SAR sea clutter amplitude distribution prediction method based on double-flow multi-task learning network

Abstract

The invention discloses an SAR sea clutter amplitude distribution prediction method based on a double-flow multi-task learning network, and relates to the technical field of synthetic aperture radar image interpretation, and the method comprises the steps: collecting high-resolution No.3 satellite wave mode SAR data, and obtaining a plurality of SAR amplitude images; constructing an SAR sea clutter amplitude distribution prediction data set by using the label corresponding to each SAR amplitude image; constructing an SAR sea clutter amplitude distribution prediction model based on a double-flow multi-task learning network; using the training set to train an SAR sea clutter amplitude distribution prediction model, and using the verification set to optimize and adjust parameters of the SAR sea clutter amplitude distribution prediction model; 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 sea clutter amplitude distribution types and distribution parameters.

Description

Technical Field

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

[0002] Sea clutter is the radar echo reflected by the sea surface and is the main interference source of target echoes in the marine environment. Synthetic aperture radar (SAR) is an active microwave imaging radar that can achieve all-weather and all-time observation and has been widely used in the field of marine remote sensing. The amplitude distribution characteristics of sea clutter directly affect the accuracy of SAR target detection. The research on the amplitude characteristics of SAR sea clutter has become an indispensable part of maritime target detection. Affected by marine environmental parameters and radar parameters, the amplitude distribution characteristics of sea clutter change continuously over time. Therefore, the amplitude distributions of sea clutter in different SAR images vary greatly. Accurately predicting the amplitude distribution type and distribution parameters of SAR images is of great significance for improving the performance of SAR target detection.

[0003] The research on the amplitude distribution of sea clutter has gone through a long development process. In the early stage, researchers mainly used the Rayleigh distribution to characterize the amplitude of sea clutter, and this model is applicable to low sea states and medium- and low-resolution radars. However, with the improvement of radar resolution, the statistical distribution of sea clutter deviates from the Rayleigh model and shows enhanced spikes and tails. Subsequently, models such as the log-normal distribution and the Weibull distribution were proposed, improving 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 spiky characteristics of sea clutter. Although there are many available sea clutter amplitude distribution models, there is no universal model that can be applied to SAR images under all marine 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 the amplitude distribution of sea clutter. It can bypass various cumbersome preprocessing steps and achieve the prediction of the amplitude distribution of sea clutter in an end-to-end manner. Many scholars have used convolutional neural networks (CNNs) to predict the amplitude distribution of sea clutter and achieved positive results. However, most of the existing deep learning-based methods for predicting the amplitude distribution of sea clutter use one-dimensional sea clutter sequence data and are not applicable to two-dimensional SAR images. The existing literature "SCA-Net: A Network Based on Multi-task Learning for Sea Clutter Amplitude Distribution Prediction of SAR Images" does not consider the influence of ocean environmental parameters (wind speed, wind direction, wave height, wave direction) on the amplitude distribution characteristics of sea clutter and cannot achieve the dynamic prediction of the amplitude distribution of sea clutter under different sea conditions.

[0005] A method for jointly predicting the amplitude distribution type and parameters of sea clutter in the prior art patent literature, based on measured X-band sea clutter data, extracts long sequence features and statistical features from the sea clutter data through feature engineering, constructs a data set, and then uses the MT1dCNN model to predict the amplitude distribution type and parameters of sea clutter. This technology has the following technical defects: (1) It is necessary to extract long sequence features and statistical features from the sea clutter data before inputting them into the neural network model, and it is impossible to achieve end-to-end prediction of the amplitude distribution of sea clutter; (2) The one-dimensional convolutional layer of this model limits the input data to only one-dimensional data and cannot achieve the prediction of the amplitude distribution of sea clutter in two-dimensional SAR images. (3) It does not consider ocean environmental parameters. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a method for predicting the amplitude distribution of SAR sea clutter based on a two-stream multi-task learning network, including: 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, use the labels corresponding to each SAR amplitude image to construct a SAR sea clutter amplitude distribution prediction data set; Step 3: Construct a SAR sea clutter amplitude distribution prediction model based on a two-stream multi-task learning network; Step 4: Use 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 use 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 in Step 1 on the SAR image to be predicted to obtain the SAR amplitude image to be predicted. 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.

