Water segmentation method for SAR remote sensing images based on deep encoding and decoding fusion network

Abstract

The present invention relates to a SAR remote sensing image water segmentation method based on a deep encoding and decoding fusion network, and belongs to the field of image segmentation technology. The method comprises the following steps: (1) constructing a data set; (2) training data annotation; (3) data enhancement; (4) constructing a deep encoding network; (5) constructing a deep decoding network; (6) constructing a codec network loss function; (7) model training and verification; (8) segmentation processing; (9) saving or outputting the segmentation results according to requirements. The model constructed by the present invention has strong generalization, high robustness, and excellent performance, which solves the drawbacks of the current water extraction technology, can realize the rapid and accurate segmentation of waters in SAR remote sensing images, and is easy to promote and apply.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and relates to a SAR remote sensing image water segmentation method, and in particular to a SAR remote sensing image water segmentation method based on a deep encoding and decoding fusion network. Background Art

[0002] Inland surface water resources are important natural resources for human survival and play a vital role in the hydrological and biochemical cycles of the earth. Realizing rapid and accurate water resource surveys is of great significance for environmental monitoring and disaster warning. Synthetic Aperture Radar (SAR) is an active microwave imaging sensor that is not affected by factors such as clouds, rain, fog, and light, and can achieve all-day and all-weather earth observation. Spaceborne SAR is mounted on a satellite platform and has the characteristics of being free from airspace restrictions, a wide observation range, and periodic revisits. It has been widely used in various fields of earth observation. In recent years, because water areas show more significant identification features in SAR remote sensing images and are easy to quickly separate from complex scenes, water area detection technology based on SAR images has been widely used and developed.

[0003] To date, a variety of SAR image water segmentation methods have been proposed, which can be divided into the following categories: threshold segmentation method, active contour model method (ACM), cluster segmentation method, and deep learning method. (1) The threshold segmentation method completes the segmentation of SAR images at the gray level by setting one or more thresholds. This method is usually regarded as one of the most commonly used methods due to its simple operation and high computational efficiency. Its disadvantages are that it is difficult to select the optimal threshold and it is easily affected by the coherent speckle noise of SAR images; (2) The SAR image water segmentation method based on ACM is mainly completed through two steps: coarse segmentation and fine segmentation. First, the approximate water boundary is obtained, and then the contour line is iteratively evolved according to the gradient information to finally obtain a more accurate water boundary. The advantage of this method is that it can obtain continuous and smooth water boundaries. The disadvantage is that the amount of calculation is large and it is highly dependent on the initial contour; (3) The cluster segmentation method is based on Euclidean distance, and the similarity between pixels and cluster centers is calculated to complete the classification of similar objects in SAR images. This method can realize unsupervised classification of SAR image objects. Its disadvantages are low efficiency in large-scale image computing and susceptibility to interference from coherent speckle noise. (4) The deep learning-based segmentation method trains the network model by manually creating a large number of training samples and manually calibrating label maps to identify water areas in the input image. The advantages of this method are strong robustness and not being easily affected by coherent speckle noise. The disadvantage is that a large number of sample sets need to be created in advance, which is labor-intensive.

[0004] Traditional water segmentation methods are generally carried out at the pixel level of the image, which can achieve water segmentation with pixel-level accuracy, and its segmentation accuracy is limited by the resolution. Remote sensing SAR waters have good segmentation effects when the landform type is simple, the water area is large, and the sea conditions are low. However, the scenes to be processed in remote sensing applications are often complex and changeable, and the deployment on the satellite platform has great limitations on the timeliness and resource occupation of the processing algorithm. The above algorithms are difficult to effectively guarantee the dual optimization of water segmentation accuracy and processing efficiency, and cannot be effectively equipped and deployed on the terminal processing platform. Therefore, how to overcome the shortcomings of the existing technology is an urgent problem to be solved in the field of image segmentation technology. Summary of the invention

[0005] The purpose of the present invention is to solve the deficiencies of the prior art and provide a SAR remote sensing image water segmentation method based on a deep encoding and decoding fusion network. First, the GF-3 satellite data in the beamforming mode is collected and divided into a training data set and a verification data set. The training data set is then annotated to obtain a label set; and the training data set is amplified to obtain an augmented data set. Then, deep encoding and decoding networks are constructed respectively to effectively identify water pixels, thereby quickly extracting water features. On this basis, operations such as deep separation convolution and dilated convolution are used to extract the feature representation of the input image under multiple resolution conditions, and a deep encoding and decoding fusion network model is constructed. Afterwards, training and verification are carried out to verify the segmentation results of the data set data and the SAR image of the image to be segmented, which proves the effectiveness of the proposed method and the feasibility of the project.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network includes the following steps:

