PolSAR two-dimensional dynamic bandwidth variational mode decomposition method based on Bayesian parameter selection, computing device and computer readable storage medium

Abstract

The invention discloses a PolSAR two-dimensional dynamic bandwidth variational mode decomposition method based on Bayesian parameter selection, a computing device and a computer readable storage medium, and the method comprises the following steps: carrying out the preprocessing of a scattering matrix of polarized SAR data, and extracting the intensity information of six polarized channels; updating each modal component through an alternating direction multiplier method, and performing two-dimensional inverse Fourier transform on each modal component; constructing a Bayesian optimization objective function, and updating an optimal evaluation value based on a Gaussian process and an expectation improvement function; outputting the optimal modal number when the maximum evaluation frequency is reached; and taking the optimal mode number as a parameter, and obtaining a final decomposition result through two-dimensional dynamic bandwidth variational mode decomposition. The method has the advantages that the two-dimensional variational mode decomposition algorithm is rewritten, so that the method has the adaptive characteristic of bandwidth limitation; secondly, a Bayesian optimization design objective function is introduced on the basis of two-dimensional dynamic bandwidth variational mode decomposition, so that a key parameter K can be adaptively selected;

Description

Technical Field

[0001] This invention relates to the field of two-dimensional variational mode decomposition technology, and in particular to a method for PolSAR two-dimensional dynamic bandwidth variational mode decomposition based on Bayesian parameter selection. Background Technology

[0002] Two-dimensional variational mode decomposition (2D-VMD) is a signal processing method based on variational principles, designed to extract multiple modal components from complex signals, and is particularly suitable for processing images and two-dimensional signals. It decomposes a two-dimensional signal into several sub-signals (modes) with specific frequencies and local features, providing an effective tool for subsequent signal analysis.

[0003] In medical image analysis, it can separate high-frequency random noise from low-frequency tissue structure information containing diagnostic value in MRI or CT images, achieving high-fidelity image denoising and edge enhancement, thereby assisting doctors in more accurate lesion identification.

[0004] In the field of remote sensing image analysis, it has become a powerful tool for processing speckle noise in polarimetric synthetic aperture radar (PolSAR) images. It can also be used for feature extraction of multispectral or hyperspectral images, improving the accuracy of land cover classification by decomposing the texture and spatial frequency information of different land features.

[0005] In traditional 2D variational mode decomposition (2D-VMD) implementations, the bandwidth penalty for each mode is fixed and requires manual setting. However, during the decomposition process, the center frequencies and energy distributions of different modes vary. A fixed bandwidth cannot simultaneously accommodate the different needs of some modes—one requiring a wider passband to preserve important high-frequency details, and the other requiring a narrower passband to eliminate interference from neighboring modes—leading to either loss of detail or mode aliasing. To address the parameter sensitivity issue faced by 2D-VMD when processing polarimetric images, traditional methods typically rely on manual experience or a grid search approach to set key parameters (number of modes and bandwidth penalty parameters). However, these methods are not only computationally expensive and labor-intensive, but also have significant uncertainties in parameter selection, easily leading to unstable mode decomposition results. This, in turn, affects the complete extraction of polarimetric information and the effectiveness of subsequent applications such as ground feature identification and classification. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a method for PolSAR two-dimensional dynamic bandwidth variational mode decomposition based on Bayesian parameter selection.

[0007] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0008] The PolSAR two-dimensional dynamic bandwidth variational mode decomposition method based on Bayesian parameter selection includes the following steps:

[0009] The scattering matrix of polarimetric SAR data is preprocessed to extract the intensity information of the six polarimetric channels;

[0010] Define the range of the number of modes K;

[0011] Initialize the maximum number of iterations for Bayesian optimization;

[0012] Calculate the bandwidth occupied by the intensity information;

[0013] Set the upper and lower bounds of the bandwidth, and initialize the Lagrange multipliers and the dynamic bandwidth penalty factor;

[0014] Each modal component is updated using the alternating direction multiplier method and then subjected to a two-dimensional inverse Fourier transform.

[0015] Update the Lagrange multipliers and dynamic bandwidth penalty factor;

[0016] Determine if the overall relative change of the modal components is less than a set threshold. If it is, output K modal components; otherwise, continue iterating.

