A POI data-based remote sensing image semantic label extraction method

Abstract

The application belongs to the technical field of space-time data analysis, and discloses a POI data-based remote sensing image semantic label extraction method, which realizes the extraction of the semantic content label of the remote sensing image by analyzing the POI data in the coverage area of the remote sensing image from the perspectives of place names and land types. Firstly, a land type label set and a place name label set as well as a density threshold and a coverage threshold of the land type label and the place name label are preset. Then, a POI data set in the coverage range of the remote sensing image is calculated and obtained, the attributes of the POI data set are subjected to Bert semantic similarity calculation with the land type label set and the place name label set, and labels with high similarity are obtained as a pre-extracted label set of the land type and the place name. Subsequently, the coverage and the density of each label in the pre-extracted label set are calculated, and labels with the coverage and the density both greater than the density threshold and the coverage threshold are screened out as the final semantic label extraction result. The method solves the problems of high cost, great difficulty and poor effect in the extraction of the semantic label in the prior art.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing data analysis, and relates to a POI data-based remote sensing image semantic label extraction method, more specifically, to a method of analyzing various ground objects in a remote sensing image through position information recorded in POI data and automatically extracting semantic labels that can represent the content of the remote sensing image. BACKGROUND

[0002] With the gradual development of various remote sensing satellites such as visible light, SAR, infrared, and the like, remote sensing image data is growing explosively. How to automatically and accurately obtain semantic content labels such as ground objects, landmarks, and targets in remote sensing images is a prerequisite for distributing and applying remote sensing data to various industries such as agriculture, forestry, animal husbandry, sideline, and fishery. For example, to provide remote sensing data useful for the business of the Hebei Water Conservancy Bureau, semantic content labels of all remote sensing image data need to be extracted, and images containing a large number of semantic content labels such as rivers and lakes are selected as data provided to the water conservancy bureau. On the contrary, if a remote sensing image without or with few water bodies is provided to the water conservancy bureau, it is likely to be of no help to its business application.

[0003] However, the current method for extracting semantic content labels of remote sensing images mainly performs target detection, ground object classification, and the like machine learning or deep learning processing on remote sensing images. This type of method requires a large amount of labeled training data, has a high cost, and still has poor final extraction accuracy, which still has deficiencies in actual application. POI data, also known as point of interest data, refers to point data in an Internet electronic map, which basically includes four attributes of name, address, coordinate, and category. Electronic map suppliers such as Gaode and Baidu and related surveying and mapping units have established a relatively complete POI data set, which has basically covered most areas in China. The name, address, and category attribute fields in the POI data implicitly contain semantic content labels such as ground classes and place names, and the coordinate field in the POI data also records the corresponding detailed position. Therefore, it is a feasible method to analyze the POI data in the coverage area of a remote sensing image to extract the semantic content labels of the remote sensing image.

[0004] However, a remote sensing image covers a large amount of POI data, and some even reach several thousand or even tens of thousands. If the attribute fields of several thousand or tens of thousands of POI data are directly used as semantic labels of remote sensing images, the semantic labels will not be representative due to the excessive amount. It is urgent to study a POI data-based remote sensing image semantic label extraction method to quickly mine semantic labels that can best represent the content of remote sensing images from the attributes of massive POI data, and realize the analysis of the content of remote sensing images by POI data-assisted personnel. SUMMARY

[0005] The present application aims at overcoming the problems of high cost and poor effect of traditional semantic content label extraction methods such as feature classification and target detection, and proposes a remote sensing image semantic label extraction method based on POI data, which extracts labels from the name, address and category attribute fields in the POI data within the coverage range of the remote sensing image, filters the label coverage and the label density, and finally obtains effective semantic content labels that can reasonably represent the content of the remote sensing image.

