A SAR image water area detection method based on a deep coding network

Abstract

The application discloses a SAR image water area detection method based on a deep coding and decoding network, and applies to the technical field of image processing and deep learning. The method comprises the following steps: collecting SAR image data, labeling the images, and constructing a sample database; building a water area detection model based on the deep coding and decoding network, including an Encoder module, a Decoder module and a self-adaptive focusing pyramid module; training and testing the water area detection model, calculating the water area detection rate and the false alarm rate, and evaluating the model performance; and applying the model to an actual SAR image water area detection task. The application can more accurately extract water area features of different scales in the SAR image, greatly improve the accuracy of the segmentation boundary, and significantly improve the precision of the water area detection; the application has stronger anti-interference ability to complex ground object interference such as mountain shadow, marsh, farmland dark spot and the like, and can be applied to SAR image detection tasks of various satellites.

Description

Technical Field

[0001] This invention relates to the fields of image processing and deep learning technology, and more specifically to a method for detecting water areas in SAR images based on deep encoding and decoding networks. Background Technology

[0002] Accurate acquisition of water body information is crucial in numerous fields, including geographic information monitoring, water resource management, and disaster early warning. Traditional methods for extracting candidate water areas are often designed based on simple, idealized scenarios. These scenarios typically feature small size, simple composition, a large proportion of water area, or significant contrast between the target and background. However, actual satellite-received scenarios are extremely complex, containing a wide variety of undefined land features, and the grayscale characteristics of these features are not significant. Furthermore, images often contain interference from complex features such as mountain shadows, swamps, and dark patches in farmland. This makes traditional methods prone to false alarms in practical applications, hindering the achievement of ideal extraction results.

[0003] While SAR satellites, with their higher orbital altitudes, can cover a wider area of ​​the Earth's surface, this also makes the scenes they encompass more complex and diverse. Faced with such complex scenes, traditional detection algorithms perform poorly in terms of robustness, failing to meet the requirements of accuracy and stability for water area detection in practical applications. A robust water area candidate region extraction algorithm is needed to adapt to complex satellite reconnaissance scenarios and achieve accurate water area detection. Therefore, how to provide a SAR image water area detection method based on deep encoding and decoding networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a method for detecting water bodies in SAR images based on deep coding and decoding networks. By optimizing the water body detection model based on deep coding and decoding networks, the method solves the problems of low detection accuracy and high false alarm rate of traditional methods in complex scenes, and achieves accurate and efficient detection of water bodies in SAR images.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for detecting water areas in SAR images based on deep encoding and decoding networks includes the following steps:

[0007] S1. Data preparation: Collect SAR image data and label the images, construct a water training and test sample database, and preprocess the data;

[0008] S2. Model Building: Build a water area detection model based on a deep encoder-decoder network, including an Encoder module, a Decoder module, and an adaptive focusing pyramid module;

[0009] S3. Model Training: The water area detection model is trained using the training set in the sample database. The network is optimized using the stochastic gradient descent algorithm, and the weights are updated through backpropagation.

[0010] S4. Model Testing and Evaluation: Test the trained water detection model using the test set in the sample database, calculate the water detection rate and false alarm rate, and evaluate the performance of the water detection model.

[0011] S5. Practical Application: Apply the water area detection model that meets the performance requirements to actual SAR image water area detection tasks.

[0012] Optionally, the Encoder module includes a MobileNet network with dilated convolutions and a spatial pyramid pooling module with dilated convolutions; the dilated convolutions control the size of the receptive field through the parameter rate, and the corresponding rate value is set according to different output_stride.

[0013] Optionally, the MobileNet network is constructed using depthwise separable convolutions, which decompose standard convolutions into a DW convolution and a PW convolution. The DW convolution applies each kernel to each channel, while the PW convolution combines the outputs of the channel convolutions. The MobileNet network uses the Adaptive-Act activation function for non-linear feature mapping.

[0014] Optionally, the water area detection model uses a batch normalization algorithm to process the input data.

