CN115082784A - Multi-temporal remote sensing image unsupervised change detection pseudo sample automatic generation method - Google Patents
Multi-temporal remote sensing image unsupervised change detection pseudo sample automatic generation method Download PDFInfo
- Publication number
- CN115082784A CN115082784A CN202210613496.4A CN202210613496A CN115082784A CN 115082784 A CN115082784 A CN 115082784A CN 202210613496 A CN202210613496 A CN 202210613496A CN 115082784 A CN115082784 A CN 115082784A
- Authority
- CN
- China
- Prior art keywords
- sample
- samples
- pseudo
- variation
- change detection
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to an automatic generation method of a pseudo sample for unsupervised change detection of a multi-temporal remote sensing image, which comprises the following steps: step 1: processing the multi-temporal image data and projecting the multi-temporal image data to a two-dimensional polar coordinate domain; step 2: generating a pseudo sample candidate region on the basis of the statistical distribution characteristics of the variation vector projection on the polar coordinate domain; and step 3: and obtaining multi-class changed and unchanged pseudo samples in the candidate region by adopting a random sample generation mode, and inputting the pseudo samples into a supervision classifier to realize binary change detection and multi-class change detection. Compared with the prior art, the method can generate the pseudo sample training sets with different confidence degrees without depending on prior information, and realizes automatic and steady unsupervised remote sensing change detection by means of a machine learning classifier.
Description
Technical Field
The invention relates to the technical field of multi-temporal remote sensing image surface change detection, in particular to an automatic generation method of a pseudo sample for unsupervised change detection of a multi-temporal remote sensing image.
Background
With the development of remote sensing technology, scholars at home and abroad have developed extensive research on the multi-temporal remote sensing image surface change detection application. The current change detection technology can be roughly divided into supervised change detection and unsupervised change detection according to whether a certain amount of prior earth surface real information is needed as a training set, wherein the unsupervised change detection can complete an automatic detection task without needing the prior information as the training set. The unsupervised change detection technology has wide application prospect under the background that the acquisition amount of remote sensing data is greatly increased and the acquisition cost of prior earth surface information is high. The existing unsupervised change detection method mainly comprises a clustering-based method, time sequence principal component analysis, iterative weighted multivariate change detection (IR-MAD), Change Vector Analysis (CVA), other special methods and the like.
Generally, the supervised change detection classifier may have better actual performance than the unsupervised change detection model, but may lack a complete change sample when it is actually applied. For such problems, the related researches have proposed the idea of generating a pseudo sample training set by using an unsupervised change detection technology and then realizing change detection by means of a machine learning classifier, such as sequence-based spectral change vector analysis (S) 2 CVA), compression change vector analysis (C) 2 VA), and the like. However, based on C 2 The VA method has a certain information loss in the compression representation process, and determining the pseudo sample generation area by only depending on a single threshold division method can cause the pseudo sample to contain incomplete change information. In the projection of the polar coordinate domain, the distribution of different variation types is relatively different, and calculating a uniform threshold value on all samples can cause the generation area of a pseudo sample of a part of variation types to be uncomfortableThe problem of matching.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an automatic generation method of a multi-temporal remote sensing image unsupervised change detection pseudo sample, which can generate pseudo sample training sets with various confidence degrees.
The purpose of the invention can be realized by the following technical scheme:
a method for automatically generating a pseudo sample for unsupervised change detection of a multi-temporal remote sensing image comprises the following steps:
step 1: processing the multi-temporal image data and projecting the multi-temporal image data to a two-dimensional polar coordinate domain;
step 2: generating a pseudo sample candidate region on the basis of the statistical distribution characteristics of the variation vector projection on the polar coordinate domain;
and step 3: and obtaining multi-class changed and unchanged pseudo samples in the candidate region by adopting a random sample generation mode, and inputting the pseudo samples into a supervision classifier to realize binary change detection and multi-class change detection.
As a preferred technical scheme, the step 1 specifically comprises the following steps:
and aiming at multi-temporal image data, sequentially carrying out image preprocessing, image registration and image subtraction to obtain a difference image, and projecting the difference image to a two-dimensional polar coordinate domain by adopting a compression change vector analysis method.
