A method for information extraction of multi-source remote sensing data

Abstract

The application relates to the technical field of multi-source remote sensing data extraction, and provides a multi-source remote sensing data information extraction method, which comprises the following steps: step 1: acquiring high-altitude image data of a target area, and extracting a plurality of matching blocks from the high-altitude image data; step 2: acquiring unmanned aerial vehicle image data of the target area, and extracting a plurality of feature blocks from the unmanned aerial vehicle image data; step 3: acquiring the measurement coordinates of each matching block in the high-altitude image data to obtain first coordinate information, and acquiring the measurement coordinates of each feature block in the unmanned aerial vehicle image data to obtain second coordinate information; and step 5: constructing a probability model of mutual registration of the matching blocks and the feature blocks for each matching group, and performing tensor decomposition on the probability model to obtain registration information corresponding to the matching blocks and the feature blocks. The method significantly improves the multi-source remote sensing data fusion precision through a grouping probability registration mechanism, and breaks through the restriction of positioning precision difference on multi-source data fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-source remote sensing data extraction, in particular to a multi-source remote sensing data information extraction method. BACKGROUND

[0002] The content of this part only provides background information related to the present application, which may not constitute prior art.

[0003] Multi-source remote sensing data refers to a heterogeneous data set obtained by cooperative observation of the same area through heterogeneous sensors (such as optical, radar, infrared, hyperspectral, etc.) or cross-platform carriers (satellite, aircraft, ground station). Through the fusion of multi-dimensional information such as spatial, spectral, temporal and radiation resolution, it can effectively make up for the inherent limitations of a single data source (such as optical images being easily disturbed by clouds and fog, and radar data lacking spectral details), significantly improve the accuracy of feature classification, dynamic monitoring capability and environmental parameter inversion reliability, and has been widely used in resource survey, disaster warning and precision agriculture fields.

[0004] However, the core challenge of multi-source data processing is the accurate matching and fusion of cross-source data. The existing scheme usually relies on geographical coordinates to directly align, and forcibly superimposes different source data at the preset position to construct a three-dimensional scene model. This method has a fundamental defect: there is a significant difference in positioning accuracy of different remote sensing systems (such as coordinate deviation of satellite images and unmanned aerial vehicle radar), and the current technology is difficult to realize adaptive matching according to the accuracy difference. As a result, effective information corresponding to space cannot be extracted from heterogeneous data; and the fusion error is transmitted to downstream applications, resulting in a serious decline in the geometric and texture accuracy of three-dimensional scene rendering. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a multi-source remote sensing data information extraction method, which can solve the technical problems in the background art.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] A multi-source remote sensing data information extraction method, comprising:

[0008] Step 1: Obtain high-altitude image data of a target area, and extract a plurality of matching blocks from the high-altitude image data;

[0009] The matching block includes at least a number of extreme points exceeding a preset threshold;

[0010] Step 2: Obtain unmanned aerial vehicle image data of the target area, and extract a plurality of feature blocks from the unmanned aerial vehicle image data;

[0011] The feature block at least includes an extreme point exceeding a preset threshold number.

[0012] Step 3: Obtain first coordinate information by obtaining the measured coordinates of each matching block in the high-altitude image data, and obtain second coordinate information by obtaining the measured coordinates of each feature block in the unmanned aerial vehicle image data;

[0013] Step 4: According to the first coordinate information and the second coordinate information, fuzzy matching is performed on the matching block and the feature block, and the matching block and the feature block whose coordinate positions are not more than a preset threshold are merged into one matching group;

[0014] Step 5: For each matching group, a probability model of mutual registration of the matching block and the feature block is constructed, and the probability model is tensor decomposed to obtain registration information corresponding to the matching block and the feature block.

[0015] The method significantly improves the fusion accuracy of multi-source remote sensing data through a grouping and probabilistic registration mechanism. First, the elevation image matching block and the unmanned aerial vehicle image feature block are dynamically grouped based on spatial proximity, avoiding the positioning error of traditional coordinate hard matching. Then, a probability model of the spatial relationship between the blocks in each group is constructed, and the optimal matching probability is solved by tensor decomposition, so that the accuracy difference between satellite elevation data (low positioning accuracy) and unmanned aerial vehicle image (high resolution but weak positioning) is converted into a calculable probability weight, realizing adaptive calibration of heterogeneous data. Finally, high-confidence registration information is output, which greatly reduces the matching error and directly improves the geometric consistency and texture reality of three-dimensional scene reconstruction, breaking through the constraints of positioning accuracy difference on multi-source data fusion.