[0007] In a preferred embodiment, in Step 2, the probability distribution histogram of the SAR amplitude image is statistically calculated, and the logarithmic cumulant algorithm is used for parameter estimation to calculate the KL divergence value: ; where g(x) and f(x) respectively represent two different probability density distributions, x represents the amplitude value of each pixel of 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.

[0008] In a preferred embodiment, Step 3 includes: Step 3.1: Construct an image feature encoder, input the SAR amplitude image into the image feature encoder to obtain a high-dimensional image feature Feature_image; Step 3.2: Construct an environmental feature encoder, input the ocean environmental parameters into the environmental feature encoder to obtain a high-dimensional parameter feature Feature_param; 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 a fused feature map Feature_fusion; Step 3.4: Construct a distribution type decoder, 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, input the fused feature map Feature_fusion into the distribution parameter decoder to obtain the prediction result of the sea clutter amplitude distribution parameters.

[0009] In a preferred embodiment, in Step 3.2, the ocean environmental parameters are input into the environmental feature encoder for channel combination to obtain a parameter matrix Input_P; the parameter matrix Input_P is passed to the first convolutional layer L1, and the spatial gradient features of different parameters are extracted using a convolutional kernel to generate a feature map F1: ; where Conv 3×3 represents a convolutional calculation with a size of 3×3; Perform batch normalization on the feature map F1, convert the distribution of the feature map F1 to a normal distribution, and obtain the normalization result BN1: ; Among them, represents the mean calculation, represents the variance calculation, is a decimal to prevent the denominator from being 0; perform activation calculation on the normalization result BN1 using the ReLU function to obtain the convolutional layer feature map Re1: ; Among them, represents the maximum value of the calculation 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: ; Among them, represents the maximum pooling operation; Successively perform calculations on multiple convolutional layers and pooling layers to obtain the high-dimensional feature map Feature_param.

[0010] 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 concatenated in the channel dimension to obtain the fused feature map Feature_fusion: ; Among them, represents the concatenate calculation.

[0011] In a preferred embodiment, the image feature encoder is alternately and serially composed of a plurality of dimensionality reduction residual blocks and a plurality of non-dimensionality reduction residual blocks, and each residual block includes a main branch, a shortcut connection branch, an addition operation, and an activation function; 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 convolutional layer, expands the number of channels, then performs batch normalization processing and activation function calculation; then passes through the second depthwise separable convolutional layer for feature extraction, and performs batch normalization processing, and then passes through the effective channel attention layer to enhance the spatial features. At the same time, the shortcut connection branch synchronously adjusts the channels and size through convolution and batch normalization processing, and the processing results of the main branch and the shortcut connection branch are added and then output the target size feature map.

[0012] In a preferred embodiment, in the non-dimensionality reduction residual block, the local features are enhanced by stacking two depthwise separable convolutional layers in the main branch, and the residual connection with identity mapping is combined to alleviate the vanishing gradient, and a high-dimensional feature map of the target size is output.

[0013] In a preferred embodiment, the environmental feature encoder is composed of multiple convolutional layers and multiple pooling layers connected in series alternately; 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, and the deep convolutional layers are used to capture the spatial correlation between wind and waves. The pooling layers are used for dimensionality reduction and enhancing translation invariance.

[0014] 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.