[0008] Step (1), constructing a data set: collecting N scenes of SAR image data, with a single scene size of no less than 10k × 10k pixels, and dividing the data into a training data set and a validation data set;

[0009] Step (2), training data annotation: read the SAR remote sensing image in the training data set, annotate the water area location information in the image, and obtain a label set;

[0010] Step (3), data enhancement: amplify the SAR remote sensing images in the training data set to obtain an augmented data set;

[0011] Step (4), constructing a deep coding network: the deep coding network includes a feature extraction network and a feature pyramid network module composed of dilated convolutions, the feature extraction network is used to extract and learn the water features in the image and generate a feature map; the feature pyramid network module composed of dilated convolutions is used to pool the extracted feature map to increase the context information summarization capability of the extracted feature map;

[0012] Step (5), constructing a deep decoding network: the deep decoding network includes a fusion module and a decoding network module, the fusion module is used to fuse feature maps of different resolutions to improve the model's detection capability for targets of different sizes, and the decoding network module is used to decode the water area location information predicted by the output network;

[0013] The feature extraction network and the feature pyramid network module composed of dilated convolution are connected to the fusion module to form a deep encoding and decoding fusion network model;

[0014] Step (6), constructing the encoding and decoding network loss function: using the cross entropy loss function to minimize the objective function for model training, and setting the loss function output node to 2 categories according to the actual task requirements of water segmentation;

[0015] Step (7), model training and verification: input the training data set and label set into the built deep codec fusion network model for training, obtain the network model architecture of the Mth generation of training, then load the trained model, and test it based on the verification data set, calculate the classification accuracy, and save the model parameters with the best accuracy;

[0016] Step (8), segmentation processing: load the optimal model parameters, input the SAR image to be segmented into the deep encoding and decoding fusion network model, obtain label information, and finally correspond the label information to the SAR image to be segmented to obtain the segmentation result of the water area image;

[0017] Step (9), save or output the segmentation results as required.

[0018] Furthermore, preferably, in step (1), SAR image data of the Gaofen-3 satellite in the spotlight mode is collected; and the data ratio between the training data set and the validation data set is 8:2.

[0019] Furthermore, preferably, in step (2), Labelme annotation software is used to select and annotate the location information of the water area in the image point by point along the edge of the water area.

[0020] Furthermore, preferably, in step (3), the amplification method includes rotation, translation, scaling, grayscale transformation and Mosaic data enhancement method.

[0021] Further, preferably, in step (4),

[0022] (4.1) Construct the first-stage deep coding network: This network consists of three residual network modules. The main branch of each residual network module uses three depth-separable convolutional layers, and the cross-layer connection branch uses one conventional convolutional layer to reduce the dimension of the input feature map and increase the channel dimension.

[0023] (4.2) Construct the second stage deep encoding network: The network consists of 4 identical separable convolutional modules, each of which consists of 3 depth-separable convolutional layers based on residual connections;

[0024] (4.3) Construct the third-stage deep encoding network: This network contains two depth-wise separable convolutional modules. The first convolutional module uses three depth-wise separable convolutional layers, and the cross-layer connection branch uses one regular convolutional layer. The second convolutional module uses a depth-wise separable convolutional layer, followed by a dilated convolutional layer.

[0025] Furthermore, preferably, the operation steps in step (5) are: taking the feature map extracted by the second-stage deep coding network constructed by the feature extraction network as a low-dimensional feature input, performing convolution processing on the feature map extracted by the third-stage deep coding network through a feature pyramid network module composed of dilated convolutions, obtaining a new feature map of the same size as the feature map extracted by the second-stage deep coding network, taking the feature map as a high-dimensional feature input, and performing channel accumulation of the two-dimensional features through a fusion module to construct a deep decoding network.

[0026] Furthermore, preferably, in step (6), the loss function output node is set to 2 categories to construct a binary cross entropy loss function.

[0027] Furthermore, preferably, the specific method of step (7) is to input the SAR remote sensing training data with label information into the deep encoding and decoding fusion network, obtain the image feature extraction result through the deep encoding network, perform pixel-by-pixel binary classification on the extracted features through the fusion module and the decoding module, and perform reverse transmission optimization on the results through the loss function to obtain the optimized model.

