CN115131274A - High-resolution remote sensing image change detection method based on multi-temporal joint decomposition - Google Patents
High-resolution remote sensing image change detection method based on multi-temporal joint decomposition Download PDFInfo
- Publication number
- CN115131274A CN115131274A CN202110805950.1A CN202110805950A CN115131274A CN 115131274 A CN115131274 A CN 115131274A CN 202110805950 A CN202110805950 A CN 202110805950A CN 115131274 A CN115131274 A CN 115131274A
- Authority
- CN
- China
- Prior art keywords
- remote sensing
- sensing image
- time phase
- temporal
- change
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
- G06T7/001—Industrial image inspection using an image reference approach
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a high-resolution remote sensing image change detection method based on multi-temporal joint decomposition, belongs to the technical field of multi-temporal remote sensing images, and solves the problems that a remote sensing image change detection method in the prior art is low in effectiveness, poor in robustness and prone to information loss. The method of the invention comprises the following steps: establishing a multi-temporal remote sensing image joint decomposition model by introducing relevant information of the previous temporal remote sensing image in the current temporal remote sensing image decomposition process; obtaining an unchanged area by utilizing the multi-temporal remote sensing image joint decomposition model; removing false alarms of the suspected change area by using a pyramid mean shift smoothing method to obtain a change area; and obtaining a change detection result according to the change area. The method is used for multi-temporal high-resolution remote sensing image change detection.
Description
Technical Field
The application relates to the technical field of multi-temporal remote sensing images, in particular to a high-resolution remote sensing image change detection method based on multi-temporal joint decomposition.
Background
With the increasing availability of multi-temporal high-resolution remote sensing images, change detection techniques have gained wide attention. However, when the spectral variability in an image is increased and the complexity of noise and a ground object background is increased, the conventional change detection method based on numerical operation, image transformation, image segmentation or post-classification often presents the problems of low effectiveness, poor robustness and the like. In addition, in the process of obtaining the difference map based on the difference value operation, an information loss phenomenon exists, and subsequent processing and application are seriously influenced. Therefore, currently, exploring a new change detection model and improving its effectiveness are still important concerns of the change detection technology.
Disclosure of Invention
The invention aims to improve the effectiveness and robustness of a remote sensing image change detection algorithm under a relatively complex background and solve the problem of information loss in the process of obtaining a difference map based on difference operation, and provides a high-resolution remote sensing image change detection method based on multi-temporal joint decomposition.
The invention is realized by the following technical scheme, and on one hand, the invention provides a high-resolution remote sensing image change detection method based on multi-temporal joint decomposition, which comprises the following steps:
acquiring a previous time phase remote sensing image and a current time phase remote sensing image;
establishing a multi-temporal remote sensing image joint decomposition model by introducing relevant information of the previous temporal remote sensing image in the current temporal remote sensing image decomposition process, wherein the relevant information comprises but is not limited to a mixing rate, a variance and a basis matrix;
obtaining the optimal solution of the base matrix and the coefficient matrix of the current time phase by using the multi-time phase remote sensing image joint decomposition model;
obtaining an unchanged area according to the optimal solution of the base matrix and the coefficient matrix of the current time phase;
obtaining a suspected change area according to the previous time phase remote sensing image, the current time phase remote sensing image and the unchanged area;
removing false alarms of the suspected change area by using a pyramid mean shift smoothing method to obtain a change area;
and obtaining a change detection result according to the change area.
Further, establishing a multi-temporal remote sensing image joint decomposition model according to the remote sensing image of the previous time phase and the remote sensing image of the current time phase, specifically:
analyzing the previous time phase remote sensing image and the current time phase remote sensing image by using Gaussian distribution to obtain the mixing rate, the variance and the basis matrix of each time phase;
constructing a likelihood function of the multi-temporal remote sensing image joint decomposition model according to the mixing ratio, the variance and the basis matrix;
obtaining a regular term of the last time phase distribution information according to the mixing rate and the variance;
acquiring the Mahalanobis distance between the current time phase subspace and the previous time phase subspace according to the basis matrix;
and obtaining the optimal solution of the mixing ratio, the variance, the base matrix and the coefficient matrix of the current time phase according to the likelihood function, the regular term and the Mahalanobis distance.