[0015] In a preferred embodiment, the distribution parameter decoder is composed of 3 fully connected layers connected in series.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention uses SAR image data and ERA5 meteorological grid data, and uses the dual-stream multi-task learning model SAR2SCAD to perform integrated prediction of the amplitude distribution type and parameters of sea clutter. Compared with the existing methods, the proposed method introduces ocean environmental parameters such as wind speed and wind direction through the dual-stream multi-task learning model, realizes cross-modal feature fusion of ocean environmental meteorological data and SAR images, effectively improves the prediction accuracy of the sea clutter amplitude distribution and the generalization ability of the model under different sea conditions. The F1 score of the sea clutter amplitude distribution type prediction is 86.6%, and the RMSE of the parameter prediction is 0.51, realizing accurate prediction of the SAR sea clutter amplitude distribution type and parameters, and providing prior knowledge and important reference for SAR ship target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is the SAR image and ERA5 data processing flow of the present invention; Figure 2 is the model structure of SAR2SCAD of the present invention; Figure 3 is the basic structure of R2, R4, R6, R8 of the present invention; Figure 4 is the basic structure of R1, R3, R5, R7 of the present invention; Figure 5 is the sea clutter amplitude distribution type prediction result of the present invention; Figure 6 is the SAR sea clutter amplitude distribution prediction result of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Step 1: Collect the wave mode SAR data of the GF-3 satellite to obtain multiple SAR amplitude images.

[0020] The wave mode SAR data of the GF-3 satellite includes a total of 3,597 single-look complex images with HH polarization. First, perform data cleaning on the SAR data, delete the images with poor quality, and obtain the SAR images containing pure sea clutter.

[0021] As Figure 1 shown, collect ERA5 data, then perform spatio-temporal matching on the filtered SAR images and ERA5 data to obtain the matched SAR data and the corresponding ERA5 data. Further, perform spatio-temporal interpolation on the ERA5 data to achieve pixel-level alignment with the SAR images. The wind speed, wind direction, wave height, and wave direction corresponding to the SAR images can be obtained from the ERA5 data, and the gridded data of wind speed, wind direction, wave height, and wave direction are obtained.

[0022] Crop the SAR images after data cleaning to a size of 1000 pixels × 1000 pixels, then calculate the amplitude of each SAR image and convert it to the TIFF image format to obtain 3,597 SAR amplitude images.

[0023] Preferably, the spatio-temporal interpolation method for ERA5 data uses the Kriging interpolation algorithm.

[0024] Step 2: Based on the multiple SAR amplitude images obtained in Step 1, use the labels corresponding to each SAR amplitude image to construct a SAR sea clutter amplitude distribution prediction dataset.

[0025] First, count the probability distribution histogram of the SAR amplitude images, denoted as P; then, select four commonly used sea clutter amplitude distribution models, including the K distribution, G0 distribution, generalized Gamma distribution, and Weibull distribution, and use these distribution models to estimate the parameters of the SAR amplitude map to obtain the shape parameters and scale parameters of each distribution model.

[0026] The shape parameters of different distribution models are respectively denoted as: μ_K, μ_G0, μ_Gamma, μ_Weibull; the scale parameters are denoted as: σ_K, σ_G0, σ_Gamma, σ_Weibull.

[0027] Substitute μ_K and σ_K into the probability density distribution function of the K-distribution to obtain the fitting result of the K-distribution, denoted as P_K; similarly, obtain the fitting results of the G0-distribution, generalized Gamma distribution, and Weibull distribution, denoted as P_G0, P_Gamma, and P_Weibull respectively.

[0028] Calculate the KL (Kullback Leibler) divergences between P_K, P_G0, P_Gamma, P_Weibull and P, denoted as KL_K, KL_G0, KL_Gamma, KL_Weibull respectively; compare the magnitudes of KL_K, KL_G0, KL_Gamma, KL_Weibull, and the distribution model with the smallest KL divergence value is the sea clutter amplitude distribution type of this SAR amplitude image, denoted as Label_class. At the same time, record the corresponding shape parameter and scale parameter of this distribution model, denoted as Label_shape and Label_scale respectively.

[0029] Perform the above operations on the 3597 SAR amplitude images obtained in step 1 in sequence to obtain Label_class, Label_shape, and Label_scale of each SAR amplitude image. Further, a SAR sea clutter amplitude distribution prediction dataset can be constructed using these SAR images and their corresponding labels, denoted as SAR_clutter. This dataset includes 3597 samples. Divide SAR_clutter into a training set Train, a validation set Val, and a test set Test according to the ratio of 8:1:1.

