A water body suspended matter spatial distribution monitoring method based on dual-polarized SAR features

Abstract

The application discloses a water body suspended matter spatial distribution monitoring method based on a dual-polarized SAR feature and belongs to the field of polarized synthetic aperture radar image processing. The application uses dual-polarized data to study the suspended silt target, aims to explore the role of the polarized data in identifying the suspended silt target. A preliminary classification map is obtained by using a new classification plane A' / alpha, A' has a more intuitive physical meaning compared with polarization entropy (H), then a likelihood ratio distance classifier capable of distinguishing weak scattering targets is used for further clustering to accurately lock the research range, and the polarized features (Shannon entropy intensity components) that are helpful to the identification of the suspended silt are analyzed, locked and used to accurately extract the suspended silt area.

Description

TECHNICAL FIELD

[0001] The application relates to a water body suspended matter spatial distribution monitoring method based on a dual-polarization SAR feature, and belongs to the field of polarized synthetic aperture radar image processing. BACKGROUND

[0002] Synthetic aperture radar (SAR) obtains measurement information by emitting microwave pulses and receiving electromagnetic signals scattered back by an observation target, and uses relevant data information processing technology to synthesize the obtained data information into a high-resolution two-dimensional earth surface reflectivity image. SAR is an active radar imaging system, does not depend on external light sources, can image day and night, and has the ability to penetrate clouds, dust, fog and other substances, and can obtain scattering data of an observation target at all times. Polarization is an inherent property of electromagnetic waves and is also a basic characteristic, which describes the shape and rotational direction of the spatial trajectory formed by the endpoint of the electric field vector as a function of time. The electric field vector of a plane electromagnetic wave can be decomposed into horizontal and vertical directions in a coordinate system, and the relative relationship between the two components constitutes the polarization mode of the plane electromagnetic wave. It is worth noting that the measured ground objects of a SAR system under different polarization states will exhibit different scattering characteristics. Single-polarization data only exist in one channel information and do not have available polarization information; full polarization has complete polarization information, but the system is complex to develop, has a small imaging width, and data is not easy to obtain; dual polarization is a compromise between the two, has available polarization information, large-width imaging, and a relatively simple data acquisition approach. Polarization data has been widely used in agriculture, military, disaster monitoring and other aspects.

[0003] Suspended sediment is an important parameter of water body, which can be used as a tracer of water movement. Suspended sediment is mostly fine sand and clay particles, which can adsorb, fix and transfer pollution elements in water to purify water quality. However, if the suspended sediment is deposited, it will affect the pollution level of river bottom. The transfer and deposition process of suspended sediment will also affect the port and navigation function of port cities and the environmental quality of coastal wetlands. The time sequence monitoring of suspended sediment is of great significance to water quality monitoring, river dam blocking and environmental protection of waterline coastal wetlands. Most of the past researches focus on suspended sediment concentration inversion and mapping. Optical and hyperspectral remote sensing is used to analyze the received light spectrum after full scattering and reflection of suspended sediment particles in water, to build a concentration inversion model for quantitative inversion of suspended sediment. Due to the wide distribution of suspended sediment, the use of a unified model for inversion of different types of water bodies may not have good inversion effect. Therefore, researchers have classified suspended sediment and used different inversion models for inversion in different types of sediment areas to improve the accuracy. Some researches have combined different sensor data to improve the accuracy of water quality parameter estimation. However, it is found that SAR data has little contribution to the estimation of water quality parameter value because the working mode of SAR and optical satellite is different. SAR signal does not penetrate into water, but is reflected from the water surface, which is only affected by the characteristics of water surface. The spectral characteristics of water body mainly show the spectral scattering information of a certain water depth and the body scattering information of the material in the water body rather than the surface scattering. In general, the existing inversion researches rarely use polarized SAR information and do not fully explore the other applications of polarized information in suspended sediment. The temporal and spatial variation of water surface roughness will interfere with the interpretation of optical data, and SAR image is sensitive to surface roughness. Therefore, the present invention does not focus on the inversion of suspended sediment concentration, but only explores the suspended sediment from the perspective of polarized SAR. Through analysis and use of effective polarized SAR features, the distribution area of suspended sediment in water body is effectively identified, which is conducive to the joint use of optical data to some extent and provides classification results to promote subsequent quantitative inversion research by using different models for inversion in different areas. SUMMARY