[0016] Further, the real-world physical size corresponding to the matching block is consistent with the real-world physical size corresponding to the feature block; and the size of the matching block is set according to the matching accuracy.

[0017] In the technical solution provided in the present application, the real-world physical size corresponding to the matching block is consistent with the real-world physical size corresponding to the feature block, which can ensure that the number of matching blocks and feature blocks in the same range is equal, facilitating one-to-one correspondence between the matching blocks and the feature blocks.

[0018] Further, the number of extreme points in the feature block and the matching block satisfies the following condition:

[0019] ;

[0020] x, y represent the center point coordinates of the feature block or the matching block, x i and y i represent the coordinates of the i-th extreme point, K represents the kernel function, n represents the total number of extreme points, and h represents the bandwidth parameter.

[0021] The distribution of extreme points in the feature block and the matching block is limited in the technical solution provided in the application, so that a sufficient number of extreme points in the matching block and the feature block can be effectively matched with each other, the accuracy of subsequent matching of the matching block and the feature block is increased, and the density of the extreme points is evaluated by using the kernel function, so that a sufficient number of extreme points in the matching block and the feature block can be matched with each other.

[0022] Further, the first coordinate information is satellite positioning information obtained by a satellite, and the second coordinate information is radar positioning information obtained by a UAV; the size of the feature block and the matching block is determined based on the minimum measurement accuracy of the satellite positioning information.

[0023] In the technical solution provided in the application, the size of the feature block and the matching block is determined based on the minimum measurement accuracy of the satellite positioning information, so that the real size of the physical world corresponding to the feature block and the matching block can be matched with each other, and the difference in the actual size between the feature block and the matching block is avoided, so that the accuracy of subsequent matching is not high.

[0024] In some possible embodiments, the acquisition manner of the matching block includes the following steps:

[0025] Step 11: The mountain boundary, the river boundary, the depression boundary and the road boundary are selected from the high-altitude image data;

[0026] Step 12: A plurality of extreme points are extracted from the river boundary, the depression boundary and the road boundary to generate a first extreme point set;

[0027] Step 13: Extreme points are extracted from the remaining area of the high-altitude image data except the mountain boundary, the river boundary, the depression boundary and the road boundary to generate a third extreme point set;

[0028] Step 14: A corresponding relationship of each extreme point in each of the first extreme point set and the third extreme point set is established;

[0029] Step 15: A plurality of matching blocks meeting the number condition of extreme points are extracted from the high-altitude image according to the first extreme point set.

[0030] The application significantly improves the accuracy and robustness of feature block matching by innovatively constructing the "matching block information" containing the multi-dimensional relative position relationship. The core is to construct the matching block meeting the requirements by using the way that a large number of extreme points exist in the special boundary position, so that after the matching of the corresponding boundary is completed, the subsequent part can be re-matched for the second time according to the corresponding relationship of the net structure, so as to ensure the matching accuracy.

[0031] Further, the acquisition manner of the feature block includes the following steps:

[0032] Step 21: picking out the mountain boundary, river boundary, depression boundary and road boundary from the unmanned aerial vehicle image data;

[0033] Step 22: extracting a plurality of extreme points from the river boundary, depression boundary and road boundary to generate a second extreme point set;

[0034] Step 23: extracting extreme points from the remaining area of the mountain boundary, river boundary, depression boundary and road boundary in the unmanned aerial vehicle image data to generate a fourth extreme point set;

[0035] Step 24: extracting a plurality of feature blocks meeting the number condition of extreme points from the elevation image according to the second extreme point set.

[0036] In the technical solution provided in the present application, the feature block extraction method is basically the same as the matching block extraction method, so when the feature block and the matching block are matched with each other, the special boundaries are matched with each other first, and then the internal features are matched with each other after the special boundaries are successfully matched with each other, which can increase the matching accuracy in practice.

