CN-122024086-A - Point supervision SAR water body extraction method and system for essential feature learning

CN122024086ACN 122024086 ACN122024086 ACN 122024086ACN-122024086-A

Abstract

The invention discloses a point supervision SAR water body extraction method and system for essential feature learning, which belong to the technical field of water body extraction of remote sensing images, and the method comprises the steps of collecting and preprocessing SAR image data, cutting the SAR image data into image slices, inputting the image slices into a water body extraction model to obtain a water body extraction result of the SAR image, and completing the point supervision SAR water body extraction; the method comprises the steps of constructing and training a water body extraction model which comprises a multi-scale feature extraction and reconstruction framework, an essential feature learning module and a loss function, wherein the essential feature learning module comprises a robust feature branch, a sensitive feature branch, a Gaussian disturbance unit and an alignment loss calculation unit. According to the method, the network fully excavates key leachable information in the SAR image under the sparse point label condition, so that the labeling cost is reduced, the effective distinction between the complex scattering characteristics in the SAR image and the noise background is realized, the accuracy and the stability of water extraction are improved, and the method has better applicability.

Inventors

REN ZHONGLE
Zang Weicheng
WANG KAI
LI WEIBIN

Assignees

陕西元翌智能科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (10)

1. The point supervision SAR water body extraction method for the essential feature learning is characterized by comprising the following steps of: s1, collecting and preprocessing SAR image data, and cutting the preprocessed SAR image into image slices; S2, inputting the image slice into a water body extraction model to obtain a water body extraction result of the SAR image; The system comprises a water body extraction model, an SAR image extraction module, a SAR image extraction module and an SAR image extraction module, wherein the water body extraction model is formed by construction and training and comprises a multi-scale feature extraction and reconstruction frame, an essential feature learning module and a loss function, wherein the essential feature learning module comprises a robust feature branch, a sensitive feature branch, a Gaussian disturbance unit and an alignment loss calculation unit; Dividing the image slice into a training set, a verification set and a test set, processing the water mask label to generate a point label, acquiring a threshold value by adopting an automatic threshold value determining method based on the back scattering intensity of the SAR image, and obtaining a false division risk region mask by utilizing threshold value division; The multi-scale feature extraction and reconstruction framework is used for extracting and fusing multi-scale semantic features of the training set image slices, the essential feature learning module is used for purifying the multi-scale semantic features to obtain essential features, the loss function comprises point supervision loss, shadow suppression loss and alignment loss, the point supervision loss is constructed based on point labels, and the shadow suppression loss is constructed based on false-division risk area masks.
2. The method for extracting the point-supervised SAR water body by the essential feature learning method according to claim 1, wherein in the step S1, SAR image data is a horizontal transmission-horizontal reception polarized image in a high-resolution three-satellite hyperfine stripe mode, preprocessing comprises multi-view processing, orthographic correction, radiation correction and bit depth conversion, and clipping adopts a sliding window strategy to clip the preprocessed SAR image into an image slice with a preset size.
3. The point supervision SAR water body extraction method based on the essential feature learning according to claim 1 is characterized in that in the step S2, the specific process of generating the point label comprises the steps of executing connected domain analysis on each type of binary masks in the water body mask label, identifying all connected regions and reserving the connected regions with the area exceeding a preset threshold, calculating Euclidean distances from all pixels in the regions to the boundary of the outer outline of the regions for each reserved connected region, generating a distance map, selecting the pixel point with the largest distance value in the distance map as the point label of the connected region, and setting the rest pixels to be in an neglected state.
4. The method for extracting the point-supervised SAR water body by the essential feature learning according to claim 1, wherein in the step S2, an optimal threshold value is automatically determined by adopting an Ojin method, the image is divided into a foreground part and a background part by traversing all possible gray threshold values in the image, the maximum inter-class variance is taken as a target, a threshold value which enables the gray average difference between the two classes to be maximum is selected from all possible gray threshold values, and the optimal threshold value is obtained, wherein the calculation formula of the inter-class variance is as follows: ; Wherein, the Is a threshold value The inter-class variance of the lower one, 、 Respectively is threshold value The proportion of the two types of pixels divided, 、 Respectively is threshold value Mathematical expectations of the gray scale of the next two types of pixels.
