images.

[0056] Preferably, the spatiotemporal interpolation method of ERA5 data uses the Kriging interpolation algorithm.

[0057] Step 2: Based on the multiple SAR amplitude images obtained in step 1, a SAR sea clutter amplitude distribution prediction dataset is constructed using the label corresponding to each SAR amplitude image.

[0058] Firstly, the probability distribution histogram of the SAR amplitude image is statistically calculated, denoted as P. Then, four commonly used sea clutter amplitude distribution models are selected, including K distribution, G0 distribution, generalized Gamma distribution and Weibull distribution. These distribution models are used to estimate the parameters of the SAR amplitude image, and the shape parameters and scale parameters of each distribution model are obtained.

[0059] The shape parameters of different distribution models are denoted as: μ_K, μ_G0, μ_Gamma, μ_Weibull; the shape parameters are denoted as: σ_K, σ_G0, σ_Gamma, σ_Weibull.

[0060] Substituting μ_K and σ_K into the probability density distribution function of the K distribution, we can obtain the fitting result of the K distribution, denoted as P_K. Similarly, we can obtain the fitting results of the G0 distribution, generalized Gamma distribution, and Weibull distribution, denoted as P_G0, P_Gamma, and P_Weibull, respectively.

[0061] Calculate the KL (Kullback Leibler) divergence of P_K, P_G0, P_Gamma, P_Weibull and P respectively, and record the results as KL_K, KL_G0, KL_Gamma, KL_Weibull respectively; compare the sizes of KL_K, KL_G0, KL_Gamma, and KL_Weibull, and take the distribution model with the smallest KL divergence value as the sea clutter amplitude distribution type of the SAR amplitude image, recorded as Label_class, and record the shape parameter and scale parameter corresponding to the distribution model, recorded as Label_shape and Label_scale respectively.

[0062] Perform the above operations on the 3597 SAR amplitude images obtained in step 1 in sequence to obtain the Label_class, Label_shape, and Label_scale of each SAR amplitude image. These SAR images and their corresponding labels can then be used to construct a SAR sea clutter amplitude distribution prediction dataset, denoted as SAR_clutter. This dataset includes 3597 samples and is divided into a training set (Train), a validation set (Val), and a test set (Test) in a ratio of 8:1:1.

[0063] The parameter estimation method used is the logarithmic cumulative value (MoLC) algorithm, and the goodness of fit evaluation indicator is the KL divergence. The calculation formula of KL divergence is as follows:

[0064] ;

[0065] in, x Represents the amplitude value of each pixel of the SAR image, g ( x )and f ( x ) represent two different probability density distributions.

[0066] Step 3: Construct a SAR sea clutter amplitude distribution prediction model based on a dual-stream multi-task learning network.

[0067] The SAR sea clutter amplitude distribution prediction model is denoted as SAR2SCAD. The model structure of SAR2SCAD is as follows: Figure 2 As shown in the figure, it consists of an image feature encoder, an environment feature encoder, a feature fusion, a distribution type decoder, and a distribution parameter decoder.

[0068] Step 3.1: Construct an image feature encoder and input the SAR amplitude image into the image feature encoder to obtain the high-dimensional image feature Feature_image.

[0069] The image feature encoder is denoted as IFE; IFE consists of eight residual blocks (R1, R2, R3, R4, R5, R6, R7, R8) connected in series to extract deep semantic features of SAR images.

[0070] The SAR amplitude image obtained in step 1 is input into IFE, and the high-dimensional semantic features of the SAR amplitude image are extracted by calculating R1, R2, R3, R4, R5, R6, R7, and R8 in sequence.

[0071] The input single-channel SAR image is first subjected to coarse-grained feature extraction by R1 to capture basic texture and edge information, and then connected in series with R2, R3, R4, R5, R6, and R7. Among them, R1, R3, R5, and R7 are dimensionality reduction residual blocks, and their structures are as follows: Figure 4 shown.