[0028] The feature pyramid network module of the present invention is composed of dilated convolutions, and specifically the convolution of the feature pyramid network module adopts dilated convolutions.

[0029] The present invention extracts effective features of SAR water images through a deep learning encoding network to improve the accuracy of water area detection; realizes low loss and effective restoration of image resolution through a deep learning decoding network, reduces false targets in water area extraction, and improves network segmentation performance.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The present invention constructs an efficient and fast feature coding network model. The model effectively solves the problem of effective extraction of SAR water areas under complex terrain environments through three-stage feature coding and pyramid fusion mechanism. While improving the segmentation accuracy, the method adopts dense connection and depth-separable convolution operations to maximize the algorithm operation efficiency.

[0032] (2) The present invention builds a deep encoding and decoding fusion network model with powerful perception and feature analysis and learning capabilities. By introducing deep separable convolution and pyramid fusion mechanisms, the feature extraction capability of the lightweight network is greatly improved. The constructed model has a simple structure, high segmentation accuracy, and low redundancy between levels, achieving an excellent balance between segmentation performance and efficiency.

[0033] (3) The present invention constructs a relatively efficient and fast method for water segmentation of SAR remote sensing images. Through the powerful feature extraction capability of the deep coding network, it can adapt to the accurate segmentation of water areas in complex and changeable processing scenarios in remote sensing applications. The size of the entire deep coding network model is 3MB, which can be effectively equipped and deployed on terminal processing platforms with strict resource constraints, and has good practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the SAR remote sensing image water segmentation method based on the deep encoding and decoding fusion network of the present invention;

[0035] Figure 2 This is the structure diagram of the deep encoding and decoding fusion network built;

[0036] Figure 3 This is an original Gaofen-3 input image;

[0037] Figure 4 Another original Gaofen-3 input image;

[0038] Figure 5 The result of processing the first scene image by the method of the present invention, the white color represents the water area;

[0039] Figure 6 The result of processing the second scene image by the method of the present invention, the white color represents the water area; DETAILED DESCRIPTION

[0040] The present invention is further described in detail below in conjunction with embodiments.

[0041] Those skilled in the art will appreciate that the following examples are only used to illustrate the present invention and should not be considered to limit the scope of the present invention. If no specific techniques or conditions are specified in the examples, the techniques or conditions described in the literature in the art or the product specifications are used. If the manufacturer of the materials or equipment used is not specified, they are all conventional products that can be purchased.

[0042] like Figure 1 , a SAR remote sensing image water segmentation method based on a deep encoding and decoding fusion network includes the following steps:

[0043] Step (1), constructing a data set: collecting N scenes of SAR image data, with a single scene size of no less than 10k × 10k pixels, and dividing the data into a training data set and a validation data set;

[0044] Step (2), training data annotation: read the SAR remote sensing image in the training data set, annotate the water area location information in the image, and obtain a label set;

[0045] Step (3), data enhancement: amplify the SAR remote sensing images in the training data set to obtain an augmented data set;

[0046] Step (4), constructing a deep coding network: the deep coding network includes a feature extraction network and a feature pyramid network module composed of dilated convolutions, the feature extraction network is used to extract and learn the water features in the image and generate a feature map; the feature pyramid network module composed of dilated convolutions is used to pool the extracted feature map to increase the context information summarization capability of the extracted feature map;

[0047] Step (5), constructing a deep decoding network: the deep decoding network includes a fusion module and a decoding network module, the fusion module is used to fuse feature maps of different resolutions to improve the model's detection capability for targets of different sizes, and the decoding network module is used to decode the water area location information predicted by the output network;

[0048] The feature extraction network and the feature pyramid network module composed of dilated convolution are connected to the fusion module to form a deep encoding and decoding fusion network model;

[0049] Step (6), constructing the encoding and decoding network loss function: using the cross entropy loss function to minimize the objective function for model training, and setting the loss function output node to 2 categories according to the actual task requirements of water segmentation;

[0050] Step (7), model training and verification: input the training data set and label set into the built deep codec fusion network model for training, obtain the network model architecture of the Mth generation of training, then load the trained model, and test it based on the verification data set, calculate the classification accuracy, and save the model parameters with the best accuracy;

[0051] Step (8), segmentation processing: load the optimal model parameters, input the SAR image to be segmented into the deep encoding and decoding fusion network model, obtain label information, and finally correspond the label information to the SAR image to be segmented to obtain the segmentation result of the water area image;

[0052] Step (9), save or output the segmentation results as required.

