An Automatic Method for Generating Pseudo Samples for Unsupervised Change Detection in Multi-Temporal Remote Sensing Images

Abstract

The present invention relates to a method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images, comprising: step 1: processing multi-temporal image data and projecting it into a two-dimensional polar coordinate domain; step 2: generating pseudo sample candidate regions based on statistical distribution characteristics of change vector projections in the polar coordinate domain; step 3: obtaining multiple types of pseudo samples with changes and no changes in the candidate regions by using a random sample generation method, and inputting them into a supervised classifier to realize binary change detection and multi-class change detection. Compared with the prior art, the present invention can generate pseudo sample training sets with different confidence levels without relying on prior information, and realizes automated and robust unsupervised remote sensing change detection with the help of a machine learning classifier.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-temporal remote sensing image surface change detection, and in particular to a method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images. Background Art

[0002] With the development of remote sensing technology, domestic and foreign scholars have conducted extensive research on the application of multi-temporal remote sensing image surface change detection. The current change detection technology can be roughly divided into supervised change detection and unsupervised change detection based on whether a certain amount of prior surface real information is required as a training set. Unsupervised change detection can complete the automatic detection task without prior information as a training set. Thanks to this, unsupervised change detection technology has broad application prospects in the context of a substantial increase in the amount of remote sensing data acquisition and the high cost of acquiring prior surface information. Existing unsupervised change detection methods mainly include clustering-based methods, time series principal component analysis, iterative weighted multivariate change detection (IR-MAD), change vector analysis (CVA) and other special methods.

[0003] Generally speaking, supervised change detection classifiers perform better than unsupervised change detection models, but they may lack complete change samples in actual applications. To address this problem, previous studies have proposed using unsupervised change detection technology to generate pseudo sample training sets, and then using machine learning classifiers to implement change detection, such as the sequential spectral change vector analysis (S 2 CVA), compressed change vector analysis (C 2 VA) etc. However, based on C 2 The VA method has a certain amount of information loss in the compression representation process. Relying only on a single threshold division method to determine the pseudo sample generation area will result in incomplete pseudo sample change information. Moreover, in the projection of the polar coordinate domain, the distribution of different change categories is also quite different. Calculating a unified threshold on all samples will also cause the pseudo sample generation area of ​​some change categories to be incompatible. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images, which can generate pseudo sample training sets with various confidence levels.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images, the method comprising:

[0007] Step 1: Process the multi-temporal image data and project it into a two-dimensional polar coordinate domain;

[0008] Step 2: Generate pseudo sample candidate regions based on the statistical distribution characteristics of the change vector projection in the polar coordinate domain;

[0009] Step 3: Use random sample generation to obtain multiple types of pseudo samples with changes and no changes in the candidate area, and input them into the supervised classifier to achieve binary change detection and multi-class change detection.

[0010] As a preferred technical solution, the step 1 is specifically as follows:

[0011] For multi-temporal image data, differential images are obtained through image preprocessing, image registration, and image subtraction, and the compressed change vector analysis method is used to project them into the two-dimensional polar coordinate domain.

[0012] As a preferred technical solution, the step 2 comprises:

[0013] The change intensity Mag and change angle Dir variables calculated based on the compressed spectrum change vector analysis are plotted in multiple fan-shaped areas in the two-dimensional polar coordinate domain. Each area represents a type of change, specifically:

[0014] The threshold T is calculated using the maximum expectation EM algorithm based on Bayesian estimation on the Mag data. 0 , divide the data set Ω into Ω C and Ω NC , in Ω C The K-means clustering algorithm is used on the Dir data to obtain K clusters, and the entire polar coordinate domain is divided into K sector areas. The boundary of each sector area on the Dir data is A i-1 and A i , i=1,2,…,K, when i=1, A 0 =0; when i=K, A K =π;

[0015] Among them, K is the number of change categories, Mag and Dir are the change intensity and change angle data obtained after the change vector is compressed, Ω is the set of all samples, and T 0 is the global rough threshold calculated by the EM algorithm based on the Mag data of Ω, Ω C and Ω NC are the changed and unchanged sample sets separated by the threshold respectively.