[0015] Optionally, the boundary-guided context information aggregation module utilizes the natural division of the category by the boundary as acquired context information to enhance intra-class consistency.

[0016] Optionally, the spatial pyramid pooling module includes one 1×1 convolutional layer, three 3×3 dilated convolutions, and one global average pooling layer. The global average pooling layer outputs image-level features. The image-level features output by the spatial pyramid pooling module are fed into a 1×1 convolutional layer and interpolated to the original size using bilinear interpolation. Features of different scales are concatenated along the channel dimension and then fed into a 1×1 convolutional layer for fusion to obtain new features as the output of the Encoder module.

[0017] Optionally, in S3, the cross-entropy function is selected as the loss function, and the derivative of the loss function is calculated using the stochastic gradient descent algorithm and the chain rule to update the network weights.

[0018] Optionally, the adaptive focusing pyramid module is positioned between the Encoder and Decoder modules. The adaptive focusing pyramid module consists of adaptive focusing units of multiple scales. Each adaptive focusing unit includes a focusing operation and a diffusion operation. The focusing operation uses an adaptive weight allocation mechanism to weight the features of different regions, highlighting the features related to water. The diffusion operation spreads the focused features spatially. The features processed by the adaptive focusing pyramid module are sent to the Decoder module and fused with the original features to jointly participate in the segmentation and prediction process.

[0019] Optionally, the Decoder module performs bilinear interpolation on the features to obtain 4x features, which are then fused with low-level features of the corresponding size in the Encoder module. First, a 1×1 convolution is used to reduce the dimensionality of the low-level features, and after feature fusion, a 3×3 convolution is used to further fuse the features. Finally, bilinear interpolation is used to obtain a segmentation prediction of the same size as the original image.

[0020] As can be seen from the above technical solution, compared with the prior art, the present invention provides a SAR image water area detection method based on a deep encoding and decoding network, which has the following beneficial effects:

[0021] 1. Improved Detection Accuracy: This invention employs an advanced Deeplab V3 network structure, combining dilated convolution, spatial pyramid pooling modules, and decoder optimization techniques. Furthermore, the adaptive focusing pyramid module performs focusing and diffusion processing on water features, enabling more accurate extraction of water features at different scales from SAR images. This significantly improves the accuracy of segmentation boundaries, thereby dramatically enhancing the accuracy of water area detection. Compared to traditional methods and methods without this module, it can more accurately identify and segment water areas in complex scenes, effectively reducing false alarms. The detection rate for water areas larger than 200 pixels can reach 85%-90%, with a false alarm rate of less than 15%.

[0022] 2. Enhanced robustness: This invention selects MobileNet as a lightweight base network and optimizes it. At the same time, it adopts a batch normalization processing algorithm and a boundary-guided context information aggregation module, combined with an adaptive focusing pyramid module, which makes the model more adaptable to complex satellite reconnaissance scenarios and has a stronger anti-interference ability against complex terrain interference such as mountain shadows, swamps, and dark spots in farmland, thus further enhancing the robustness of the model.

[0023] 3. Wide adaptability: The water area detection method of the present invention can be applied to SAR image detection tasks of various satellites, such as GF3 B / C series satellites. It can achieve stable and efficient water area detection on images of different resolutions and in various complex scenarios, and has broad application prospects.

[0024] 4. Data accumulation and support: In practical applications, a database of water training and testing samples can be formed by accumulating and annotating a large amount of SAR satellite water data, laying the foundation for subsequent research and applications in related fields. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a flowchart of the SAR image water area detection method based on a deep encoding and decoding network according to the present invention;

[0027] Figure 2 This is a schematic diagram of the water area detection model based on a deep codec network according to the present invention;

[0028] Figure 3 This is a schematic diagram of the spatial pyramid pooling module of the present invention;

[0029] Figure 4 This is a schematic diagram of the adaptive focusing pyramid module of the present invention;

[0030] Figure 5 This is the HV polarization geometry correction diagram in an embodiment of the present invention;

