Dynamic flood situation monitoring algorithm based on SAR (Synthetic Aperture Radar) time sequence feature optimization

Abstract

The invention relates to the technical field of flood situation monitoring, in particular to a dynamic flood situation monitoring algorithm based on SAR (Synthetic Aperture Radar) time sequence feature optimization, which comprises the following steps of: acquiring flood season Sentinel-1SAR sample data of a plurality of target areas; processing the sample data, and constructing an initial feature set; tags are given to part of sample data, a training set is formed, and the tag types comprise a normal water body, a submerged water body and a non-water body; training an identification model by using the training set; determining an optimal feature number; and based on the optimal feature subset, taking the optimal feature subset as input, and using the trained recognition model to output a spatial-temporal change result of the target region. According to the method, feature joint optimization selection is carried out from the polarization dimension and the time dimension, high-precision flood season water body monitoring is realized by using few and effective optimization features, the accuracy and reliability of flood situation dynamic monitoring are improved, and automatic, rapid and effective monitoring of large-scale water body range spatio-temporal information is realized.

Description

Technical Field

[0001] The present application relates to the technical field of flood monitoring, and in particular to a dynamic flood monitoring algorithm based on SAR time series feature optimization. Background Art

[0002] SAR can be used to conduct all-weather water monitoring in the target area and issue early warnings for possible flood hazards. In existing technologies, the monitoring accuracy is not high due to the influence of SAR image speckle noise and image water accuracy.

[0003] Therefore, there is a need for a monitoring algorithm that can accurately identify the flood situation in the target area. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present application provides a flood dynamic monitoring algorithm based on SAR time series feature optimization. First, multi-phase processing and interference processing are performed on the flood season Sentinel-1SAR data to obtain the backscattering coefficient and interference coherence coefficient corresponding to the flood season time series, and then the initial feature set is constructed in combination with the SDWI polarization index. Then, sample selection is performed, and the sample is tested in combination with high-definition images and related flood data of the same period. Feature optimization is then performed from the polarization dimension and the time dimension, and the optimal number of features is determined by the distance between classes within and between classes. Finally, based on the optimal feature subset, the random forest classification method is used to obtain the spatiotemporal change results of the water body, and the accuracy is evaluated. By jointly optimizing the selection of features from the polarization dimension and the time dimension, high-precision flood season water body monitoring is achieved using a small number of effective optimized features, which improves the accuracy and reliability of dynamic flood monitoring and realizes automatic, rapid and effective monitoring of spatiotemporal information of water bodies at a large scale.

[0005] The above application objectives of this application are achieved through the following technical solutions:

[0006] A flood dynamic monitoring algorithm based on SAR time series feature optimization includes the following steps:

[0007] Acquire Sentinel-1SAR sample data during flood season in several target areas;

[0008] Process the Sentinel-1SAR sample data of the target area during the flood season to construct an initial feature set;

[0009] Label the Sentinel-1SAR sample data of some target areas during the flood season to form a training set. The label types include normal water bodies, submerged water bodies, and non-water bodies.

[0010] Use the training set to train the recognition model;

[0011] Determine the optimal number of features;

[0012] Based on the optimal feature subset, the optimal feature subset is used as input and the trained recognition model is used to output the spatiotemporal change results of the target area.

[0013] Optionally, the recognition model is a random forest classification model.

[0014] Optionally, the SDWI polarization index is calculated as follows:

[0015] K SDWI =ln(10*VV*VH)

[0016] Wherein: VV and VH represent VV and VH polarization images respectively.

[0017] Optionally, the method further includes the following steps:

[0018] Label the Sentinel-1SAR sample data in some target areas during the flood season to form a validation set;

[0019] The validation set is used to verify the accuracy of the trained recognition model.

[0020] Optionally, the Sentinel-1SAR sample data of the target area during the flood season is processed to construct an initial feature set including:

[0021] Multi-temporal processing and interferometric processing are performed on the Sentinel-1SAR sample data of the target area during the flood season to obtain the backscattering coefficient and interferometric coherence coefficient corresponding to the flood season time series;

[0022] Calculate the SDWI polarization index of Sentinel-1SAR sample data in the target area during the flood season;

[0023] The initial feature set is constructed based on the SDWI polarization index, backscattering coefficient and interference coherence coefficient of Sentinel-1SAR sample data in the target area during the flood season.