As a preferred technical solution, the step 2 comprises:
the variation intensity Mag and variation angle Dir variables obtained by compressed spectrum variation vector analysis and calculation are drawn in a plurality of fan-shaped areas in a two-dimensional polar coordinate domain, and each area represents a type of variation, specifically:
calculating threshold T on Mag data by using maximum expected EM algorithm based on Bayesian estimation 0 Divide the dataset omega into omega C And Ω NC At Ω C The Dir data is subjected to K-means clustering algorithm to obtain K clustering clusters, the whole polar coordinate domain is divided into K sector areas, and the boundary of each sector area on the Dir data is A i-1 And A i I is 1,2, …, K, when i is 1, a 0 0; when i is K, A K ＝π；
Wherein K is the number of the change types, Mag and Dir are respectively the change intensity and change angle data obtained by compressing the change vector, omega is the set of all samples, T is the number of the change types 0 Global coarse threshold, omega, obtained for EM algorithm calculation based on omega's Mag data C And Ω NC The set of changed and unchanged samples that are sorted for the threshold, respectively.
As a preferable technical solution, the step 2 further comprises:
and further dividing each sector area, and dividing three different types of areas according to the confidence coefficient of the sample label so as to ensure that the generated pseudo sample has higher reliability.
As a preferred technical solution, the specific method for dividing three different types of regions according to the confidence of the sample label in step 2 is as follows:
for the sample set omega in each sector i At its intensity variable Mag i Calculating the threshold value by using EM algorithm
Bisecting a region into variations
And do not change
Set of samples, denoted as
The Mag data of (A) is recorded as
Computing a set of samples
Frequency histogram statistical characteristics based on the determination
Rationality of distribution of inner samples:
wherein the content of the first and second substances,
and
are respectively based on
Calculating the obtained median and mode by the frequency histogram statistical data;
and
are respectively
Maximum and minimum values of;
a rationality threshold boundary representing a distribution of samples;
the peak features of the sample frequency histogram statistics in the sector area are reflected, including the sample intensity of changeUpwardly gathered location information;
for change sample set
Computing
When in use
Then, it is determined that the samples within the set are not reasonably distributed;
for a set of invariant samples
Calculating out
When in use
It is determined that the distribution of samples within the set is not reasonable.
As a preferred technical solution, the step 2 specifically comprises:
for is to
Threshold calculation based on variation intensity data, respectively
And
respectively record as
Dividing the sector polar coordinate domain of each category into non-variable confidence pseudo sample selection areas
Invariant dummy sample supplemental region
Non-candidate region
Changing dummy sample supplemental regions
And change high confidence pseudo-sample
Selecting five areas of the area, specifically:
wherein t is a dependent variable and represents a threshold to be determined; l represents a sample set
Mag range of
And
a difference of (d); n is the number of samples in the sample set;
representing a high confidence boundary in the set, ω ═ C,
ω＝NC，
N t indicating that the Mag value in the sector is t and
the number of samples within the range;
the value is a parameter for measuring the overall distribution of the threshold t and the sample set, and comprises the clustering position and the concentration degree information of the frequency histogram statistics.
As a preferred technical solution, in the step 2:
for unreasonably distributed sample sets, adaptive based on frequency distribution histogram feature calculation
The value ω ═ { C, NC }, specifically:
wherein, alpha and P C Is a constant parameter representing the sample distribution model fit;
the parameters reflecting the distribution specifically include:
N τ ＝0.75N
computing what satisfies the above constraints on the Mag histogram statistics of the set of unreasonably distributed samples
N τ Indicates the Mag value in the sample set is
Number of samples in the range.