[0016] As a preferred technical solution, the step 2 further comprises:

[0017] Each sector area is further divided into three different types of areas according to the confidence of the sample label to ensure that the generated pseudo samples have high reliability.

[0018] As a preferred technical solution, the specific method of dividing the three different types of areas according to the sample label confidence in step 2 is:

[0019] For each sample set Ω in the fan-shaped area i , in which the intensity variable Mag i The threshold is calculated using the EM algorithm Divide the region into two parts and no change The sample set is denoted as The Mag data is recorded as

[0020] Calculate the sample set Frequency histogram statistical characteristics, based on this judgment Reasonable distribution of internal samples:

[0021]

[0022] ω=C

[0023] ω=NC

[0024] in, and Based on The median and mode obtained by frequency histogram statistics calculation; and They are The maximum and minimum values ​​of ; Represents the reasonableness threshold boundary of the sample distribution; It reflects the peak characteristics of the sample frequency histogram statistics in the fan-shaped area, including the location information of the sample clustering in the direction of the changing intensity;

[0025] For the set of changing samples calculate when When , it is determined that the sample distribution in the set is unreasonable;

[0026] For the invariant sample set calculate when , it is determined that the sample distribution in the set is unreasonable.

[0027] As a preferred technical solution, the step 2 is specifically as follows:

[0028] right The threshold calculation based on the change intensity data is and Recorded as Divide the sector polar coordinate domain of each category into unchanged high confidence pseudo sample selection areas Unchanged pseudo sample supplement area Non-candidate area Change pseudo sample supplement area and variation of high confidence pseudo samples Select five areas, specifically:

[0029]

[0030]

[0031] Among them, t is the dependent variable, which represents the threshold to be determined; L represents the sample set The Mag range is and The difference between ; N is the number of samples in the sample set; represents the high confidence boundary in the set, ω=C, ω=NC, N t Indicates that the Mag value in the sector area is between t and The number of samples in the range; The value is a parameter that measures the threshold t and the overall distribution of the sample set, including the clustering position and concentration information of the frequency histogram statistics.

[0032] As a preferred technical solution, in step 2:

[0033] For sample sets with unreasonable distribution, adaptive Value, ω={C,NC}, specifically:

[0034]

[0035] Among them, α and P C is a constant parameter representing the sample distribution model fit; is a parameter that reflects the distribution, specifically:

[0036]

[0037] N τ =0.75N

[0038]

[0039] The Mag histogram statistics of the sample set with unreasonable distribution are calculated to meet the above constraints. N τ Indicates that the Mag value in the sample set is The number of samples in the range.

[0040] As a preferred technical solution, in step 2:

[0041] For a set of changing samples with unreasonable distribution That is, ω=C, calculate Specifically:

[0042]

[0043]

[0044] For an invariant sample set with unreasonable distribution That is, ω=NC, calculate Specifically:

[0045]

[0046]

[0047] exist and The thresholds are calculated separately and Obtain an unchanged high confidence pseudo sample selection area in this fan-shaped polar coordinate domain Dir∈[A i-1 ,A i ]} and change the high confidence pseudo sample selection area Dir∈[A i-1 ,A i ]}.

[0048] As a preferred technical solution, the step 2 further comprises:

[0049] In the fan-shaped area, for a sample set with reasonable distribution, and The same strategy is used to calculate the thresholds for the two sample candidate areas. For samples with unreasonable distribution, Select the samples with the smallest or largest Mag value within the interval of the high confidence pseudo sample selection area with the same number of samples as the sample number in the high confidence pseudo sample selection area to form the unchanged or changed pseudo sample supplementary samples, and calculate the unchanged pseudo sample supplementary area Dir∈[A i-1 ,A i ]} and change pseudo sample supplement area Dir∈[A i-1 ,A i ]}; The remaining part of the fan-shaped area is the non-candidate area Dir∈[A i-1,A i ]}.