[0072] Among them, the structure of the 8 residual blocks includes a main branch, a shortcut connection branch, an addition operation, and an activation function. Except for the difference in the parameter settings of the convolution layer, the main branch includes depth-wise separable convolution, batch normalization, activation function, depth-wise separable convolution, batch normalization, and effective channel attention.

[0073] Taking R1 as an example, the main branch reduces the input size from 1000×1000 to 500×500 through the first 3×3 depthwise separable convolution layer (step 2, padding 1), and expands the number of channels from 1 to 64. Batch normalization and activation function calculation are then performed, and the features are further refined through the second 3×3 depthwise separable convolution, and batch normalization is performed. After calculation of the effective channel attention layer, the spatial feature extraction capability of the model is enhanced. At the same time, the shortcut connection branch synchronously adjusts the channel and size through 1×1 convolution (step 2) and batch normalization. The results of the main branch and the shortcut connection branch are added to output a 500×500×64 feature map.

[0074] Similarly, the feature map size output by R3 is 250×250×128, the feature map size output by R5 is 125×125×256, and the feature map size output by R7 is 62×62×512.

[0075] The non-dimensionality reduction residual blocks R2, R4, R6, and R8 keep the input and output sizes and dimensions consistent, and the structure is as follows Figure 3 As shown in the figure, two 3×3 depthwise separable convolutions in the main branch are stacked to enhance local feature expression. The residual connection with the identity mapping effectively alleviates the vanishing gradient. The IFE method uses four dimensionality reduction steps to gradually compress the spatial resolution and increase the channel dimension. It integrates the multi-scale semantic information of the SAR image and ultimately outputs a 62×62×512 high-dimensional feature map. The high-dimensional image features are recorded as Feature_Image.

[0076] Preferably, the activation function is the Leaky ReLU function, and the convolution uses depthwise separable convolution, which significantly reduces the computational cost and parameter quantity while ensuring model performance by decoupling feature learning of spatial and channel dimensions; the channel attention module uses effective channel attention. Compared with traditional channel attention, it directly interacts with channel features through global average pooling and one-dimensional convolution, avoiding the information loss problem caused by dimensionality reduction in the traditional attention mechanism.

[0077] The calculation formula of the Leaky ReLU function is as follows:

[0078] ;

[0079] Among them, a represents the value of each pixel in the feature map.

[0080] Step 3.2: Construct an environmental feature encoder and input the ocean environment parameters into the environmental feature encoder to obtain high-dimensional parameter features Feature_param.

[0081] The environmental feature encoder, denoted as RFE, consists of four convolutional layers (L1, L2, L3, L4) and four pooling layers (P1, P2, P3, P4) alternating in series. It extracts features from the input ocean environmental parameters and produces a high-dimensional feature map. The convolutional layers consist of shallow and deep convolutional layers. Shallow convolutional layers extract the local spatial gradient features of each parameter, while deep convolutional layers capture more complex patterns, such as spatial correlations between wind and waves, and extract more abstract features. Pooling layers reduce dimensionality, enhance translation invariance, and preserve important features.

[0082] The specific calculation process is as follows:

[0083] The ocean environment parameters (wind speed, wind direction, wave height, and wave direction) are input into the RFE. First, they are channel-wise combined to obtain a 1000×1000×4 parameter matrix, denoted as Input_P. Input_P is then passed to the first convolutional layer, L1, which uses a 3×3 convolution kernel to extract the spatial gradient features of different parameters and explore the spatial correlation of different parameters. The resulting feature map is denoted as F1, with a dimension of 1000×1000×64:

[0084] ;

[0085] Among them, Conv 3×3 Indicates a convolution calculation of size 3×3.

[0086] Furthermore, the generated feature map F1 is batch normalized to convert the distribution of the feature map F1 into a normal distribution to solve the internal covariate transfer problem. The result of batch normalization is recorded as BN1:

[0087] ;

[0088] in, represents mean calculation, represents variance calculation, To prevent decimals with a denominator of 0.