[0030] The parameter estimation method used is the method of logarithmic cumulants (MoLC) algorithm, and the goodness-of-fit evaluation index uses the KL divergence. The calculation formula of the KL divergence is as follows: ; where, x represents the amplitude value of each pixel of the SAR image, g ( x ) and f ( x ) represent two different probability density distributions respectively.

[0031] Step 3: Construct a SAR sea clutter amplitude distribution prediction model based on a two-stream multi-task learning network.

[0032] Denote the SAR sea clutter amplitude distribution prediction model as SAR2SCAD. The model structure of SAR2SCAD is as Figure 2 shown, and it consists of an image feature encoder, an environmental feature encoder, feature fusion, a distribution type decoder, and a distribution parameter decoder.

[0033] Step 3.1: Construct an image feature encoder. Input the SAR amplitude image into the image feature encoder to obtain high-dimensional image features Feature_image.

[0034] Denote the image feature encoder as IFE; IFE is composed of 8 residual blocks (R1, R2, R3, R4, R5, R6, R7, R8) connected in series, which is used to extract the deep semantic features of the SAR image. Input the SAR amplitude image obtained in Step 1 into IFE, and successively perform calculations through R1, R2, R3, R4, R5, R6, R7, R8 to extract the high-dimensional semantic features of the SAR amplitude image.

[0035] The input single-channel SAR image first performs coarse-grained feature extraction through R1 to capture basic texture and edge information, and then successively connects R2, R3, R4, R5, R6, R7 in series. Among them, R1, R3, R5, R7 are dimensionality reduction residual blocks, and their structures are as Figure 4 shown.

[0036] Among them, the structures of the 8 residual blocks include a main branch, a shortcut connection branch, an addition operation, and an activation function. Except for the different parameter settings of the convolutional layers in the main branch, it includes depthwise separable convolution, batch normalization, activation function, depthwise separable convolution, batch normalization, and effective channel attention.

[0037] Taking R1 as an example, the main branch reduces the input size from 1000×1000 to 500×500 and expands the number of channels from 1 to 64 through the first 3×3 depthwise separable convolutional layer (stride 2, padding 1), then performs batch normalization processing and activation function calculation, further refines the features through the second 3×3 depthwise separable convolution, and performs batch normalization processing, and then passes through the calculation of the effective channel attention layer to enhance the spatial feature extraction ability of the model. At the same time, the shortcut connection branch synchronously adjusts the channels and dimensions through 1×1 convolution (stride 2) and batch normalization processing. After the results of the main branch and the shortcut connection branch are added, a feature map of 500×500×64 is output.

[0038] Similarly, the size of the feature map output by R3 is 250×250×128, the size of the feature map output by R5 is 125×125×256, and the size of the feature map output by R7 is 62×62×512.

[0039] The non-dimensionality reduction residual blocks R2, R4, R6, R8 keep the input and output sizes and dimensions consistent, and the structures are as Figure 3As shown in the figure, two 3×3 depth-separable convolutions of the main branch are stacked to enhance the local feature expression ability, and the residual connection combined with the identity mapping effectively alleviates the gradient disappearance. IFE gradually compresses the spatial resolution and improves the channel dimension through four-dimensionality reduction, integrates the multi-scale semantic information of the SAR image, and finally outputs a 62×62×512 high-dimensional feature map. The high-dimensional image feature is recorded as Feature_Image.

[0040] Preferably, the activation function is the Leaky ReLU function, and the convolution uses the depthwise separable convolution, which greatly reduces the computational cost and the number of parameters while ensuring the model performance by decoupling the feature learning of the spatial and channel dimensions; the channel attention module uses the effective channel attention, which directly interacts with the channel features through global average pooling and one-dimensional convolution compared to the traditional channel attention, avoiding the information loss problem caused by dimensionality reduction in the traditional attention mechanism.

[0041] The calculation formula of the Leaky ReLU function is as follows: ; Among them, a represents the value of each pixel in the feature map.