[0037] Step 4 includes the following steps:

[0038] Step 41: obtaining the mountain boundary, river boundary, depression boundary and road boundary in the aerial image data, and taking the mountain boundary, river boundary, depression boundary and road boundary as the matching area;

[0039] Step 42: dividing the matching area into a plurality of matching grids based on a preset grid, and the size of the matching grid is set according to a preset threshold;

[0040] Step 43: merging the matching block and the feature block in the reduced dimension matching grid into one matching group.

[0041] In the technical solution provided in the present application, the preset grid is used for dividing the matching group, which has high accuracy in practice and can ensure that the matching block and the feature block in different areas are matched with each other.

[0042] Further, step 5 includes the following steps:

[0043] Step 51: taking the similarity of the feature block and the matching block as the first matching element, taking the similarity of the texture features of the feature block and the matching block as the second matching element, and taking the ratio of the distance of the feature block relative to the reference object and the distance of the matching block relative to the reference object as the third matching element; the reference object is a specific physical marker set in the target area in advance;

[0044] Step 52: establishing a probability model based on the first matching element, the second matching element and the third matching element; and performing tensor decomposition on the probability model to reduce the model dimension.

[0045] The application significantly improves the matching accuracy and efficiency by innovative multi-dimensional matching element fusion and model optimization strategy, introduces spatial context information: on the basis of traditional similarity (first matching element) and texture similarity (second matching element), a new feature block and matching block relative to the direction difference of the common reference (third matching element) is added, which fully utilizes the implicit spatial relationship features, making the matching more comprehensive and robust.

[0046] Further, step 52 specifically includes the following steps:

[0047] Step 521: Obtain all feature blocks a and matching blocks b in the matching group, calculate the first matching element, the second matching element and the third matching element for each feature block a and matching block b, and convert the first matching element, the second matching element and the third matching element into the first matching probability, the second matching probability and the third matching probability;

[0048] Step 522: Based on the first matching probability, the second matching probability and the third matching probability, a probability model of the feature blocks and the matching blocks in the matching group is constructed;

[0049] 523: Tensor decomposition is performed on the probability model to reduce the model dimension.

[0050] In the technical scheme provided by the application, the model dimension can be effectively reduced by tensor decomposition of the probability model, the information density of the extracted information is increased, and the difficulty of fusing high-altitude image data and unmanned aerial vehicle image data is reduced.

[0051] Further, step 522 includes the following steps:

[0052] Step 5221: Set initial parameters;

[0053] , B= , , , ;

[0054] Wherein, A represents the first extreme point set of the matching block, B represents the third extreme point set of the matching block, respectively represent the first extreme point, the second extreme point and the Mth extreme point in the first extreme point set, respectively represent the first extreme point, the second extreme point and the Mth extreme point in the second extreme point set; , respectively represent the first matching probability, the second matching probability and the third matching probability of and ; ​

[0055] Step 5222: constructing a probability model :

[0056] .

[0057] Step 523 includes the following steps:

[0058] Step 5231: performing slice decomposition on the probability model :

[0059] , slice 1;

[0060] , slice 2;

[0061] , slice 3;

[0062] Step 5232: decomposing the probability model into a low-rank representation according to the slice results:

[0063] ;

[0064] wherein R represents a decomposition rank, represents a feature block factor vector, represents a matching block factor vector, a matching type factor vector, and r represents a summation index for traversing all decomposition ranks;

[0065] Step 5233: reconstructing the probability model to obtain a tensor-decomposed probability model :

[0066] .

[0067] In the technical solution provided in the present application, through the rank index r, the tensor decomposition compresses the high-dimensional matching relationship into a low-dimensional latent space while retaining the key matching patterns, thereby realizing efficient and interpretable matching modeling, and enabling the mining of the implicit connections between data and providing greater value for subsequent data utilization. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 is a flowchart of an information extraction method for multi-source remote sensing data. DETAILED DESCRIPTION

[0069] For the purposes of the present application, the technical solutions and advantages thereof, the technical solutions of the present application will be clearly and completely described in conjunction with specific embodiments below. The same reference signs in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0070] Compared with the embodiments shown in the drawings, the feasible implementation solutions within the protection scope of the present application can have fewer components, other components not shown in the drawings, different components, differently arranged components, or differently connected components, etc. In addition, two or more components in the drawings can be implemented in a single component, or a single component shown in the drawings can be implemented as a plurality of separate components.