[0024] Optionally, determining the optimal number of features comprises the following steps:

[0025] Optimize features from the polarization dimension and time dimension;

[0026] The intra-class and inter-class distances are used to determine the number of optimal feature subsets.

[0027] Optionally, in the feature optimization from the polarization dimension and the time dimension, feature selection from the polarization dimension is as follows:

[0028]

[0029] Among them, NEHS is the standard entropy histogram similarity obtained by normalizing the entropy histogram similarity EHS, S is the structural similarity index SSIM, and β is a constant;

[0030] The EHS calculation is as follows:

[0031]

[0032] Where M and N are the length and width of the image, and T is the column vector of the image expansion dimension, whose shape can be expressed as 1×(M·N);

[0033] SSIM is calculated as follows:

[0034]

[0035] Among them, S(x,y) is the structural similarity measure of the statistical similarity between image x and image y, μx and μy are the statistical average values ​​of pixels in image x and image y respectively. are the statistical variances of the pixels of image x and image y respectively, and c1 and c2 are two constants.

[0036] Optionally, in the feature optimization from the polarization dimension and the time dimension, feature selection from the time dimension is as follows:

[0037]

[0038] Among them, ED is the Euclidean distance, SCS is the spectral correlation similarity, and ED and SCS are calculated as follows:

[0039]

[0040] Where x and y represent two different samples, respectively. μ and σ represent the mean and standard deviation of the samples.

[0041] In summary, this application has the following beneficial technical effects:

[0042] By jointly optimizing the selection of features from the polarization dimension and the time dimension respectively, high-precision flood season water body monitoring can be achieved using a few and effective optimization features, which improves the accuracy and reliability of dynamic flood monitoring and realizes automatic, rapid and effective monitoring of spatiotemporal information of water bodies at a large scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of an implementation method of the present application. DETAILED DESCRIPTION

[0044] The present application is further described in detail below in conjunction with the accompanying drawings.

[0045] In order to more clearly understand the technical solutions presented in the embodiments of the present application, a brief introduction is first given to monitoring flood disasters in the prior art.

[0046] Flood disasters are one of the most frequent, widespread, and most devastating natural disasters in the world. Large-scale, long-term, and dynamic flood monitoring can accurately collect and feedback spatiotemporal information on flood disasters, which is of great significance for flood warning, rapid response, and risk prevention and control.

[0047] Synthetic Aperture Radar (SAR) images are not affected by clouds, fog, rain and snow, and can monitor water changes during the flood season all day and all weather. They have the advantages of no safety hazards, large inspection coverage area, rapid response and economy. Commonly used flood monitoring methods mainly include difference feature analysis method and classification result comparison method. The former is affected by SAR image speckle noise and has limited accuracy. The latter is mainly affected by water body accuracy. The single difference feature analysis method and classification result comparison method are difficult to meet the accuracy requirements of large-scale and long-term flood season dynamic monitoring.

[0048] In order to solve the above technical problems, the embodiment of the present application provides a flood dynamic monitoring algorithm based on SAR time series feature optimization, including the following steps:

[0049] S101. Obtain Sentinel-1SAR sample data of several target areas during flood season;

[0050] S102, processing Sentinel-1SAR sample data of the target area during the flood season to construct an initial feature set;

[0051] S103, assigning labels to Sentinel-1SAR sample data of some target areas during the flood season to form a training set, wherein the label types include normal water bodies, submerged water bodies, and non-water bodies;

[0052] S104, training a recognition model using a training set;

[0053] S105, determining the optimal number of features;

[0054] S106. Based on the optimal feature subset, the optimal feature subset is used as input, and the trained recognition model is used to output the spatiotemporal change results of the target area.

[0055] The following is a further introduction based on specific usage scenarios.