[0050] As a preferred technical solution, the step 3 is specifically as follows:

[0051] Firstly, in the high-confidence pseudo sample candidate area of ​​the entire polar coordinate domain, according to the random sampling strategy, p% of the samples in the area are selected as the unchanged and changed pseudo sample training sets, and input into the supervised classifier for classification to obtain the preliminary binary change detection results;

[0052] Secondly, the unchanged samples in the preliminary results are removed, and the remaining samples are used to generate pseudo sample training sets in high confidence and supplementary areas of each sector polar coordinate domain representing a type of change according to the same random sampling strategy, and then input into the supervised classifier to obtain more refined binary change detection results;

[0053] Finally, in the change samples, pseudo sample training sets are randomly generated from each sector polar coordinate domain for multi-class change detection, and finally multi-class change detection results are obtained.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] Effectively generate pseudo sample training sets with various confidence levels: The method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images in the present invention uses the statistical distribution characteristics of change vector projection to automatically generate pseudo sample candidate areas, and then uses random sample generation to obtain multiple types of changed and unchanged pseudo samples in the candidate areas. Pseudo sample training sets with different confidence levels can be generated without relying on prior information, and automated and robust unsupervised remote sensing change detection can be achieved with the help of machine learning classifiers. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of the process of the method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images in an embodiment of the present invention;

[0057] Figure 2 It is a false color composite image of Landsat 5TM image in a certain study area in an embodiment of the present invention;

[0058] in Figure 2 (a) is the original image of the previous phase. Figure 2 (b) is the original image at the later stage. Figure 2 (c) is the difference image. Figure 2 (d) is the ground surface change detection reference truth map;

[0059] Figure 3 It is a schematic diagram of the pseudo sample generation area division process in the polar coordinate domain of a certain study area sample in an embodiment of the present invention;

[0060] in, Figure 3 (a) is a schematic diagram of using EM to divide the threshold based on all sample Mags. Figure 3 (b) is the fan-shaped area division diagram of the three change categories in polar coordinates. Figure 3 (c) is the intensity frequency distribution histogram of the fan-shaped polar coordinate domain representing the quarry change categories. Figure 3 (d) is a histogram of intensity frequency distribution in the sector polar coordinate domain representing the fire area change category. Figure 3 (e) is a histogram of the intensity frequency distribution of the fan-shaped polar coordinate domain representing the water area change category. Figure 3 (f) is the pseudo sample generation area division diagram in the polar coordinate domain;

[0061] Figure 4 This is a comparison diagram of the change detection results of a certain study area in an embodiment of the present invention;

[0062] in, Figure 4 (a) is the preliminary binary change detection result diagram. Figure 4 (b) is the result of fine binary change detection. Figure 4 (c) is the result diagram of multi-class change detection. Figure 4 (d) is S 2 Multi-class change detection result diagram obtained by CVA method, Figure 4 (e) is C 2 Figure 2. Multi-class change detection results obtained by the VA method. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0064] The term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present application. In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "upper", "lower", "left", "right", "top", "bottom", etc. is based on the orientation or positional relationship shown in the accompanying drawings, which is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present application. In addition, the terms "first" and "second" are used only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" may include one or more of the features explicitly or implicitly. Moreover, the terms "first", "second", etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here.

[0065] Example 1

[0066] Figure 1 This is a flowchart of a method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images provided in an embodiment of the present application. This application provides method operation steps as described in the embodiment or flowchart, but based on routine or non-creative labor, more or fewer operation steps may be included. The order of steps listed in the embodiment is only one of many step execution orders and does not represent the only execution order. The method should be implementable by software and / or hardware. Please refer to Figure 1 , the method may include:

[0067] (1) Compressed change vector polar coordinate projection

[0068] After preprocessing the image data of a certain study area before and after the time phase, the differential image is obtained through registration and difference calculation, and the image data is projected into the two-dimensional polar coordinate domain using the method of compressed change vector.

