CN-121685571-B - SAR image flood area extraction method and system based on multi-source feature fusion

CN121685571BCN 121685571 BCN121685571 BCN 121685571BCN-121685571-B

Abstract

The invention provides a SAR image flood area extraction method and system based on multi-source feature fusion, which belong to the technical field of remote sensing disaster monitoring, wherein the extraction method is applied to a system comprising an image processing module, a model training module and a flood area extraction module, and specifically comprises the steps of preprocessing SAR images, extracting corresponding polarized physical features and time sequence abnormal features, constructing a flood learning sample by combining preset tag data, and training and learning a U-shaped network fused with a residual error network to obtain a pre-training model; and outputting a flood area mask map corresponding to the target SAR image based on the flood area recognition model. According to the invention, the migration model is introduced on the basis of fusing the multidimensional features, the U-shaped network architecture is combined and improved, the identification and extraction precision of the flood area is effectively improved, the risk of false detection due to missing detection is avoided, and the suitability is higher.

Inventors

YU QIAN
SHEN WENYUAN
CHEN XIANG
HONG DAOJIAN
ZHANG QIAN
YU DACHENG
NING JUNBAO
LI HUI
WANG XIAOFEI
FU XIAOTIAN
YANG QIAN
XU CHUHANG
ZHANG XUEPENG
RAO HANYU
JIN YIJIONG

Assignees

国网浙江省电力有限公司台州供电公司
国网浙江省电力有限公司信息通信分公司

Dates

Publication Date: 20260508
Application Date: 20260211

Claims (9)

1. The SAR image flood area extraction method based on multi-source feature fusion is characterized by comprising the following steps of: Preprocessing the SAR image, extracting polarized physical characteristics and time sequence abnormal characteristics based on the preprocessed SAR image, and constructing a flood learning sample by combining preset tag data; Training and learning the U-shaped network fused with the residual error network through a flood learning sample to obtain a pre-training model; Based on transfer learning, adjusting a pre-training model by combining a target scene corresponding to a target SAR image to obtain a flood area identification model; outputting a flood area mask map corresponding to the target SAR image based on the flood area recognition model; The SAR image extraction polarization physical characteristics and time sequence abnormal characteristics based on preprocessing comprise the following steps: Acquiring single-view complex data of the preprocessed SAR image, constructing a dual-polarized scattering matrix based on the single-view complex data, and converting the scattering matrix into a target vector; Calculating a corresponding coherence matrix based on the target vector, and carrying out eigenvalue decomposition on the coherence matrix to obtain polarization physical characteristics; Acquiring ground distance detection data in the preprocessed SAR image in a historical time phase, taking the ground distance detection data as a historical baseline data set, and calculating a mean value and a standard deviation of a backscattering coefficient of each pixel in the ground distance detection data; Based on the preprocessed SAR image of the current time phase, calculating the time sequence anomaly score of each pixel in the preprocessed SAR image of the current time phase by combining the calculated scattering coefficient mean value and standard deviation, and identifying time sequence anomaly characteristics by combining a time sequence anomaly score threshold value, When the timing anomaly score is below a timing anomaly score threshold, a timing anomaly characteristic is identified.
2. The method for extracting the flooding area of the SAR image based on the multi-source feature fusion according to claim 1, wherein said preprocessing the SAR image comprises: Performing radiation correction on the SAR image by combining the absolute scaling factor and the offset; Geometrically correcting the SAR image after radiation correction through the track parameters and the ground control points of the SAR image; and filtering the geometrically corrected SAR image to obtain a preprocessed SAR image.
3. The SAR image flood area extraction method based on multi-source feature fusion of claim 1, wherein the polarized physical features comprise at least polarized entropy, scattering angle and anisotropy.
4. The method for extracting the flooding domain of the SAR image based on the multi-source feature fusion according to claim 1, wherein said constructing the flooding study sample by combining the preset tag data comprises: the method comprises the steps of taking a pre-marked flood disaster distribution map as preset tag data, matching polarized physical characteristics and time sequence abnormal characteristics with corresponding pixels of the preset tag data according to pixel positions, and generating corresponding characteristic tag pair data according to a matching result; and cutting the characteristic label data into flood learning samples with preset pixel sizes.
5. The method for extracting the SAR image flooding domain based on the multi-source feature fusion according to claim 1, wherein the training and learning the U-shaped network fused with the residual network by using the flooding learning sample to obtain the pre-training model comprises: Preprocessing a flood learning sample, and splitting the flood learning sample into a training set and a verification set according to a preset proportion; initializing a U-shaped network fusing a residual error network and setting training parameters; Based on training parameters, randomly extracting batches of samples from a training set in the process of each training iteration, inputting the samples into a U-shaped network of a fusion residual error network for forward propagation, calculating an average loss value of each batch of samples, and updating network parameters through a back propagation algorithm; and calculating network performance indexes through the verification set, updating network parameters based on the network performance indexes until the iteration ending condition in the training parameters is met, and obtaining the pre-training model.
6. The method for extracting the flood area of the SAR image based on the multi-source feature fusion according to claim 5, wherein the steps of obtaining the flood area identification model based on the migration learning and combining the target scene corresponding to the target SAR image to adjust the pre-training model comprise the following steps: Acquiring a historical SAR image of a target scene corresponding to the target SAR image, preprocessing the historical SAR image, and constructing a fine adjustment training set; And adjusting training parameters, freezing encoder low-layer weights in the pre-training model, and training and learning the pre-training model through a fine-tuning training set to obtain a flood area identification model.
7. The method for extracting the flood area of the SAR image based on the multi-source feature fusion according to claim 1, wherein the outputting the flood area mask map corresponding to the target SAR image based on the flood area identification model comprises: preprocessing a target SAR image, and extracting corresponding polarization physical characteristics and time sequence abnormal characteristics; Inputting the extracted polarization physical characteristics and time sequence abnormal characteristics into a flood area identification model, and outputting a corresponding flood attribution probability map; and carrying out binarization processing on the flood attribution probability map to obtain a corresponding flood area mask map.
8. A multi-source feature fusion-based SAR image flood area extraction system for executing the multi-source feature fusion-based SAR image flood area extraction method of any one of claims 1 to 7, comprising: The image processing module is used for preprocessing the SAR image, extracting polarization physical characteristics and time sequence abnormal characteristics based on the preprocessed SAR image, and constructing a flood learning sample by combining preset tag data; The model training module is used for training and learning the U-shaped network fused with the residual error network through the flood learning sample to obtain a pre-training model, and based on transfer learning, the pre-training model is adjusted by combining a target scene corresponding to the target SAR image to obtain a flood area identification model; and the flood area extraction module is used for outputting a flood area mask map corresponding to the target SAR image based on the flood area identification model.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the SAR image flood area extraction method based on multi-source feature fusion of any one of claims 1 to 7 when executing the program.