[0053] The preferred solution is that in step (1), the GF-3 satellite SAR image data in the spotlight mode is collected; the data ratio in the training data set and the verification data set is 8:2.

[0054] The preferred solution is that in step (2), Labelme annotation software is used to select and annotate the location information of the water area in the image point by point along the edge of the water area.

[0055] In a preferred embodiment, in step (3), the amplification methods include rotation, translation, scaling, grayscale transformation and Mosaic data enhancement methods.

[0056] Preferably, in step (4),

[0057] (4.1) Construct the first-stage deep coding network: This network consists of three residual network modules. The main branch of each residual network module uses three depth-separable convolutional layers, and the cross-layer connection branch uses one conventional convolutional layer to reduce the dimension of the input feature map and increase the channel dimension.

[0058] (4.2) Construct the second stage deep encoding network: The network consists of 4 identical separable convolutional modules, each of which consists of 3 depth-separable convolutional layers based on residual connections;

[0059] (4.3) Construct the third-stage deep encoding network: This network contains two depth-wise separable convolutional modules. The first convolutional module uses three depth-wise separable convolutional layers, and the cross-layer connection branch uses one regular convolutional layer. The second convolutional module uses a depth-wise separable convolutional layer, followed by a dilated convolutional layer.

[0060] In a preferred embodiment, the operation steps in step (5) are as follows: the feature map extracted by the second-stage deep coding network constructed by the feature extraction network is used as a low-dimensional feature input, and the feature map extracted by the third-stage deep coding network is convolved through a feature pyramid network module composed of a hole convolution to obtain a new feature map of the same size as the feature map extracted by the second-stage deep coding network, and the feature map is used as a high-dimensional feature input, and the two-dimensional features are channel-accumulated through a fusion module to construct a deep decoding network.

[0061] In the preferred embodiment, in step (6), the output node of the loss function is set to 2 categories, and a binary cross entropy loss function is constructed.

[0062] The preferred solution, the specific method of step (7) is to input the SAR remote sensing training data with label information into the deep encoding and decoding fusion network, obtain the image feature extraction result through the deep encoding network, perform pixel-by-pixel binary classification on the extracted features through the fusion module and the decoding module, and perform reverse transmission optimization on the results through the loss function to obtain the optimized model.

[0063] Application Examples

[0064] Reference Figure 1 and Figure 2 , the specific steps for implementing the present invention are further described.

[0065] The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network includes the following steps:

[0066] Step 1, build a data set: collect 100 scenes of Gaofen-3 satellite SAR image data in the spotlight mode, with a single scene size of no less than 10k×10k pixels and a resolution better than 10 meters;

[0067] Step 2: Training data annotation: read the SAR remote sensing images in the training data set, annotate the water area location information in the image, and obtain the label set;

[0068] Step 3, data enhancement processing: Perform sample enhancement processing on the small batch of original input images. The main processing methods involved are: conventional enhancement strategies such as rotation, translation, scaling, grayscale transformation, and Mosaic enhancement strategy. After enhancement processing, the sample space is expanded from hundreds of scenes to tens of thousands of scenes, which can meet the needs of network training. Among them, the Mosaic data enhancement method randomly scales and splices 4 input images to form 1 enhanced image;

[0069] Step 4, construct a deep coding network: stack multiple groups of convolution, pooling, and activation layers to construct a deep coding network structure. The network consists of a feature extraction network and a feature pyramid network module composed of dilated convolutions. The feature extraction network is responsible for extracting and learning the water features in the image. The feature pyramid network module composed of dilated convolutions is used to increase the context information summarization ability of the extracted high-dimensional feature graph. The specific construction of each network is as follows;

[0070] Step 1: Construct the first-stage deep coding network: This network consists of three residual network modules. The main branch of each residual network module adopts three depth separation convolution layers. The structure of the convolution layer is a 4-dimensional vector. The first two dimensions represent the size of the convolution kernel, the third dimension represents the number of convolution kernels, and the last dimension represents the step size of the convolution kernel. The cross-layer connection branch uses a conventional convolution layer to reduce the dimension of the input feature map and increase the dimension of the channel to ensure that the dimension of the feature map after convolution is consistent with the dimension of the output feature map of the main branch.

[0071] Step 2: construct a second-stage deep coding network: the network consists of four identical separation convolution modules, each of which consists of three depth-separable convolution layers based on residual connections. In order to improve the feature mapping relationship between the separation convolution modules, the present invention proposes a separation convolution network based on dense connections, which connects the output feature maps of all the previous network modules in parallel through channels, and uses the paralleled feature maps as the input of the lower network.