[0089] The ReLU function is used to activate the normalized result BN1 to enhance the nonlinear expression ability of the model and obtain the convolutional layer feature map Re1:

[0090] ;

[0091] in, Indicates the maximum value of the calculated variable.

[0092] Then, the convolutional layer feature map Re1 output by the first convolutional layer L1 is input to the first pooling layer P1. P1 uses a pooling kernel with a size of 3×3 and a stride of 2 to calculate Re1. The purpose is to downsample the feature map, achieve dimensionality reduction of the feature map, reduce the number of parameters, and remove redundant features while retaining key features. After P1 calculation, the resulting pooling layer feature map is recorded as Pool1. The calculation process can be expressed as:

[0093] ;

[0094] in, Represents the maximum pooling operation.

[0095] According to the above formula, the calculation of L2, P2, L3, P3, L4, and P4 layers is performed in sequence to obtain the high-dimensional feature map extracted by the environmental feature encoder RFE, which is recorded as Feature_param.

[0096] RFE extracts high-level semantic features of these ocean environment parameters through convolution operations.

[0097] The number of input and output channels for L1 is 4 and 64 respectively, the number of input and output channels for L2 is 64 and 128 respectively, the number of input and output channels for L3 is 128 and 256 respectively, and the number of input and output channels for L4 is 256 and 512 respectively. The convolution kernel size for L1, L2, L3, and L4 is all 3×3, with a stride of 1, and the activation function is the ReLU function. P1, P2, P3, and P4 are all max pooling layers, with a pooling kernel size of 3×3 and a stride of 2.

[0098] Step 3.3, perform cross-modal feature fusion on the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param to obtain the fused feature map Feature_fusion;

[0099] The high-dimensional image features Feature_image extracted by IFE and the high-dimensional parameter features Feature_param extracted by RFE are input into the cross-modal feature fusion module. First, Feature_image and Feature_param are spatially aligned, and then Feature_image and Feature_param are spliced ​​in the channel dimension to obtain the fused feature map, which is recorded as Feature_fusion.

[0100] ;

[0101] in, Indicates concatenate calculation.

[0102] Step 3.4: Construct a distribution type decoder and input the fused feature map Feature_fusion into the distribution type decoder to obtain the prediction result of the sea clutter amplitude distribution type.

[0103] The distribution type decoder is denoted as DC. It consists of a global average pooling layer, a flattening layer, and a fully connected layer (F1) in series. DC takes Feature_fusion as input and performs global average pooling and flattening in sequence. Finally, it passes through F1 to obtain the predicted sea clutter amplitude distribution type. The number of nodes in F1 is 4.

[0104] Step 3.5: Construct a distribution parameter decoder and input the fused feature map Feature_fusion into the distribution parameter decoder to obtain the prediction results of the sea clutter amplitude distribution parameters.

[0105] The distributed parameter decoder is denoted as DP. It consists of three fully connected layers connected in series, denoted as F2, F3, and F4. DP takes Feature_fusion as input and passes it through F2, F3, and F4 in sequence, ultimately obtaining the predicted sea clutter amplitude distribution parameters. The number of nodes in F2, F3, and F4 is 1024, 512, and 2, respectively.

[0106] Through the above steps, the construction of IFE, RFE, DC, and DP is completed. The IFE and RFE are feature fused and then connected in series with DC and DP respectively to complete the construction of the SAR2SCAD model.

[0107] Step 4: Use the Train of the SAR_clutter dataset constructed in step 2 to train SAR2SCAD, and use Val to optimize and adjust the model parameters.

[0108] During the training process, the model is trained using the classic Adam optimizer. The loss function consists of two parts: the type prediction task uses the classic cross entropy as the loss function, and the parameter prediction task uses Smooth-L1 as the loss function. The uncertainty weight loss is used to assign weights to the losses of the distribution type prediction task and the distribution parameter prediction task, and their weights are dynamically adjusted during the training process.

[0109] After the training is completed, the sea clutter amplitude distribution prediction model is obtained, which is recorded as Trained_model.

[0110] Step 5: Test and evaluate the trained SAR sea clutter amplitude distribution prediction model Trained_model.

