Multi-temporal remote sensing image change detection method for change-oriented feature extraction

Abstract

The application discloses a kind of multi-temporal remote sensing image change detection methods for change-oriented map patch extraction, by spatial normalization processing to remote sensing image, eliminate the influence of phenological change on the change detection accuracy of land cover type, so that change vector analysis method is not limited by the condition that remote sensing image phase needs to be the same season, can select any phase remote sensing image within the scale of year to extract change map patch, can timely and accurately detect the surface change condition. By migrating the previous phase unchanged sample to train the current phase classifier, and then reclassifying the land cover type only for the detected change pixel in the current phase, the workload of change map patch extraction is reduced, and the classification efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection and classification, and in particular to a multi-temporal remote sensing image change detection method for change patch extraction. BACKGROUND

[0002] Traditional remote sensing image change patch extraction based on remote sensing technology adopts a change information extraction method of scene-by-scene classification or manual visual interpretation. Scene-by-scene classification selects training samples for each image to train a classifier to classify the new land cover types changed in the image. Since training samples are obtained by manual visual interpretation for each image, scene-by-scene classification and manual visual interpretation have the problems of large workload, low work efficiency and low automation, which limit the rapidity of change patch extraction.

[0003] In recent years, with the continuous deepening of remote sensing image change detection research, the extraction of change patches of land cover types in remote sensing images by means of change detection algorithms has gradually become an important means of geographic condition monitoring. The accuracy of change patch extraction based on change detection algorithms mainly depends on the accuracy of change detection. Phenological change of land cover (especially vegetation cover) in remote sensing images is one of the factors affecting the accuracy of change detection. In order to eliminate pseudo changes caused by phenological change, most change detection algorithms usually select images of the same season for change detection, such as change vector analysis method or LandTrendr algorithm, which usually selects quasi-annual images of the same season in different years for change detection. However, the observation frequency of one image per year cannot realize the change detection at the intra-annual scale, or when the disturbance event occurs after the observation time of the year, the disturbance occurrence time will be incorrectly labeled as the next year, which prolongs the extraction cycle of change patches and cannot ensure the timeliness of land cover information. Increasing the observation frequency of remote sensing images will increase the workload of change patch extraction by change detection algorithms, so increasing the observation frequency is not an arbitrary behavior. Therefore, how to increase the observation frequency of multi-temporal remote sensing images, eliminate pseudo changes caused by phenological change, ensure the accurate and timely acquisition of intra-annual land cover information and extract change patches using any period of remote sensing images, and ensure the timeliness of land cover information are technical problems to be solved in the technical field. SUMMARY

[0004] The present application aims to realize the timely and accurate identification of change information of remote sensing images at the intra-annual scale, and realizes the accurate extraction of change patches in images by eliminating pseudo changes caused by phenological change and other technical means, and provides a multi-temporal remote sensing image change detection method for change patch extraction.

[0005] To achieve this purpose, the present application adopts the following technical solutions:

[0006] The application provides a multi-temporal remote sensing image change detection method for change-oriented feature extraction, comprising the following steps:

[0007] S1, determining the classification and change detection features of each remote sensing image collected at the same observation area at time t i+1 and t i , wherein t i is the time of the previous remote sensing image collection of t i+1 ;

[0008] S2, constructing a remote sensing classification sample set at time t i , and making a classification base map for change detection of the second remote sensing image collected at time t i+1 ;

[0009] S3, calculating the SVI value of each pixel in the second remote sensing image and the first remote sensing image collected at the same observation area at time t i , to obtain the first spatial normalization image and the second spatial normalization image corresponding to the first remote sensing image and the second remote sensing image respectively;

[0010] S4, using the first spatial normalization image and the second spatial normalization image to perform pixel change detection and result correction on the second remote sensing image;

[0011] S5, training the classifier at time t i+1 using the background area with no change in pixels after the correction in step S4 as a sample;

[0012] S6, classifying the target area with changed pixels in the second remote sensing image by using the classifier to obtain the changed land cover type, and then fusing the first classification result of the target area in the second remote sensing image and the second classification result of each background area in the second remote sensing image in the first remote sensing image into the land cover type map at time t i+1 , to update the classification base map corresponding to time t i made in step S2.