[0042] Step 3.2: Construct an environmental feature encoder, input the ocean environmental parameters into the environmental feature encoder, and obtain a high-dimensional parameter feature Feature_param.

[0043] The environmental feature encoder is denoted as RFE; RFE consists of 4 convolutional layers (L1, L2, L3, L4) and 4 pooling layers (P1, P2, P3, P4) connected alternately in series to extract features of the input ocean environment parameters and obtain a high-dimensional feature map. The convolutional layer includes shallow convolutional layers and deep convolutional layers. The shallow convolutional layer is used to extract the local spatial gradient features of each parameter, while the deep convolutional layer is used to capture more complex patterns, such as the spatial correlation between wind and waves, and the extracted features are more abstract. The role of the pooling layer is to reduce dimensionality and enhance translation invariance, retaining important features.

[0044] The specific calculation process is as follows: The ocean environment parameters (wind speed, wind direction, wave height, wave direction) are input into RFE respectively. First, they are combined by channels to obtain a parameter matrix of 1000×1000×4, which is recorded as Input_P. Then Input_P is passed to the first convolution layer L1, and a convolution kernel of size 3×3 is used for calculation to extract the spatial gradient features of different parameters and mine the spatial correlation of different parameters. The generated feature map is recorded as F1, and the dimension of F1 is 1000×1000×64: ; Among them, Conv 3×3Represents a 3×3 convolution calculation.

[0045] Further, batch normalization is performed on the generated feature map F1 to convert the distribution of the feature map F1 into a normal distribution to solve the problem of internal covariate shift. The result of batch normalization is denoted as BN1: ; Among them, Represents the mean calculation, Represents the variance calculation, To prevent decimals with a denominator of 0.

[0046] The ReLU function is used to perform activation calculation on the normalized result BN1 to enhance the non-linear expression ability of the model, and the convolutional layer feature map Re1 is obtained: ; Among them, Represents the maximum value of the calculation variable.

[0047] Then, the convolutional layer feature map Re1 output by the first convolutional layer L1 is input into 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 perform downsampling on the feature map, achieve dimensionality reduction of the feature map, reduce the number of parameters, and at the same time remove redundant features and retain key features. After the calculation of P1, the obtained pooling layer feature map is denoted as Pool1, and the calculation process can be expressed as: ; Among them, Represents the maximum pooling operation.

[0048] According to the above formula, the calculations of the L2, P2, L3, P3, L4, and P4 layers are performed in sequence, and the high-dimensional feature map extracted by the environmental feature encoder RFE can be obtained, denoted as Feature_param.

[0049] RFE extracts the high-level semantic features of these marine environmental parameters through convolution operations.

[0050] Among them, the number of input channels and output channels of L1 are 4 and 64 respectively, the number of input channels and output channels of L2 are 64 and 128 respectively, the number of input channels and output channels of L3 are 128 and 256 respectively, and the number of input channels and output channels of L4 are 256 and 512 respectively; the convolutional kernel sizes of L1, L2, L3, and L4 are all 3×3, the stride is 1, and the activation function is the ReLU function. P1, P2, P3, and P4 are all maximum pooling layers, and the size of the pooling kernel is 3×3 and the stride is 2.

[0051] 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; Input the high-dimensional image feature Feature_image extracted by IFE and the high-dimensional parameter feature Feature_param extracted by RFE into the cross-modal feature fusion module. First, align Feature_image and Feature_param spatially, and then perform concatenation processing on Feature_image and Feature_param in the channel dimension to obtain the fused feature map, denoted as Feature_fusion; ; where, denotes the concatenate calculation.

[0052] Step 3.4: Construct a distribution type decoder, input the fused feature map Feature_fusion into the distribution type decoder, and obtain the prediction result of the sea clutter amplitude distribution type.

[0053] Denote the distribution type decoder as DC; the distribution type decoder is composed of a global average pooling layer, a flattening layer, and a fully connected layer (F1) in series. DC takes Feature_fusion as the input, performs global average pooling and flattening processing in sequence, and finally passes through F1 to obtain the prediction result of the sea clutter amplitude distribution type. Among them, the number of nodes in F1 is 4.