As a preferred technical solution, in the step 2:
sample set for unreasonable distributed variation
I.e., ω ═ C, calculated
The method specifically comprises the following steps:
invariant sample set for unreasonable distributions
I.e. ω is NC, calculate
The method specifically comprises the following steps:
in that
And
respectively calculate the threshold value
And
obtaining an invariant confidence pseudo sample selection region in the sector polar coordinate domain
Dir∈[A i-1 ,A i ]Area for selecting high confidence sample
Dir∈[A i-1 ,A i ]}。
As a preferable technical solution, the step 2 further comprises:
within the sector, for a reasonably distributed set of samples
And
calculating the threshold value of the two sample candidate areas by adopting the same strategy again
For an unreasonably distributed sample set, the method is
Selecting the sample with the same number of Mag values as the samples in the high-confidence pseudo sample selection area in the interval range to form the constant or variable pseudo sample supplement sample, and calculating the constant pseudo sample supplement area
Dir∈[A i-1 ,A i ]Fill area for sum-change dummy samples
Dir∈[A i-1 ,A i ]}; the remaining part of the sector area is a non-candidate area
Dir∈[A i-1 ,A i ]}。
As a preferred technical solution, the step 3 specifically comprises:
firstly, selecting p% of samples in the region as an invariant and variant pseudo sample training set in a high-confidence pseudo sample candidate region of the whole polar coordinate region according to a random sample selection strategy, and inputting the samples into a supervision classifier for classification to obtain a preliminary binary variation detection result;
secondly, removing the invariant samples in the preliminary result, generating a pseudo sample training set by the residual samples in each high-confidence and supplementary area of the sector polar coordinate domain representing a type of variation according to the same random sampling strategy, and inputting the pseudo sample training set into a supervision classifier to obtain a more refined binary variation detection result;
and finally, randomly generating a pseudo sample training set from each sector polar coordinate field in the variation sample to perform multi-class variation detection, and finally obtaining multi-class variation detection results.
Compared with the prior art, the invention has the following beneficial effects:
effectively generating a pseudo sample training set of various confidences: the automatic generation method of the pseudo sample for the unsupervised change detection of the multi-temporal remote sensing image automatically generates a pseudo sample candidate region by adopting the statistical distribution characteristics of the change vector projection, then obtains pseudo samples with various changes and invariance in the candidate region by adopting a random sample generation mode, can generate pseudo sample training sets with different confidence degrees without depending on prior information, and realizes automatic and stable unsupervised remote sensing change detection by means of a machine learning classifier.
Drawings
FIG. 1 is a schematic flow chart of a method for automatically generating a pseudo sample for unsupervised change detection of a multi-temporal remote sensing image according to an embodiment of the present invention;
FIG. 2 is a pseudo-color composite of Landsat 5TM images in a study area according to an embodiment of the present invention;
wherein fig. 2(a) is a front time phase original image, fig. 2(b) is a rear time phase original image, fig. 2(c) is a difference image, and fig. 2(d) is a ground surface variation detection reference true value image;
FIG. 3 is a schematic diagram of a process of dividing a pseudo sample generation region of a sample in a research area in a polar coordinate domain according to an embodiment of the present invention;
wherein, fig. 3(a) is a schematic diagram of using EM to divide a threshold based on all samples Mag, fig. 3(b) is a sector area division diagram of three variation categories in polar coordinates, fig. 3(c) is a sector polar coordinate domain intensity frequency distribution histogram representing a quarry variation category, fig. 3(d) is a sector polar coordinate domain intensity frequency distribution histogram representing a fire area variation category, fig. 3(e) is a sector polar coordinate domain intensity frequency distribution histogram representing a water area variation category, and fig. 3(f) is a pseudo sample generation area division diagram in polar coordinate domain;
FIG. 4 is a graph comparing the results of change detection in a certain area of interest in an embodiment of the present invention;
wherein, fig. 4(a) is a preliminary binary change detection result diagram, fig. 4(b) is a fine binary change detection result diagram, fig. 4(c) is a multi-class change detection result diagram, and fig. 4(d) is S 2 The result chart of the multi-class change detection obtained by the CVA method is shown in FIG. 4(e) as C 2 And (3) obtaining a multi-class change detection result graph by the VA method.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.
Reference herein to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic may be included in at least one implementation of the present application. In the description of the present application, it is to be understood that the terms "upper", "lower", "left", "right", "top", "bottom", and the like, indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, are only for convenience in describing the present application and simplifying the description, and do not indicate or imply that the referred devices or elements must have a specific orientation, be constructed and operated in a specific orientation, and thus, should not be construed as limiting the present application. Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or to implicitly indicate the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. Moreover, the terms "first," "second," and the like are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are capable of operation in sequences other than those illustrated or described herein.
Example 1
Fig. 1 is a flowchart of an automatic generation method for a multi-temporal remote sensing image unsupervised change detection pseudo sample in the embodiment of the present application. The present application provides method steps as described in the examples or flowcharts, but may include more or fewer steps based on conventional or non-inventive labor. The order of steps recited in the embodiments is merely one manner of performing the steps in a multitude of orders and does not represent the only order of execution. The method should be implementable in software and/or hardware. Referring to fig. 1, the method may include:
(1) compressed change vector polar projection
The image data of a certain research area are preprocessed in front and rear time phases, then are registered and differenced to obtain a difference image, and the image data are projected to a two-dimensional polar coordinate domain by using a method of compressing a change vector.