[0013] Preferably, step S1 specifically comprises the following steps:

[0014] S11, calculating the vegetation index of each type of ground object sample in the remote sensing image;

[0015] S12, constructing a corresponding ground object sample combination for each two different types of ground object samples in the remote sensing image, and then selecting the ground object sample combination with the maximum JM distance value as the best vegetation index combination as the classification and change detection features of the remote sensing image based on the JM distance measurement.

[0016] As preferred, in step S11, for the ground object sample type being vegetation, a vegetation index is obtained after normalization by the following formula (1):

[0017]

[0018] For the ground object sample type being water body, a vegetation index is obtained after normalization by the following formula (2):

[0019]

[0020] For the ground object sample type being building, a vegetation index is obtained after normalization by the following formula (3):

[0021]

[0022] For the ground object sample type being soil, a vegetation index is obtained after normalization by the following formula (4):

[0023]

[0024] In formulas (1)-(4), v 近红外 represents a characteristic value of the remote sensing image in a near-infrared band;

[0025] v 红 represents a characteristic value of the remote sensing image in a red band in a visible light band;

[0026] v 绿 represents a characteristic value of the remote sensing image in a green band in a visible light band;

[0027] v 短波红外 represents a characteristic value of the remote sensing image in a short-wave infrared band.

[0028] As preferred, in step S2, the remote sensing classification sample set at time t i is constructed and the classification base map changed by the second remote sensing image collected at time t i+1 The method for making the classification base map changed by the second remote sensing image collected at time t

[0029] A1, a multi-type ground object sample collection is performed on the observation area at time t i by using a stratified random sampling method;

[0030] A2, the first remote sensing image collected at time t i containing each type of ground object sample collected in step A1 is checked by using a manual method to verify the ground object sample type of each ground object sample in the first remote sensing image at time t i-1 corresponding to time t i-1The classifier is trained by using the remote sensing classification sample set constructed at the time t, and the accuracy of the classification result of the ground object sample in the first remote sensing image is verified. i The remote sensing classification sample set constructed at the time t;

[0031] A3, using the remote sensing classification sample set corresponding to t i The classifier corresponding to t i The time t is trained.

[0032] A4, using the trained classifier corresponding to t i The time t is used to classify the ground object sample in the first remote sensing image, and the classification result is used as the classification map for change detection of the second remote sensing image collected at the time t. i+1 The classification map.

[0033] As preferred, in step A3, the method for training the classifier specifically includes the following steps:

[0034] A31, set the "random number" field of the remote sensing classification sample set to a floating point number in the interval of 0-1, select the classification sample with a random number field less than 0.7 in the field as the training sample, and the remaining samples as the verification sample to complete the splitting of the remote sensing classification sample set.

[0035] A32, superimpose the training sample and the classification feature set constructed based on the spectral band and vegetation index features of the remote sensing image, and extract the corresponding classification feature value of the training sample.

[0036] A33, use the random forest machine learning module in python 3.7.0 tool to train the training sample with classification feature value, and obtain the classifier.

[0037] As preferred, in step S3, the SVI value of the pixel in the remote sensing image is calculated by the following formula (5):

[0038]

[0039] In formula (5), VI center represents the vegetation index value of the center pixel in the spatial window of the preset size;

[0040] VI median represents the average value of the vegetation index values of each pixel in the spatial window with the same ground cover type as the center pixel.

[0041] As preferred, in step S4, the method for detecting pixel change of the second remote sensing image includes the following steps:

[0042] S41, calculating the change intensity AG of the second remote sensing image at time t i+1 compared with time t i in formula (6),

[0043]

[0044] In formula (6), denotes the first SVI value of the jth pixel in the first remote sensing image;

[0045] denotes the second SVI value of the jth pixel in the second remote sensing image;

[0046] n denotes the number of pixels in the second remote sensing image;

[0047] S42, judging whether AG is greater than a preset change intensity threshold,

[0048] if yes, determining that the jth pixel in the second remote sensing image is a change pixel;

[0049] if no, determining that the jth pixel in the second remote sensing image is an unchanged pixel.