[0071] Unless otherwise defined, the technical terms or scientific terms used herein should be understood as the usual meanings understood by those of ordinary skill in the art to which the present application belongs. The "first", "second", and similar words used in the specification and claims of the present application do not represent any order, quantity, or importance, but are only used to distinguish different components. Similarly, "one" or "a" and similar words do not necessarily represent a quantity limitation. "Up", "down", and the like are only used to represent a relative positional relationship, which may change accordingly when the absolute position of the described object changes.

[0072] Embodiment 1:

[0073] Reference Figure 1 The first embodiment of the present application discloses a multi-source remote sensing data information extraction method, comprising:

[0074] Step 1: Obtain high-altitude image data of a target area, and extract a plurality of matching blocks from the high-altitude image data; the matching blocks at least include a plurality of extreme points exceeding a preset threshold.

[0075] The matching block acquisition method includes the following steps:

[0076] Step 11: Select the mountain boundary, river boundary, depression boundary, and road boundary from the high-altitude image data.

[0077] Load the digital elevation model (DEM) data of the target area, and extract the mountain boundary, river boundary, depression boundary, and road boundary through a terrain analysis tool set.

[0078] Step 12: Extract a plurality of extreme points from the river boundary, depression boundary, and road boundary to generate a first extreme point set.

[0079] The extreme value point extraction algorithm provided in Embodiment 2 is used to extract extreme value points in the mountain boundary, the river boundary, the depression boundary, and the road boundary, and the number of the extreme value points is as large as possible to increase the accuracy of the feature region description, so that a first extreme value point set is obtained, which describes the feature information of the mountain boundary, the river boundary, the depression boundary, and the road boundary.

[0080] Step 13: Extreme value points are extracted from the remaining regions of the mountain boundary, the river boundary, the depression boundary, and the road boundary in the aerial image data, and a third extreme value point set is generated.

[0081] Similarly, extreme value points also need to be extracted from the remaining regions of the non-mountain boundary, the river boundary, the depression boundary, and the road boundary, and these extreme value points are not involved in matching at first, and are collected together as the third extreme value point set.

[0082] Step 14: The corresponding relationship of each extreme value point in the first extreme value point set and the third extreme value point set is established.

[0083] The corresponding relationship of each extreme value point in the first extreme value point set and the third extreme value point set is established, and the corresponding relationship is mainly the positional relationship, that is, after the positions of all extreme value points in the first extreme value point set are described on the blank map, the positions of the remaining extreme value points can be restored on the blank map according to the corresponding relationship.

[0084] Step 15: A plurality of matching blocks meeting the number condition of extreme value points are extracted from the aerial image data according to the first extreme value point set. The matching block is not a pixel point, but a specific region, which contains a plurality of pixel points and extreme value points. Based on the matching of the matching blocks with a large area, the size difference between the satellite image data and the unmanned aerial vehicle image data can be well measured.

[0085] Step 2: Obtain the unmanned aerial vehicle image data of the target region, and extract a plurality of feature blocks from the unmanned aerial vehicle image data. The feature blocks at least include a plurality of extreme value points exceeding a preset threshold number;

[0086] The feature block acquisition method includes the following steps:

[0087] Step 21: The mountain boundary, the river boundary, the depression boundary, and the road boundary are selected from the unmanned aerial vehicle image data;

[0088] Step 22: A plurality of extreme value points are extracted from the river boundary, the depression boundary, and the road boundary, and a second extreme value point set is generated;

[0089] Step 23: Extreme value points are extracted from the remaining regions of the mountain boundary, the river boundary, the depression boundary, and the road boundary in the unmanned aerial vehicle image data, and a fourth extreme value point set is generated;

[0090] Step 24: Extracting a plurality of feature blocks meeting the quantity condition of extreme points from the elevation image according to the second extreme point set.