(2) Pseudo sample candidate region construction
Calculating a threshold T on the Mag data using a Bayesian estimation based maximum Expectation (EM) algorithm 0 Divide the dataset omega into omega C And Ω NC At Ω C The Dir data is subjected to K-means clustering algorithm to obtain K clustering clusters, the whole polar coordinate domain is divided into K sector areas, and the boundary of each sector area on the Dir data is A i-1 And A i I is 1,2, …, K, note that when i is 1, a 0 0; when i is K, A K ＝π。
For the sample set omega in each sector i In Mag i Calculating thresholds on the data using EM algorithm
Splitting samples into a set of variation samples
And invariant sample set
Record as
The Mag data of (A) is recorded as
Calculating a set of samples using the following formula
Frequency histogram statistical characteristics based on the determination
The distribution of the inner samples is reasonable.
Wherein the content of the first and second substances,
and
are respectively based on
Calculating the obtained median and mode by the frequency histogram statistical data;
and
are respectively
Maximum and minimum values of;
a rationality threshold boundary representing a distribution of samples;
reflecting the peak value characteristics of the sample frequency histogram statistics in the sector area, wherein the peak value characteristics comprise the position information of the samples gathered in the direction of the variation intensity;
for change sample set
Computing
When in use
Then, the samples in the set are judged to be distributed unreasonably;
for a set of invariant samples
Computing
When in use
It is determined that the distribution of samples within the set is not reasonable.
To pair
Threshold calculation based on variation intensity data using the following formulas, respectively
And
respectively record as
Dividing each sector polar coordinate domain into non-variable confidence pseudo sample selection areas
Invariant dummy sample supplemental region
Non-candidate region
Changing dummy sample supplemental regions
And change high confidence pseudo-sample
Five areas are selected.
Wherein t is a dependent variable and represents a threshold to be determined; l represents a sample set
Mag range of
And
a difference of (d); n is the number of samples in the sample set;
representing a high confidence boundary in the set, ω ═ C,
ω＝NC，
N t indicating that the Mag value in the sector is t and
the number of samples within the range;
the value is a parameter for measuring the overall distribution of the threshold t and the sample set, and comprises the clustering position and the concentration degree information of frequency histogram statistics, so that a reasonable threshold can be effectively determined. For a well-distributed set of variation samples
For a well-distributed invariant sample set
For an unreasonably distributed sample set, the adaptive is calculated based on frequency distribution histogram features using the following formula
The value ω ═ C, NC }.
Wherein, alpha and P C Is a constant parameter representing the model fit of the sample distribution, α is 1.25, P C ＝1/3；
The parameters reflecting the distribution specifically include:
N τ ＝0.75N
computing what satisfies the above constraints on the Mag histogram statistics of the set of unreasonably distributed samples
N τ Indicates the Mag value in the sample set is
Number of samples in the range.
Sample set for unreasonable distributed variation
I.e., ω ═ C, calculate
The method specifically comprises the following steps:
invariant sample set for unreasonable distributions
I.e. ω is NC, calculate
The method specifically comprises the following steps:
in that
And
respectively calculate the threshold value
And
obtaining an invariant confidence pseudo sample selection region in the sector polar coordinate domain
Dir∈[A i-1 ,A i ]Area for selecting high confidence sample
Dir∈[A i-1 ,A i ]}。
Within the sector, for a reasonably distributed set of samples
And
calculating the threshold value of the two sample candidate areas by adopting the same strategy again
For an unreasonably distributed sample set, the method is
Selecting the sample with the same number of Mag values as the samples in the high-confidence pseudo sample selection area in the interval range to form the constant or variable pseudo sample supplement sample, and calculating the constant pseudo sample supplement area
Dir∈[A i-1 ,A i ]Fill area with changed pseudo samples
Dir∈[A i-1 ,A i ]}; the remaining part of the sector area is a non-candidate area
Dir∈[A i-1 ,A i ]}。
(3) Pseudo sample generation and change detection
And generating a constant and variable pseudo sample training set in the high-confidence pseudo sample selection area of the whole polar coordinate area according to a strategy of randomly selecting 10 percent of samples, inputting the training set into a support vector machine for classification, and obtaining a preliminary binary change detection result.