[0004] Since few researches have explored the role of polarized SAR data in suspended sediment area monitoring, the main purpose of the present invention is to provide a water body suspended matter spatial distribution monitoring method based on dual-polarized SAR features, which is applied to the dynamic monitoring of suspended sediment by polarized SAR data in dual-polarized VV+VH mode, and explores the effectiveness of polarized SAR data in suspended sediment target identification.

[0005] The H / alpha likelihood ratio distance classification method is beneficial to the classification of the easily confused weak scattering target ground objects on the polarimetric SAR image, such as the water body, the road, the shadow and the low backscattering coefficient value target such as the bare soil. However, the calculation process of the polarization feature polarization entropy (H) parameter obtained based on the feature decomposition is relatively complex and the physical meaning is not clear. In view of this point, the application proposes a new classification plane A' / alpha plane. On the basis of the A' / alpha preliminary classification, the likelihood ratio distance clustering is carried out, and finally the effective monitoring of the suspended sediment target is carried out in combination with the intensity component feature of the Shannon entropy which has obvious distinguishing effect on the suspended sediment target.

[0006] The technical scheme of the application specifically comprises the following contents:

[0007] 1. Obtaining C2 matrix: obtaining the dual polarization covariance matrix C2 through specific sentinel data processing software for data preprocessing, for subsequent experimental classification.

[0008] 2. A' / alpha classification plane parameter research: using the polarization entropy H value obtained by the feature decomposition of the C2 matrix in step (1) as a reference, using the anisotropy A, an A' parameter having an approximate linear relationship with H is proposed, and the physical meaning is more intuitive.

[0009] 3. A' / alpha classification: according to the parameter A' obtained in step (2), the segmentation threshold for dividing the low entropy, medium entropy and high entropy range is re-determined, and the 9-class target recognition classification is carried out in combination with alpha.

[0010] 4. Likelihood ratio distance clustering: on the basis of the preliminary 9-class distribution map obtained in step (3), the likelihood ratio distance clustering is carried out, and the water body, the bare soil and the water body surface containing the suspended sediment area which is not too turbid are effectively classified, so as to reduce the research range in step (5).

[0011] 5. Joint Shannon entropy intensity component suspended sediment accurate identification: according to the classification result of step (4), the area for further research is selected, and the K-means iterative clustering is carried out by using the Shannon entropy intensity component feature, so as to further effectively distinguish the relatively clear water body area and the suspended sediment area in the water body range, so as to obtain the final suspended sediment monitoring result.

[0012] In step (1), the existing data processing software is used to preprocess the original data and export the covariance matrix C2 required for subsequent experiments. The covariance matrix C2 is calculated by the conjugate product of the target vector. Taking the VV+VH dual polarization mode as an example, the target vector k is defined as:

[0013] k=[<S vv ><S vh >] T(1)

[0014] where S vv and S vh represent the scattering information of linear polarization state of dual-polarization data VV, VH, respectively, and <...> denotes the spatial statistical average of random scattering medium in isotropic case. The covariance matrix is defined as:

[0015]

[0016] where * denotes the complex conjugate.

[0017] Natural objects in nature can generally maintain the symmetry of the flight direction, so the co-polarization component and cross-polarization component are not related, and and then equation (2) can be expressed as:

[0018]

[0019] where,

[0020]

[0021]

[0022] The covariance matrix is a semi-positive definite Hermite matrix, so the covariance matrix C2 can be decomposed into the weighted sum of two mutually orthogonal covariance matrices, and the expression is as follows:

[0023]

[0024] where λ i and e i respectively represent the real eigenvalues and eigenvectors of the covariance matrix C2, C i all represent independent covariance matrices with rank 1, and λ i respectively represent a scattering mechanism, and the corresponding eigenvalue λ i represents the intensity of the scattering mechanism. In order to analyze the disorder of the scattering process of the object, the scattering entropy is introduced, which is defined as:

[0025]

[0026] where, λ1, λ2 are two eigenvalues of the target decomposition, and when the polarization entropy H is 0, it means that the system is in a completely polarized state; when one of the eigenvalues is larger and the other is smaller, H can be ignored; when H is smaller, the system is close to a completely polarized state; when H is larger, the system is close to a completely non-polarized state, and at this time the two eigenvalues are close; when the value is 1, the system is in a completely non-polarized state, the polarization information is 0, and the target scattering completely represents the randomness of the target.