[0031] Figure 6 The truth graph in the embodiments of the present invention

[0032] Figure 7 This is a binary image of the detection result in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] This invention discloses a method for detecting water areas in SAR images based on a deep codec network, such as... Figure 1 As shown, it includes the following steps:

[0035] S1. Data preparation: Collect SAR image data and label the images, construct a water training and test sample database, and preprocess the data;

[0036] S2. Model Building: Build a water area detection model based on a deep encoder-decoder network, such as... Figure 2 As shown, it includes the Encoder module, the Decoder module, and the adaptive focusing pyramid module;

[0037] S3. Model Training: The water area detection model is trained using the training set in the sample database. The network is optimized using the stochastic gradient descent algorithm, and the weights are updated through backpropagation.

[0038] S4. Model Testing and Evaluation: Test the trained water detection model using the test set in the sample database, calculate the water detection rate and false alarm rate, and evaluate the performance of the water detection model.

[0039] S5. Practical Application: Apply the water area detection model that meets the performance requirements to actual SAR image water area detection tasks.

[0040] In this embodiment of the invention, the data collected by S1 should cover different scenarios, such as pure land, large bodies of water, mountainous areas, urban areas, and land-sea junctions, etc., including HV polarization geometry correction maps of secondary satellite imagery of water areas, as shown in the example. Figure 5 As shown, the collected images are labeled, with water areas marked as 1 (represented in black in the output binary image) and land areas marked as 0 (represented in white in the output binary image), thus constructing a water area training and testing sample database. The labeling process must be strictly carried out according to a unified standard to ensure the accuracy and consistency of the labeling. The ground truth labeling results are as follows: Figure 6 As shown, after the annotation is completed, the data is preprocessed, including image cropping, normalization and other operations, so that the data meets the requirements of network input.

[0041] In S4, the test results are compared with the pre-set indicator requirements to evaluate whether the model's performance meets the needs of practical applications. If the model performance does not meet the requirements, the model parameters can be further adjusted, the network structure optimized, or the amount of training data increased, and the training and testing can be repeated until the model performance meets the requirements.

[0042] like Figure 7 As shown, the model outputs a binary map of the extracted water body, providing accurate water information support for related fields such as geographic information monitoring, water resource management, and disaster early warning.

[0043] Furthermore, the Encoder module includes a MobileNet network with dilated convolutions and a spatial pyramid pooling module with dilated convolutions, such as... Figure 3As shown, dilated convolution controls the size of the receptive field through the parameter rate, and the corresponding rate value is set according to different output_stride.

[0044] Furthermore, the MobileNet network is constructed using depthwise separable convolutions, decomposing standard convolutions into a DW convolution and a PW convolution. The DW convolution applies each kernel to each channel, while the PW convolution combines the outputs of the channel convolutions. The MobileNet network uses the Adaptive-Act activation function for non-linear feature mapping.

[0045] The Adaptive-Act function dynamically adjusts the activation threshold and slope based on the features of the input data. Specifically, the Adaptive-Act function introduces an adaptive mechanism to perform local statistical analysis on the input feature map, automatically determining the most suitable activation parameters based on the feature distribution in different regions. In regions of the image with rich and rapidly changing features, the function uses a larger slope and threshold to fully activate neurons and retain more detailed information; while in regions with relatively flat features and less change, the slope and threshold are appropriately reduced to avoid information loss due to overactivation. This adaptive characteristic allows the Adaptive-Act function to better adapt to different types of input data, further improving the model's ability to extract and express complex image features. The expression is as follows:

[0046] ;

[0047] In the formula, The dynamic threshold represents the non-linear interval threshold of the activation function, which is dynamically determined by the mean absolute value of the local region (such as a 3×3 neighborhood) where the input x is located. The expression is as follows:

[0048] ;

[0049] In the formula, This is the scaling factor (default 1.2). To prevent the use of tiny constants with a threshold of 0 (default 1e-5);

[0050] The strong activation coefficient is calculated using the following formula:

[0051] ;

[0052] when Effective immediately, through the Sigmoid transformation of local variance ( Dynamically enhance activation strength; the greater the feature fluctuation, the stronger the activation strength. The closer it is to 2.