[0056] First, step S101 is performed to obtain Sentinel-1SAR sample data of several target areas during the flood season according to the detection position and flood season time of the target area;

[0057] S102. Process the Sentinel-1SAR sample data of the target area during the flood season to construct an initial feature set, which includes:

[0058] Multi-temporal processing and interferometric processing are performed on the Sentinel-1SAR sample data of the target area during the flood season to obtain the backscattering coefficient and interferometric coherence coefficient corresponding to the flood season time series. Multi-temporal processing may include pre-processing such as data import, cropping, multi-view processing, image registration, multi-temporal filtering, geocoding and radiation calibration. De GrandiSpatio-Temporal Filtering is used for multi-temporal filtering. In the dynamic monitoring of flood conditions, the automatic detection of flood time series can analyze the temporal and spatial variation of flood conditions from a macro perspective.

[0059] Calculate the SDWI polarization index of Sentinel-1SAR sample data in the target area during the flood season:

[0060] K SDWI =ln(10*VV*VH)

[0061] Where: VV and VH represent VV and VH polarization images respectively, Sentinel-1SAR data are VV and VH dual-polarization data, and the SDWI polarization index can effectively distinguish water bodies from non-water bodies;

[0062] The initial feature set is constructed based on the SDWI polarization index, backscattering coefficient and interference coherence coefficient of Sentinel-1SAR sample data in the target area during the flood season;

[0063] S103, assigning labels to Sentinel-1SAR sample data of some target areas during the flood season to form a training set, wherein the label types include normal water bodies, submerged water bodies, and non-water bodies, and the labels of the sample points can be assigned based on high-definition images of the same period, relevant flood data, etc.;

[0064] S104, using the training set, taking the initial feature set of Sentinel-1SAR samples in the flood season of the target area as input, and taking the labels corresponding to the Sentinel-1SAR sample data in the flood season of the target area as supervision, training the recognition model, wherein the recognition model is a random forest classification model. Of course, labels may also be assigned to the Sentinel-1SAR sample data in the flood season of some target areas to form a verification set, and the verification set is used to verify the accuracy of the trained recognition model;

[0065] S105, feature optimization is performed from the polarization dimension and the time dimension, and the optimal number of features is determined by using the intra-class and inter-class distances. In the feature optimization from the polarization dimension and the time dimension, feature selection is performed from the polarization dimension as follows:

[0066]

[0067] Among them, NEHS is the standard entropy histogram similarity obtained by normalizing the entropy histogram similarity EHS, S is the structural similarity index SSIM, and β is a constant used to give different weights to the two measurement methods. The SSIM that can reflect the image structure and the image entropy EHS that represents the grayscale distribution of the image are combined to consider the similarity between images more comprehensively.

[0068] The EHS calculation is as follows:

[0069]

[0070] Among them, M and N are the length and width of the image, T is the column vector of the image expansion dimension, and its shape can be expressed as 1×(M·N). The role of image entropy in similarity measurement is to measure the information content and difference between two images, so as to determine whether they are similar;

[0071] SSIM is calculated as follows:

[0072]

[0073] Among them, S(x,y) is the structural similarity measure of the statistical similarity between image x and image y, μx and μy are the statistical average values ​​of pixels in image x and image y respectively. are the statistical variances of the pixels of image x and image y respectively, c1 and c2 are two constants;

[0074] In feature optimization from the polarization dimension and the time dimension, feature selection from the time dimension is as follows:

[0075]

[0076] Among them, ED is the Euclidean distance, SCS is the spectral correlation similarity, which reflects the shape similarity of the curve. SSV comprehensively evaluates the above indicators. ED and SCS are calculated as follows:

[0077]

[0078] Where x and y represent two different samples. μ and σ represent the mean and standard deviation of the samples.