[0111] Using Test as input, the trained_model obtained in step 4 is used to perform model testing, resulting in the sea clutter amplitude distribution prediction results. Classic evaluation index calculation methods are then used to calculate the distribution type prediction accuracy and distribution parameter prediction accuracy. Table 1 shows the sea clutter amplitude distribution prediction accuracy of the SAR2SCAD model and other classic deep learning algorithms.

[0112] Table 1 Prediction accuracy of sea clutter amplitude distribution of different models

[0113]

[0114] Table 1 shows that the proposed SAR2SCAD model achieves an F1 score of 86.6% for distribution type prediction. The MAE and RMSE for distribution parameter prediction are 0.25 and 0.51, respectively. The SAR2SCAD model achieves the best accuracy among all methods. Experimental results demonstrate that the SAR2SCAD model achieves the best accuracy in predicting the type and parameters of SAR sea clutter amplitude distribution.

[0115] Figure 5 The confusion matrix of the sea clutter amplitude distribution type prediction results is shown. K, G0, Gamma, and Weibull represent K distribution, G0 distribution, generalized Gamma distribution, and Weibull distribution, respectively. It can be seen that all types can be predicted relatively accurately.

[0116] Step 6: The SAR image to be predicted is processed as in step 1, and then input into the trained SAR sea clutter amplitude distribution prediction model SAR2SCAD to obtain the prediction results of the sea clutter amplitude distribution type and distribution parameters.

[0117] Figure 6 An example of using SAR2SCAD to predict the amplitude distribution of SAR sea clutter is given. As can be seen from the figure, the predicted PDF is very close to the original distribution histogram, which proves the effectiveness of the SAR sea clutter amplitude distribution prediction method proposed in this paper.

[0118] The sea clutter prediction results include distribution type and distribution parameters, where the distribution parameters include shape parameters and scale parameters. Substitute the shape parameters and scale parameters into the probability density distribution function (PDF) of the distribution and draw Figure 6 The blue curve of SAR image can be used to obtain the sea clutter amplitude distribution of SAR image.

[0119] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. The embodiments should therefore be considered illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be encompassed therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network, characterized in that: include: Step 1: collect GF-3 satellite wave mode SAR data and obtain multiple SAR amplitude images through data processing; Step 2: Based on the multiple SAR amplitude images obtained in step 1, a SAR sea clutter amplitude distribution prediction dataset is constructed using the labels corresponding to each SAR amplitude image. Step 3: Build a SAR sea clutter amplitude distribution prediction model based on a dual-stream multi-task learning network, including: Step 3.1: Construct an image feature encoder and input the SAR amplitude image into the image feature encoder to obtain the high-dimensional image feature Feature_image. Step 3.2: Construct an environmental feature encoder, input the ocean environmental parameters into the environmental feature encoder, and obtain high-dimensional parameter features Feature_param; Step 3.3, performing cross-modal feature fusion on the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param to obtain a fused feature map Feature_fusion; Step 3.4: Construct a distribution type decoder and input the fused feature map Feature_fusion into the distribution type decoder to obtain the prediction result of the sea clutter amplitude distribution type; Step 3.5: Construct a distribution parameter decoder and input the fused feature map Feature_fusion into the distribution parameter decoder to obtain the prediction results of the sea clutter amplitude distribution parameters; Step 4: using the training set of the SAR sea clutter amplitude distribution prediction data set constructed in step 2 to train the SAR sea clutter amplitude distribution prediction model, and using the validation set to optimize and adjust the parameters of the SAR sea clutter amplitude distribution prediction model; Step 5: test and evaluate the trained SAR sea clutter amplitude distribution prediction model; Step 6: Perform the data processing of step 1 on the SAR image to be predicted to obtain the SAR amplitude image to be predicted, and input the SAR amplitude image to be predicted into the trained SAR sea clutter amplitude distribution prediction model to obtain the prediction results of the sea clutter amplitude distribution type and distribution parameters.