And removing the invariant samples in the preliminary result, generating a pseudo sample training set by the residual samples in each high-confidence and supplementary area representing a class of changed fan-shaped polar coordinate domain according to the same random sampling strategy, and inputting the pseudo sample training set into a support vector machine to obtain a finer binary change detection result.
And finally, randomly generating a pseudo sample training set from each sector polar coordinate domain in the variation sample to perform multi-class variation detection, and finally obtaining multi-class variation detection results.
Example 2
The experimental data adopts 30m resolution Landsat-5 satellite remote sensing data. The image size is 300 × 412 pixels. Anterior and posterior phase images were acquired at 1995-9 and 1996-7 months, respectively, as shown in fig. 2(a) and 2(b) (band 3, 2, 1 synthesis).
The differential image of the study area is shown in fig. 2(c) (band 3, 2, 1 synthesis), and the area has three types of surface variations: the detailed information of the amplification changes of the quarry, the forest fire area and the water area is shown in table 1, and the ground surface change detection reference truth value is shown in fig. 2 (d).
TABLE 1 Change detection study area Change Category data
| Categories | Number of pixels (Pixel) | Colour(s) |
| Quarry | 214 | Red wine |
| Fire area | 2414 | Green |
| Water area | 7480 | Blue (B) |
| Area of no change | 113492 | Black (black) |
The experimental results are as follows:
1. generating region partitioning results by pseudo samples
Using the original image S 2 The method of CVA projects to a two-dimensional polar coordinate domain and calculates a threshold value using the EM method based on Mag data, as shown in fig. 3 (a). Threshold value T 0 (T 0 39.7798) into a set of variant samples omega C And a set of invariant samples Ω NC And (c) clustering the Dir data of the change sample set by using a K-means method to obtain three cluster clusters, and dividing a sector polar coordinate domain representing three change categories according to the cluster clusters, as shown in fig. 3 (b). Wherein C is 1 Representing a quarry change class, C 2 Representing fire change class, C 3 Representing the class of water area variations.
FIG. 3(C), FIG. 3(d) and FIG. 3(e) are based on C 1 、C 2 And C 3 The result obtained by performing frequency histogram statistics on the Mag data of the sample in the sector polar coordinate domain is shown in a schematic diagram, and the black dotted line in the diagram represents a threshold value obtained by using an EM method on the basis of the Mag data of the sample set. At C 1 Area (quarry change class) and C 2 In the frequency histogram of the region (fire change type), because the number of changed samples is less than that of unchanged samples, obvious change sample aggregation characteristics are difficult to form, and C 3 The regions (water area change class) are opposite, and the frequency histogram of the regions shows obvious two gathered peak characteristics. Influenced by two factors of less changed sample number and less obvious aggregation characteristic, and is based on C 1 、C 2 The threshold value calculated for the samples in the region is clearly not accurate enough, which introduces a lot of errors for the subsequent selection of pseudo samples. Therefore, it is necessary to determine the reasonableness of the threshold value based on the frequency distribution histogram feature and perform classification processing. The results of the data rationality determination are shown in table 2.
Finally, as shown in fig. 3(f), the regions divided in the polar coordinate domain are restricted in the range of the dummy sample selection region in the proposed method for the classes with unreasonable distribution, thereby ensuring that a dummy sample set with high reliability can be generated.
TABLE 2 rationality judgment results
2. Change detection result
The change detection method provided by the invention can be divided into three steps in the generation of a pseudo sample training set and the classification based on a machine learning method: preliminary binary change detection, fine binary change detection, and multi-class change detection. The selection of the pseudo sample set adopts a strategy of randomly selecting samples in a region, the number of the selected samples is in proportion to the total number of the samples in the region, generally 10%, and the specific information of the pseudo sample set is shown in table 3.