[0017] Calculate the mean and minimum JS divergence of the K modal components;

[0018] Calculate the reconstructed mean square error of the K modal components;

[0019] Construct a Bayesian optimization objective function and update the optimal evaluation value based on the Gaussian process and the desired improvement function;

[0020] Output the optimal number of modes when the maximum number of evaluations is reached;

[0021] Using the optimal number of modes as a parameter, the final decomposition result is obtained through two-dimensional dynamic bandwidth variational mode decomposition.

[0022] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for PolSAR two-dimensional dynamic bandwidth variational mode decomposition based on Bayesian parameter selection.

[0023] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for PolSAR two-dimensional dynamic bandwidth variational mode decomposition based on Bayesian parameter selection.

[0024] The beneficial effects of this invention are:

[0025] This invention modifies the two-dimensional variational mode decomposition algorithm to give it adaptive characteristics with bandwidth constraints. Secondly, it introduces Bayesian optimization on the basis of two-dimensional dynamic bandwidth variational mode decomposition to maximize the discriminativeness between the decomposed modes and the difference between the reconstructed and original data. The objective function is designed to allow for adaptive selection of the key parameter K. Attached Figure Description

[0026] Figure 1 This is a system flowchart of the present invention;

[0027] Figure 2 The optimal result image for modality selection of a synthesized image with 5 different modalities;

[0028] Figure 3 This is a comparison chart of the decomposition results of the present invention with those of traditional 2D-VMD and precise component tuning;

[0029] Figure 4 This is the optimal result image for intensity image modality selection on the Radarsar-2 dataset according to the present invention;

[0030] Figure 5 This is a diagram showing the decomposition results of this invention on the Radarsar-2 dataset. Detailed Implementation

[0031] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0032] This invention focuses on the feature extraction problem of Polarimetric Synthetic Aperture Radar (PolSAR) data, aiming to efficiently decompose and extract multimode scattering features from raw polarimetric echoes. Traditional methods for two-dimensional variational mode decomposition (2D-VMD) address the parameter sensitivity issue when processing polarimetric images. These methods typically rely on manual experience or grid search to set key parameters (number of modes and bandwidth penalty parameters). However, these approaches are not only computationally expensive and labor-intensive, but also suffer from significant uncertainty in parameter selection, leading to unstable mode decomposition results and consequently affecting the complete extraction of polarimetric information and subsequent practical applications such as ground feature identification and classification.

[0033] To solve the above problems, the present invention adopts the following technical solution, and the system flowchart is as follows. Figure 1 As shown:

[0034] 1) Data preprocessing

[0035] In polarimetric SAR data, the scattering matrix contains attributes such as the target's phase, energy, and polarization, and can comprehensively describe the scattering characteristics resulting from the interaction between electromagnetic waves and the target. The scattering matrix S can be expressed as:

[0036]

[0037] Among them, S i,j The polarization channel information is represented by (i,j = H,V), which represent the polarization modes of the received and transmitted signals, respectively. H and V correspond to horizontal and vertical polarization, respectively. In single-view mode, the second-order statistical properties of PolSAR data can be expressed as the coherence matrix T3:

[0038]

[0039] According to the reciprocity principle, we have S HV =S VH T 12 =T 21 T 13 =T 31 T 23 =T 32 Therefore, the polarization characteristics of a target can be represented by the following six elements:

[0040] T 11 =<|S HH +S VV | 2 >

[0041] T 12 =<(S HH +S VV (S) HH -S VV ) * >

[0042] T 13 =<2(S) HH +S VV )S HV * >

[0043] T 22 =<|S HH -S VV | 2 >

[0044] T 23 =2 HV (S HH +S VV ) * >

[0045] T 33 =4 <S​HV 2 >

[0046] The intensity information of these six elements can be represented as TDB. i,j =log|T i,j This representation can be used to analyze the radiation characteristics of a target. However, these features often contain a large amount of noise information. To reduce the impact of noise interference on feature extraction, this invention is used to decompose these six intensity information components, effectively separating multimodal features and noise features.