[0091] The extraction manner of the feature block in step 2 is the same as that of the matching block in step 1, only the object is different, step 2 is to deal with unmanned aerial vehicle image data, unmanned aerial vehicle image data is the influence data of aerial photography by a man-machine. Relatively speaking, unmanned aerial vehicle image data is clearer and can represent more image information, but the flight height of the unmanned aerial vehicle is unstable and stable elevation information cannot be obtained. Generally, the image data collected by the unmanned aerial vehicle is used as material information and is rendered to the aerial image data. The specific rendering manner is not described herein. The present application is mainly used for determining the corresponding relationship between the unmanned aerial vehicle image data and the aerial image data, that is, how to accurately render the unmanned aerial vehicle image data to the aerial image data is solved.

[0092] The above gives the acquisition manner of the matching block and the feature block, but the following requirements need to be met when acquiring the matching block and the feature block:

[0093] (1) The physical size of the real world corresponding to the matching block is consistent with the physical size of the real world corresponding to the feature block; the size of the matching block is set according to the matching accuracy.

[0094] (2) The number of extreme points in the feature block and the matching block meets the following condition:

[0095] ;

[0096] x, y represent the center point coordinates of the feature block or the matching block, x i and y i represent the coordinates of the i-th extreme point, K represents the kernel function, n represents the total number of extreme points, and h represents the bandwidth parameter.

[0097] For condition 1, it is only necessary to set reasonably according to the measurement accuracy of the aerial image data and the unmanned aerial vehicle image data. For example, in the present application, the size of the matching block and the feature block in the real physical world is set to 10 meters 10 meters. In the remaining manner, it can be set according to the demand.

[0098] For condition 2, it is necessary to increase the collection performance of the extreme points, and the present application mainly collects the extreme points of the boundary of the mountain body, the boundary of the river, the boundary of the depression and the boundary of the road, which has obvious topographic change characteristics and is easier to form extreme points.

[0099] Step 3: Obtaining the first coordinate information by acquiring the measurement coordinates of each matching block in the aerial image data, and obtaining the second coordinate information by acquiring the measurement coordinates of each feature block in the unmanned aerial vehicle image data.

[0100] The first coordinate information and the second coordinate information are actually the coordinates of the matching block and the feature block in the physical world. In the present scheme, the coordinate information is understood as the longitude and latitude coordinates of the center points of the matching block and the feature block.

[0101] Step 4: According to the first coordinate information and the second coordinate information, the matching block and the feature block are matched, and the matching block and the feature block with a distance not exceeding a preset threshold are merged into one matching group.

[0102] Within a certain range, the matching block and the feature block correspond to each other, but the corresponding relationship will have a comparison error due to the measurement offset. Therefore, in one matching group, the matching block and the feature block can correspond to each other, but the corresponding relationship will have a certain offset. Therefore, the present scheme divides the target region into a plurality of matching groups. Each matching group performs corresponding matching of the feature block and the matching block. Relatively, the overall number of matching blocks and feature blocks that need to be considered during matching can be reduced, and the accuracy of matching can be ensured.

[0103] Step 4 includes the following steps:

[0104] Step 41: Obtain the mountain boundary, river boundary, depression boundary and road boundary in the aerial image data, and take the mountain boundary, river boundary, depression boundary and road boundary as the matching region;

[0105] Step 42: Divide the matching region into a plurality of matching grids based on a preset grid, and the size of the matching grid is set according to a preset threshold;

[0106] Step 43: Merge the matching block and the feature block in the dimension-reduced matching grid into one matching group.

[0107] The matching group is actually a matching of the feature block and the matching block in a larger range. For example, the size of the feature block and the matching block in the real world is 10 meters 10 meters, and the size of each grid is 100 meters 100 meters. Therefore, there are 100 matching groups and 100 feature groups that need to be matched in one matching group.

[0108] Step 5: For each matching group, a probability model of mutual registration of the matching block and the feature block is constructed, and the probability model is tensor decomposed to obtain registration information of one-to-one correspondence of the matching block and the feature block.

[0109] There are various ways to mutually register the feature block and the matching block in the matching group, and each way is theoretically possible. Therefore, the present application provides a probability model of mutual matching of the feature block and the matching block. After obtaining the probability model, the subsequent fusion of the aerial image data and the unmanned aerial vehicle image data has higher precision.

[0110] Step 5 comprises the following steps:

[0111] Step 51: taking the similarity between the feature block and the matching block as a first matching element, taking the similarity between the texture features of the feature block and the matching block as a second matching element, and taking the ratio of the distance of the feature block relative to a reference object and the distance of the matching block relative to the reference object as a third matching element; the reference object is a specific physical marker preset in the target region.