[0027] The effective range of the introduced average scattering angle a corresponds to the continuous change of the scattering mechanism, and when the values are 0°, 45°, and 90°, they respectively represent the surface scattering of geometric optics, single scattering of anisotropic particle cloud or dipole scattering, and double scattering of a metal surface. The parameter is expressed as:

[0028]

[0029] The value of the introduced anisotropy A can represent the proportion of two scattering mechanisms, and is expressed as:

[0030]

[0031] The polarization entropy H decomposed by H / a in the above-mentioned step (2) has a complicated expression, and its actual meaning cannot be intuitively understood. Therefore, by referring to related literature and multiple experiments, we use A to propose an A' parameter to optimize and improve the problem. A' is expressed as:

[0032]

[0033] When λ1> λ2, formula (10) is expressed as:

[0034]

[0035] When λ1< λ2, formula (10) is expressed as:

[0036]

[0037] The meaning of the parameter A' is more intuitive and clear. As can be seen from formula (11) and formula (12), its value is consistent with the value range of H, which is between 0 and 1. Since the eigenvalue can represent the strength of the scattering mechanism, we can intuitively see that when λ1= λ2, A' takes a value of 1, which represents that the weights of the two decomposed scattering mechanisms are the same, and the scattering mechanism of the scattering unit is complex. When one eigenvalue is much larger than the other, the value of A' tends to 0, and at this time, the scattering mechanism of the scattering unit has only one dominant mechanism, and the scattering mechanism is relatively simple.

[0038] In step (3), the parameter A' and a are used for 9-class region division. The segmentation threshold for distinguishing low, medium, and high entropy regions of A' is selected according to the corresponding relationship curve of A' and H of the data set. Compared with full polarization data, dual polarization data carries less information of one polarization channel, and cannot completely distinguish the 9 categories only according to the division range, and the scattering type of the scattering unit at the division boundary of H / a is prone to misclassification. Therefore, after completing the preliminary division of the 9-class region, a likelihood ratio distance classifier needs to be used for further clustering to obtain a more accurate classification result.

[0039] The likelihood ratio distance classifier used in step (4) has been proven to be more effective in classifying weak scattering targets (water, bare soil, asphalt road and shadow, etc.) in polarimetric data than the Wishart classifier. It updates the class of a pixel using the likelihood distance, which is defined as follows:

[0040] d(C1,C2) = (v1 + v2) ln |C| - v1 ln |C1| - v2 ln |C2| (13)

[0041] where, v i represents the number of sample pixels used to estimate the covariance matrix C i .

[0042] From the classification result of step (4), we can further narrow down the study area by selecting the classified water and suspended sediment regions. By observing the available data's dual-polarization characteristics, we can find that the Shannon entropy intensity component can effectively display the surface geometric characteristics of suspended sediment material within the water body. Therefore, this feature value is used to further automatically and accurately divide the material within the water body. Shannon entropy is only related to the determinant of the covariance matrix (coherence matrix), which contains rich polarization information, and is represented as follows:

[0043] SE = log(π 2 e 2 |C2|) = SE I + SE P (14)

[0044] where SE I and SE P represent the intensity component related to the total backscatter power and the polarization component related to the Barakat degree of polarization, respectively. SE I is expressed as:

[0045]

[0046] Compared with water, the backscatter value of the suspended sediment surface is higher. This is used in step (5) to further distinguish between clear water and sediment-laden water. The total number of classes of the pixel points in the study area further narrowed down in step (4) is used as the number of clustering classes for K-means clustering, and the average SE I value of each class of pixel points is used as the initial clustering center for one iteration to obtain the final suspended sediment detection map.