Description

SAR image flood area extraction method and system based on multi-source feature fusion Technical Field The invention relates to the technical field of remote sensing disaster monitoring, in particular to a SAR image flood area extraction method and system based on multisource feature fusion. Background The existing flood area extraction mode mostly uses optical images as areas to extract data sources, but the optical images are easily shielded by common weather conditions such as overcast and rainy and cloud layers in the flood area due to the fact that the optical images are imaged by means of visible light and near infrared bands, clear ground object information is difficult to obtain through interference of the cloud and rain, and flood disaster monitoring accuracy is low easily. Synthetic Aperture Radar (SAR) images can effectively penetrate through a cloud and rain barrier, and the surface coverage characteristics can be reflected through the differences of the backscattering coefficients of different ground features, so that the SAR images become a main data source for flood area extraction gradually. The current flood area extraction method based on SAR images mainly comprises a traditional threshold segmentation method, a single feature machine learning method and a U-shaped network segmentation method. The traditional threshold segmentation method is highly dependent on the bimodal distribution characteristic of the SAR image backward scattering coefficient gray level histogram, and has high misjudgment rate and low extraction precision in a complex scene. The single-feature machine learning rule only depends on the single feature of the backscattering coefficient, so that the physical difference between the temporary flood area and other areas with larger humidity is easily confused, and the flood area is easily missed. Although the U-shaped network segmentation method can improve the feature extraction capability through deep learning, overcomes the problems existing in the traditional threshold segmentation method and the single feature machine learning method, improves the extraction precision, and avoids the problem of missed detection, most of the U-shaped network segmentation method has single input features, has insufficient distinguishing capability for complex scattering mechanisms, easily has the problem of gradient disappearance when the depth of a network is increased, has unstable model training process and lower segmentation precision, cannot adapt to the feature differences of different scenes, and can obviously reduce the extraction precision when a model trained in a certain area is directly transferred to other areas for application. Disclosure of Invention The invention aims to overcome the defects of low extraction precision, weak anti-interference capability and poor flooding capability of a flooding area when the U-shaped network is utilized to carry out SAR influence on the extraction of the flooding area, provides a SAR image flooding area extraction method and a system based on multi-source feature fusion, and distinguishes a flooding area from confusing ground objects by fusing polarized physical features and time sequence abnormal features, the problem of false detection due to missing detection is avoided, the problem of gradient disappearance of a deep network is solved by improving a U-shaped network architecture, the boundary segmentation precision of a flood area is improved, the adaptation capability of a model under different distribution characteristic scenes is improved by migration learning, the cross-area monitoring precision is ensured, and the high-precision flood area extraction is realized. The invention aims at realizing the following technical scheme: The SAR image flood area extraction method based on multi-source feature fusion comprises the following steps: Preprocessing the SAR image, extracting polarized physical characteristics and time sequence abnormal characteristics based on the preprocessed SAR image, and constructing a flood learning sample by combining preset tag data; Training and learning the U-shaped network fused with the residual error network through a flood learning sample to obtain a pre-training model; Based on transfer learning, adjusting a pre-training model by combining a target scene corresponding to a target SAR image to obtain a flood area identification model; And outputting a flood area mask map corresponding to the target SAR image based on the flood area recognition model. Further, the preprocessing the SAR image includes: Performing radiation correction on the SAR image by combining the absolute scaling factor and the offset; Geometrically correcting the SAR image after radiation correction through the track parameters and the ground control points of the SAR image; and filtering the geometrically corrected SAR image to obtain a preprocessed SAR image. Further, the extracting the polarization physical feature and the time