[0112] The first matching element is the similarity between the feature block and the matching block, which mainly converts all the pixels in the feature block and the matching block into vector representations, and then calculates the similarity between the two vectors. The similarity between the feature block and the matching block measures the global similarity between the two, but lacks detailed features.

[0113] The second matching element is mainly the similarity between the texture features of the feature block and the matching block, that is, the texture information is extracted from the feature block, and the texture information is extracted from the matching block, and the similarity obtained by comparing the two texture information is obtained.

[0114] The extraction method of the texture information is as follows:

[0115] The texture information is R.

[0116] .

[0117] Where i and j represent the horizontal and vertical coordinates of the pixel points of the feature block or the matching block, and N represents the maximum gray value, is the frequency of the feature block or the matching block .

[0118] The third matching element is the ratio of the distance of the feature block relative to a reference object and the distance of the matching block relative to the reference object. The distance here is the vector distance, that is, the length information with direction. The reference object is a pre-set landmark, such as a ruler.

[0119] The third matching element can measure the corresponding positions of the feature block and the matching block in the real world. When the third matching element is close to 1, it means that the matching probability of the two is higher.

[0120] Step 52: establishing a probability model based on the first matching element, the second matching element and the third matching element, and performing tensor decomposition on the probability model to reduce the dimension of the model.

[0121] Further, step 52 specifically comprises the following steps:

[0122] Step 521: Obtain all feature blocks a and matching blocks b in the matching group, and calculate the first matching element, the second matching element and the third matching element for each feature block a and matching block b, and convert the first matching element, the second matching element and the third matching element into the first matching probability, the second matching probability and the third matching probability.

[0123] When the first matching element and the second matching element are converted into the first matching probability and the second matching probability, they only need to be multiplied by the preset weight parameters respectively.

[0124] The closer the third matching element is to 1, the higher the matching probability of the matching block and the feature block is. Therefore, when the third matching element is converted into the third matching probability, the following formula needs to be used:

[0125] Step 522: Construct the probability model of the feature blocks and the matching blocks in the matching group based on the first matching probability, the second matching probability and the third matching probability.

[0126] Further, step 522 includes the following steps:

[0127] Step 531: Obtain all feature blocks a and matching blocks b in the matching group, and calculate the first matching element, the second matching element and the third matching element for each feature block a and matching block b, and convert the first matching element, the second matching element and the third matching element into the first matching probability, the second matching probability and the third matching probability.

[0128] Step 532: Construct the probability model of the feature blocks and the matching blocks in the matching group based on the first matching probability, the second matching probability and the third matching probability.

[0129] Specifically, step 532 includes the following steps:

[0130] Step 5321: Set the initial parameters;

[0131] , B= , , , ;

[0132] wherein A represents the first extreme point set of the matching block, B represents the third extreme point set of the matching block, respectively represent the first extreme point, the second extreme point and the Mth extreme point in the first extreme point set, respectively represent the first extreme point, the second extreme point and the Mth extreme point in the second extreme point set; , , respectively represent and a first matching probability, a second matching probability, and a third matching probability;

[0133] Step 5322: constructing a probability model :

[0134] ;

[0135] After the probability model is constructed, the data dimension of the probability model is too high, and it is difficult to perform feature fusion subsequently. It is necessary to reduce the dimension of the probability model.

[0136] 533: performing tensor decomposition on the probability model to reduce the dimension of the model.

[0137] Step 533 includes the following steps:

[0138] Step 531: performing slice decomposition on the probability model :

[0139] Slice 1;

[0140] Slice 2;

[0141] Slice 3;

[0142] Step 532: according to the slice result, decomposing the probability model into a low-rank representation:

[0143] ;

[0144] wherein R represents a decomposition rank, represents a feature block factor vector, represents a matching block factor vector, a matching type factor vector, and r represents a summation index for traversing all decomposition ranks.

[0145] Specifically, is used to describe the latent representation of the feature block, is used to describe the latent representation of the matching block, is used to describe the matching type weight.

[0146] Step 533: Step 5233: reconstructing the probability model to obtain a tensor-decomposed probability model :

[0147] .