[0050] As a preferred, the change intensity threshold is set by using the maximum inter-class variance method, specifically comprising steps of:

[0051] S421, calculating the total number of pixels n in the second remote sensing image;

[0052] S422, calculating the probability P j of each change intensity value appearing in the histogram, wherein j denotes the change intensity value of the jth pixel;

[0053] S423, by traversing all possible threshold values x, calculating the intra-class variance and the inter-class variance, and finding the threshold value corresponding to the maximum inter-class variance:

[0054] S424, setting the threshold value x corresponding to the maximum inter-class variance calculated as the change intensity threshold.

[0055] As a preferred, the method for correcting the change detection result of the pixels in the second remote sensing image is:

[0056] when determining that the center pixel in the spatial window is a change pixel, further judging whether the neighborhood pixels of the center pixel are unchanged pixels,

[0057] if yes, correcting the center pixel as an unchanged pixel;

[0058] if no, maintaining the center pixel as a change pixel.

[0059] As preferred, in step S6, the method for classifying the target region with changed pixels in the second remote sensing image by the classifier corresponding to the time t i+1 The method for classifying the target region with changed pixels in the second remote sensing image by the classifier corresponding to the time t

[0060] S61, extracting the changed pixels according to the SVI value from the change detection grid result respectively;

[0061] S62, using the changed pixels as a mask to crop the second remote sensing image to obtain a changed region remote sensing image;

[0062] S63, determining whether there is a changed pixel in the classified sample patch in the changed region remote sensing image by using a spatial intersection tool, and removing the sample patch with the changed pixel inside;

[0063] S64, splitting the processed sample patch into a training set and a validation set, and using a random forest machine learning module to train the classifier;

[0064] S65, using the trained classifier to classify the remote sensing image of the changed region.

[0065] 11. The multi-temporal remote sensing image change detection method for change-oriented patch extraction according to claim 1, further comprising the step of:

[0066] S7, setting i=i+1, and then returning to step S1 until the change patch classification detection of the n remote sensing images collected at different time points for the same observation area is completed, i=1, 2, …, n.

[0067] The present application has the following beneficial effects:

[0068] 1. By performing spatial normalization processing on the remote sensing image, the influence of phenological change on the change detection accuracy of land cover type is eliminated, so that the change vector analysis method is not limited by the condition that the time phases of the remote sensing images are the same season, and any remote sensing image at an arbitrary time phase within the annual scale can be selected for change patch extraction, so that the ground change can be detected in a timely and accurate manner.

[0069] 2. By training the current time phase classifier by migrating the unchanged sample of the previous time phase, and then reclassifying the land cover type only for the detected changed pixels at the current time phase, the workload of change patch extraction is reduced, and the classification efficiency and the classification automation degree are improved. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. Obviously, the drawings described below are only some of the embodiments of the present application, and for those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

[0071] Figure 1 is the implementation step diagram of the multi-temporal remote sensing image change detection method for change-oriented feature extraction provided by the embodiment of the present application;

[0072] Figure 2 is the technical roadmap of the multi-temporal remote sensing image change detection provided by the embodiment;

[0073] Figure 3 is an example diagram for change detection of multi-temporal remote sensing images. DETAILED DESCRIPTION

[0074] The technical solutions of the present application will be further illustrated by specific embodiments in combination with the drawings.

[0075] Among them, the drawings are only used for exemplary illustration, and the representation is only a schematic diagram, not a physical diagram, and cannot be understood as a limitation on the present patent; in order to better illustrate the embodiments of the present application, some components of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some known structures and their descriptions in the drawings can be omitted.

[0076] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for exemplary illustration, and cannot be understood as a limitation on the present patent, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0077] In the description of the present application, unless otherwise explicitly specified and limited, if the term "connection" and the like indicating the connection relationship between components appears, the term should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication or interaction relationship between two components. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0078] The multi-temporal remote sensing image change detection method for change-oriented patch extraction provided by the embodiment of the present application comprises the steps of: Figure 1 and Figure 2 as shown in the drawings, comprising the steps of:

[0079] S1, determining the classification and change detection features of each remote sensing image collected at the same observation area at t i+1 and t i moments, and the determination method comprises the steps of:

[0080] S11, calculating the vegetation index of each type of ground object sample in the remote sensing image.