Further, the last time phase remote sensing image and the current time phase remote sensing image are analyzed by using gaussian distribution to obtain a mixing ratio, a variance and a basis matrix of each time phase, and the method specifically comprises the following steps:
the distribution form of the remote sensing image of each time phase can be expressed as
Where xt represents the image with index t in the multi-temporal image, zt is a hidden variable subject to a polynomial distribution, i and j represent the position index on the image t,
and
respectively, the mixing ratio and variance of the gaussian distributions, NGd the number of the gaussian distributions,
and
representing a base matrix and a coefficient matrix, m and n representing the width and height of the image, l representing the number of bands, s < min (m × n, l) representing the UVT rank, u i And v i The ith vector representing U and V, respectively;
according to the concept of conjugate priors, the variance and mixing ratio of each phase can be expressed as
Wherein
inv-Gamma and dir respectively represent inverse Gamma distribution and dirichlet distribution;
the basis matrix for each phase can be expressed as
Wherein
Is a positive definite matrix, and lambda is a control parameter.
Further, the constructing a likelihood function of the multi-temporal remote sensing image joint decomposition model according to the mixing ratio, the variance and the basis matrix specifically comprises:
the likelihood function of the multi-temporal remote sensing image joint decomposition model can be expressed as
L(Π t ,∑ t ,v t ,U t )＝-lnp(x t |Π t ,∑ t ,v t ,U t )
+R F (Π t ,∑ t )+R B (U t )
Wherein
Further, the obtaining the regular term of the last time phase distribution information according to the mixing ratio and the variance specifically includes:
the regular term of the last time phase distribution information can be expressed as
Further, the obtaining, according to the basis matrix, a mahalanobis distance between the current time phase subspace and the previous time phase subspace specifically includes:
the mahalanobis distance of the current-phase subspace from the last-phase subspace can be expressed as
Further, the step of obtaining a suspected change area according to the previous time phase remote sensing image, the current time phase remote sensing image and the unchanged area specifically includes:
performing difference operation on the previous time phase remote sensing image and the current time phase remote sensing image and the unchanged area respectively to obtain a first suspected change area and a second suspected change area;
and carrying out summation operation on the first suspected change area and the second suspected change area to obtain a suspected change area.
Further, the removing the false alarm of the suspected change area by using the pyramid mean shift smoothing method to obtain the change area specifically includes:
performing drift balance processing on each pixel on the suspected change image by using a pyramid mean shift smoothing method to obtain a change area, specifically:
step 2, calculating the sum of color vectors of all pixel points on the suspected change image relative to the circle center in the five-dimensional space sphere to obtain a total vector;
step 3, when the circle center is not coincident with the end point of the total vector, moving the circle center to the end point of the total vector, constructing a five-dimensional space sphere by taking the color space radius and the physical space radius as radii, and then repeating the step 2;
if the circle center is coincident with the end point of the total vector, performing step 4;
step 4, updating the color value of a certain pixel point on the suspected change image to the pixel value corresponding to the color vector end point of the circle center;
and 5, repeating the steps 1 to 4 until each pixel point on the suspected change image completes color value updating, and further obtaining the change area.
Further, the obtaining of the change detection result according to the change area specifically includes:
and obtaining a change detection result by adopting a fuzzy C mean value clustering method (FLICM) according to the change region.
In another aspect, the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the steps of the method for detecting changes in high-resolution remote sensing images based on multi-temporal joint decomposition as described above when the processor runs the computer program stored in the memory.
The invention has the beneficial effects that: according to the method, from the perspective of multi-time-phase data combined processing, a space-time correlation model is constructed by introducing the last time-phase noise correlation information and space correlation information in the current time-phase decomposition process, the model fully considers the statistical stability of an unchanged area in time phase from the perspective of statistical learning, the effectiveness and robustness of a remote sensing image change detection algorithm in a relatively complex background are further improved, and the problem that an information loss phenomenon exists in the process of obtaining a difference diagram based on difference value operation is solved.
In addition, in order to further improve the contrast between the changed region and the unchanged region and remove irrelevant texture information, a pyramid mean shift smoothing method is introduced, so that the precision of change detection is improved.
The method lays a new technical foundation for subsequent damage effect evaluation, disaster evaluation, scene target integrated dynamic monitoring and other applications, and provides technical inspiration for the applications related to the technical field.