[0079] Choosing the appropriate number of input features has a great impact on the performance of the model. By calculating the intra-class distance and inter-class distance under different numbers of feature combinations, the number of optimal feature subsets is determined. For a category Ci, the average distance between all sample points within it can be expressed as,

[0080]

[0081] Where ni is the number of samples in category Ci, xj and xk are any two samples in category Ci, and d(xj,xk) is the distance between samples xj and xk;

[0082] For two different categories Ci and Cj, the average distance between all sample points can be expressed as:

[0083]

[0084] Where ni and nj are the number of samples in categories Ci and Cj respectively, xj and xk are any two samples in categories Ci and Cj respectively, and d(xj,xk) represents the distance between samples xj and xk;

[0085] S106. Based on the optimal feature subset, the optimal feature subset is used as input, and the trained recognition model is used to output the spatiotemporal change results of the target area.

[0086] In general, the embodiments of the present application achieve high-precision flood season water body monitoring by jointly optimizing and selecting features from the polarization dimension and the time dimension respectively, using a small number of effective optimization features, thereby improving the accuracy and reliability of dynamic flood monitoring, and realizing automatic, rapid and effective monitoring of spatiotemporal information of water bodies at a large scale.

[0087] The embodiments of this specific implementation method are all preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Therefore, all equivalent changes made based on the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A flood dynamic monitoring algorithm based on SAR time series feature optimization, characterized in that: include: Acquire Sentinel-1SAR sample data during flood season in several target areas; Process the sample data and construct the initial feature set; Assigning labels to some sample data to form a training set, wherein the label types include normal water bodies, submerged water bodies, and non-water bodies; Use the training set to train the recognition model; Determine the optimal number of features; Based on the optimal feature subset, the optimal feature subset is used as input and the trained recognition model is used to output the spatiotemporal change results of the target area.

2. According to claim 1, a flood dynamic monitoring algorithm based on SAR time series feature optimization is characterized in that: The SDWI polarization index is calculated as follows: K SDWI =ln(10*VV*VH) Wherein: VV and VH represent VV and VH polarization images respectively.

3. The flood dynamic monitoring algorithm based on SAR time series feature optimization according to claim 1 is characterized in that: The recognition model is a random forest classification model.

4. The flood dynamic monitoring algorithm based on SAR time series feature optimization according to claim 3 is characterized in that: Also includes: Label the Sentinel-1SAR sample data in some target areas during the flood season to form a validation set; The validation set is used to verify the accuracy of the trained recognition model.

5. The flood dynamic monitoring algorithm based on SAR time series feature optimization according to claim 1 is characterized in that: Process the sample data and construct the initial feature set including: Multi-temporal processing and interferometric processing are performed on the Sentinel-1SAR sample data of the target area during the flood season to obtain the backscattering coefficient and interferometric coherence coefficient corresponding to the flood season time series; Calculate the SDWI polarization index of the sample data; The initial feature set is constructed based on the SDWI polarization index, backscattering coefficient and interference coherence coefficient of the sample data.

6. The flood dynamic monitoring algorithm based on SAR time series feature optimization according to claim 1 is characterized in that: Determining the optimal number of features involves: Optimize features from the polarization dimension and time dimension; The intra-class and inter-class distances are used to determine the number of optimal feature subsets.

7. The flood dynamic monitoring algorithm based on SAR time series feature optimization according to claim 6 is characterized in that: Feature selection from polarization dimension: Among them, NEHS is the standard entropy histogram similarity obtained by normalizing the entropy histogram similarity EHS, S is the structural similarity index SSIM, and β is a constant; The EHS calculation is as follows: Where M and N are the length and width of the image, and T is the column vector of the image expansion dimension, whose shape can be expressed as 1×(M·N); SSIM is calculated as follows: Among them, S(x,y) is the structural similarity measure of the statistical similarity between image x and image y, μx and μy are the statistical average values ​​of pixels in image x and image y respectively. are the statistical variances of the pixels of image x and image y respectively, and c1 and c2 are two constants.

8. The flood dynamic monitoring algorithm based on SAR time series feature optimization according to claim 7 is characterized in that: Feature selection from the time dimension: Among them, ED is the Euclidean distance, SCS is the spectral correlation similarity, and ED and SCS are calculated as follows: Here, x and y represent two different samples, μ and σ represent the mean and standard deviation of the samples.

Images