[0081] Suppose that n multi-scene Landsat remote sensing images are observed at t1, t2, …, t n moments in time sequence. The remote sensing image has at least one type of ground object sample as recorded in Table a.

[0082] Table a

[0083]

[0084]

[0085]

[0086] In this embodiment, the visible light band (red, green, blue), near-infrared band, and short-wave infrared band are used as the classification features of the ground object samples, and the vegetation index is calculated for each remote sensing image to classify the ground object samples. Table b shows part of the names and calculation methods of the vegetation index.

[0087] Table b

[0088]

[0089] In Table b, “near-infrared”, “red”, “green”, and “short-wave infrared” in the calculation formula respectively refer to the near-infrared band characteristic value, the red band characteristic value in the visible light band, the green band characteristic value in the visible light band, and the short-wave infrared band characteristic value of the remote sensing image.

[0090] In Table b, “vegetation” in the normalized vegetation index refers to trees, bamboo forests, shrubs, and grasses in Table a.

[0091] S12, constructing a corresponding ground object sample combination for each pair of different types of ground object samples in the remote sensing image, and then selecting the ground object sample combination corresponding to the best vegetation index combination expressed by the maximum JM distance value as the classification and change detection features of the remote sensing image based on the JM distance measurement.

[0092] Assuming that the remote sensing image has four types of ground object samples, i.e., vegetation, water body, building, and soil, the combinations of two different types of ground object samples include: the first combination of vegetation and water body, the second combination of vegetation and building, the third combination of vegetation and soil, the fourth combination of water body and building, the fifth combination of water body and soil, and the sixth combination of building and soil. Then, the JM distance value of the vegetation index of the two elements in each combination is calculated, and the ground object sample combination with the maximum JM distance value is taken as the classification and change detection feature of the remote sensing image. The JM distance measurement is a prior method, and therefore, how to calculate the JM distance value of the vegetation index of the two elements in the ground object sample combination is not specifically described here. In this embodiment, the separability of different types of ground object samples is tested by JM distance measurement.

[0093] After determining the classification and change detection feature of each remote sensing image by step S1, the multi-temporal remote sensing image change detection method for change graph extraction provided in this embodiment proceeds to step:

[0094] S2, constructing a remote sensing classification sample set at t i time, and making a land class and vegetation cover type map for change detection of the second remote sensing image collected at t i+1 time (defined as a classification map);

[0095] With the existing classification products (such as a university based on all available Landsat data on GEE, constructing spatio-temporal features, combining a random forest classifier to obtain a classification result, proposing a post-processing method containing temporal filtering and logical reasoning to obtain 30-meter global CLCD land cover data per year, a space agency based on Sentinel-1 and Sentinel-2 data using the deep learning AI land classification model of Impact Observatory to obtain a global 10m land use product, etc.), a stratified random sampling method is used to obtain different land class and vegetation cover type samples in the observation area, and t i time high-resolution images and high-spatial-resolution images are used to randomly sample and check the classification samples by artificial visual interpretation to ensure the accuracy of the labels, and finally a spatially balanced, representative, and high-quality classification sample set is established.

[0096] Specifically, in step S2, the remote sensing classification sample set at t i time is constructed and the classification map for the second remote sensing image collected at t i+1 time is made, which includes the following steps:

[0097] A1, using a stratified random sampling method to obtain different land class and vegetation cover type samples in the observation area at t iAt the moment, the ground object samples are collected in the observation area in the existing local class and vegetation cover type products, such as the sample collection of the arbor forest land with code 0301 in the above table a in the observation area. The method of collecting ground object samples by stratified random sampling method is as follows: the arbor forest pixels in the existing classification product are extracted, a random point tool is used to randomly generate a sampling point, it is ensured that the selection of each sample point is independent, and each point has the same selection probability.

[0098] A2, the first remote sensing image collected at t i The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t i-1 The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t i-1 The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t i The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t

[0099] A3, using the remote sensing classification sample set corresponding to t i The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t i The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t i The accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t

[0100] A31, the classification sample data set is set to "random number" field, and the floating point number in the interval of 0-1 is assigned. The classification samples with random number field less than 0.7 in the field are selected as training samples, and the remaining samples are verified as verification samples to complete the splitting of the classification sample set.