TABLE 3 pseudo sample training set information
The preliminary binary change detection result is shown in fig. 4(a), and because the pseudo sample training set is generated randomly in the high confidence pseudo sample selection area, the sample set contains less information although the confidence is higher, and the training set does not relate to the part where the changed and unchanged samples are easy to be confused. The final result is that the detection omission rate of the preliminary binary change detection result is low, but the false detection is more. And removing the constant samples detected in the preliminary binary result, and randomly selecting samples from the remaining variable samples according to the same strategy to form a pseudo sample training set. Based on the fine binary result, after the invariant samples are removed, a pseudo sample training set is randomly generated in the remaining samples and multi-class variation detection is performed, and the final result is as shown in fig. 4 (b). The experimental results adopt two indexes of Overall Accuracy (OA for short) and Kappa Coefficient (K for short) for quality evaluation, and because the method has random factors, the experiment is repeated for 10 times, and the specific results are shown in table 4 (the detection result shown in fig. 4 is obtained by experiment number 1).
TABLE 4 evaluation of test results
3. The method is compared with a partial unsupervised change detection result graph for analysis
The experiment also compares the S based on the original S 2 CVA and C 2 The VA two methods divide areas to randomly generate the result of a pseudo sample training set, samples with the same number as the experiment are randomly selected as the training set, 10 groups of experiments are carried out, and the quality evaluation result is shown in a table 5.
TABLE 5 evaluation of comparative experiment results
The detection result obtained by the method of the present embodiment is shown in FIG. 4(c), which uses the original S 2 CVA method obtainingThe results of detection are shown in FIG. 4(d), using C 2 The results of the detection obtained by the VA method are shown in FIG. 4 (e). Original S 2 CVA method and C 2 In the VA method, a processing method of integrally dividing the threshold is adopted, and the threshold calculated for the data distribution type with less sample amount, non-centralized distribution in a polar coordinate domain and serious confusion of the changed and unchanged samples is unreasonable, so that a large number of unchanged pixels can be wrongly divided, such as the variation class of a quarry. In the invention, a self-adaptive pseudo sample candidate region division method based on the statistical characteristics of the frequency histogram of the projection domain of the change vector is used, and a high-confidence pseudo sample selection region, a supplementary selection region and a non-candidate region with serious confusion are more finely divided from the self distribution of data, so that a high-quality pseudo sample training set can be effectively obtained. The experiment comparison analysis proves that the method has higher overall precision in change detection application, and particularly has obvious advantages in distinguishing small sample amount change classes and inhibiting the influence of error samples on a pseudo sample training set.
While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications and substitutions can be easily made by those skilled in the art within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A method for automatically generating a pseudo sample for unsupervised change detection of a multi-temporal remote sensing image is characterized by comprising the following steps of:
step 1: processing the multi-temporal image data and projecting the multi-temporal image data to a two-dimensional polar coordinate domain;
step 2: generating a pseudo sample candidate region on the basis of the statistical distribution characteristics of the variation vector projection on the polar coordinate domain;
and step 3: and obtaining multi-class changed and unchanged pseudo samples in the candidate region by adopting a random sample generation mode, and inputting the pseudo samples into a supervision classifier to realize binary change detection and multi-class change detection.
2. The method for automatically generating the pseudo sample for the unsupervised change detection of the multi-temporal remote sensing image according to claim 1, wherein the step 1 specifically comprises:
aiming at multi-temporal image data, a difference image is obtained by sequentially carrying out image preprocessing, image registration and image subtraction, and is projected to a two-dimensional polar coordinate domain by adopting a compression change vector analysis method.
3. The method for automatically generating the unsupervised change detection pseudo sample of the multi-temporal remote sensing image according to claim 1, wherein the step 2 comprises:
the variation intensity Mag and variation angle Dir variables obtained by compressed spectrum variation vector analysis and calculation are drawn in a plurality of fan-shaped areas in a two-dimensional polar coordinate domain, and each area represents a type of variation, specifically:
calculating threshold T on Mag data by using maximum expected EM algorithm based on Bayesian estimation 0 Divide the dataset omega into omega C And Ω NC At Ω C The Dir data is subjected to K-means clustering algorithm to obtain K clustering clusters, the whole polar coordinate domain is divided into K sector areas, and the boundary of each sector area on the Dir data is A i-1 And A i I is 1,2, …, K, when i is 1, a 0 0; when i is K, A K ＝π；
Wherein K is the number of the variation types, Mag and Dir are variation intensity and variation angle data obtained by compressing the variation vector, omega is the set of all samples, T is the number of the variation types 0 Global coarse threshold, omega, obtained for EM algorithm calculation based on omega's Mag data C And Ω NC The set of changed and unchanged samples assigned to the threshold, respectively.