[0047] 2) Set the range of the number of patterns: K∈[2,10]

[0048] 3) Initialize the maximum number of Bayesian iterations: MaxObjectiveEvaluations = 20

[0049] 4) Calculate the occupied bandwidth for polarimetric SAR intensity information: BW 2D =r high -r low

[0050] Among them: BW 2D The calculation method references the approach used to calculate the bandwidth occupied by a one-dimensional signal. First, the characteristic component TDB is calculated. i,j The power spectrum is obtained by performing a Fourier transform on (x,y):

[0051] S i,j (f x ,f y )=Γ{TDB i,j (x,y)}

[0052] P i,j (f x ,f y )=|S i,j (f x ,f y )| 2

[0053] Here, f x and f y For two-dimensional frequency coordinates, S i,j (f x ,f y ) and P i,j (f x ,f y ) are TDB i,j Two-dimensional Fourier transform and power spectral density of (x,y). In the frequency domain plane (f x ,f y Define radial distance And take its maximum value as r max, the interval [0, r max Divide the data into M equal segments, where M = max(H). x H y ), H x and H y This represents the number of sampling points in the horizontal (x) and vertical (y) directions. The dividing point is denoted as:

[0054] 0 = r0 < r1 < ... < r M =r max

[0055] The m-th segment (r) m-1 ≤r<r m Radial energy E m Defined as:

[0056]

[0057] This yields the cumulative energy ratio C. m :

[0058]

[0059] This method defines the occupied bandwidth of a signal as the range covering 99% of the total spectral energy. Therefore, the difference between the low-frequency boundary radius where the accumulated energy reaches 0.5% and the high-frequency boundary radius where the accumulated energy reaches 99.5% is used as the two-dimensional occupied bandwidth. Hence, r low =min{r m |C m ≥0.005},r high =min{r m |C m ≥0.995}.

[0060] 5) Set the upper and lower bounds of bandwidth:

[0061] 6) Initialize the Lagrange multipliers for the upper and lower bandwidth bounds of the mode components, and the dynamic bandwidth penalty.

[0062] λ α =λ β =1

[0063] dynamic_alpha = 1 + λ α -λ β

[0064] Where: λ α and λ β These are the Lagrange multipliers for the upper and lower bounds of the bandwidth, respectively, and dynamic_alpha represents the dynamic bandwidth.

[0065] 7) Update modal components using dynamic bandwidth penalty

[0066]

[0067] in: For the frequency domain representation of the k-th mode, For two-dimensional Hilbert transform, For the Fourier transform of the input data, λ t (w) is the Lagrange multiplier for the data fidelity term. Let be the center frequency of the k-th mode.

[0068] The above equation is the frequency domain update of the modal components under the alternating direction multiplier method. The spatial domain mode components can be obtained by performing a two-dimensional inverse Fourier transform on it.

[0069] 8) Update the Lagrange multipliers of the upper and lower bounds of the bandwidth of the updated mode components, as well as the dynamic bandwidth penalty.

[0070]

[0071] dynamic_alpha = 1 + λ α -λ β

[0072] in, and For λ α and λ β The corresponding update step size is set here as follows:

[0073] 9) When the overall relative change of all modes is less than a specified threshold, the algorithm is considered to have converged, its calculation results are sufficiently stable, and the iteration process is terminated. That is:

[0074]

[0075] in: and These are the solution results for the k-th mode at the (t+1)th and t-th iterations, respectively.

[0076] If the above relationship is satisfied, proceed to step 10; otherwise, return to step 7 and proceed to the next iteration update.

[0077] 10) Output K mode components

[0078] 11) Calculate the mean and minimum values ​​of the JS divergence of the K pattern components.

[0079] JS mean =mean({JS(p,q)|{p,q}∈C})

[0080] Where: JS(p,q) represents the set of JS divergences for all p≠q in set P, and mean represents the average value of the set.

[0081] JS min =min({JS(p,q)|{p,q}∈C})

[0082] Where: JS(p,q) represents the set of JS divergences for all p≠q in set P, and min represents taking the minimum value in the set.

[0083] 12) Calculate the mean square error of the reconstruction of the K mode components.

[0084]

[0085] 13) Calculate the objective function of Bayesian optimization.

[0086] EvaF = γ(JS) mean +JS min )-ηMSE

[0087] Here, γ and η are weighting coefficients used to balance the relationship between decomposition independence and reconstruction accuracy.

[0088] 14) Update the current optimal evaluation function by combining the Gaussian process and the EI selection function.

[0089] 15) Determine whether the maximum number of evaluations of the objective function of Bayesian optimization has been reached. If not, return to step 12; otherwise, output the optimal modal parameter values.

[0090] 16) Input the optimal number of modes into the 2D-DB-VMD decomposition to obtain the final result.