[0101] A32, the training sample is superimposed with the classification feature set constructed based on the spectral band and vegetation index features obtained from the remote sensing image, and the corresponding classification feature value of the training sample is extracted.

[0102] A33, the random forest machine learning module in python 3.7.0 tool is used to train the training sample with classification feature value.

[0103] A4, the ground object sample classification of the first remote sensing image is completed by using the trained random forest classifier, and the classification result is used as the classification base map for change detection of the second remote sensing image collected at t i+1

[0104] Through step S2, the accuracy of the ground object sample classification result of the first remote sensing image is verified by using the classifier trained by the remote sensing classification sample set constructed at t i+1 ​After the production of the classification base map of the second remote sensing image collected at the time t for change detection, the multi-temporal remote sensing image change detection method for change map extraction provided by the embodiment enters the step of:

[0105] S3, calculating t i+1 The second SVI value of each pixel in the second remote sensing image collected at the time t for the observation area is obtained to obtain the second space normalized image corresponding to the second remote sensing image, and t i The first SVI value of each pixel in the first remote sensing image collected at the time t for the same observation area is solved to obtain the first space normalized image corresponding to the first remote sensing image. i+1 > t i , t i is the last remote sensing image collection time of t i+1 ;

[0106] The SVI value of the pixel in the remote sensing image is calculated by the following formula (5):

[0107]

[0108] In formula (5), VI center represents the vegetation index value of the center pixel in the spatial window of the preset size;

[0109] VI median represents the average value of the vegetation index values of each pixel in the spatial window with the same surface coverage type as the center pixel.

[0110] Because the classification characteristics of different types of ground objects have inconsistent dimensions, at the same time, the phenological change of the ground class in the remote sensing images collected at the two times t i+1 and t i , atmospheric condition difference and other factors will interfere with the change detection result. In order to solve this problem, the spatial normalization operation is performed on the vegetation index value of each pixel in the remote sensing image to eliminate the interference of the above factors on the change detection.

[0111] The specific process of spatial normalization is as follows: first, set a spatial window of a certain size with any pixel as the center pixel, and the window size is determined according to the spatial resolution of the remote sensing image, such as setting a spatial window of 15x15 pixels to ensure that there are not less than 30 pixels with the same surface coverage type as the center pixel in the window to calculate VI median . If the number of the same type of pixels in the window is less than 30, then gradually increase the window with a step of "2" until the requirement is met.

[0112] Then, select the pixels with the same surface coverage type as the center pixel from the window to calculate VI median . VImedian The calculation method is as follows: Using the attribute extraction tool, extract pixels with the same land cover type within the window, and calculate the average value of the vegetation index pixels at their corresponding locations to obtain the VI. median .

[0113] Finally, the SVI value of the center pixel is calculated using the above formula (5).

[0114] Get t i+1 and t i After obtaining the second spatially normalized image and the first spatially normalized image corresponding to each time point, the multi-temporal remote sensing image change detection method for extracting changed patches provided in this embodiment proceeds to the following steps:

[0115] S4, perform pixel change detection and result correction on the second remote sensing image;

[0116] The method for detecting pixel changes in a second remote sensing image includes the following steps:

[0117] S41, the second remote sensing image at t is calculated using the following formula (6). i+1 Comparison of time t i The intensity of the change at time ΔG,

[0118]

[0119] In formula (6), This represents the first SVI value of the j-th pixel in the first remote sensing image;

[0120] This represents the second SVI value of the j-th pixel in the second remote sensing image;

[0121] n represents the number of pixels in the second remote sensing image.

[0122] The larger ||ΔG|| is, the more the same pixel appears in t. i+1 Comparison of time t i The greater the difference in time, the greater the likelihood of a change in land cover type. Therefore, a threshold can be set, and the intensity of ||ΔG|| can be used to determine whether a pixel has changed.

[0123] In this embodiment, the maximum inter-class variance (MOV) method is used to adaptively divide the change intensity into two parts: background (invariant pixels) and target (changing pixels). The larger the inter-class variance between the background and the target, the greater the difference between the two parts constituting the remote sensing image, and the smaller the probability of misclassification. That is, the change and non-change are optimally separated, realizing the change detection of the remote sensing image. Therefore, the MOV method can separate changing pixels with the minimum probability of misclassification, avoiding the subjectivity and randomness of manually setting thresholds.