[0072] Step 3: Construct the third-stage deep encoding network: This network contains two depth-wise separable convolutional modules. The first convolutional module is similar to the residual module structure described in (3a), using three depth-wise separable convolutional layers, and the cross-layer connection branch uses one conventional convolutional layer. The second convolutional module combines the depth-wise separable convolutional layer and the dilated convolutional layer, and the dilation rates of the dilated convolution are (1, 2, 4) respectively. In addition, the layers of the second separable convolutional module do not use residual connections, but directly use a cascade method.

[0073] Step 5, constructing a deep decoding network: the deep decoding network includes a fusion module and a decoding network module, the fusion module is used to fuse feature maps of different resolutions, and the decoding network module is used to decode the water area location information predicted by the output network;

[0074] The feature extraction network and the feature pyramid network module composed of dilated convolution are connected to the fusion module to form a deep encoding and decoding fusion network model;

[0075] The operation steps in step (5) are as follows: the feature map extracted by the second-stage deep coding network constructed by the feature extraction network is used as a low-dimensional feature input, and the feature map extracted by the third-stage deep coding network is convolved through a feature pyramid network module composed of dilated convolutions to obtain a new feature map of the same size as the feature map extracted by the second-stage deep coding network, and the feature map is used as a high-dimensional feature input, and the two-dimensional features are channel-accumulated through a fusion module to construct a deep decoding network.

[0076] Step 1: Use the constructed second-stage deep encoding network to extract low-dimensional feature maps, and perform convolution on the extracted feature maps to reduce the number of channels of the feature maps;

[0077] Step 2: extract feature maps through the third-stage deep coding network, and use a feature pyramid network module composed of dilated convolutions to convolve the extracted feature maps to obtain multiple groups of feature maps of the same size as the feature maps extracted by the second-stage deep coding network. In this embodiment, three groups are set;

[0078] Step 3: The feature map extracted by the feature extraction network and the feature map output by the feature pyramid network module are connected in parallel by channel dimension superposition, and convolution is used to extract features and reduce the channel dimension of the combined feature map to obtain the final feature map;

[0079] Step 4: Finally, the feature map obtained in step 3 is upsampled by 4 times using bilinear interpolation to obtain a feature map with the same size as the input image, and each pixel in the feature map is judged by the softmax function to confirm whether the pixel is a water pixel (if the softmax value is less than 0.5, it is a non-water area and marked as 0; if the value is greater than 0.5, it is a water pixel and marked as 1), and the final segmentation result can be obtained;

[0080] Step 6: Construct the encoding and decoding network loss function: Use the cross entropy loss function as the optimization objective function, and set the loss function output node to 2 categories according to the actual task requirements of water segmentation. The objective function is shown as follows: (no formula)

[0081]

[0082] Among them, M is the number of categories, N is the number of samples, L is the cross entropy loss objective function, y ic is the sign function (0 or 1), p ic is the predicted probability that the observed sample i belongs to category c.

[0083] Step 7, model training and verification: Input the training data set and label set into the built deep codec fusion network model, train the segmentation network, obtain the network model architecture of the Mth generation of training, and save the optimal model parameters for training;

[0084] Step 8, test the deep codec fusion network model: load the optimal model parameters, set the network parameters, use the deep codec fusion network to process the input image, obtain the label information, and finally correspond the label information to the SAR image to be segmented to obtain the segmentation result of the water area image;

[0085] Step 9: Save or output the segmentation results as required.

[0086] The effect of the present invention is further described below in conjunction with simulation experiments:

[0087] 1. Simulation experiment conditions:

[0088] The hardware platform of the simulation experiment of the present invention is: the processor is Intel i7 4710HQ CPU, the main frequency is 2.5GHz, the memory is 8GB, the graphics card is NVIDIA GTX 2080GPU, the memory is 12GB.

[0089] The software platform of the simulation experiment of the present invention is: Ubuntu 16.04.

[0090] 2. Simulation content and results analysis:

[0091] All data in this simulation experiment are actual remote sensing data, mainly based on the beamforming mode data of the Gaofen-3 satellite, with a resolution of 1 meter. The method of the present invention takes an average of 5 seconds to process a 10000×10000 pixel image, which has high timeliness. Figure 3 and Figure 4 This is a local area captured from two views of the Gaofen-3 satellite data, and the image size is 4096×4096 pixels.