[0053] The weak activation coefficient is calculated using the following formula:

[0054] ;

[0055] when It takes effect immediately, adjusting the activation intensity inversely through local variance, in areas with smooth features. Approaching 0.5 reduces noise amplification.

[0056] Bneck is the core module of MobileNet-v3 and a fundamental module of the network. It implements channel-separable convolution, SE channel attention mechanism, and residual connections. The SE module pools the channels of the obtained feature matrix to obtain one-dimensional elements corresponding to each channel. Then, it passes through a fully connected layer and uses the ReLU activation function to reduce the number of channels in the expanded layer to 1 / 4 of the original number. Next, a second fully connected layer maintains the same number of channels as before pooling and uses the H-signal activation function. Its role is to automatically learn the importance of each feature channel, thereby enhancing useful features and suppressing features that are not useful for the current task.

[0057] Furthermore, the water area detection model employs a batch normalization algorithm to process the input data.

[0058] Define the batch size as m, and the data contained within the batch size as B = x{x1, x2, ..., x}. m The entire processing can be viewed as performing a linear transformation on the data to obtain y1, y2, ..., y3. m The input y to the next layer is derived by a specific method, which involves the standard deviation and variance of the data in the batch size, the normalized value, and the constant ε used to ensure stability when calculating the variance, as well as the parameters γ and β that need to be learned during the batch normalization process.

[0059] Furthermore, the boundary-guided context information aggregation module utilizes the natural division of the category by the boundary as acquired context information to enhance intra-class consistency.

[0060] In this embodiment of the invention, given a feature map A∈RC1×H×W generated by the backbone network and a boundary feature map B∈RC2×H×W generated by the multi-scale boundary extraction module, they are processed by two convolutional layers to obtain two new feature maps {A1,B1}∈RC×H×W. The features are then stretched to RC×N, where N=H×W is the number of pixels. The transposed and stretched A1 and B1 are then multiplied by a matrix and a softmax function is applied to obtain the boundary semantic similarity feature F. Finally, the contextual information along the boundary in A is aggregated into internal pixels to obtain a feature map D enhanced by boundary context aggregation. By using the boundary-guided contextual information aggregation module, intra-class consistency can be better guaranteed in the event of ambiguity, improving the accuracy of water area detection.

[0061] Furthermore, the spatial pyramid pooling module includes one 1×1 convolutional layer, three 3×3 dilated convolutions, and one global average pooling layer. The global average pooling layer outputs image-level features. The image-level features output by the spatial pyramid pooling module are fed into a 1×1 convolutional layer and interpolated to the original size using bilinear interpolation. Features of different scales are concatenated together along the channel dimension and then fed into a 1×1 convolutional layer for fusion, resulting in new features that serve as the output of the Encoder module.

[0062] In this embodiment of the invention, output_stride=16, and the rate of the three 3×3 dilated convolutions is (6, 12, 18); if output_stride=8, the rate is doubled. The number of output channels for these 1×1 convolutional layers and the three 3×3 dilated convolutions are all 256, and all contain a Batch Normalization (BN) layer to ensure network stability and training efficiency. Image-level features are fed into the 1×1 convolutional layers and output 256 channels, which are then fused to obtain a new 256-channel feature.

[0063] Furthermore, in S3, the cross-entropy function is selected as the loss function, and the derivative of the loss function is calculated using the stochastic gradient descent algorithm and the chain rule to update the network weights.

[0064] Define Y = {y1, y2, ..., y} n} represents each pixel of the image, D = x{x1, x2, ..., x3}. n The corresponding tag category, where y i =1 indicates that the pixel belongs to the water area, y i =0 represents that the pixel belongs to the non-water area (0 < i < n, where n is the number of pixels in the image); the model output is the probability that each pixel belongs to the water area. The training process is equivalent to maximizing the probability that the pixel belongs to the correct category. Therefore, the cross-entropy function is chosen as the loss function.