[0124] The method of setting the change detection threshold by using the maximum between-class variance method specifically comprises the following steps:

[0125] S421, calculating the total number of pixels n.

[0126] S422, calculating the probability P of each change intensity value in the histogram j , j represents the change intensity value of the jth pixel.

[0127] S423, by traversing all possible thresholds x, calculating the within-class variance and the between-class variance, and finding the threshold corresponding to the maximum between-class variance:

[0128] S424, setting the threshold x corresponding to the maximum between-class variance calculated as the change intensity threshold.

[0129] Considering the spatial continuity feature of the change, in order to eliminate the false changes caused by various noises, in the embodiment, for the pixels determined as unchanged pixels, a method of result correction for unchanged pixels is provided, specifically:

[0130] When the center pixel in the spatial window is determined as a changed pixel, it is further determined whether the neighborhood pixels of the center pixel are all unchanged pixels,

[0131] if yes, the center pixel is corrected as an unchanged pixel;

[0132] if no, the determination of the center pixel as a changed pixel is maintained.

[0133] The neighborhood pixels of the center pixel are, for example, the pixels arranged above, below, left, right, top-left, bottom-left, top-right and bottom-right of the center pixel.

[0134] After the change detection and result correction of the pixels of the second remote sensing image collected at the t i+1 moment are completed, as shown in FIG. 6, the multi-temporal remote sensing image change detection method provided by the embodiment facing the change patch extraction enters the step of: Figure 1

[0135] S5, taking the background area of the changed pixels 2 corrected in step S4 as a sample to train the classifier at the t i+1 moment (also preferably a random forest classifier), the method of training the classifier at the t i+1 moment is the same as the method of training the classifier at the t i moment, and thus is not described again.

[0136] ​S6, the classifier trained in step S5 classifies the target region in which the pixel in the second remote sensing image has changed, to obtain the changed land cover type, and then fuses the first classification result of the target region in the second remote sensing image and the second classification result of each background region in the second remote sensing image corresponding to the second classification result of the target region in the first remote sensing image into a t i+1 time land cover type map, to update the classification base map corresponding to the t i time in step S2;

[0137] the classifier corresponding to the t i+1 time classifies the target region in which the pixel in the second remote sensing image has changed.

[0138] S61, the changed pixels are extracted from the change detection grid result according to the SVI value.

[0139] S62, the second remote sensing image is cropped by taking the changed pixels as a mask to obtain a change region remote sensing image

[0140] S63, it is judged by a spatial intersection tool whether there is a changed pixel in the classification sample patch, and the sample patch with the changed pixel inside is removed; it should be noted that if the sample patch has a changed pixel, it means that the land cover type corresponding to the changed pixel is only known in the first time phase. In the subsequent time phase, the land cover type becomes unknown due to the change. In the experiment of the present application, the sample patch needs to be migrated to the next time phase for continuous use. In order to avoid the influence of the unknown land cover type on the classification, the patch containing the changed pixel needs to be deleted. Only the patch with unchanged pixels is migrated. In summary, the appearance of a changed pixel means that the pixel in the sample patch has changed, and the only standard for whether the sample in the present application can be used as a training sample of the random forest classifier in the next time phase is whether it has changed.

[0141] S64, the processed sample patch is split into a training set and a validation set, and a random forest machine learning module is used to train the classifier.

[0142] S65, the trained classifier is used for classification of the remote sensing image of the change region.

[0143] The method for fusing the first classification result and the second classification result is briefly described as follows:

[0144] The classification base map is cropped by taking the changed pixel as a mask to obtain the classification result of the unchanged region. Subsequently, the classification result corresponding to the change region is inlaid with the classification result of the unchanged region, and the fusion and update of the classification result are completed.