[0148] The technical scheme provided in the application can more intuitively indicate the corresponding relationship between the high-altitude image data and the unmanned aerial vehicle image data after reconstructing the joint probability, thereby providing a basis for subsequent rendering of the three-dimensional model.

[0149] In the technical scheme provided in the application, the matching relationship between the first extreme point set and the second extreme point set is represented by a probability model after tensor decomposition After the corresponding relationship between the first extreme point set and the second extreme point set is determined, the corresponding relationship between the third extreme point set and the fourth extreme point set is also determined.

[0150] In this way, the application extracts the correlation characteristics between the two data from the high-altitude image data and the unmanned aerial vehicle image data by using the probability model, thereby providing a basis for rendering the high-altitude image data from the subsequent unmanned aerial vehicle image data.

[0151] Embodiment 2: When performing image registration, the application needs to extract extreme points and needs to extract as many extreme points as possible, and for this purpose, the application provides the following technical scheme:

[0152] Calculate the gradient amplitude and gradient direction of each pixel point in the image to be extracted, wherein the gradient amplitude includes a gradient value of 0°, a gradient value of 45°, a gradient value of 90°, and a gradient value of 135°.

[0153] ; ;

[0154] ; .

[0155] Correspondingly, the calculation formulae of the gradient amplitude of the pixel point and the gradient direction of the pixel point are as follows:

[0156] ;

[0157] ;

[0158] wherein U(x, y) is the gradient amplitude of the pixel point (x, y), a(x, y) is the gradient direction of the pixel point (x, y), is the gradient value of 0°, is the gradient value of 45°, is the gradient value of 90°, is the gradient value of 135°.

[0159] All the pixel points of the image to be extracted are traversed one by one, and if the gradient amplitude of the pixel point is a local maximum value in the gradient direction, the pixel point is taken as an extreme point.

[0160] Compared with the existing extraction algorithm, four directions are selected for extracting extreme points in the application, so that more extreme points can be extracted, and the accuracy of the registration of high-altitude image data and unmanned aerial image data is increased.