5. The method for extracting the point-supervised SAR water body by essential feature learning according to claim 1, wherein in S2, the calculation formula for obtaining the misclassification risk area mask by using the threshold division is as follows: ; Wherein, the For pixels in SAR images Is used for the gray-scale value of (c), Pixels in mask for misclassification risk area Is used for the value of (a) and (b), Is the optimal threshold.
6. The point supervision SAR water body extraction method for the essential feature learning according to claim 1 is characterized in that in S2, a multi-scale feature extraction and reconstruction framework comprises an encoder and a feature fusion and reconstruction unit, the encoder adopts ResNe deep convolutional neural network, a plurality of layers of multi-scale semantic features are extracted through layer-by-layer convolution, normalization and downsampling operation, the feature fusion and reconstruction unit starts from the deepest layer features, the multi-scale feature integration from bottom to top is achieved through step up sampling, channel splicing and convolution layer laminating integration, and after the fused semantic features are purified through an essential feature learning module, a water body probability map is obtained through output convolution layer and Sigmoid activation function processing.
7. The point supervision SAR water body extraction method for the essential feature learning according to claim 1 is characterized in that in S2, an essential feature learning module comprises a robust feature branch, a sensitive feature branch, a Gaussian disturbance unit and an alignment loss calculation unit, the specific working process comprises the steps that the robust feature branch adopts a shallow ConvBNReLU structure formed by cavity convolution to extract structural features, the sensitive feature branch stacks three residual modules to extract local gradient change and high-frequency detail features, the Gaussian disturbance unit respectively adds Gaussian noise interference to the robust feature and the sensitive feature, the alignment loss calculation unit normalizes the feature added with noise, calculates cos similarity of the normalized feature, and finally obtains alignment loss according to the similarity, and the calculation formula of the alignment loss is as follows: ; Wherein, the 、 The normalized robust features and sensitive features respectively, Is that And Cos similarity of (c).
8. The method for extracting a point-supervised SAR water body for learning essential features according to claim 1, wherein in S2, the point-supervised loss is a cross entropy loss, and the calculation formula is as follows: ; Wherein, the Is the first The true label of the individual pixels is that, Is the first The probability that a pixel predicts as a body of water, Is the effective pixel number.
9. The method for extracting a point-supervised SAR water body for learning essential features according to claim 1, wherein in S2, the shadow suppression loss is obtained by punishing an erroneous high water body response in a low gray shadow region, and the calculation formula is as follows: ; Wherein, the Pixels predicted for model Is the probability of a body of water, At the level of the minimum value of the total number of the components, Pixels in mask for misclassification risk area Is used for the value of (a) and (b), For the sample level indication quantity, Is a numerical stable term.
10. The point supervision SAR water body extraction system for the essential feature learning is used for realizing the point supervision SAR water body extraction method for the essential feature learning according to any one of claims 1 to 9, and is characterized by comprising a data acquisition and processing module and an SAR image water body extraction module, wherein: The data acquisition and processing module is used for collecting and preprocessing SAR image data and cutting the preprocessed SAR image into image slices; The SAR image water body extraction module is used for inputting the image slice into the water body extraction model to obtain a water body extraction result of the SAR image; The system comprises a water body extraction model, an SAR image extraction module, a SAR image extraction module and an SAR image extraction module, wherein the water body extraction model is formed by construction and training and comprises a multi-scale feature extraction and reconstruction frame, an essential feature learning module and a loss function, wherein the essential feature learning module comprises a robust feature branch, a sensitive feature branch, a Gaussian disturbance unit and an alignment loss calculation unit; Dividing the image slice into a training set, a verification set and a test set, processing the water mask label to generate a point label, acquiring a threshold value by adopting an automatic threshold value determining method based on the back scattering intensity of the SAR image, and obtaining a false division risk region mask by utilizing threshold value division; The multi-scale feature extraction and reconstruction framework is used for extracting and fusing multi-scale semantic features of the training set image slices, the essential feature learning module is used for purifying the multi-scale semantic features to obtain essential features, the loss function comprises point supervision loss, shadow suppression loss and alignment loss, the point supervision loss is constructed based on point labels, and the shadow suppression loss is constructed based on false-division risk area masks.