[0145] S7, set u=i+1, and then return to step S1 until the change patch classification detection and extraction of n scenes of remote sensing images collected at different time for the same observation area is completed, i=1, 2, …, n. The change patch detection result is as shown in Figure 3

[0146] In summary, by performing spatial normalization processing on the remote sensing image, the influence of the phenological change on the change detection precision of the land cover type is eliminated, the change vector analysis method is not limited by the condition that the time phases of the remote sensing image are the same season, the remote sensing image of any time phase within the annual scale can be selected to extract the change patch, and the land surface change can be detected in time and accurately. By migrating the unchanged sample of the previous time phase to train the current time phase classifier, and then reclassifying the land cover type of the detected change pixel in the current time phase, the workload of the change patch extraction is reduced, and the classification efficiency and the classification automation degree are improved.

[0147] It should be noted that the above specific embodiments are only the preferred embodiments of the present application and the technical principles applied. Those skilled in the art should understand that various modifications, equivalent replacements, changes, etc. can also be made to the present application. However, as long as these changes do not deviate from the spirit of the present application, they should be within the protection scope of the present application. In addition, some terms used in the present application specification and claims are not limited, but only for the convenience of description.​

Claims

1. A method for detecting changes in multi-temporal remote sensing images for extracting changing patches, characterized in that, Including the following steps: S1, determine t i+1 and t i The classification and change detection characteristics of each remote sensing image collected in the same observation area at any time, t i For t i+1 The moment the previous remote sensing image was acquired; S2, construct t i The remote sensing classification sample set at time t, and the production of the data for t i+1 A classification base map for change detection of second-generation remote sensing images acquired at different times; S3, for the second remote sensing image and in t i For the first remote sensing image acquired at the same observation area, the SVI value of each pixel in the image is calculated to obtain the first spatially normalized image and the second spatially normalized image corresponding to the first remote sensing image and the second remote sensing image, respectively. The SVI value of the pixel in the remote sensing image is calculated by the following formula: In the formula, The vegetation index value represents the center cell of a spatial window of a preset size. This represents the average vegetation index value of all pixels within the spatial window that have the same land cover type as the central pixel; When the land cover type of the central pixel is vegetation, its vegetation index value is obtained by normalization using the following formula: Official (1) in, This represents the characteristic value of the remote sensing image in the near-infrared band. This represents the characteristic value of the red band in the visible light band of the remote sensing image; S4, using the first spatially normalized image and the second spatially normalized image, perform pixel change detection and result correction on the second remote sensing image; S5, using the unchanged background region after step S4 as samples to train t i+1 A time-based classifier; S6, the classifier is used to classify the target areas in the second remote sensing image where pixels have changed, to obtain the changed land cover type. Then, the first classification result of the target areas in the second remote sensing image and the second classification result of each background area in the second remote sensing image corresponding to the first remote sensing image are merged into t. i+1 A land cover type map at time t, to correspond to the t map created in step S2. i The classification base map is updated at any given time.

2. The method for detecting changes in multi-temporal remote sensing images oriented towards extracting changing patches according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11, Calculate the vegetation index for each category of land cover sample in the remote sensing image; S12, construct corresponding land cover sample combinations for each pair of different categories of land cover samples in the remote sensing image, and then select the land cover sample combination with the maximum JM distance value as the optimal vegetation index combination based on the JM distance metric as the classification and change detection feature of the remote sensing image.

3. The multi-temporal remote sensing image change detection method for extracting changing patches according to claim 2, characterized in that, In step S11, for land cover samples of water type, the vegetation index is obtained after normalization using the following formula (2): Official (2) For land cover samples of building type, the vegetation index is obtained by normalization using the following formula (3): Official (3) For soil samples, the vegetation index is obtained by normalization using the following formula (4): Official (4) In formulas (2)-(3), This represents the characteristic value of the green band in the visible light band of the remote sensing image; This represents the characteristic value of the remote sensing image in the shortwave infrared band.