[0054] Step 3.5: Construct a distribution parameter decoder, input the fused feature map Feature_fusion into the distribution parameter decoder, and obtain the prediction result of the sea clutter amplitude distribution parameter.

[0055] Denote the distribution parameter decoder as DP; the distribution parameter decoder is composed of 3 fully connected layers in series, and the 3 fully connected layers are denoted as F2, F3, and F4 respectively. DP takes Feature_fusion as the input and passes through F2, F3, and F4 in sequence, and finally obtains the prediction result of the sea clutter amplitude distribution parameter. Among them, the number of nodes in F2, F3, and F4 are 1024, 512, and 2 respectively.

[0056] The construction of IFE, RFE, DC, and DP is completed through the above steps. Perform feature fusion on IFE and RFE, and then connect them in series with DC and DP respectively, that is, complete the construction of the SAR2SCAD model.

[0057] 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.

[0058] During the training process, use the classical Adam optimizer to train the model. The loss function consists of two parts. The classical cross-entropy is used as the loss function for the type prediction task, and Smooth-L1 is used as the loss function for the parameter prediction task. Use the uncertainty weight loss to assign weights to the losses of the distribution type prediction task and the distribution parameter prediction task, and dynamically adjust their weights during the training process.

[0059] After the training is completed, a sea clutter amplitude distribution prediction model is obtained, denoted as Trained_model.

[0060] Step 5: Test and evaluate the trained SAR sea clutter amplitude distribution prediction model Trained_model.

[0061] Use Test as the input, and use Trained_model obtained in Step 4 to perform model testing to obtain the sea clutter amplitude distribution prediction results. Then, adopt the classical evaluation index calculation method to calculate the distribution type prediction accuracy and the distribution parameter prediction accuracy respectively. Table 1 shows the sea clutter amplitude distribution prediction accuracies of the SAR2SCAD model and other classical deep learning algorithms.

[0062] Table 1 Sea clutter amplitude distribution prediction accuracies of different models

[0063] Table 1 shows that the F1 score of the distribution type prediction result of the SAR2SCAD model proposed in the present invention is 86.6%. The MAE of the distribution parameter prediction result is 0.25, and the RMSE is 0.51. The accuracy of the SAR2SCAD model is the best among all methods. The experimental results show that the SAR2SCAD model has the best accuracy in predicting the type and parameters of the SAR sea clutter amplitude distribution.

[0064] Figure 5 The confusion matrix showing the sea clutter amplitude distribution type prediction results is presented. K, G0, Gamma, and Weibull represent the K distribution, G0 distribution, generalized Gamma distribution, and Weibull distribution respectively. It can be seen that all types can be predicted relatively accurately.

[0065] Step 6: Perform the processing of Step 1 on the SAR image to be predicted, and then input it 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.

[0066] Figure 6 An example of using SAR2SCAD to predict the amplitude distribution of SAR sea clutter is given. It can be seen from the figure that the predicted PDF is very close to the original distribution histogram, proving the effectiveness of the SAR sea clutter amplitude distribution prediction method proposed by the present invention.

[0067] The sea clutter prediction results include the distribution type and distribution parameters, where the distribution parameters include the shape parameter and the scale parameter. Substitute the shape parameter and the scale parameter into the probability density distribution function (PDF) of the distribution, and draw the Figure 6 blue curve of, and the amplitude distribution of the sea clutter in the SAR image can be obtained.