[0047] Compared with the prior art, the new classification plane A' / alpha used in the application effectively solves the problems of complicated calculation of polarization entropy H parameter in H / alpha calculation and unclear expression of physical meaning, so that the speed of parameter calculation is improved to a certain extent, and the expression mode of parameter eigenvalue is conducive to directly judging the proportion of a certain scattering mechanism. The role of polarimetric SAR information in suspended sediment identification is explored, and the dual-polarization characteristic information is compared and analyzed, and SE I The characteristics are used as the last division basis of the flow, and the results prove that the identification effect of the water body suspended sediment is good, which will be conducive to subsequent quantitative research.

[0048] In summary, the application aims at the phenomenon that few studies on the role of polarimetric SAR data in suspended sediment area monitoring in the past, and a better suspended sediment area identification effect is realized by using the new classification plane and polarization characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A curve graph of the relationship between the parameter A' and the polarization entropy H.

[0050] Figure 2 The same data A' / alpha and H / alpha classification chart.

[0051] Figure 3 It is an A' / alpha likelihood ratio distance classification chart, (a) is a scatter plot, and (b) is a likelihood ratio clustering effect chart.

[0052] Figure 4 The SE I characteristics of the clear water body and the suspended sediment research area.

[0053] Figure 5 The result chart of the suspended sediment area of time series monitoring.

[0054] Figure 6 The flowchart of the method. DETAILED DESCRIPTION

[0055] The application will be described in detail below in combination with the drawings and examples.

[0056] Figure 6The overall method flow of the present application is consistent with the foregoing summary. The method proposed by the present application is verified on the Sentinel-1 dual-station estuary time series data set. The time period for obtaining the time series data is between April and July 2022, and a total of 3 Sentinel-1 time series data are used on April 21, May 3 and May 15, and the size of the image is about 2186x2281, and the polarization mode is VV+VH. The existing SNAP and PolSARpro software are used for data preprocessing, mainly including orbit correction, radiation calibration, multi-view, terrain correction, and generation of C2 matrix, polarization filtering, etc.

[0057] According to formula (10), the parameter A' is calculated, compared with the calculation of polarization entropy H, for a data set with a size of 2353x2281, using a computer with an Intel Core i5 processor and a CPU frequency of 2.42 GHz and a memory of 16 GB, the pixel-by-pixel calculation time using Matlab can be shortened by about 5.625 s. We calculated the A' parameter map of each time series data set used in the experiment, Figure 1 The curve relationship diagram of the parameter A' and the polarization entropy H of the dual-polarization data on May 3 is shown, and the curve forms of the other data sets are similar, and it can be observed that A' and H present an approximately linear relationship, and the segmentation threshold for determining the low-entropy, medium-entropy and high-entropy regions of the data is changed from the original 0.5 and 0.9 to 0.4691 and 0.795 on the relationship curve, and the segmentation threshold is used in step (3) to classify A' / α combined with A' and α. Figure 2 (a), (b), (c) and (d) respectively show the scatter plot and classification diagram of H / α and A' / α under the same data, and it can be seen that the classification results of the two classification methods are basically consistent. Table 1 below shows the classification confusion matrix of the dual-polarization VV+VH mode, the vertical direction represents the H / α classification result, and the horizontal direction represents the A' / α classification information. The H / α classification result is used as the true value to compare the predicted value of the A' / α classification result. Among them, the higher the UA value represents the higher the matching degree of the A' / α classification result and the H / α classification true class label, and the smaller the probability of misclassification, the PA value represents the part of the A' / α classification result consistent with the H / α classification true class, and the overall accuracy OA can be calculated from the confusion matrix as 99.99%, and the Kappa coefficient is 0.9999, which shows that the classification results of the A' / α plane and the H / α plane are almost completely consistent, Figure 2The classification results shown in (c) and (d) are also visually similar. By observing the UA and PA values in Table 1, it can be found that the classification results of each pixel unit in the A' / a plane and the H / a plane in the data set are almost consistent, both reaching an accuracy of 99.98%, which shows that the A' / a plane can better replace the H / a plane to classify the scattering mechanism categories of the pixel units, and the combination of the expressions (11) and (12) of the A' feature can directly show the meaning it represents, which shows that the construction effect of the A' / a plane is excellent.