[0069] (2) Construction of pseudo sample candidate regions

[0070] The threshold T is calculated using the maximum expectation (EM) algorithm based on Bayesian estimation on the Mag data. 0 , divide the data set Ω into Ω C and Ω NC , in Ω C The K-means clustering algorithm is used on the Dir data to obtain K clusters, and the entire polar coordinate domain is divided into K sector areas. The boundary of each sector area on the Dir data is A i-1 and Ai , i=1,2,…,K, note that when i=1, A 0 =0; when i=K, A K =π.

[0071] For each sample set Ω in the fan-shaped area i , in Mag i The EM algorithm is used to calculate the threshold on the data Divide the samples into a set of changing samples and the unchanged sample set Recorded as The Mag data is recorded as Use the following formula to calculate the sample set Frequency histogram statistical characteristics, based on this judgment The distribution of internal samples is reasonable.

[0072]

[0073] ω=C

[0074] ω=NC

[0075] in, and Based on The median and mode obtained by frequency histogram statistics calculation; and They are The maximum and minimum values ​​of ; Represents the reasonableness threshold boundary of the sample distribution; It reflects the peak characteristics of the sample frequency histogram statistics in the fan-shaped area, including the location information of the sample clustering in the direction of the changing intensity;

[0076] For the set of changing samples calculate when When , it is determined that the sample distribution in the set is unreasonable;

[0077] For the invariant sample set calculate when , it is determined that the sample distribution in the set is unreasonable.

[0078] right The following formulas are used to calculate the threshold value based on the change intensity data. and Recorded as Divide each sector polar coordinate domain into unchanged high confidence pseudo sample selection areas Unchanged pseudo sample supplement area Non-candidate area Change pseudo sample supplement area and variation of high confidence pseudo samples Select Region Five Regions.

[0079]

[0080]

[0081] Among them, t is the dependent variable, which represents the threshold to be determined; L represents the sample set The Mag range is and The difference between ; N is the number of samples in the sample set; represents the high confidence boundary in the set, ω=C, ω=NC, N t Indicates that the Mag value in the sector area is between t and The number of samples in the range; The value is a parameter that measures the threshold t and the overall distribution of the sample set, including the clustering position and concentration information of the frequency histogram statistics, which can effectively determine a reasonable threshold. For a reasonably distributed set of invariant samples For sample sets with unreasonable distribution, the following formula is used to calculate the adaptive Value ω={C,NC}.

[0082]

[0083] Among them, α and P C is a constant parameter representing the sample distribution model fitting, α = 1.25, P C =1 / 3; is a parameter that reflects the distribution, specifically:

[0084]

[0085] N τ =0.75N

[0086]

[0087] The Mag histogram statistics of the sample set with unreasonable distribution are calculated to meet the above constraints. N τ Indicates that the Mag value in the sample set is The number of samples in the range.

[0088] For a set of changing samples with unreasonable distribution That is, ω=C, calculate Specifically:

[0089]

[0090]

[0091] For an invariant sample set with unreasonable distribution That is, ω=NC, calculate Specifically:

[0092]

[0093]

[0094] exist and The thresholds are calculated separately and Obtain an unchanged high confidence pseudo sample selection area in this fan-shaped polar coordinate domain Dir∈[A i-1 ,A i ]} and change the high confidence pseudo sample selection area Dir∈[A i-1 ,A i ]}.

[0095] In the fan-shaped area, for a sample set with reasonable distribution, and The same strategy is used to calculate the thresholds for the two sample candidate areas. For samples with unreasonable distribution, Select the samples with the smallest or largest Mag value within the interval of the high confidence pseudo sample selection area with the same number of samples as the sample number in the high confidence pseudo sample selection area to form the unchanged or changed pseudo sample supplementary samples, and calculate the unchanged pseudo sample supplementary area Dir∈[A i-1 ,A i ]} and change pseudo sample supplement area Dir∈[A i-1 ,A i ]}; The remaining part of the fan-shaped area is the non-candidate area Dir∈[A i-1 ,A i ]}.