[0068] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for predicting the amplitude distribution of SAR sea clutter based on a dual-stream multi-task learning network, characterized in that, Including: Step 1: Collect the wave-mode SAR data of the GF-3 satellite, and obtain multiple SAR amplitude images through data processing; Step 2: Based on the multiple SAR amplitude images obtained in Step 1, use the labels corresponding to each SAR amplitude image to construct a SAR sea clutter amplitude distribution prediction dataset; Step 3: Construct a SAR sea clutter amplitude distribution prediction model based on a two-stream multi-task learning network; Step 4: Use the training set of the SAR sea clutter amplitude distribution prediction dataset constructed in Step 2 to train the SAR sea clutter amplitude distribution prediction model, and use 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 in 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, characterized in that, In Step 2, the probability distribution histogram of the SAR amplitude image is statistically analyzed, the parameter estimation is performed using the logarithmic cumulant algorithm, and the KL divergence value is calculated: ； where \(g(x)\) and \(f(x)\) represent two different probability density distributions respectively, x represents the amplitude value of each pixel in the SAR image; the distribution model with the minimum 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 a dual-stream multi-task learning network according to claim 1, characterized in that Step 3 includes: Step 3.1: Construct an image feature encoder, input the SAR amplitude image into the image feature encoder, and obtain a 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 a high-dimensional parameter feature Feature_param; 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 a fused feature map Feature_fusion; Step 3.4: Construct a distribution type decoder, input the fused feature map Feature_fusion into the distribution type decoder, and obtain the prediction result of the sea clutter amplitude distribution type; Step 3.5: Construct a distribution parameter decoder, input the fused feature map Feature_fusion into the distribution parameter decoder, and obtain the prediction result of the sea clutter amplitude distribution parameter.

4. The SAR sea clutter amplitude distribution prediction method based on the dual-stream multi-task learning network according to claim 3, characterized in that, In Step 3.2, the ocean environmental parameters are input into the environmental feature encoder for channel combination to obtain a parameter matrix Input_P; the parameter matrix Input_P is passed to the first convolutional layer L1, and the spatial gradient features of different parameters are extracted using the convolutional kernel to generate a feature map F1: ； Among them, Conv 3×3 represents a convolution calculation with a size of 3×3; Perform batch normalization on the feature map F1 to convert the distribution of the feature map F1 into a normal distribution to obtain a normalized result BN1: ； Among them, represents mean calculation, represents variance calculation, is a decimal number to prevent the denominator from being 0; the ReLU function is used to perform activation calculation on the normalization result BN1 to obtain the convolutional layer feature map Re1: ； Among them, represents the maximum value of the calculation variable; The convolutional layer feature map Re1 of the first convolutional layer L1 is input into the first pooling layer P1 to obtain a pooling layer feature map Pool1: ； Among them, represents the maximum pooling operation; Perform calculations of multiple convolutional layers and pooling layers in sequence to obtain a high-dimensional feature map Feature_param.

5. The SAR sea clutter amplitude distribution prediction method based on a two-stream multi-task learning network according to claim 3, 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 concatenated in the channel dimension to obtain the fused feature map Feature_fusion: ； Among them, represents the concatenate calculation.

6. The SAR sea clutter amplitude distribution prediction method based on the dual-stream multi-task learning network according to claim 3, characterized in that, The image feature encoder is composed of multiple alternating series-connected dimensionality reduction residual blocks and multiple non-dimensionality reduction residual blocks. Each residual block includes a main branch, a shortcut connection branch, an addition operation, and an activation function.

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

8. The SAR sea clutter amplitude distribution prediction method based on a two-stream multi-task learning network according to claim 6, characterized in that In the non-dimensionality reduction residual block, the local features are enhanced by stacking two depthwise separable convolutional layers in the main branch, and the residual connection of the identity mapping is combined to alleviate the vanishing gradient, and a high-dimensional feature map of the target size is output.

9. The SAR sea clutter amplitude distribution prediction method based on the dual-stream multi-task learning network according to claim 3, wherein, The environmental feature encoder is composed of multiple alternating series-connected convolutional layers and multiple pooling layers; 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, and the deep convolutional layers are used to capture the spatial correlation between wind and waves. The pooling layers are used for dimensionality reduction and enhancing translational invariance.

10. The SAR sea clutter amplitude distribution prediction method based on the dual-stream multi-task learning network according to claim 3, wherein The distribution type decoder is composed of a global average pooling layer, a flattening layer, and a fully connected layer connected in series.

Images