4. The multi-temporal remote sensing image change detection method for extracting changing patches according to claim 1, characterized in that, In step S2, t is constructed. i The remote sensing classification sample set at time t and the production of the pair t i+1 The method for modifying the classification base map based on the second remote sensing image acquired at different times includes the following steps: A1, using a stratified random sampling method, at t i At any given time, samples of various types of ground features are collected from the observation area; A2, for t i The first remote sensing images, which contain various types of land cover samples collected in step A1, are manually verified. i-1 Time corresponding to t i-1 The classifier, trained on the remote sensing classification sample set constructed at each time point, is used to assess the accuracy of the classification results of ground cover samples in the first remote sensing image, and the ground cover samples that pass the verification are added to the classifier for t. i The remote sensing classification sample set constructed at that time; A3, using the corresponding t i The remote sensing classification sample set at time t is used to train the corresponding t i A time-based classifier; A4, using the corresponding t after training is complete. i The classifier at time t classifies ground feature samples from the first remote sensing image and uses the classification result as the basis for classifying the ground feature samples from the first remote sensing image. i+1 The classification base map is used for change detection of the second remote sensing image acquired at different times.

5. The multi-temporal remote sensing image change detection method for extracting changing patches according to claim 4, characterized in that, Step A3, the method for training the classifier specifically includes the following steps: A31, assign a floating-point number in the range of 0-1 to the "random number" field of the remote sensing classification sample set, select classification samples with random number segments less than 0.7 in this field as training samples, and the remaining samples as validation samples to complete the splitting of the remote sensing classification sample set. A32, the training samples are superimposed with the classification feature set constructed based on the spectral bands and vegetation index features obtained from remote sensing images to extract the corresponding classification feature values ​​for the training samples. A33, the classifier is obtained by training training samples with classification feature values ​​using the Random Forest machine learning module in Python 3.7.

0.

6. The method for detecting changes in multi-temporal remote sensing images for extracting changing patches according to claim 1, characterized in that, In step S4, the method for detecting pixel changes in the second remote sensing image includes the following steps: S41, the second remote sensing image at t is calculated using the following formula (6). i+1 Comparison of time t i The intensity of the change at time ΔG, Official (6) In formula (6), This represents the first SVI value of the j-th pixel in the first remote sensing image; This represents the second SVI value of the j-th pixel in the second remote sensing image; n represents the number of pixels in the second remote sensing image; S42, determine whether ΔG is greater than the preset change intensity threshold. If so, then the j-th pixel in the second remote sensing image is determined to be a changed pixel; If not, then the j-th pixel in the second remote sensing image is determined to be an invariant pixel.

7. The multi-temporal remote sensing image change detection method for extracting changing patches according to claim 6, characterized in that, Setting the change intensity threshold using the Otsu's method specifically includes the following steps: S421, Calculate the total number of pixels n in the second remote sensing image; S422, Calculate the probability P of each change intensity value appearing in the histogram. j , j represents the change intensity value of the j-th pixel; S423, by iterating through all possible thresholds x, calculate the within-class variance and between-class variance, and find the threshold corresponding to the maximum between-class variance: S424, set the threshold x corresponding to the calculated maximum inter-class variance as the change intensity threshold.

8. The method for detecting changes in multi-temporal remote sensing images for extracting changing patches according to claim 1, characterized in that, The method for correcting the pixel change detection results of the second remote sensing image is as follows: When the central pixel in the spatial window is determined to be a changing pixel, it is further determined whether the neighboring pixels of the central pixel are invariant pixels. If so, the central pixel is modified to an invariant pixel; If not, then the central pixel remains a variable pixel.

9. The method for detecting changes in multi-temporal remote sensing images oriented towards extracting changing patches according to claim 1, characterized in that, In step S6, corresponding to t i+1 The method for classifying target regions in the second remote sensing image whose pixels have changed at any given time using the classifier includes the following steps: S61, extract the changed pixels from the change detection raster results according to the SVI value; S62, using the changed pixels as a mask, the second remote sensing image is cropped to obtain the changed area remote sensing image; S63, use a spatial intersection tool to determine whether there are changed pixels in the classified sample patches of the changed area remote sensing image, and remove sample patches with changed pixels inside. S64, the processed sample patches are split into training and validation sets, and the random forest machine learning module is used to train the classifier; S65 uses the trained classifier to classify ground features in remote sensing images of changed areas.

10. The method for detecting changes in multi-temporal remote sensing images for extracting changing patches according to claim 1, characterized in that, It also includes the following steps: S7, set i = i + 1, then return to step S1 until the classification and detection of change patches of n remote sensing images collected at different times for the same observation area is completed, i = 1, 2, ..., n.

Images