2. The SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network according to claim 1 is characterized in that: In step 2, the probability distribution histogram of the SAR amplitude image is statistically analyzed, and the logarithmic cumulant algorithm is used to perform parameter estimation and calculate the KL divergence value: Where g(x) and f(x) represent two different probability density distributions, respectively, and x represents the amplitude value of each pixel in the SAR image. The distribution model with the smallest KL divergence value is taken as the sea clutter amplitude distribution type of the SAR amplitude image.

3. The SAR sea clutter amplitude distribution prediction method based on dual-stream multi-task learning network according to claim 1 is characterized in that: In step 3.2, the ocean environment parameters are input into the environmental feature encoder for channel combination to obtain the parameter matrix Input_P. The parameter matrix Input_P is passed to the first convolutional layer L1, and the convolution kernel is used to extract the spatial gradient features of different parameters to generate the feature map F1: F1=Conv 3×3 (Input_P); Among them, Conv 3×3 Indicates a convolution calculation of size 3×3; Perform batch normalization on the feature map F1, convert the distribution of the feature map F1 into a normal distribution, and obtain the normalized result BN1: Among them, m(·) represents the mean calculation, σ 2 (·) represents the variance calculation, and ε is a decimal to prevent the denominator from being 0. The ReLU function is used to activate the normalized result BN1 to obtain the convolutional layer feature map Re1: Re1=max{0,BN1}; Among them, max(·) represents the maximum value of the calculated variable; The convolutional layer feature map Re1 of the first convolutional layer L1 is input into the first pooling layer P1 to obtain the pooling layer feature map Pool1: Pool1 = maxpool(Re1); Among them, maxpool(·) represents the maximum pooling operation; Perform multiple convolutional layers and pooling layers in sequence to obtain a high-dimensional feature map Feature_param.

4. The SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network according to claim 1 is characterized in that: In step 3.3, the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param are spatially aligned, and the high-dimensional image feature Feature_image and the high-dimensional parameter feature Feature_param are spliced ​​in the channel dimension to obtain a fused feature map Feature_fusion: in, Indicates concatenate calculation.

5. The SAR sea clutter amplitude distribution prediction method based on dual-stream multi-task learning network according to claim 1 is characterized in that: The image feature encoder is composed of a plurality of dimension-reduced residual blocks and a plurality of non-dimensionality-reduced residual blocks connected in series alternately, and each residual block includes a main branch, a shortcut connection branch, an addition operation and an activation function.

6. The SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network according to claim 5 is characterized in that: In the dimensionality reduction residual block, the main branch reduces the input size of the SAR amplitude image through the first depthwise separable convolution layer, expands the number of channels, and then performs batch normalization and activation function calculation; it then undergoes feature extraction and batch normalization through the second depthwise separable convolution layer, and then enhances the spatial features through the effective channel attention layer. At the same time, the shortcut connection branch synchronously adjusts the channel and size through convolution and batch normalization. The processing results of the main branch and the shortcut connection branch are added to output the target size feature map.

7. The SAR sea clutter amplitude distribution prediction method based on dual-stream multi-task learning network according to claim 5 is characterized in that: In the non-dimensionality reduction residual block, local features are enhanced by stacking two depth-wise separable convolutional layers of the main branch, and the residual connection with identity mapping is combined to alleviate gradient disappearance and output a high-dimensional feature map of the target size.

8. The SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network according to claim 1 is characterized in that: The environmental feature encoder consists of multiple convolutional layers and multiple pooling layers connected alternately in series; the convolutional layers include shallow convolutional layers and deep convolutional layers. The shallow convolutional layers are used to extract the local spatial gradient features of each parameter, while the deep convolutional layers are used to capture the spatial correlation between wind and waves. The pooling layers are used to reduce dimensionality and enhance translation invariance.

9. The SAR sea clutter amplitude distribution prediction method based on a dual-stream multi-task learning network according to claim 1, characterized in that: The distribution type decoder is composed of a global average pooling layer, a flattening layer and a fully connected layer in series.

Images