Drawings
In order to more clearly explain the technical solution of the present application, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious to those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flow chart of a high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to the present application;
FIG. 2 is an original image of a previous time-phase remote sensing image;
FIG. 3 is an original image of a current time phase remote sensing image;
FIG. 4 is a graph of true values of the change detection results;
FIG. 5 is a diagram showing the result of PCA _ Kmeans method change detection;
FIG. 6 is a diagram of a change detection result of a Gabor wavelet hierarchical clustering method;
FIG. 7 is a diagram of the change detection result of the Gabor wavelet similarity measure method;
FIG. 8 is a diagram of variation detection results of a multi-class statistical learning method;
FIG. 9 is a diagram of the results of the collaborative segmentation method change detection;
FIG. 10 is a diagram of a change detection result of the high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to the present application;
FIG. 11 is a comparison of the results of different detection methods.
Detailed Description
The flow chart of the high-resolution remote sensing image change detection method based on multi-temporal joint decomposition provided by the invention is shown in figure 1.
The first embodiment is as follows: the method for detecting the change of the high-resolution remote sensing image based on the multi-temporal joint decomposition comprises the following specific processes:
acquiring a previous time phase remote sensing image and a current time phase remote sensing image;
establishing a multi-temporal remote sensing image joint decomposition model by introducing relevant information of the previous temporal remote sensing image in the current temporal remote sensing image decomposition process, wherein the relevant information comprises but is not limited to a mixing rate, a variance and a basis matrix;
the traditional single matrix decomposition strategy usually utilizes the information of the image itself to obtain a spatial representation matrix thereof, and the distribution form of noise needs to be assumed in the decomposition process. In general, the gaussian mixture distribution can be used as a general approximator and used in the fitting process of noise and completing the matrix decomposition task of a single image. However, the single matrix decomposition is performed on each time phase image separately, and the association of the data layers cannot be established, so that the acquisition of the change information cannot be completed. The invention is based on the point, researches a novel multi-temporal joint decomposition method, introduces a target function into a regular term of noise information of a previous time-phase remote sensing image and a regular term of subspace distribution information, wherein the noise information comprises a mixing rate and a variance of the previous time-phase remote sensing image, the subspace distribution information comprises a basis matrix and a coefficient matrix of the previous time-phase remote sensing image, and joint modeling is carried out on the time-phase information so as to obtain an unchanged area.
Obtaining the optimal solution of the base matrix and the coefficient matrix of the current time phase by using the multi-time phase remote sensing image joint decomposition model;
obtaining an unchanged area according to the optimal solution of the base matrix and the coefficient matrix of the current time phase;
obtaining a suspected change area according to the previous time phase remote sensing image, the current time phase remote sensing image and the unchanged area;
removing false alarms of the suspected change area by using a pyramid mean shift smoothing method to obtain a change area;
on the basis of modeling of a multi-temporal joint decomposition model, the method carries out differential operation on the original data of the multi-temporal remote sensing image and the decomposed spatial representation matrix (unchanged area) to obtain the suspected changed area. The suspected change pixels (pixels of the suspected change area) have a higher grayscale value than the unchanged pixels (pixels of the unchanged area). However, due to spectral variations existing in the multi-temporal data acquisition process, the accuracy of subsequent variation information extraction is seriously affected by the existence of the variations. Therefore, a spectral constraint method is designed and proposed, which makes full use of the regional spectral consistency, improves the contrast between the changed region and the unchanged region, and removes the false alarm caused by the spectral variability to a certain extent.
And obtaining a change detection result according to the change area.
Finally, a fuzzy C-means clustering method (FLICM) can be adopted to obtain the change position, obtain the change detection result and complete the change detection.
The invention establishes an association mechanism which cannot be established by a single decomposition method, provides a new technical basis for multi-temporal data space-time association, and further provides a model basis for subsequently utilizing the model to complete change detection.