[0006] The technical scheme adopted by the present application is:

[0007] A remote sensing image semantic label extraction method based on POI data, comprising the following steps:

[0008] (1) A set of land class label set {landclass1, landclass2,..., landclass m} and a set of land name label set {landname1, landname2,..., landname m} are designed in advance, m is the number of labels; at the same time, the land class label density threshold, the land class label coverage threshold, the land name label density threshold and the land name label coverage threshold are set in advance;

[0009] (2) The latitude and longitude range of the remote sensing image to be extracted and the coordinate data in the POI data are intersected to obtain the POI data set {Point1, Point2,..., Point n} corresponding to the coverage range of the remote sensing image, n is the number of POI data, wherein the kth POI data Point k is represented as: Point k ={name k , address k , coordinate k , category k};

[0010] (3) Select one label landclass i from the land class label set, calculate the approximate degree of landclass i and the name attribute and category attribute in each data of the POI data set by using Bert, and combine the POI data with an approximate degree greater than zero into the approximate POI set of landclass i ; at the same time, select one label landname i from the land name label set, calculate the approximate degree of landname iBased on the similarity of the name and address attributes in each data point in the POI dataset, POI data with a similarity greater than zero are grouped into a landname. i The approximate set of POIs; where 1≤i≤m;

[0011] (4) If landclass i If the approximate set of POIs is empty, it indicates that the land class i It is not a semantic label corresponding to the remote sensing image, if landclass i If the set of approximate POIs is not empty, then landclass i The land use pre-extracted labels belonging to the corresponding remote sensing image are used to obtain the set of land use pre-extracted labels for the remote sensing image: {landclasslabel1, landclasslabel2, ..., landclasslabel}. l}, l≤m; if landname i If the approximate set of POIs is empty, it indicates that the landname i It is not a semantic label corresponding to the remote sensing image, if the landname i If the set of approximate POIs is not empty, then landname i The pre-extracted place name labels belonging to the corresponding remote sensing image are used to obtain the set of pre-extracted place name labels for the remote sensing image: {landnamelable1,landnamelable2, ...,landnamelable...} t}, t≤m;

[0012] (5) Select one label, landclasslabel, from the pre-extracted label set of land categories in sequence. i ,1≤i≤l, will landclasslable i The outer envelope of all data in the approximate POI set is calculated, and the outer envelope area is obtained. The outer envelope area is then divided by the remote sensing image coverage area to obtain the landclass label. i Land class label coverage, and landclass label i The landclass label is obtained by dividing the number of data points in the approximate POI set by the area of ​​the outer envelope. i The density of land category tags; simultaneously, select one tag, landnamelabel, from the pre-extracted tag set of place names. i ,1≤i≤t, will be landnamelable i The outer envelope of all data in the approximate POI set is calculated, and the outer envelope area is obtained. The outer envelope area is then divided by the remote sensing image coverage area to obtain the landname label.i The coverage rate of place name tags, and the landname label i The landname label is obtained by dividing the number of data points in the approximate POI set by the area of ​​the outer envelope. i The density of place name tags;

[0013] (6) Sequentially add the landclass labels i The land class label coverage and density are compared with preset land class label density thresholds and land class label coverage thresholds, respectively. If both are greater than the thresholds, the corresponding land class label is retained. i Otherwise, landclasslabel i It is not used as the land cover label for the corresponding remote sensing image; at the same time, the land class label is added sequentially. i The place name label coverage and place name label density are compared with preset place name label density thresholds and place name label coverage thresholds, respectively. If both are greater than the thresholds, the corresponding place name label (landclass label) is retained. i Otherwise, landclasslabel i Not used as place name labels for the corresponding remote sensing image;

[0014] (7) Merge the remaining land category labels in the land category pre-extracted label set and the remaining place name labels in the place name pre-extracted label set to obtain the semantic labels of the remote sensing image.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] This invention proposes a method for extracting semantic tags from remote sensing images based on POI data. It extracts semantic content tags from remote sensing images by analyzing POI data within the coverage area of ​​the remote sensing image from the perspectives of place name and land category. This solves the problems of high cost, difficulty and poor effect of extracting semantic tags through machine learning or deep learning such as target detection and land feature classification. Attached Figure Description

[0017] Figure 1 This is the overall process framework design diagram of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0019] Figure 1 is a principle flow framework chart of a specific implementation of the POI data based remote sensing image semantic label extraction method of the present application.