Description

Point supervision SAR water body extraction method and system for essential feature learning Technical Field The invention belongs to the technical field of water extraction in remote sensing images, and particularly relates to a point supervision SAR water extraction method and system for essential feature learning. Background The water body extraction in the remote sensing image has important invention significance in civil fields and the like, and has wide application prospect in aspects of disaster prevention and reduction, water resource regulation and control, ecological environment monitoring and the like. With the continuous development of remote sensing imaging technology, the observation means is continuously abundant. The synthetic aperture radar (SYNTHETIC APERTURE RADAR, SAR) is one of key technologies in the remote sensing field, and by virtue of the advantages of active microwave imaging, the limitation of the traditional optical remote sensing due to factors such as weather, illumination, cloud cover and the like can be effectively overcome, and all-weather and all-day earth observation can be realized. Therefore, SAR data has become an important source of telemetry information that is currently not an alternative. The water extraction based on SAR images gradually becomes an important direction in the remote sensing invention. The task aims at accurately classifying each pixel in the image, rapidly identifying the water body area, and providing key data support for flood disaster monitoring, water resource management, ecological wetland protection, environment assessment and the like. The method is particularly suitable for dynamic water monitoring scenes because the earth surface information can be stably obtained even under complex weather and cloud cover conditions due to the all-day and all-weather imaging capability of the SAR. In flood disasters, the rapid and accurate extraction of water distribution has important significance for emergency rescue, and can be used for judging a flooding range, identifying risk areas and supporting emergency dispatch. After disaster, the water body change information can also provide reference for infrastructure damage evaluation, agricultural disaster statistics and recovery strategy formulation. In addition, the water body extraction has important application value in the aspects of daily water environment monitoring, river and lake management, water resource scheduling and the like. With the rapid development of deep learning technology, a segmentation method based on a neural network has shown excellent performance in SAR water extraction tasks. However, the existing method relies on a large number of pixel-level full-supervision labels for training. The pixel-level labeling of SAR images requires professional personnel to interpret pixel by pixel, has high cost and huge time consumption, and is difficult to meet the actual requirements of rapid processing of large-scale and multi-temporal remote sensing data. Thus, weakly supervised learning is increasingly a viable alternative to alleviating fully supervised labeling pressure. The weak supervision labels are various in forms, including image-level labels, frame labels, point labels and the like, and can provide enough supervision signals for the model while remarkably reducing the labeling cost. In various weak supervision forms, the point labels are particularly suitable for scenes with dense ground features and complex boundaries in SAR images and water body extraction tasks due to the advantages of convenience in labeling, low cost, clear space information and the like. Compared with the image-level label, the point label not only can provide the approximate space position of the target, but also can provide stronger space priori and positioning capability, and compared with the frame label, the point label has smaller labeling workload and lower ambiguity and is more in line with SAR data characteristics. Therefore, the weak supervision learning method based on the point labels has remarkable advantages and wide development space in the SAR image water body extraction task. A point supervision SAR water body extraction method based on essential feature learning and shadow suppression loss. In the present, china patent CN119516395A discloses a deep learning-based lake water body extraction method, a deep learning-based lake water body extraction system, a deep learning-based lake water body extraction storage medium and an electronic device, wherein the method utilizes the polarization characteristic difference of lake water bodies and other ground objects in polarization decomposition to construct a multi-level deep learning water body extraction model, and on the basis, the accurate extraction of water boundary information is further realized through the characteristic learning and optimization process. In Chinese patent, CN118865141A discloses a deep learning-based SAR water body extractio