[0091] Two-dimensional dynamic bandwidth variational mode decomposition: The core idea of ​​2D-VMD is to seek a set of modes {u} by using the sum of the squared gradients of the bandwidths of each mode as a penalty target, while ensuring reconstruction accuracy. k} and its center frequency {w k This allows each mode to focus on a specific frequency band (with limited bandwidth) while still collectively reconstructing the original signal. Its mathematical expression is as follows:

[0092]

[0093] Here, u AS,k It is an analytic signal defined according to the Hilbert transform, w k For mode u k The corresponding center frequency, the square of the norm is defined as: α k This represents the bandwidth penalty strength for the corresponding mode k.

[0094] In traditional 2D-VMD implementations, for each mode α k Its bandwidth penalty strength is fixed and needs to be set manually. However, during the decomposition process, the center frequencies and energy distributions of different modes vary. A fixed bandwidth cannot satisfy the needs of some modes that require a wider passband to retain important high-frequency details, or some modes that require a narrower passband to eliminate interference from adjacent modes, resulting in either loss of details or mode aliasing. Therefore, to solve this problem, a dynamic bandwidth adjustment mechanism is introduced on the original basis. By using the sum of the squared gradients of the bandwidths of each mode as the bandwidth estimate of each mode, and setting an upper and lower bound for the bandwidth of each mode, the variational constraint problem of 2D-VMD can be rewritten as:

[0095]

[0096] here, 0 ≤ α ≤ β is the lower bound for each modal bandwidth estimate. The upper bound for the bandwidth estimation of each modality, where BW 2D It is TDB i,j The two-dimensional bandwidth occupied, where K is the total number of modes.

[0097] BW 2D The calculation method references the approach used to calculate the bandwidth occupied by a one-dimensional signal. First, the characteristic component TDB is calculated. i,j The power spectrum is obtained by performing a Fourier transform on (x,y):

[0098] S i,j (f x ,f y )=Γ{TDB i,j (x,y)}

[0099] P i,j (f x ,f y )=|S i,j (f x ,f y )| 2

[0100] Here, f x and f y For two-dimensional frequency coordinates, S i,j (f x ,f y ) and P i,j (f x ,f y ) are TDB i,j Two-dimensional Fourier transform and power spectral density of (x,y). In the frequency domain plane (f x ,f y Define radial distance And take its maximum value as r max , the interval [0, r max Divide the data into M equal segments, where M = max(H). x H y ), H x and H y This represents the number of sampling points in the horizontal (x) and vertical (y) directions. The dividing point is denoted as:

[0101] 0 = r0 < r1 < ... < r M =r max

[0102] The m-th segment (r) m-1 ≤r<r m Radial energy E m Defined as:

[0103]

[0104] This yields the cumulative energy ratio C. m :

[0105]

[0106] This method defines the occupied bandwidth of a signal as the range covering 99% of the total spectral energy. Therefore, the difference between the low-frequency boundary radius where the accumulated energy reaches 0.5% and the high-frequency boundary radius where the accumulated energy reaches 99.5% is used as the two-dimensional occupied bandwidth.

[0107] BW 2D =r high -r low

[0108] Where r low =min{r m |C m ≥0.005},r high =min{r m |C m ≥0.995}.

[0109] To transform the 2D-DB-VMD constrained variational problem into an unconstrained form, the augmented Lagrange equation combining the quadratic penalty term and Lagrange multiplier constraints is as follows:

[0110]

[0111] Where, λ α and λ β These are the Lagrange multipliers corresponding to α and β, respectively. The problem is solved using the alternating direction multiplier method, achieving convergence through iterative solutions.

[0112] Bayesian Optimization for Adaptive Selection of Mode Number: Building upon the theoretical framework of two-dimensional variational mode decomposition with dynamic bandwidth constraints, this invention focuses on solving the key technical problem of adaptive optimization of the number of modes. Addressing the shortcomings of existing technologies that require manual pre-setting of the number of modes and lack adaptability, this invention innovatively introduces a Bayesian optimization framework to achieve automatic selection of the optimal number of modes. The core technical challenge lies in designing an objective optimization function that can accurately evaluate the decomposition quality.