[0020] In the present embodiment, the POI data based remote sensing image semantic label extraction method as shown in Figure 1 includes the following steps:

[0021] (1) A set of land class labels {landclass1, landclass2, …, landclass m} and a set of land name labels {landname1, landname2, …, landname m} are designed in advance, and m is the number of labels; wherein the land class labels include playground, farmhouse, shopping mall, scenic spot, lake, etc., and the land name labels include Yellow Crane Tower, Yellow Mountain, Haidian District, Hengshui Lake, etc.; at the same time, the land class label density threshold, the land class label coverage threshold, the land name label density threshold and the land name label coverage threshold are set in advance;

[0022] (2) Each POI data contains name, address, coordinates and category, and the kth POI data is represented as Point k . The latitude and longitude range of the remote sensing image to be extracted and the coordinate data in the POI data are intersected to obtain the POI data set {Point1, Point2, …, Point m} in the coverage range of the remote sensing image, wherein Point k can be represented as:

[0023] Point k ={name k , address k , coordinates k , category k};

[0024] (3) One label landclass i in the land class label set is selected in turn, and Bert is used to calculate the approximate degree of landclass i and the name attribute and the category attribute in each data of the POI data set, and the POI data with the approximate degree greater than zero is combined into the approximate POI set of landclass i ; at the same time, one label landname i in the land name label set is selected in turn, and Bert is used to calculate the approximate degree of landname iBased on the similarity of the name and address attributes in each data point in the POI dataset, POI data with a similarity greater than zero are grouped into a landname. i The approximate set of POIs; where 1≤i≤m;

[0025] (4) If landclass i If the approximate set of POIs is empty, it indicates that the land class i It is not a semantic label corresponding to the remote sensing image, if landclass i If the set of approximate POIs is not empty, then landclass i The land use pre-extracted labels belonging to the corresponding remote sensing image are used to obtain the set of land use pre-extracted labels for the remote sensing image: {landclasslabel1, landclasslabel2, ..., landclasslabel}. l}, l≤m; if landname i If the approximate set of POIs is empty, then it indicates that the landname i It is not a semantic label corresponding to the remote sensing image, if the landname i If the set of approximate POIs is not empty, then landname i The pre-extracted place name labels belonging to the corresponding remote sensing image are used to obtain the set of pre-extracted place name labels for the remote sensing image: {landnamelable1,landnamelable2, ...,landnamelable...} t}, t≤m;

[0026] (5) Select one label, landclasslabel, from the pre-extracted label set of land categories in sequence. i ,1≤i≤l, will landclasslable i The outer envelope of all data in the approximate POI set is calculated, and the outer envelope area is obtained. The outer envelope area is then divided by the remote sensing image coverage area to obtain the landclass label. i Land class label coverage, and landclass label i The landclass label is obtained by dividing the number of data points in the approximate POI set by the area of ​​the outer envelope. i The density of land category tags; simultaneously, select one tag, landnamelabel, from the pre-extracted tag set of place names. i ,1≤i≤t, will be landnamelable i The outer envelope of all data in the approximate POI set is calculated, and the outer envelope area is obtained. The outer envelope area is then divided by the remote sensing image coverage area to obtain the landname label.i landnamelable i landnamelable i landnamelable

[0027] (6)landclasslable i landclasslable i landclasslable i landclasslable i landclasslable i landclasslable i landclasslable

[0028] (7)landclasslable

[0029] Although the above describes the specific embodiments of the present application in detail, so as to facilitate the understanding of the present application by those skilled in the art, it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, as long as various changes are within the spirit and scope of the present application defined and determined by the appended claims, all the application and creation utilizing the concept of the present application are within the protection.