4. The method for automatically generating the unsupervised change detection pseudo sample of the multi-temporal remote sensing image according to claim 3, wherein the step 2 further comprises:
and further dividing each sector area, and dividing three different types of areas according to the confidence coefficient of the sample label so as to ensure that the generated pseudo sample has higher reliability.
5. The method for automatically generating the pseudo sample for the unsupervised change detection of the multi-temporal remote sensing image according to claim 4, wherein the specific method for dividing three different types of regions according to the confidence of the sample label in the step 2 comprises the following steps:
for the sample set omega in each sector i In its intensity variable Mag i Calculating the threshold value T by using EM algorithm 1 i Bisecting a region into variations
And do not change
Set of samples, recorded as
The Mag data of (A) is recorded as
Computing a set of samples
Frequency histogram statistical characteristics based on the determination
Rationality of distribution of inner samples:
wherein the content of the first and second substances,
and
are respectively based on
Calculating the obtained median and mode by the frequency histogram statistical data;
and
are respectively
Maximum and minimum values of;
a rationality threshold boundary representing a distribution of samples;
reflecting the peak value characteristics of the sample frequency histogram statistics in the sector area, wherein the peak value characteristics comprise the position information of the samples gathered in the direction of the variation intensity;
for change sample set
Computing
When in use
Then, it is determined that the samples within the set are not reasonably distributed;
for a set of invariant samples
Computing
When in use
It is determined that the distribution of samples within the set is not reasonable.
6. The method for automatically generating the pseudo sample for the unsupervised change detection of the multi-temporal remote sensing image according to claim 5, wherein the step 2 specifically comprises:
for is to
Threshold calculation based on variation intensity data, respectively
And
respectively record as
Will each beOne category of sector polar coordinate domain is divided into non-variable confidence pseudo sample selection areas
Invariant dummy sample supplemental region
Non-candidate region
Changing dummy sample supplemental regions
And change high confidence pseudo-sample
Selecting five areas of the area, specifically:
wherein t is a dependent variable and represents a threshold to be determined; l represents a sample set
Mag range of
And with
A difference of (d); n is the number of samples in the sample set;
representing a high confidence boundary in the set, ω ═ C,
ω＝NC，
N t indicating that the Mag value in the sector is t and
the number of samples within the range;
the value is a parameter for measuring the overall distribution of the threshold t and the sample set, and comprises the clustering position and the concentration degree information of the frequency histogram statistics.
7. The method for automatically generating the unsupervised change detection pseudo sample of the multi-temporal remote sensing image according to claim 6, wherein in the step 2:
for unreasonably distributed sample sets, adaptive based on frequency distribution histogram feature calculation
The value ω ═ { C, NC }, specifically:
wherein, alpha and P C Is a constant parameter representing the sample distribution model fit;
the parameters reflecting the distribution specifically include:
N τ ＝0.75N
computing what satisfies the above constraints on the Mag histogram statistics of the set of unreasonably distributed samples
N τ Indicates the Mag value in the sample set is
Number of samples in the range.