[0092] from Figure 5 and Figure 6 It can be seen that the SAR water area extraction method of the present invention has good extraction effect for a variety of complex scenes, high extraction accuracy, high processing efficiency, and good practical value. Through experimental comparison, the segmentation accuracy of the technology of the present invention is improved by 30% and 25% respectively compared with the current traditional canny-based edge segmentation method and region growing method segmentation method.

[0093] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A SAR remote sensing image water segmentation method based on a deep encoding and decoding fusion network is characterized by: The steps include: Step (1), constructing a data set: collecting N scenes of SAR image data, with the size of each scene not less than 10k × 10k pixels, and dividing the data into a training data set and a validation data set; Step (2), training data annotation: read the SAR remote sensing image in the training data set, annotate the water area location information in the image, and obtain the label set; Step (3), data enhancement: amplify the SAR remote sensing images in the training data set to obtain an augmented data set; Step (4), constructing a deep coding network: the deep coding network includes a feature extraction network and a feature pyramid network module composed of dilated convolutions, the feature extraction network is used to extract and learn the water features in the image and generate a feature map; the feature pyramid network module composed of dilated convolutions is used to pool the extracted feature map; Step (5), constructing a deep decoding network: the deep decoding network includes a fusion module and a decoding network module, the fusion module is used to fuse feature maps of different resolutions, and the decoding network module is used to decode the water area location information predicted by the output network; The feature extraction network and the feature pyramid network module composed of dilated convolution are connected to the fusion module to form a deep encoding and decoding fusion network model; Step (6), constructing the encoding and decoding network loss function: using the cross entropy loss function to minimize as the objective function for model training; Step (7), model training and verification: input the training data set and label set into the built deep codec fusion network model for training, obtain the network model architecture of the Mth generation of training, then load the trained model and test it based on the verification data set, calculate the classification accuracy, and save the model parameters with the best accuracy; Step (8), segmentation processing: load the optimal model parameters, input the SAR image to be segmented into the deep encoding and decoding fusion network model, obtain label information, and finally correspond the label information to the SAR image to be segmented to obtain the segmentation result of the water area image; Step (9), saving or outputting the segmentation results according to requirements; In step (4), (4.1) Construct the first-stage deep encoding network: This network consists of three residual network modules. The main branch of each residual network module uses three depth-wise separable convolutional layers, and the cross-layer connection branch uses one conventional convolutional layer to reduce the dimension of the input feature map and increase the channel dimension. (4.2) Construct the second stage deep encoding network: The network consists of 4 identical separable convolutional modules, each of which consists of 3 depth-separable convolutional layers based on residual connections; (4.3) Construct the third-stage deep encoding network: This network contains two depth-separable convolutional modules. The first convolutional module uses three depth-separable convolutional layers, and the cross-layer connection branch uses one conventional convolutional layer; the second convolutional module uses a depth-separable convolutional layer, followed by a dilated convolutional layer in series; The specific method of step (7) is to input the SAR remote sensing training data with label information into the deep encoding and decoding fusion network, obtain the image feature extraction result through the deep encoding network, perform pixel-by-pixel binary classification on the extracted features through the fusion module and the decoding module, and perform reverse transmission optimization on the results through the loss function to obtain the optimal model.

2. The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network according to claim 1 is characterized by: In step (1), the SAR image data of the Gaofen-3 satellite in the spotlight mode is collected; the data ratio between the training data set and the validation data set is 8:

2.

3. The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network according to claim 1 is characterized in that: In step (2), use Labelme annotation software to select and annotate the location information of the water area in the image point by point along the edge of the water area.

4. The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network according to claim 1 is characterized in that: In step (3), the amplification methods include rotation, translation, scaling, grayscale transformation and Mosaic data enhancement method.

5. The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network according to claim 1 is characterized in that: The operation steps in step (5) are as follows: the feature map extracted by the second-stage deep coding network constructed by the feature extraction network is used as a low-dimensional feature input, and the feature map extracted by the third-stage deep coding network is convolved through a feature pyramid network module composed of dilated convolutions to obtain a new feature map of the same size as the feature map extracted by the second-stage deep coding network, and the feature map is used as a high-dimensional feature input, and the two-dimensional features are channel-accumulated through a fusion module to construct a deep decoding network.

6. The SAR remote sensing image water segmentation method based on deep encoding and decoding fusion network according to claim 1 is characterized by: In step (6), the loss function output node is set to 2 categories, and a binary cross entropy loss function is constructed.

Images