[0058] Table 1 VVVH mode classification confusion matrix

[0059]

[0060] After obtaining the A' / a preliminary classification map of step (3), further clustering is performed using the likelihood ratio distance formula to more effectively obtain a more accurate physical scattering mechanism classification map, which is beneficial to further narrowing the research area in step (5) and effectively analyzing the display effect of the polarization feature in different water body material targets. Figure 3 The scatter plot of (a) shows that the classification of the scatter points of each category after the likelihood ratio distance classification changes greatly, and most of them change in the adjacent category area, which shows that the boundary pixel confusion of the fixed boundary of the H / a (A' / a) classification plane has a certain degree of improvement. Figure 3 (b) shows the effect of the likelihood ratio clustering, and it can be clearly seen that compared with the initial A' / a classification map, further clustering using the likelihood ratio distance can obtain a classification effect map with clear categories, and the category attributes of each pixel point are more clear, which is more conducive to subsequent experiments.

[0061] The Shannon entropy in step (5) is defined as the sum of the intensity (SE I ) and polarization (SE P ) components by Morio et al., and by observing Figure 4 , it can be seen that the Shannon entropy intensity component SE I of the suspended sediment area is relatively clear, and the SE I value of the relatively clear water body is larger. Using this feature, we can effectively distinguish the water body area containing suspended sediment from the relatively clear water body area again, and distinguish them by using the 1-time K-means iteration adaptive method. In K-means clustering, the average SE I value of each pixel point in the same classification region label in the research area selected in step (4) is taken as the initial clustering center, and after calculating the discrimination distance of each pixel point from the clustering center, the pixel point is classified into the smallest distance class.Figure 5 The final data classification results are shown in the last column, Figure 5 (a), (b) and (c) represent the intensity components of the VV channel of each data set, (d), (e) and (f) represent the final classification results of the corresponding data, (g), (h) and (i) represent the contrast results of H / αWishart clustering, and (j), (k) and (l) represent the contrast results of H / A / αWishart clustering. It can be observed that the color representation of the water body and the suspended sediment area is not fixed, which indicates that a fixed scattering mechanism cannot be used to describe both at this time series data. However, they can still be effectively distinguished, and the suspended sediment area is contained in the water body range. In (d), the magenta color represents the suspended sediment area, and the blue-violet color represents the relatively clear water body area; in (e), the black color represents the suspended sediment area, and the dark blue color represents the relatively clear water body area; in (f), the dark blue color represents the suspended sediment area, and the green color represents the relatively clear water body. (g)-(i) represent the H / αWishart results of each time series. The contrast experimental results can clearly show that the classification categories of (g)-(i) are relatively monotonous, there is confusion between different categories, and there is no good effect for the recognition of the suspended sediment area in the water body; (j)-(l) represent the H / A / αWishart clustering results of each time series. It has a more clear category division than the H / αWishart result, but the recognition effect for the suspended sediment target is also relatively poor. In summary, the recognition effect of the suspended sediment obtained by the experimental process of the present application is relatively good.