Suppose that
An image representing a phase in a multi-phase sequence of images, where m and n represent the width and height of the image and l represents the number of bands. The generic matrix decomposition form of the image can be expressed as:
wherein
And
representing the basis matrix and coefficient matrix, s represents UV T The low-rank property of (a) is,
represents L P The number of the norm is calculated,
is an indication matrix. According to the above formula, each pixel can be decomposed into
Wherein x ij Is the pixel value of image x at the ith row and jth column. u. of i And v i The ith vectors representing U and V, respectively. s < min (m × n, l) denotes UV T Is determined. e.g. of the type ij Represents one sample point of the noise profile p (e). Since the gaussian mixture model has general approximation capability for arbitrary distributions, it can be used in the matrix decomposition task to approximate the noise distribution p (e). The noise distribution can be expressed as a Gaussian mixture distribution according to the expression form
Wherein
And
indicating the mixing ratio and variance of the gaussian distribution. N is a radical of Gd Indicates the number of gaussian distributions. According to the characteristics of mixed Gaussian distribution, the distribution parameters and the matrix space characterization parameters, namely the mixing ratio, the variance, the basis matrix and the coefficient matrix, can be solved by using the maximum likelihood criterion and the EM algorithm. To obtain the optimal solution, the likelihood function needs to be constructed according to the maximum likelihood rule, which can be tabulatedShown as
According to the basic principle of the single-time phase image matrix decomposition method, information such as a mixing rate, a variance, a basis matrix, a coefficient matrix and the like corresponding to the current time phase can be obtained, and the current image is subjected to matrix decomposition by using a formula (1). However, the matrix decomposition method only considers the data characteristics of the matrix decomposition method and does not consider the space-time correlation among multi-time phase data. Therefore, the object of the present invention is to construct a new matrix decomposition model under which the matrix decomposition process of each image considers the information of the previous phase in addition to the self data characteristics. This information includes distribution information and subspace information that can be estimated. To accomplish this, the idea of temporal regularization may be introduced. The temporal regularization is a process of constraining current temporal information processing by means of regularization. The information of the previous time phase is introduced into the current time phase in a regular term mode, so that the decomposition of the current time phase can be restricted, and the purpose of associating the time phase information is achieved.
The second embodiment is as follows: in this embodiment, further describing the first embodiment, the establishing a multi-temporal remote sensing image joint decomposition model according to the remote sensing image of the previous time phase and the remote sensing image of the current time phase specifically includes:
analyzing the last time phase remote sensing image and the current time phase remote sensing image by using Gaussian distribution to obtain the mixing rate, the variance and the basis matrix of each time phase, wherein the method specifically comprises the following steps:
assuming that the distribution form of each time phase satisfies the Gaussian mixture distribution, the distribution form of each time phase image can be expressed as
Wherein x is t Representing images with index t in multi-temporal images。Z t Is a hidden variable subject to a polynomial distribution. i and j represent position indices on the image t. According to the concept of conjugate priors, the variance and mixing ratio will satisfy a certain distribution form, which can be expressed as
Wherein
inv-Gamma and dir denote the inverse Gamma distribution and the dirichlet distribution, respectively. These profiles will inherently contain noise related information of the last phase image. Due to the above-described distribution properties, the subspace information may be represented as
Wherein
Is a positive definite matrix, and lambda is a control parameter.
Constructing a likelihood function of the multi-temporal remote sensing image joint decomposition model according to the mixing ratio, the variance and the basis matrix;
the mathematical form of the related information of the remote sensing image of the last time phase can be assumed to be theta t-1 The related information comprises mixing rate, variance, base matrix and coefficient matrix, and the posterior probability distribution can be expressed as Bayesian posterior probability criterion
p(Π t ,∑ t ,v t ,U t |x t ,θ t-1 )∝p(x t |Π t ,∑ t ,v t ,U t )
p(Π t |θ t-1 )p(∑ t |θ t-1 )p(U t |θ t-1 )p(v t ) (9)
Similar to the single image decomposition method, the likelihood function in the multi-temporal data joint decomposition can be expressed as
L(Π t ,∑ t ,v t ,U t )＝-lnp(x t |Π t ,∑ t ,v t ,U t )
+R F (Π t ,∑ t )+R B (U t ) (10)
Wherein
The item utilizes the data characteristic of the current time phase to ensure that the updated parameters can be suitable for the decomposition of the current time phase image.
Obtaining a regular term of the last time-phase distribution information according to the mixing rate and the variance;
the RF represents a regular term of the last time phase distribution information, and its parameters will be used to constrain the decomposition of the new time phase image in the update process, so as to implement the time phase correlation of the noise information.
Acquiring the Mahalanobis distance between the current time phase subspace and the previous time phase subspace according to the basis matrix;
RB represents the relationship of spatial statistical characteristics between multi-temporal images, is the mahalanobis distance between the current-temporal subspace and the last-temporal subspace, and is used to implement temporal correlation of subspace information.