[0096] (3) Pseudo sample generation and change detection

[0097] In the high confidence pseudo sample selection area of ​​the entire polar coordinate domain, a 10% random sampling strategy is used to generate unchanged and changed pseudo sample training sets, which are input into the support vector machine for classification to obtain preliminary binary change detection results.

[0098] The unchanged samples in the preliminary results are removed, and the remaining samples are used to generate pseudo sample training sets in high confidence and supplementary areas of each sector polar coordinate domain representing a type of change according to the same random sampling strategy, and input into the support vector machine to obtain more refined binary change detection results.

[0099] Finally, in the change samples, pseudo sample training sets are randomly generated from each sector polar coordinate domain for multi-class change detection, and finally multi-class change detection results are obtained.

[0100] Example 2

[0101] The experimental data uses 30m resolution Landsat-5 satellite remote sensing data. The image size is 300×412 pixels. The before and after phase images were acquired in September 1995 and July 1996, respectively. Figure 2 (a) and Figure 2 As shown in (b) (bands 3, 2, and 1 are synthesized).

[0102] The differential image of the study area is as follows Figure 2 As shown in (c) (band 3, 2, 1 synthesis), there are three types of surface changes in this area: quarry, forest fire area and water amplification change. The detailed information is shown in Table 1. The reference value of surface change detection is as follows: Figure 2 (d) as shown.

[0103] Table 1 Change category data of the change detection study area

[0104] category Number of pixels (pixels) color quarry 214 red Fire area 2414 green Waters 7480 blue No change area 113492 black

[0105] Experimental results:

[0106] 1. Pseudo sample generation area division results

[0107] Use S 2 The CVA method is projected into a two-dimensional polar coordinate domain, and the threshold is calculated based on the Mag data using the EM method, such as Figure 3 (a) shows the threshold T 0 (T 0 =39.7798) divides the samples into a set of changing samples Ω C and the invariant sample set Ω NC , K-means method is used to cluster the Dir data of the change sample set to obtain three clusters, and the fan-shaped polar coordinate domains representing the three change categories are divided according to the clusters, as shown in Figure 3 (b) is shown. Where C1 represents the quarry change class, C 2 Represents the fire change class, C 3 Represents water area change category.

[0108] Figure 3 (c) Figure 3 (d) and Figure 3 (e) are based on C 1 , C 2 and C 3 Schematic diagram of the results obtained by performing frequency histogram statistics on the Mag data of samples in the sector polar coordinate domain. The black dotted line in the figure represents the threshold value obtained using the EM method based on the Mag data of the sample set. 1 Area (quarry change class) and C 2 In the frequency histogram of the area (fire change class), the number of change samples is less than that of unchanged samples, so it is difficult to form obvious change sample aggregation characteristics. 3 The frequency histogram of the area (water change category) shows two obvious peak characteristics of clustering. Affected by the small number of change samples and the lack of obvious clustering characteristics, the C 1 , C 2 The threshold calculated for samples in the region is obviously not accurate enough, which will bring a lot of errors to the subsequent pseudo sample selection. Therefore, it is necessary to judge the rationality of the threshold based on the frequency distribution histogram characteristics and process them by type. The results of the rationality judgment of the data are shown in Table 2.

[0109] The final area divided in the polar coordinate domain is as follows Figure 3 As shown in (f), for categories with unreasonable distribution, the proposed method suppresses the range of pseudo sample selection areas to ensure that a high-confidence pseudo sample set can be generated.

[0110] Table 2 Reasonableness judgment results

[0111]

[0112] 2. Change Detection Results

[0113] The change detection method proposed in the present invention can be divided into three steps in the generation of pseudo sample training sets and classification based on machine learning methods: preliminary binary change detection, fine binary change detection and multi-class change detection. The selection of pseudo sample sets adopts the strategy of randomly selecting samples in the region. The number of selected samples is proportional to the total number of samples in the region, generally 10%. The specific information of the pseudo sample set is shown in Table 3.