[0113] Based on the essence of signal decomposition technology, the ideal decomposition result should meet the following technical requirements: the spectra of each modal component should not overlap, ensuring the independence of the decomposition; mode overlap caused by over-decomposition should be avoided; and high-precision signal reconstruction capability should be maintained. Based on the above technical criteria, this invention comprehensively considers Jensen-Shannon (JS) divergence and reconstruction mean square error (MSE) as two key evaluation indicators as components of the objective function. JS divergence is used to quantify the spectral separation between each modal component, ensuring the independence and completeness of the decomposition; reconstruction mean square error directly reflects the fidelity of the decomposition result, ensuring the complete preservation of the original signal information. In this invention, Jensen-Shannon divergence and reconstruction mean square error (MSE) can be expressed in the patent specification as follows:

[0114] (1) Calculation of Jensen–Shannon divergence

[0115] For the k-th modal signal u k (x,y), its normalized power spectral distribution p k (n) is:

[0116]

[0117] Among them, U k (n) is u k The two-dimensional Fourier transform of (x,y), where M and N are the length and width, respectively. Let the normalized power distribution set be: P = {p1, p2, ..., p...} K Let the mixture distribution of any two elements p, q ∈ P be: The corresponding JS divergence is:

[0118]

[0119] Where: JS(p||q) is the JS divergence of the elements p and q, p(n) and q(n) are the normalized power spectral distributions of p and q respectively, and m p,q (n) represents the mixed power spectrum distribution of p and q.

[0120] As can be seen from the above formula, the formula is strictly symmetric, i.e., JS(p||q)=JS(q||p). Therefore, let the set of all distinct distribution pairs be: C={{p,q}|p,q∈P,p≠q}, then the minimum divergence is:

[0121] JS min =min({JS(p,q)|{p,q}∈C})

[0122] Where: JS(p,q) represents the set of JS divergences for all p≠q in set P, and min represents taking the minimum value in the set.

[0123] The average divergence between different modes is:

[0124] JS mean =mean({JS(p,q)|{p,q}∈C})

[0125] Where: JS(p,q) represents the set of JS divergences for all p≠q in set P, and mean represents the average value of the set.

[0126] (2) Calculation of reconstruction mean squared error (MSE)

[0127] Let the original signal be TDB. i,j (x,y), the sum of the reconstructions of each modal component is:

[0128]

[0129] The reconstruction mean square error is defined as:

[0130]

[0131] Where H and W represent the length and width of the input data, TDB i,j (x,y) represents the intensity information of PolSAR, where i and j represent different polarization modes. This represents the reconstructed features after decomposition.

[0132] By organically integrating these two complementary indicators, this invention constructs a comprehensive objective optimization function:

[0133] EvaF = γ(JS) mean +JS min )-ηMSE

[0134] Wherein, γ and η are the weighting coefficients of JS divergence and reconstruction mean square error, respectively. In this invention, they are equally important, so γ = η = 0.5.

[0135] After determining the objective function, a Bayesian optimization framework is used to efficiently search for the optimal number of modes. The core idea of ​​this method is as follows: First, a surrogate model is constructed to approximate the computationally expensive objective function; then, an acquisition function is used to determine the next most promising evaluation point, thereby approximating the global optimum with the fewest evaluations.

[0136] In selecting the surrogate model, the Gaussian Process (GP), as a powerful nonparametric model, was chosen due to its high flexibility, ability to effectively avoid overfitting, and ability to provide uncertainty estimates. For the acquisition function, this study adopted the classic Expected Improvement (EI) strategy. The EI strategy achieves an excellent balance between "exploration" (searching in regions of high uncertainty) and "exploration" (searching near the current optimal solution), thus efficiently guiding the model to converge to the global optimum.

[0137] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method of PolSAR two-dimensional dynamic bandwidth variational modal decomposition based on Bayesian parameter selection, characterized in that, The method comprises the following steps: The scattering matrix of the polarimetric SAR data is preprocessed to extract intensity information of six polarization channels; A range of the number of modes K is set; The maximum number of iterations of Bayesian optimization is initialized; An occupied bandwidth of the intensity information is calculated; An upper limit and a lower limit of the bandwidth are set, and a Lagrange multiplier and a dynamic bandwidth penalty factor are initialized; Each mode component is updated by an alternating direction multiplier method, and a two-dimensional inverse Fourier transform is performed on the mode component; The Lagrange multiplier and the dynamic bandwidth penalty factor are updated; It is judged whether a total relative change of the mode components is less than a set threshold, and if yes, the K mode components are output, otherwise, iteration is continued; A mean value and a minimum value of JS divergence of the K mode components are calculated; A reconstruction mean square error of the K mode components is calculated; A Bayesian optimization objective function is constructed, and an optimal evaluation value is updated based on a Gaussian process and an expected improvement function; When the maximum number of evaluations is reached, an optimal number of modes is output; The optimal number of modes is taken as a parameter, and a final decomposition result is obtained by two-dimensional dynamic bandwidth variational mode decomposition.