[0065] In this embodiment of the invention, appropriate training parameters, such as learning rate and number of iterations, are selected. During the training process, close attention is paid to the changes in the loss function and the performance of the model on the test set, such as the detection rate and false alarm rate. If the model is found to be overfitting or underfitting, the training parameters or network structure are adjusted in a timely manner to ensure that the model can stably and effectively learn the features of water areas in SAR images.

[0066] Furthermore, the adaptive focusing pyramid module is positioned between the Encoder and Decoder modules, such as... Figure 4 As shown, the adaptive focusing pyramid module consists of multiple adaptive focusing units of various scales. Each adaptive focusing unit includes a focusing operation and a diffusion operation. The focusing operation uses an adaptive weight allocation mechanism to weight the features of different regions, highlighting the features related to water. The diffusion operation spreads the focused features spatially. The features processed by the adaptive focusing pyramid module are fed into the Decoder module and fused with the original features to jointly participate in the segmentation prediction process, further improving the segmentation accuracy of complex water scenes and the detection capability of small water areas.

[0067] In this embodiment of the invention, the focusing operation utilizes an attention mechanism to calculate the importance weight of each pixel position in different channels, and then performs weighted processing on the feature map, thereby enhancing regions closely related to water features and reducing interference from irrelevant information. The diffusion operation, through careful design of the convolution kernel size and stride, allows feature information to diffuse within local regions, enabling adjacent pixels to share features, thus better capturing the overall shape and texture features of the water. Focusing and diffusion units of different scales can capture details of water areas of different sizes, comprehensively improving the understanding and expression of water features from micro to macro perspectives.

[0068] Furthermore, the Decoder module performs bilinear interpolation on the features to obtain 4x features, which are then fused with low-level features of the corresponding size in the Encoder module. First, a 1×1 convolution is used to reduce the dimensionality of the low-level features, and after feature fusion, a 3×3 convolution is used to further fuse the features. Finally, bilinear interpolation is used to obtain a segmentation prediction of the same size as the original image.

[0069] In this embodiment of the invention, the model can be further optimized and adjusted according to different application scenarios and needs. For example, for water area detection in a specific region, more data from that region can be collected specifically for training, improving the model's adaptability and detection accuracy for that region. Simultaneously, combining other auxiliary information, such as geographic information data and meteorological data, can further enhance the accuracy and reliability of water area detection.

[0070] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for water area detection of SAR image based on deep coding network, characterized in that, Comprise the following steps: S1, data preparation: collect SAR image data and label the image, build a water area training test sample database, and preprocess the data; S2, model building: build a water area detection model based on a deep coding and decoding network, including an Encoder module, a Decoder module, and a self-adaptive focusing pyramid module; S3, model training: use the training set in the sample database to train the water area detection model, use the stochastic gradient descent algorithm to optimize the network, and update the weights through back propagation; S4, model testing and evaluation: use the test set in the sample database to test the trained water area detection model, calculate the water area detection rate and false alarm rate, and evaluate the performance of the water area detection model; S5, practical application: apply the water area detection model that meets the performance requirements to the actual SAR image water area detection task; The adaptive focusing pyramid module is arranged between the Encoder module and the Decoder module, the adaptive focusing pyramid module is composed of adaptive focusing units of multiple scales, each adaptive focusing unit includes a focusing operation and a diffusion operation; the focusing operation assigns weights to features in different regions through an adaptive weight distribution mechanism, highlighting features related to water areas; the diffusion operation diffuses and propagates the focused features in space, and the features processed by the adaptive focusing pyramid module are sent to the Decoder module to fuse with the original features and participate in the segmentation prediction process together; The focusing operation calculates the importance weight of each pixel position in different channels using the attention mechanism, and then performs weighted processing on the feature map, so that the region closely related to the water feature is enhanced and the interference of irrelevant information is reduced; the diffusion operation diffuses the feature information in the local area through careful design of the convolution kernel size and step, so that adjacent pixels can share features, thereby better capturing the overall shape and texture features of the water area; the focusing and diffusion units of different scales capture the details of water areas of different sizes, from micro to macro, comprehensively improving the understanding and expression ability of water features.