[0114] Table 3 Pseudo sample training set information

[0115]

[0116] Preliminary binary change detection results are shown in Figure 4 (a), because the pseudo sample training set is randomly generated in the high-confidence pseudo sample selection area, although the sample set has a high confidence, it contains less information. The training set does not involve the parts that are easily confused with the changing and unchanged samples. The final result is that the missed detection rate of the preliminary binary change detection results is very low, but there are many false detections. Remove the unchanged samples detected in the preliminary binary results, and randomly select samples from the remaining changing samples according to the same strategy to form a pseudo sample training set. Based on the refined binary results, after removing the unchanged samples, randomly generate a pseudo sample training set from the remaining samples and perform multi-class change detection. The final result is as follows Figure 4 (b). The experimental results are evaluated by using two indexes: Overall Accuracy (OA) and Kappa Coefficient (K). Due to the random factors in the method, the experiment is repeated 10 times. The specific results are shown in Table 4 ( Figure 4 The test results shown are those obtained from Experiment No. 1).

[0117] Table 4 Experimental results evaluation

[0118]

[0119] 3. Comparative analysis of this method and some unsupervised change detection results

[0120] The experiment also compared the original S 2 CVA and C 2 The results of randomly generating pseudo sample training sets by dividing the area using the two VA methods. The same number of samples as in the above experiments were randomly selected as training sets, and 10 groups of experiments were conducted. The quality evaluation results are shown in Table 5.

[0121] Table 5 Comparative experimental results evaluation

[0122]

[0123] The detection results obtained using the method described in this embodiment are shown in FIG. Figure 4 As shown in (c), using the original S 2 The test results obtained by CVA method are shown in Figure 4 (d) Using C 2 The test results obtained by VA method are shown in Figure 4 (e) Original S 2 CVA method and C 2The VA method adopts an overall division threshold processing method. For data distribution types with small sample size, non-concentrated distribution in the polar coordinate domain, and serious confusion between change and unchanged samples, the calculated threshold is unreasonable, which may cause a large number of unchanged pixels to be misclassified, such as the quarry change class. The present invention uses an adaptive pseudo-sample candidate area division method based on the statistical characteristics of the frequency histogram in the change vector projection domain. Starting from the distribution of the data itself, it divides the high-confidence pseudo-sample selection area, the supplementary selection area, and the non-candidate area with serious confusion more finely, and can effectively obtain a high-quality pseudo-sample training set. The experimental comparative analysis proves that the proposed method has higher overall accuracy in change detection applications, especially in distinguishing small sample size change classes and suppressing the influence of erroneous samples on the pseudo-sample training set.