Claims

1. A method for extracting semantic tags from remote sensing images based on POI data, characterized in that, Includes the following steps: (1) Pre-design a set of land class labels {landclass1,landclass2,...,landclass...} m } and a set of place name tags {landname1,landname2,...,landname} m }, where m is the number of tags; at the same time, the thresholds for land category tag density, land category tag coverage, place name tag density, and place name tag coverage are preset; (2) Find the intersection of the latitude and longitude range of the remote sensing image from which semantic labels are to be extracted and the coordinate data in the POI data to obtain the POI data set {Point1, Point2, ..., Point...} within the coverage area of ​​the corresponding remote sensing image. n }, where n is the number of POI data points, and the k-th POI data point is... k Represented as: Point k ={name k ,address k ,coordinate k ,category k }; (3) Select one label from the set of land class labels in sequence. i BERT is used to calculate the landclass separately. i Based on the similarity of the name and category attributes in each data point in the POI dataset, POI data with a similarity greater than zero are grouped into a landclass. i A set of approximate POIs; simultaneously, select one label, "landname," from the set of place name labels in sequence. i BERT is used to calculate the landname separately. i Based on the similarity of the name and address attributes in each data point in the POI dataset, POI data with a similarity greater than zero are grouped into a landname. i The approximate set of POIs; where 1≤i≤m; (4) If landclass i If the approximate set of POIs is empty, it indicates that the land class i It is not a semantic label corresponding to the remote sensing image, if landclass i If the set of approximate POIs is not empty, then landclass i The land use pre-extracted labels belonging to the corresponding remote sensing image are used to obtain the set of land use pre-extracted labels for the remote sensing image: {landclasslabel1, landclasslabel2, ..., landclasslabel}. l }, l≤m; if landname i If the approximate set of POIs is empty, then it indicates that the landname i It is not a semantic label corresponding to the remote sensing image, if the landname i If the set of approximate POIs is not empty, then landname i The pre-extracted place name labels belonging to the corresponding remote sensing image are used to obtain the set of pre-extracted place name labels for the remote sensing image: {landnamelable1,landnamelable2, ...,landnamelable...} t }, t≤m; (5) Select one label, landclasslabel, from the pre-extracted label set of land categories in sequence. i ,1≤i≤l, will landclasslable i The outer envelope of all data in the approximate POI set is calculated, and the outer envelope area is obtained. The outer envelope area is then divided by the remote sensing image coverage area to obtain the landclass label. i Land class label coverage, and landclass label i The landclass label is obtained by dividing the number of data points in the approximate POI set by the area of ​​the outer envelope. i The density of land category tags; simultaneously, select one tag, landnamelabel, from the pre-extracted tag set of place names. i ,1≤i≤t, will be landnamelable i The outer envelope of all data in the approximate POI set is calculated, and the outer envelope area is obtained. The outer envelope area is then divided by the remote sensing image coverage area to obtain the landname label. i The coverage rate of place name tags, and the landname label i The landname label is obtained by dividing the number of data points in the approximate POI set by the area of ​​the outer envelope. i The density of place name tags; (6) Sequentially add the landclass labels i The land class label coverage and density are compared with preset land class label density thresholds and land class label coverage thresholds, respectively. If both are greater than the thresholds, the corresponding land class label is retained. i Otherwise, landclasslabel i It is not used as the land cover label for the corresponding remote sensing image; at the same time, the land class label is added sequentially. i The place name label coverage and place name label density are compared with preset place name label density thresholds and place name label coverage thresholds, respectively. If both are greater than the thresholds, the corresponding place name label (landclass label) is retained. i Otherwise, landclasslabel i Not used as place name labels for the corresponding remote sensing image; (7) Merge the remaining land category labels in the land category pre-extracted label set and the remaining place name labels in the place name pre-extracted label set to obtain the semantic labels of the remote sensing image.

Images