2. The method of claim 1, wherein, The range of the number of modes K is set as K∈[2, 10].

3. The method of claim 1, wherein, The calculation formula of the occupied bandwidth is: BW 2D = r high -r low where: BW 2D The calculation of the one-dimensional signal occupation bandwidth is used as a reference. First, the Fourier transform is performed on the characteristic component TDB i,j (x,y) to obtain its power spectrum: S i,j (f x ,f y ) = Γ{TDB i,j (x,y)} P i,j (f x ,f y )=|S i,j (f x ,f y )| 2 where f x and f y are two-dimensional frequency coordinates, S i,j (f x ,f y ) and P i,j (f x ,f y ) are the two-dimensional Fourier transform and power spectral density of TDB i,j (x,y), respectively; the radial distance is defined in the frequency domain plane (f x ,f y ) and its maximum value is r max , and the interval [0, r max ] is equally divided into M segments, M = max(H x ,H y ), H x and H y are the number of sampling points in the horizontal x and vertical y directions; the dividing points are denoted as: 0＝r0＜r1＜…＜r M ＝r max r m-1 ≤ r < r m ) the radial energy E m is defined as: The cumulative energy ratio C is thus obtained m : The method defines the occupied bandwidth of the signal as covering the 99% interval of the total spectrum energy, so the difference between the low frequency boundary radius where the cumulative energy reaches 0.5% and the high frequency boundary radius where the cumulative energy reaches 99.5% is taken as the two-dimensional occupied bandwidth, so r low = min{r m | C m ≥ 0.005}, r high = min{r m | C m ≥ 0.995}.

4. The method of claim 1, wherein, The bandwidth upper bound and bandwidth lower bound are: Initialize the Lagrange multiplier as and the dynamic bandwidth penalty factor as: λ α = λ β = 1 dynamic_alpha = 1 + λ α -λ β where: λ α and λ β are Lagrange multipliers for the upper and lower bandwidth bounds, respectively, and dynamic_alpha is the dynamic bandwidth.

5. The method of claim 1, wherein, The formula for updating each mode component is: wherein: is a frequency domain representation of the kth modality, is a two-dimensional Hilbert transform, is a Fourier transform of the input data, λ t (w) is a Lagrange multiplier for the data fidelity term, is a center frequency of the kth modality.

6. The method of claim 1, wherein, The formula for updating the Lagrange multiplier and the dynamic bandwidth penalty factor is: dynamic_alpha = 1 + λ α -λ β in, and For λ α and λ β The corresponding update step size is set here as follows:

7. The method of claim 1, wherein, The formula for judging the total relative change of the mode components is: wherein: and are the solution results of the kth modality at the t+1th and tth time, respectively.

8. The method of claim 1, wherein, The minimum value and the mean value of the JS divergence of the K mode components can be derived from the following formula: For the kth modality signal u k (x, y), whose normalized power spectral distribution p k (n) is: where U k (n) is u k the two-dimensional Fourier transform of (x, y), M and N are the length and width; let the set of normalized power distribution be: P = {p1, p2, …, p K For any two elements p, q ∈ P, let their mixed distribution be: The corresponding JS divergence is then: where: JS(p||q) is the JS divergence of the sought elements p and q, p(n) and q(n) are the normalized power spectral distributions of p and q, respectively, m p,q (n) is the mixed power spectral distribution of p and q; As can be seen from the above formula, the formula is strictly symmetric, that is, JS(p||q)=JS(q||p), therefore, all different distribution pairs are recorded as C={{p,q}|p,q∈Ρ,p≠q}, and the minimum divergence is: JS min = min({JS(p, q) | {p, q} ∈ C}) Wherein: JS(p,q) represents a JS divergence set of all p≠q in the set P, and min represents a minimum value in the set; The average divergence between different modes is: JS mean = mean({JS(p, q) | {p, q} e C}) Wherein: JS(p,q) represents a JS divergence set of all p≠q in the set P, and mean represents an average value in the set; The formula for calculating the reconstruction mean square error of the K mode components is: where H and W represent the length and width of the input data, TDB i,j (x, y) represents the intensity information of the PolSAR, i and j represent different polarization modes, represents the reconstructed feature after decomposition; The formula of the Bayesian optimization objective function is: EvaF = γ (JS mean + JS min ) - η MSE Wherein, γ and η are weight coefficients of the JS divergence and the reconstruction mean square error respectively, in the present application, both are equally important, therefore γ=η=0.

5.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the method of any one of claims 1-8 when executing the program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-8.

Images