[0124] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for automatically generating pseudo samples for unsupervised change detection in multi-temporal remote sensing images, characterized in that: The pseudo sample automatic generation method comprises: Step 1: Process the multi-temporal image data and project it into a two-dimensional polar coordinate domain; Step 2: Generate pseudo sample candidate regions based on the statistical distribution characteristics of the change vector projection in the polar coordinate domain; The change intensity Mag and change angle Dir variables calculated based on the compressed spectrum change vector analysis are plotted in multiple fan-shaped areas in the two-dimensional polar coordinate domain. Each area represents a type of change, specifically: The maximum expectation EM algorithm based on Bayesian estimation is used on the Mag data to calculate the threshold T0 and divide the data set Ω into Ω C and Ω NC , in Ω C The K-means clustering algorithm is used on the Dir data to obtain K clusters, and the entire polar coordinate domain is divided into K sector areas. The boundary of each sector area on the Dir data is A i-1 and A i , i=1,2,…,K, when i=1, A0=0; when i=K, A K =π; Where K is the number of change categories, Mag and Dir are the change intensity and change angle data obtained after the change vector is compressed, Ω is the set of all samples, T0 is the global rough threshold calculated by the EM algorithm based on the Mag data of Ω, Ω C and Ω NC are the change and unchanged sample sets separated by the threshold respectively; Each sector area is further divided into three different types of areas based on the confidence of the sample label, including: For each sample set Ω in the fan-shaped area i , in which the intensity variable Mag i The threshold T1 is calculated using the EM algorithm. i Divide the region into two parts and no change The sample set is denoted as The Mag data is recorded as Calculate the sample set Frequency histogram statistical characteristics, based on this judgment Reasonable distribution of internal samples: in, and Based on The median and mode obtained by frequency histogram statistics calculation; and They are The maximum and minimum values ​​of ; Represents the reasonableness threshold boundary of the sample distribution; It reflects the peak characteristics of the sample frequency histogram statistics in the fan-shaped area, including the location information of the sample clustering in the direction of the changing intensity; For the set of changing samples calculate when When , it is determined that the sample distribution in the set is unreasonable; For the invariant sample set calculate when When , it is determined that the sample distribution in the set is unreasonable; right The threshold calculation based on the change intensity data is and Recorded as Divide the sector polar coordinate domain of each category into unchanged high confidence pseudo sample selection areas Unchanged pseudo sample supplement area Non-candidate area Change pseudo sample supplement area and variation of high confidence pseudo samples There are five selected areas, namely, high-confidence pseudo sample selection area, pseudo sample supplement area and non-candidate area: Among them, t is the dependent variable, which represents the threshold to be determined; L represents the sample set The Mag range is and The difference between ; N is the number of samples in the sample set; represents the high confidence boundary in the set, ω=C, ω=NC, N t Indicates that the Mag value in the sector area is between t and The number of samples in the range; The value is a parameter that measures the threshold t and the overall distribution of the sample set, including the clustering position and concentration information of the frequency histogram statistics; For sample sets with unreasonable distribution, adaptive Value, ω={C,NC}, specifically: Among them, α and P C is a constant parameter representing the sample distribution model fit; is a parameter that reflects the distribution, specifically: N τ =0.75N The Mag histogram statistics of the sample set with unreasonable distribution are calculated to meet the above constraints. N τ Indicates that the Mag value in the sample set is The number of samples in the range; For a set of changing samples with unreasonable distribution That is, ω=C, calculate Specifically: For an invariant sample set with unreasonable distribution That is, ω=NC, calculate Specifically: exist and The thresholds are calculated separately and Obtain an unchanged high confidence pseudo sample selection area in this fan-shaped polar coordinate domain And change high confidence pseudo sample selection area In the fan-shaped area, for a sample set with reasonable distribution, and The same strategy is used to calculate the thresholds for the two sample candidate areas. For a sample set with unreasonable distribution, Select the samples with the smallest or largest Mag value within the interval of the high confidence pseudo sample selection area with the same number of samples as the sample number in the high confidence pseudo sample selection area to form the unchanged or changed pseudo sample supplementary samples, and calculate the unchanged pseudo sample supplementary area and change pseudo sample supplement area The remaining part of the fan-shaped area is the non-candidate area Step 3: Use random sample generation to obtain multiple types of pseudo samples with changes and no changes in the candidate area, and input them into the supervised classifier to achieve binary change detection and multi-class change detection.

2. The method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images according to claim 1, characterized in that: The step 1 is specifically as follows: For multi-temporal image data, differential images are obtained through image preprocessing, image registration, and image subtraction, and the compressed change vector analysis method is used to project them into the two-dimensional polar coordinate domain.

3. The method for automatically generating pseudo samples for unsupervised change detection of multi-temporal remote sensing images according to claim 1, characterized in that: The step 3 is specifically as follows: Firstly, in the high-confidence pseudo sample candidate area of ​​the entire polar coordinate domain, according to the random sampling strategy, p% of the samples in the area are selected as the unchanged and changed pseudo sample training sets, and input into the supervised classifier for classification to obtain the preliminary binary change detection results; Secondly, the unchanged samples in the preliminary results are removed, and the remaining samples are used to generate pseudo sample training sets in high confidence and supplementary areas of each sector polar coordinate domain representing a type of change according to the same random sampling strategy, and then input into the supervised classifier to obtain more refined binary change detection results; Finally, in the change samples, pseudo sample training sets are randomly generated from each sector polar coordinate domain for multi-class change detection, and finally multi-class change detection results are obtained.

Images