Claims

1. A method for monitoring spatial distribution of water body suspended matter based on dual-polarization SAR features, characterized in that: Comprise the following contents: Step (1), acquisition Matrix: The dual-polarized covariance matrix is obtained by preprocessing the data through specific sentinel data processing software , for subsequent experimental classification; Step (2) Classification plane parameter study: using the parameters in step (1) Polarization entropy obtained by eigenvalue decomposition of a matrix Using the value as a reference, the anisotropy degree is utilized. To propose a Parameters with an approximately linear relationship ; Step (3), Classification: parameters obtained according to step (2) , re-determine the segmentation threshold for dividing the low-entropy, medium-entropy and high-entropy ranges, and jointly perform 9-class target recognition classification; Step (4), likelihood ratio distance clustering: on the basis of the preliminary 9-class distribution map obtained in step (3), the likelihood ratio distance clustering is carried out, water, bare soil and water surface with less turbidity of suspended sediment area in the SAR image are preliminarily and effectively classified, and the research range is reduced for step (5); Step (5), accurate identification of suspended sediment combined with shannon entropy intensity component: according to the classification result of step (4), the further researched area is selected, and the K-means iterative clustering is carried out by using the shannon entropy intensity component feature, the relatively clear water area and the suspended sediment area in the water body are further effectively distinguished, so as to obtain the final suspended sediment monitoring result; Covariance matrix is calculated by the conjugate product of the target vector, VV+VH in dual-polarized mode, the target vector is defined as : （1）， where and The components represent the scattering information of the dual-polarized data VV, VH linear polarization states, respectively, denotes the spatial statistical average of the random scattering medium in the case of isotropy; the covariance matrix is defined as: （2）， where * denotes complex conjugation; is the target vector; Utilizing It is proposed Parameter optimization improves, is expressed as: （10）， wherein when Formula (10) is represented as: （11）， When Formula (10) is expressed as: （12）， From formula (11), formula (12), it can be known that The value of the value range is consistent with The value range is consistent, between 0 and 1.

2. The method according to claim 1, wherein the method is characterized by: In step (1), the original data is pre-processed using existing data processing software and the covariance matrix required for subsequent experiments is exported ; The co-polarized component and the cross-polarized component are not correlated, and = 0 and = 0, then equation (2) is expressed as: （3）， Wherein, （4）， （5）， The covariance matrix is a Hermitian matrix that is positive semi-definite, so the covariance matrix is decomposed into a weighted sum of two mutually orthogonal covariance matrices, expressed as follows: （6）， where and denote the real eigenvalues and eigenvectors of the covariance matrix respectively, denote two independent covariance matrices of rank 1, and denote two scattering mechanisms with corresponding eigenvalues denote the intensity of the scattering mechanisms; to analyze the disorder of the scattering process of ground objects, a scattering entropy is introduced, defined as （7）， wherein , , are two eigenvalues of the target decomposition, the polarization entropy is 0, the system is in a fully polarized state; is the average scattering angle; The effective range of the introduced average scattering angle corresponds to the continuous variation of the scattering mechanism, taking values between when, respectively, represent the geometric optics surface scattering, single scattering by a dipole scatterer or anisotropic cloud of particles and the double scattering from a metallic surface; it is expressed as: （8）， introduced anisotropy The value of the introduced anisotropy degree represents the proportion of the two scattering mechanisms, which is expressed as: （9）。 3. The method according to claim 2, wherein the method is characterized by: In step (3) using the parameters With 9 categories of regional division, The segmentation threshold of low, medium and high entropy region is selected according to the corresponding relationship curve of data set And The clustering is carried out by using likelihood ratio distance classifier after the preliminary division of 9 categories of regions is completed, and the classification result is obtained. The likelihood ratio distance classifier used in step (4) is as follows: （13）， wherein , represents the number of sample pixels used for estimating the covariance matrix ​ 4. The method according to claim 3, wherein the method is characterized by: Through the classification result of step (4), the dual polarization features of the data are selected by the water and suspended sediment area labels, the surface geometric characteristics of the suspended sediment material in the water body are effectively displayed by the shannon entropy intensity component, and the material in the water body is automatically and accurately divided; The shannon entropy is only related to the determinant of the covariance matrix, contains polarization information, and is expressed as follows: （14）， wherein, and represent the intensity component related to the total backscattered power and the polarization component related to the Barakat degree of polarization, respectively; is represented as: （15）， The total number of categories existing in the pixel points in the research area further reduced in step (4) is taken as the number of clustering categories of K-means clustering, and the average value of the pixel points of each category is taken as the initial clustering center to perform one iteration to obtain the final suspended sediment detection map. value of the pixel points of each category is taken as the initial clustering center to perform one iteration to obtain the final suspended sediment detection map.

Images