2. The SAR image water area detection method based on deep coding network according to claim 1, characterized in that, The Encoder module includes a MobileNet network with a hollow convolution and a spatial pyramid pooling module with a hollow convolution; the hollow convolution controls the size of the receptive field through the parameter rate, and sets the corresponding rate value according to different output_stride.

3. The SAR image water area detection method based on deep coding network according to claim 2, characterized in that, The MobileNet network uses a depth separable convolution to build, which decomposes a standard convolution into a DW convolution and a PW convolution, the DW convolution applies each convolution kernel to each channel, and the PW convolution is used to combine the output of the channel convolution; the MobileNet network uses an Adaptive-Act activation function for nonlinear mapping of features; The Adaptive-Act function can dynamically adjust the activation threshold and slope according to the characteristics of the input data, specifically, the Adaptive-Act function performs local statistical analysis on the input feature map by introducing an adaptive mechanism, automatically determines the most suitable activation parameters according to the feature distribution in different regions, and the expression is as follows: ； In the formula, is a dynamic threshold value, representing the threshold value of the nonlinear interval of the activation function, determined dynamically by the absolute value mean of the local region where the input x is located. The expression is as follows: ； wherein is a scaling factor, is a small constant to prevent threshold values of 0; For the strong activation coefficient, the calculation formula is as follows: ； When takes effect, the dynamic enhancement activation strength is transformed by a Sigmoid transformation of the local variance , and the greater the characteristic fluctuation is, the closer to 2 ; is a weak activation coefficient, and the calculation formula is as follows: ； When takes effect, the activation strength is inversely adjusted by local variance.

4. The SAR image water area detection method based on deep coding network according to claim 1, characterized in that, The water area detection model adopts a batch normalization processing algorithm to process input data.

5. The SAR image water area detection method based on deep coding network according to claim 1, characterized in that, The boundary-guided context information aggregation module utilizes the natural division of the boundary within the class as the obtained context information to enhance the intra-class consistency; the input of the boundary-guided context information aggregation module is a feature map A and a boundary feature map B, two new feature maps A1 and B1 are obtained through two convolutional layer processing, matrix multiplication is performed on the transposed and stretched A1 and B1, and a softmax function is applied to obtain boundary semantic similarity features F, the context information along the boundary in A is aggregated into internal pixels to obtain a feature map D enhanced by boundary context aggregation.

6. The SAR image water area detection method based on deep coding network according to claim 2, characterized in that, The spatial pyramid pooling module includes one 1x1 convolutional layer, three 3x3 atrous convolutions, and one global average pooling layer, the global average pooling layer outputs image-level features, the image-level features output by the spatial pyramid pooling module are input into one 1x1 convolutional layer, and are bilinearly interpolated to the original size, features of different scales are concatenated in the channel dimension, and are input into a 1x1 convolution for fusion to obtain new features as the output of the encoder module.

7. The SAR image water area detection method based on deep coding network according to claim 1, characterized in that, In S3, the cross-entropy function is selected as the loss function, the loss function derivative is calculated through the stochastic gradient descent algorithm and the chain rule, and the network weight is updated.

8. The SAR image water area detection method based on deep coding network according to claim 1, characterized in that, The decoder module performs bilinear interpolation on the features to obtain 4x features, and then fuses the features with low-level features of the corresponding size in the encoder module; first, a 1x1 convolution is used to reduce the dimension of the low-level features, then the features are fused, and then a 3x3 convolution is used to further fuse the features, and finally, bilinear interpolation is performed to obtain a segmentation prediction with the same size as the original picture.

Images