8. The method for automatically generating the pseudo sample for unsupervised change detection of the multi-temporal remote sensing image according to claim 7, wherein in the step 2:
sample set for unreasonable distributed variation
I.e., ω ═ C, calculated
The method specifically comprises the following steps:
invariant sample set for unreasonable distributions
I.e. ω is NC, calculate
The method comprises the following specific steps:
in that
And
respectively calculate the threshold value
And
obtaining an invariant confidence pseudo sample selection region in the sector polar coordinate domain
And varying high confidence pseudo sample selection areas
9. The method for automatically generating the unsupervised change detection pseudo sample of the multi-temporal remote sensing image according to claim 8, wherein the step 2 further comprises:
within the sector, for a reasonably distributed set of samples
And
calculating the threshold value of the two sample candidate areas by adopting the same strategy again
For an unreasonably distributed sample set, the method is
Selecting the sample with the same number of Mag values as the samples in the high-confidence pseudo sample selection area in the interval range to form the constant or variable pseudo sample supplement sample, and calculating the constant pseudo sample supplement area
And a change dummy sample supplement region
The remaining part of the sector area is a non-candidate area
10. The method for automatically generating the pseudo sample for the unsupervised change detection of the multi-temporal remote sensing image according to claim 1, wherein the step 3 specifically comprises:
firstly, selecting p% of samples in the region as an invariant and variant pseudo sample training set in a high-confidence pseudo sample candidate region of the whole polar coordinate region according to a random sample selection strategy, and inputting the samples into a supervision classifier for classification to obtain a preliminary binary variation detection result;
secondly, removing the invariant samples in the preliminary result, generating a pseudo sample training set by the residual samples in each high-confidence and supplementary area of the sector polar coordinate domain representing a type of variation according to the same random sampling strategy, and inputting the pseudo sample training set into a supervision classifier to obtain a more refined binary variation detection result;
and finally, randomly generating a pseudo sample training set from each sector polar coordinate domain in the variation sample to perform multi-class variation detection, and finally obtaining multi-class variation detection results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210613496.4A CN115082784A (en) | 2022-05-31 | 2022-05-31 | Multi-temporal remote sensing image unsupervised change detection pseudo sample automatic generation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210613496.4A CN115082784A (en) | 2022-05-31 | 2022-05-31 | Multi-temporal remote sensing image unsupervised change detection pseudo sample automatic generation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115082784A true CN115082784A (en) | 2022-09-20 |
Family
ID=83250266
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210613496.4A Pending CN115082784A (en) | 2022-05-31 | 2022-05-31 | Multi-temporal remote sensing image unsupervised change detection pseudo sample automatic generation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115082784A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115795255A (en) * | 2022-09-21 | 2023-03-14 | 深圳大学 | Method, device, medium and terminal for detecting time series change of wetland |
-
2022
- 2022-05-31 CN CN202210613496.4A patent/CN115082784A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115795255A (en) * | 2022-09-21 | 2023-03-14 | 深圳大学 | Method, device, medium and terminal for detecting time series change of wetland |
| CN115795255B (en) * | 2022-09-21 | 2024-03-26 | 深圳大学 | Method, device, medium and terminal for detecting time sequence change of wetland |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10891511B1 (en) | Human hairstyle generation method based on multi-feature retrieval and deformation | |
| CN109522908B (en) | Image significance detection method based on region label fusion | |
| CN108765465B (en) | Unsupervised SAR image change detection method | |
| CN107229917B (en) | A kind of several remote sensing image general character well-marked target detection methods based on iteration cluster | |
| CN107330875B (en) | Water body surrounding environment change detection method based on forward and reverse heterogeneity of remote sensing image | |
| CN107230188B (en) | Method for eliminating video motion shadow | |
| CN103914705B (en) | Hyperspectral image classification and wave band selection method based on multi-target immune cloning | |
| CN105740915B (en) | A kind of collaboration dividing method merging perception information | |
| CN106971158B (en) | A kind of pedestrian detection method based on CoLBP symbiosis feature Yu GSS feature | |
| CN102360503B (en) | SAR (Specific Absorption Rate) image change detection method based on space approach degree and pixel similarity | |
| CN109409438B (en) | Remote sensing image classification method based on IFCM clustering and variational inference | |
| CN104657980A (en) | Improved multi-channel image partitioning algorithm based on Meanshift | |
| CN111009005A (en) | Scene classification point cloud rough registration method combining geometric information and photometric information | |
| CN108846845B (en) | SAR image segmentation method based on thumbnail and hierarchical fuzzy clustering | |
| CN110458812A (en) | A kind of similar round fruit defects detection method based on color description and sparse expression | |
| CN117576079A (en) | Industrial product surface abnormality detection method, device and system | |
| CN115690086A (en) | Object-based high-resolution remote sensing image change detection method and system | |
| CN102081740B (en) | 3D image classification method based on scale invariant features | |
| CN110310263B (en) | SAR image residential area detection method based on significance analysis and background prior | |
| CN115311449A (en) | Weak supervision image target positioning analysis system based on class reactivation mapping chart | |
| CN105787505A (en) | Infrared image clustering segmentation method combining sparse coding and spatial constraints | |
| CN115082784A (en) | Multi-temporal remote sensing image unsupervised change detection pseudo sample automatic generation method | |
| CN107346549B (en) | Multi-class change dynamic threshold detection method utilizing multiple features of remote sensing image | |
| CN113657216A (en) | Method for separating tree crown and wood point of tree in point cloud scene based on shape characteristics | |
| CN109508674A (en) | Airborne lower view isomery image matching method based on region division |
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 |