Obtaining an optimal solution of a mixing ratio, a variance, a base matrix and a coefficient matrix of the current time phase according to the likelihood function, the regular term and the mahalanobis distance, specifically:
the EM algorithm can be used to perform parameter estimation according to the matrix decomposition principle under the constraints of likelihood functions and gaussian mixture distribution given in (10). The EM algorithm estimates the above parameters in two steps, maximizing the utilization of the iteration and calculating expectation. In solving the expectation, the expectation of the posterior probability thereof can be expressed as:
to solve the optimal solution, the partial derivatives of the objective function can be solved
The closed type solution (the optimal solution of the mixing rate, the variance, the base matrix and the coefficient matrix of the current time phase) is obtained by calculation
Wherein
The third concrete implementation mode: in this embodiment, further describing the first embodiment, the obtaining of the suspected change area according to the previous time phase remote sensing image, the current time phase remote sensing image and the unchanged area specifically includes:
performing difference operation on the previous time phase remote sensing image and the current time phase remote sensing image and the unchanged area respectively to obtain a first suspected change area and a second suspected change area;
and carrying out summation operation on the first suspected change area and the second suspected change area to obtain a suspected change area.
The fourth concrete implementation mode: in this embodiment, to further explain the first embodiment, the removing false alarm of the suspected change area by using the pyramid mean shift smoothing method to obtain the change area specifically includes:
performing drift balance processing on each pixel on the suspected change image by using a pyramid mean shift smoothing method to obtain a change area, specifically:
step 2, calculating the sum of color vectors of all pixel points on the suspected change image relative to the circle center in the five-dimensional space sphere to obtain a total vector;
step 3, when the circle center is not coincident with the end point of the total vector, moving the circle center to the end point of the total vector, constructing a five-dimensional space sphere by taking the color space radius and the physical space radius as radii, and then repeating the step 2;
if the circle center is coincident with the end point of the total vector, performing step 4;
step 4, updating the color value of a certain pixel point on the suspected change image to the pixel value corresponding to the color vector end point of the circle center;
and 5, repeating the steps 1 to 4 until each pixel point on the suspected change image completes color value updating, and further obtaining the change area.
According to the modeling and optimizing method of the multi-temporal joint decomposition model, the optimal solution of noise-related distribution parameters and subspace parameters can be obtained, and a suspected change area is obtained. It should be noted that, in the whole model, the statistical optimization may result in many outlier pixels due to the spectral variability not being accurately fitted, and the focus of the spatio-temporal regularization method is also the statistical characteristics and spatial information of the image, and the overall utilization rate of the spectral information is relatively low, so that the suspected change region may contain many false changes. To solve this problem, a pyramid mean shift smoothing method may be used to improve the accuracy of change detection. Mean shift smoothing is a method for finding local extrema in the density distribution of a set of data, and can be used in the processes of image filtering, video tracking, image segmentation, etc. The basic principle is that for a given sample, one of the samples is selected to define a circular area or a square area for a central point, the centroid of the sample in the circular area or the square area, namely the point with the maximum density, is calculated, and the iteration process is continuously executed by taking the point as the center until the point is converged finally. The method is applied to restrict the suspected change result, and the suspected change result mainly comprises two parameters of color space radius and physical space radius in the process. The constraint process comprises the following steps: firstly, selecting any point on a suspected change image as a circle center, and taking two parameters (color space radius and physical space radius) as a radius to be used as a 5-dimensional space sphere; then, in the built space sphere, the sum of the color vectors of all the sample points (i.e. all the pixel points on the suspected change image) relative to the central point (the center of the circle) is calculated to obtain a total vector, the central point of the iterative space (5-dimensional space sphere) is moved to the end point of the total vector, and the sum of the vectors of all the sample points in the spherical space is calculated again, so that iteration is carried out until the end point of the sum of the vectors in the last space sphere is the central point of the space sphere, the color value of the original point (i.e. the center of the circle selected at the beginning) is updated to the color vector of the central point, i.e. the pixel value corresponding to the end point of the color vector, and then the drift smoothing process is completed.
Finally, a drift smoothing process is completed for each pixel point on the suspected change image, so that the purpose of removing the false alarm of the suspected change area is achieved, and the accuracy of extracting the subsequent change information is improved.
The fourth concrete implementation mode: in this embodiment, further describing the first embodiment, the obtaining of the change detection result according to the change area specifically includes:
and obtaining a change detection result by adopting a fuzzy C mean value clustering method (FLICM) according to the change region, wherein the fuzzy C mean value clustering method (FLICM) is a common means for marking the change detection result.
In order to illustrate the effectiveness of the detection method, 1 group of high-resolution remote sensing data is utilized to qualitatively and quantitatively compare the detection method provided by the invention with the existing methods such as typical arithmetic operation, image transformation, image segmentation, significance analysis and the like, and the experimental result shows that the method provided by the invention has better change detection performance.