[0161] The preferred embodiments of the present application have been described above, but the present application is not limited to the above, and various modifications and changes can be made by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for information extraction of multi-source remote sensing data, characterized in that, The method comprises the following steps: Step 1: obtaining high-altitude image data of a target area, and extracting a plurality of matching blocks from the high-altitude image data; The matching blocks comprise at least a number of extreme points exceeding a preset threshold value; Step 2: obtaining unmanned aerial vehicle image data of the target area, and extracting a plurality of feature blocks from the unmanned aerial vehicle image data; The feature blocks comprise at least a number of extreme points exceeding a preset threshold value; Step 3: obtaining measurement coordinates of each matching block in the high-altitude image data to obtain first coordinate information, and obtaining measurement coordinates of each feature block in the unmanned aerial vehicle image data to obtain second coordinate information; Step 4: performing fuzzy matching on the matching blocks and the feature blocks according to the first coordinate information and the second coordinate information, and merging the matching blocks and the feature blocks with a distance not exceeding a preset threshold value into one matching group; Step 5: constructing a probability model of mutual registration of the matching blocks and the feature blocks for each matching group, and performing tensor decomposition on the probability model to obtain registration information corresponding to the matching blocks and the feature blocks; Step 5 comprises the following steps: Step 51: taking similarity of the feature blocks and the matching blocks as a first matching element, taking similarity of texture features of the feature blocks and the matching blocks as a second matching element, and taking a ratio of distances of the feature blocks relative to a reference object and distances of the matching blocks relative to the reference object as a third matching element; the reference object is a specific physical marker arranged in the target area in advance; Step 52: establishing a probability model based on the first matching element, the second matching element and the third matching element; and performing tensor decomposition on the probability model to reduce the dimension of the model; Step 52 specifically comprises the following steps: Step 521: obtaining all the feature blocks A and the matching blocks B in the matching group, calculating the first matching element, the second matching element and the third matching element for each of the feature blocks A and the matching blocks B, and converting the first matching element, the second matching element and the third matching element into a first matching probability, a second matching probability and a third matching probability; Step 522: constructing a probability model of the feature blocks and the matching blocks in the matching group based on the first matching probability, the second matching probability and the third matching probability; Step 522 comprises the following steps: Step 5221: setting initial parameters; 、B= 、 、 、 ； Wherein, A represents the first extreme point set of the matching block, B represents the third extreme point set of the matching block, respectively represent the first extreme point, the second extreme point and the Mth extreme point in the first extreme point set, respectively represent the first extreme point, the second extreme point and the Mth extreme point in the second extreme point set; 、 、 respectively represent and the first matching probability, the second matching probability and the third matching probability of Step 5222: Constructing the probability model : ； Step 523: performing tensor decomposition on the probability model to reduce the dimension of the model; Step 523 comprises the following steps: Step 5231 : performing slice decomposition on the probability model performing slice decomposition: Slice 1 ; slice 2; Slice 3; Step 5232: Based on the slice results, the probabilistic model is decomposed into a low-rank representation: ； where R denotes the decomposition rank, denotes the characteristic block factor vector, denotes the matching block factor vector, the matching type factor vector, r denotes the summation index for iterating over all decomposition ranks; Step 5233: obtaining the probability model reconstructing to obtain the probability model after tensor decomposition : ； The first matching element is similarity of the feature blocks and the matching blocks, all the pixels in the feature blocks and the matching blocks are converted into vector representation, and then similarity of the two vectors is calculated; The second matching element is similarity of texture features of the feature blocks and the matching blocks, texture information is extracted from the feature blocks, texture information is extracted from the matching blocks, and similarity obtained by comparing the two texture information is obtained; The third matching element is a ratio of distances of the feature blocks relative to a reference object and distances of the matching blocks relative to the reference object, the distances are vector distances. 2.The method of claim 1, wherein, The real-world physical size corresponding to the matching blocks is consistent with the real-world physical size corresponding to the feature blocks; the size of the matching blocks is set according to matching accuracy. 3.The method of claim 2, wherein, The number of extreme points in the feature blocks and the matching blocks satisfies the following condition: ； x, y represent the coordinates of the center point of the feature block or the matching block, x i and y i represent the coordinates of the i-th extreme point, K represents the kernel function, n represents the total number of extreme points, and h represents the bandwidth parameter. 4.The method of claim 1, wherein, The first coordinate information is satellite positioning information obtained by the satellite, and the second coordinate information is radar positioning information obtained by the unmanned aerial vehicle; the size of the feature block and the matching block is determined based on the minimum measurement accuracy of the satellite positioning information. 5.The method of claim 1, wherein, The acquisition method of the matching block comprises the following steps: Step 11: selecting the mountain boundary, river boundary, depression boundary and road boundary from the high-altitude image data; Step 12: extracting a plurality of extreme points from the river boundary, depression boundary and road boundary to generate a first extreme point set; Step 13: extracting extreme points from the remaining area of the high-altitude image data except the mountain boundary, river boundary, depression boundary and road boundary to generate a third extreme point set; Step 14: establishing a corresponding relationship of each extreme point in each of the first extreme point set and the third extreme point set; Step 15: extracting a plurality of matching blocks meeting the quantity condition of extreme points from the high-altitude image according to the first extreme point set; the corresponding relationship is a positional relationship. 6.The method of claim 5, wherein, The acquisition method of the feature block comprises the following steps: Step 21: selecting the mountain boundary, river boundary, depression boundary and road boundary from the unmanned aerial vehicle image data; Step 22: extracting a plurality of extreme points from the river boundary, depression boundary and road boundary to generate a second extreme point set; Step 23: extracting extreme points from the remaining area of the unmanned aerial vehicle image data except the mountain boundary, river boundary, depression boundary and road boundary to generate a fourth extreme point set; Step 24: extracting a plurality of feature blocks meeting the quantity condition of extreme points from the high-altitude image according to the second extreme point set. 7.The method of claim 5, wherein, Step 4 comprises the following steps: Step 41: acquiring the mountain boundary, river boundary, depression boundary and road boundary in the high-altitude image data, and taking the mountain boundary, river boundary, depression boundary and road boundary as a matching area; Step 42: dividing the matching area into a plurality of matching grids based on a preset grid, and the size of the matching grid is set according to a preset threshold; Step 43: merging the matching block and the feature block in the reduced dimension matching grid into one matching group.

Images