The high resolution image data was acquired by Quick Bird satellites between 2013 and 2014 at a port in the united states. The image had a spatial resolution of 2.4m, a number of bands of 4, a spectral range of 0.45-0.9 microns, and a cut-off size of 600 x 600. The data mainly comprises changes of ships, docks, oil depots, vegetation, infrastructure and the like, and analysis of the changes of the group of data can reveal part of change rules of the ports in one year. The corresponding real ground object image is obtained by manual marking, the original image of the previous time phase remote sensing image and the original image of the current time phase remote sensing image are respectively shown in fig. 2 and 3, and the true value image of the change detection result is shown in fig. 4.
The comparison method adopted in the experimental process comprises 5 methods of PCA _ Kmeans method, Gabor wavelet hierarchical clustering method, Gabor wavelet similarity measurement method, multiclass statistical learning method and collaborative segmentation method. As shown in fig. 5 to 10, the change detection results obtained under different detection methods are respectively shown,
fig. 5 is a graph of the change detection result of the PCA _ Kmeans method, as shown in fig. 5, the result contains more false alarms, and the real change position can hardly be determined, which verifies that the validity of the high-score data is limited by the pixel-based method due to the richness of the detail information in the high-score image.
Fig. 6 and 7 are a Gabor wavelet hierarchical clustering method change detection result graph and a Gabor wavelet similarity measurement method change detection result graph, respectively, which are detection results obtained based on a spatial feature analysis method. The Gabor filter is a feature extraction method widely applied in image processing, and is often used as one of spatial feature extractors in the change detection of high-score data. As can be seen from the detection results, the method effectively reduces the false alarm generated by the pixel-level method, can detect relatively accurate change positions, but the existence of spectral variability limits the effectiveness of the method on the data set.
Fig. 8 is a diagram of a variation detection result of a multi-class statistical learning method, as shown in fig. 8, different from the above-mentioned pixel level and spatial feature extraction method, the multi-class statistical learning method analyzes data characteristics from the perspective of overall data distribution, taking complexity of noise distribution into consideration, and thus the result shown in fig. 8 shows that the method has a certain robustness to noise. However, the statistical analysis strategy destroys the spatial relationship and texture features of the pixels in the original data, and the problem of limited spatial utilization rate in the processing process causes the integrity of the detection result to be further improved.
Fig. 9 is a diagram of a change detection result of the collaborative segmentation method, and as shown in fig. 9, in terms of integrity, the collaborative segmentation can obtain a change position and an accurate boundary thereof at a low false alarm. However, due to the influence of the selection of the segmentation scale, different layer scale change regions are difficult to be considered at the same time, and partial scale change is lost.
Fig. 10 is a graph of the change detection result of the high-resolution remote sensing image change detection method based on multi-temporal joint decomposition, and as shown in fig. 10, it can be found that the detection method of the present application obtains the highest detection accuracy. The main reason is that the multi-temporal joint decomposition model makes full use of the representation space information and the noise distribution information of the image of each temporal phase and avoids the problem of information loss caused by direct difference operation in the original space.
Fig. 11 quantitatively analyzes quantitative Accuracy indexes (Accuracy analysis) of different detection methods (PCA _ Kmeans, Gabor wavelet hierarchical clustering, Gabor wavelet similarity measurement, multi-class statistical learning and collaborative segmentation methods) from the perspective of Accuracy rate (precision), recall rate (recall) and F score (F-score), and shows that the detection method provided by the present invention obtains the highest detection effect. Under the condition of ensuring low false alarm rate and low omission factor, the precision (precision), recall (call) and F score (F-score) are all about 90%. The conclusion shows the effectiveness of the high-resolution remote sensing image change detection method based on multi-temporal joint decomposition from the quantitative point of view.
An embodiment of the present invention further provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor runs the computer program stored in the memory, the method performs the steps of the method for detecting changes in high-resolution remote sensing images based on multi-temporal joint decomposition.
Claims (10)
1. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition is characterized by comprising the following steps of:
acquiring a previous time phase remote sensing image and a current time phase remote sensing image;
establishing a multi-temporal remote sensing image joint decomposition model by introducing relevant information of the previous temporal remote sensing image in the current temporal remote sensing image decomposition process, wherein the relevant information comprises but is not limited to a mixing rate, a variance and a basis matrix;
obtaining the optimal solution of the base matrix and the coefficient matrix of the current time phase by using the multi-time phase remote sensing image joint decomposition model;
obtaining an unchanged area according to the optimal solution of the base matrix and the coefficient matrix of the current time phase;
obtaining a suspected change area according to the previous time phase remote sensing image, the current time phase remote sensing image and the unchanged area;
removing false alarms of the suspected change area by using a pyramid mean shift smoothing method to obtain a change area;
and obtaining a change detection result according to the change area.
2. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 1, characterized in that: and establishing a multi-temporal remote sensing image joint decomposition model according to the remote sensing image of the previous time phase and the remote sensing image of the current time phase, specifically:
analyzing the previous time phase remote sensing image and the current time phase remote sensing image by using Gaussian distribution to obtain the mixing rate, the variance and the basis matrix of each time phase;
constructing a likelihood function of the multi-temporal remote sensing image joint decomposition model according to the mixing ratio, the variance and the basis matrix;
obtaining a regular term of the last time phase distribution information according to the mixing rate and the variance;
acquiring the Mahalanobis distance between the current time phase subspace and the previous time phase subspace according to the basis matrix;
and obtaining the optimal solution of the mixing ratio, the variance, the base matrix and the coefficient matrix of the current time phase according to the likelihood function, the regular term and the Mahalanobis distance.
3. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 2, characterized in that: the method comprises the following steps of analyzing the previous time phase remote sensing image and the current time phase remote sensing image by utilizing Gaussian distribution to obtain the mixing rate, the variance and the basis matrix of each time phase, and specifically comprises the following steps:
the distribution form of the remote sensing image of each time phase can be expressed as
Wherein x t Representing the image of index t, z, in a multi-temporal image t Is a hidden variable subject to a polynomial distribution, i and j represent the indices of positions on the image t,
and
respectively representing the mixing ratio and variance of the Gaussian distribution, N Gd The number of the gaussian distributions is represented,
and
representing a base matrix and a coefficient matrix, m and n representing the width of the imageAnd height, l denotes the number of bands, s < min (m × n, l) denotes UV T Rank of (u) i And v i The ith vector representing U and V, respectively;
according to the concept of conjugate priors, the variance and mixing ratio of each phase can be expressed as
Wherein
inv-Gamma and dir respectively represent inverse Gamma distribution and dirichlet distribution;
the basis matrix for each phase can be expressed as
Wherein
Is a positive definite matrix, and lambda is a control parameter.
4. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 2, characterized in that: and constructing a likelihood function of the multi-temporal remote sensing image joint decomposition model according to the mixing ratio, the variance and the basis matrix, specifically:
the likelihood function of the multi-temporal remote sensing image joint decomposition model can be expressed as
L(Π t ,∑ t ,v t ,U t )＝-lnp(x t |Π t ,∑ t ,v t ,U t )+R F (Π t ,∑ t )+R B (U t )
Wherein
5. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 2, characterized in that: the obtaining of the regular term of the last time phase distribution information according to the mixing ratio and the variance specifically includes:
the regular term of the last time phase distribution information can be expressed as
6. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 2, characterized in that: the obtaining of the mahalanobis distance between the current time phase subspace and the previous time phase subspace according to the basis matrix specifically includes:
the mahalanobis distance of the current-phase subspace from the last-phase subspace can be expressed as
7. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 1, characterized in that: the method comprises the following steps of obtaining a suspected change area according to a previous time phase remote sensing image, a current time phase remote sensing image and an unchanged area, and specifically comprises the following steps:
performing difference operation on the previous time phase remote sensing image and the current time phase remote sensing image and the unchanged area respectively to obtain a first suspected changed area and a second suspected changed area;
and carrying out summation operation on the first suspected change area and the second suspected change area to obtain a suspected change area.
8. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 1, characterized in that: the removing the false alarm of the suspected change area by using the pyramid mean shift smoothing method to obtain the change area specifically comprises the following steps:
performing drift balance processing on each pixel on the suspected change image by using a pyramid mean shift smoothing method to obtain a change area, specifically:
step 1, constructing a five-dimensional space sphere by taking a certain pixel point on a suspected change image as a circle center and taking a color space radius and a physical space radius as radii;
step 2, calculating the sum of color vectors of all pixel points on the suspected change image relative to the circle center in the five-dimensional space sphere to obtain a total vector;
step 3, when the circle center is not coincident with the end point of the total vector, moving the circle center to the end point of the total vector, constructing a five-dimensional space sphere by taking the color space radius and the physical space radius as radii, and then repeating the step 2;
if the circle center is coincident with the end point of the total vector, performing step 4;
step 4, updating the color value of a certain pixel point on the suspected change image to the pixel value corresponding to the color vector end point of the circle center;
and 5, repeating the steps 1 to 4 until each pixel point on the suspected change image completes color value updating, and further obtaining the change area.
9. The high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to claim 1, characterized in that: the obtaining of the change detection result according to the change area specifically includes:
and obtaining a change detection result by adopting a fuzzy C mean value clustering method (FLICM) according to the change region.
10. A computer device, characterized by: the device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored by the memory, the processor executes the high-resolution remote sensing image change detection method based on multi-temporal joint decomposition according to any one of claims 1-9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110805950.1A CN115131274A (en) | 2021-07-16 | 2021-07-16 | High-resolution remote sensing image change detection method based on multi-temporal joint decomposition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110805950.1A CN115131274A (en) | 2021-07-16 | 2021-07-16 | High-resolution remote sensing image change detection method based on multi-temporal joint decomposition |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115131274A true CN115131274A (en) | 2022-09-30 |
Family
ID=83375180
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110805950.1A Pending CN115131274A (en) | 2021-07-16 | 2021-07-16 | High-resolution remote sensing image change detection method based on multi-temporal joint decomposition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115131274A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116894841A (en) * | 2023-09-08 | 2023-10-17 | 山东天鼎舟工业科技有限公司 | Visual detection method for quality of alloy shell of gearbox |
-
2021
- 2021-07-16 CN CN202110805950.1A patent/CN115131274A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116894841A (en) * | 2023-09-08 | 2023-10-17 | 山东天鼎舟工业科技有限公司 | Visual detection method for quality of alloy shell of gearbox |
| CN116894841B (en) * | 2023-09-08 | 2023-11-28 | 山东天鼎舟工业科技有限公司 | Visual detection method for quality of alloy shell of gearbox |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Touati et al. | An energy-based model encoding nonlocal pairwise pixel interactions for multisensor change detection | |
| CN109859245B (en) | Multi-target tracking method and device for video target and storage medium | |
| JP4849464B2 (en) | Computerized method of tracking objects in a frame sequence | |
| Kim et al. | Color–texture segmentation using unsupervised graph cuts | |
| CN111311647B (en) | Global-local and Kalman filtering-based target tracking method and device | |
| Dao et al. | Improving hyperspectral image segmentation by applying inverse noise weighting and outlier removal for optimal scale selection | |
| US11587249B2 (en) | Artificial intelligence (AI) system and methods for generating estimated height maps from electro-optic imagery | |
| US11238307B1 (en) | System for performing change detection within a 3D geospatial model based upon semantic change detection using deep learning and related methods | |
| Ouma et al. | Comparison of machine learning classifiers for multitemporal and multisensor mapping of urban LULC features | |
| US11747468B2 (en) | System using a priori terrain height data for interferometric synthetic aperture radar (IFSAR) phase disambiguation and related methods | |
| US11636649B2 (en) | Geospatial modeling system providing 3D geospatial model update based upon predictively registered image and related methods | |
| Zhang et al. | Saliency-driven oil tank detection based on multidimensional feature vector clustering for SAR images | |
| Tian et al. | The image segmentation based on optimized spatial feature of superpixel | |
| Okorie et al. | Region-based image registration for remote sensing imagery | |
| CN115131274A (en) | High-resolution remote sensing image change detection method based on multi-temporal joint decomposition | |
| CN113191996A (en) | Remote sensing image change detection method and device and electronic equipment thereof | |
| CN108805186B (en) | SAR image circular oil depot detection method based on multi-dimensional significant feature clustering | |
| US11816793B2 (en) | Geospatial modeling system providing 3D geospatial model update based upon iterative predictive image registration and related methods | |
| CN109409375B (en) | SAR image semantic segmentation method based on contour structure learning model | |
| Leduc | Feature space optimization prior to fuzzy image classification | |
| Maia et al. | Visual object tracking by an evolutionary self-organizing neural network | |
| Yang et al. | AAE-Dpeak-SC: A novel unsupervised clustering method for space target ISAR images based on adversarial autoencoder and density peak-spectral clustering | |
| Meng et al. | Object-oriented building extraction from high-resolution imagery based on fuzzy SVM | |
| US11302071B1 (en) | Artificial intelligence (AI) system using height seed initialization for extraction of digital elevation models (DEMs) and related methods | |
| CN116402999B (en) | SAR (synthetic aperture radar) instance segmentation method combining quantum random number and deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |