CN109063754B - Remote sensing image multi-feature joint classification method based on OpenStreetMap - Google Patents

Remote sensing image multi-feature joint classification method based on OpenStreetMap Download PDF

Info

Publication number
CN109063754B
CN109063754B CN201810792087.9A CN201810792087A CN109063754B CN 109063754 B CN109063754 B CN 109063754B CN 201810792087 A CN201810792087 A CN 201810792087A CN 109063754 B CN109063754 B CN 109063754B
Authority
CN
China
Prior art keywords
remote sensing
class
samples
classification
sensing image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201810792087.9A
Other languages
Chinese (zh)
Other versions
CN109063754A (en
Inventor
卢其楷
万太礼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN201810792087.9A priority Critical patent/CN109063754B/en
Publication of CN109063754A publication Critical patent/CN109063754A/en
Application granted granted Critical
Publication of CN109063754B publication Critical patent/CN109063754B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images

Abstract

The invention discloses a remote sensing image multi-feature classification method based on OpenStreetMap, which extracts various spatial features from a remote sensing image to describe the characteristics of different aspects of the image, removes object boundary confusion pixels by adopting morphological corrosion filtering, performs intra-class clustering analysis aiming at each class, removes error marks, obtains an accurate classification sample, combines the obtained image features and the classification sample, combines a support vector machine classifier, calculates the probability that each pixel belongs to different classes under different features, combines the classification reliability to realize the weighted fusion of the class probabilities corresponding to the different features, and completes the remote sensing image classification. According to the invention, by mining the spatial information of the remote sensing image, various classical spatial feature extraction algorithms are effectively integrated, the separability of different classes can be enhanced, and more accurate classification results of each class and higher classification precision can be obtained.

Description

Remote sensing image multi-feature joint classification method based on OpenStreetMap
Technical Field
The invention relates to the technical field of remote sensing image processing, in particular to a multi-feature joint classification method, and especially relates to a remote sensing image multi-feature joint classification joint method based on OpenStreetMap.
Background
With the development of WEB 2.0, the public can not only acquire knowledge from the internet, but also upload own knowledge and experience to the internet to share information with other people. Among them, Information with a Geographical position contributed by the public is called spontaneous geographic Information (VGI). Among them, openstreetmap (osm) is one of the most typical VGI projects, which aims to draw an open-source free world map by gathering public power and finally serve the public. The public can record the track as a measuring means by using a portable GPS device, and can also draw a map by using a Bing image as a base map according to the knowledge of the local area.
In recent years, researchers have applied OSM data to remote sensing image classification and achieved a good effect. The remote sensing image is an important data source for researching earth surface coverage and sensing earth changes, and can reflect the coverage condition of the earth surface. With the development of imaging elements and technologies, the spatial resolution of remote sensing images is higher and higher, more detailed earth surface spatial structure information is provided, and more accurate data support is provided for earth observation. The classification technology is combined with training samples and image features, can automatically judge the class information of unlabelled samples on the images, and is an important means for interpreting remote sensing image information. However, the method of labeling samples by field investigation is time and labor consuming and does not allow a sufficient amount of training samples to be obtained quickly. The OSM can provide a large number of training samples for remote sensing image classification, can effectively reduce the workload of experts, and provides a new way for rapidly acquiring the remote sensing image classification samples.
The existing method extracts classification samples from the labeled information in the OSM data and classifies images according to the spectral characteristics of pixels. However, the data quality of the OSM is greatly different in different areas due to the data acquisition mode of the OSM. In some regions with high public engagement, especially in some developed countries, OSM data is already quite abundant, and the quality of the data is comparable to that produced by professional organizations. However, limited by the participation degree of the public, the data volume of most regions is limited, certain errors exist in the aspects of spatial position and category attribute, and the OSM directly used as a training sample is difficult to adapt to complex remote sensing image scenes. Meanwhile, the current remote sensing image classification method based on the OSM is usually based on the spectral characteristics of the image, and the abundant spatial structure information of the remote sensing image cannot be effectively utilized. Due to the complexity of the attributes of the ground objects of the remote sensing images, the classification mode based on the image spectrum is difficult to accurately predict the pixel classes. The characteristics of the spatial distribution of the ground objects on the image are excavated, the spatial information is obtained, the defects of spectral characteristics can be effectively made up, and the classification precision of the remote sensing image is favorably improved. However, the spatial structure characteristics of different ground features are difficult to be reflected comprehensively by a single spatial feature, and the advantages of different features can be organically combined by a multi-feature combined classification mode to obtain a better result than a single feature. The current multi-feature fusion mode is usually to form a feature input classifier with higher dimensionality through vector superposition for interpretation, and the contribution of different features to a classification result is not considered, so that the classification precision is not favorably improved further.
Disclosure of Invention
The invention provides a remote sensing image multi-feature combined classification method based on OpenStreetMap, aiming at the problems in the prior art, the method extracts feature mark information from an OSM for generating classification samples, adopts morphological processing and sample clustering means to reduce false marks in the OSM, obtains accurate and reliable classification samples, extracts various features from the remote sensing image to describe the spatial distribution of the features, obtains classification results corresponding to different features, outputs and calculates the reliability of different classification results according to class probability, weights and fuses the classification results corresponding to the multiple features, and completes the classification of the remote sensing image.
The technical scheme of the invention is a remote sensing image multi-feature classification method based on OpenStreetMap, which comprises the following steps:
1) registering OpenStreetMap (OSM) with the remote sensing image to be classified, selecting a vector layer corresponding to the interested category from the OSM according to the actual classification requirement, and converting the vector layer into a raster image;
2) extracting a plurality of different spatial features of the remote sensing image to be classified, and describing the distribution characteristics of the ground features of the remote sensing image to be classified from a plurality of aspects;
3) performing morphological processing on the OSM grid images corresponding to the categories to obtain a sample of removing target edge pixels and reserving center pixels;
4) performing cluster analysis on the samples of each category according to the characteristics of the pixels by using a clustering algorithm, removing the samples which are marked by errors according to the distribution of the clustering centers of each subclass and the number of the samples in each subclass,
5) classifying the remote sensing images to be classified by using the acquired samples and using various different spatial features respectively in combination with a support vector machine, and acquiring the probability that each pixel belongs to different classes under different spatial features;
6) and calculating the reliability of each classification result according to the output class probability, and performing weighted fusion on the output results of the plurality of classification results on the basis of the reliability, thereby obtaining the final classification result.
Further, in the step 1), the spatial resolution of the obtained raster image is consistent with the spatial resolution of the remote sensing image to be classified.
Further, in the step 2), diversified spatial texture information of the remote sensing image is extracted by utilizing Gray-L evel Co-occurrence Matrix (G L CM), spatial structure information of the remote sensing image is extracted by utilizing Morphological Profiles (MPs), and a complex scene of the remote sensing image, including a house Index, a shadow Index and a vegetation Index, is extracted by utilizing Multi-Index features (MIF), and is combined with complementary information of the 3 different spatial features, so that the problem that the spatial information of the image is difficult to be comprehensively expressed by a single spatial Feature is solved.
Further, in step 3), obtaining the sample of the retained central pixel of the removed target edge pixel is implemented as follows,
each raster image can be regarded as a two-class label map, labels belonging to the class are 1, labels not belonging to the class are 0; in order to remove the boundary region of the target, morphological corrosion operation is performed on each of the two types of marked graphs, and the formula is as follows:
Figure BDA0001735156590000047
wherein the content of the first and second substances,
Figure BDA0001735156590000048
is a corrosion operator, BxRepresenting the structural element B with the origin at position x on the image, the result of the erosion is a new image composed of all positions x where the structural element B can be completely contained in a.
Further, in step 4), a k-means algorithm is adopted to perform cluster analysis on the samples of each category, and the calculation flow of the k-means algorithm is as follows:
step 4.1, randomly selecting M seed points from the sample set as initial clustering centers;
step 4.2, calculating the distance from all samples to each clustering center in the feature space, and if the sample is closest to the mth clustering center, classifying the sample into a point group corresponding to the mth clustering center;
step 4.3, based on the obtained M point groups, recalculating the center of each point group, and using the recalculated center as a clustering center of the next iteration;
and 4.4, repeating the steps 4.3 and 4.4 until the clustering center is not changed or the iteration stop condition is reached, and clustering the sample set into M subclasses.
Further, in step 4), the elimination of the incorrectly marked samples is realized as follows,
counting the number of samples belonging to each cluster subclass, and deleting a cluster subclass when the distance between the center of the cluster subclass in the feature space and the centers of other subclasses is too large and the number of samples belonging to the subclass is too small, specifically as follows: if the number of samples in the ith category is SiClustering into M subclasses through k-means; the number of samples of the token class is
Figure BDA0001735156590000041
The distance between the center of the subclass and the nearest center of the subclass is
Figure BDA0001735156590000042
Number of samples in subclass of this class
Figure BDA0001735156590000043
Mean value of SuiStandard deviation is SsiOf variable quantity
Figure BDA0001735156590000044
Has a mean value of DuiStandard deviation is DsiIf a subclass satisfies
Figure BDA0001735156590000045
And is
Figure BDA0001735156590000046
The subclass is considered invalid and the samples in the subclass are deleted.
Further, the step 5) is realized as follows,
supposing that the remote sensing image to be classified needs to be classified into K ground object types, selecting a certain type of sample from the samples obtained in the step 4) as a positive sample, taking all other training samples as negative samples, constructing K classifiers, and training through the K SVM classifiers and a certain characteristic of the sample to obtainObtaining K classified hyperplanes in high-dimensional space, for pixel x, obtaining its distance from every classified hyperplane, and using dk(x) Representing its distance from the k-th class hyperplane, to calculate the probability that the pixel belongs to class k,
Figure BDA0001735156590000051
wherein A iskControlling the rate of decay of probability with distance, BkControl dk(x) Probability intercept at the origin, two parameters obtained by minimizing the mean square error between the original mark of the sample and the class output;
and substituting different characteristics of the samples into the K SVM classifiers for training, thereby obtaining the probability that each pixel belongs to different categories under different characteristics.
Further, the step 6) is realized as follows,
(a) the formula for calculating the reliability of the classification result according to the class probability output result obtained by different spatial features is as follows:
Figure BDA0001735156590000052
wherein K represents the number of categories defined in the image,
Figure BDA0001735156590000053
outputting corresponding to the probability after descending order; r (x) ranges from 0 to 1, with larger values of R (x) indicating higher classification reliability;
(b) and (3) using the classification certainty as the weight of different classification results, fusing class probability data results corresponding to different spatial features, and acquiring weighted probability output:
Figure BDA0001735156590000054
wherein the content of the first and second substances,
Figure BDA0001735156590000055
represents the weighted probability that the pixel x belongs to the class k, and satisfies
Figure BDA0001735156590000061
And Rf(x) Respectively representing the probability of the pixel x belonging to the class k and the corresponding classification reliability in the case of using the spatial feature f;
(c) for each pixel, the class with the highest corresponding probability is selected as its label:
Figure BDA0001735156590000062
where C (x) is the class label for pixel x.
Compared with the prior art, the invention has the following advantages:
1) according to the method, through morphological processing and clustering algorithm, error marks in the OSM are reduced, and the accuracy of classification samples obtained from the OSM is improved.
2) According to the invention, by mining the spatial information of the remote sensing image, various classical spatial feature extraction algorithms are effectively integrated, the separability of different classes can be enhanced, and more accurate classification results of each class and higher classification precision can be obtained.
3) According to the method, the advantages of various characteristics can be effectively concentrated through a probability weighted fusion algorithm, the weighted fusion of the classification results of the images under different characteristics is realized, and the reliability of different characteristics on the discrimination pixels is integrated.
Drawings
Fig. 1 is a flowchart of a method implementation of an embodiment of the invention.
Detailed Description
The technical scheme of the invention is further specifically described by the following embodiments and the accompanying drawings.
Example (b):
step 1, data preprocessing. And selecting the OSM data of the corresponding area according to the geographic position coordinates of the remote sensing image. The OSM data has a ground object type system, and if the image classification has the requirements of K ground object types (buildings, vegetation, roads and the like), data corresponding to the image classification type is extracted from the OSM to obtain K OSM ground object image layers of different types. The extracted initial sample retains the vector data format of the OSM and needs to be converted into a raster image format consistent with the remote sensing image.
The method selects three classical spatial features, namely Gray-L evel Co-occurrence Matrix (G L CM), Morphological Profiles (MPs) and Multi-exponential features (MIF), and an extraction algorithm describes the spatial distribution characteristics of the ground features to make up for the defect of single spectral Feature.
G L CM is a video diversified spatial texture information extraction mode based on spatial symbiotic matrix, using different gray scale quantization levels to measure the spectral similarity of video context, and using multiple spatial symbiotic distances to describe local gray scale symbiotic relations in different directions, the spatial distribution of different ground objects in the video corresponds to different gray scale symbiotic matrices, the obtained statistical characteristics can describe the texture and structural characteristics of different ground objects, G L CM can represent the frequency of the specific spatial relation of the pixel with specific gray scale value, and the statistical quantity obtained from the gray scale symbiotic matrix can reflect the pixel distribution and texture information in the region, and can be represented as fGLCM(w, m), where w and m represent the size and texture statistics of the local window, respectively.A contrast and homogeneity are selected as G L CM statistical measures to describe the texture characteristics of the image.A contrast reflects the difference in gray levels of neighboring pixels, and a homogeneity reflects the change in gray level information in the local region, the statistical measures are expressed as follows:
contrast ratio:
Figure BDA0001735156590000071
s.t.k＝|i-j|
homogeneity:
Figure BDA0001735156590000072
wherein g (i, j) represents the co-occurrence probability of pixels with gray levels i and j, k is the absolute value of the gray level difference, NgRepresenting the grey scale level in the image.
The MPs is a feature algorithm capable of reflecting the morphology of a target structure, which is proposed based on mathematical morphology. The mathematical morphology operation can filter dark targets and bright targets with small areas on the image while keeping the spatial structure attribute of the ground feature. Therefore, based on the morphological open reconstruction and the morphological closed reconstruction, the spatial structure information of the morphological contour feature description target of the image can be extracted. Due to the multi-scale property of the ground features on the remote sensing images, a series of morphological operators with different sizes are usually adopted to extract the features of the ground features in practical application. Therefore, by using morphological operators with different sizes to perform the open and closed reconstruction of the base image, MPs can be obtained:
Figure BDA0001735156590000081
Figure BDA0001735156590000082
wherein, γλ(I) And phiλ(I) Each represents a morphological open/close operation result obtained when a structural operator having a size λ is used for the base image I. It is noted that when the morphological structure operator size is set to 0, the acquired MP is equivalent to the original image.
MIF aims at describing complex scenes with lower-dimensional exponential features with semantic information instead of high-dimensional low-level features. The index features used include house index, shadow index, and vegetation index.
House index: the house on the image is usually adjacent to the shadow, so that the house has a higher luminance value than its surrounding area. Meanwhile, compared with other natural ground objects, the house is regular in shape and size and is easier to recognize. Therefore, the house index can be obtained by establishing a connection between morphological operation and the house space spectral characteristics. Firstly, calculating the maximum response value of each pixel in the visible light wave band as the corresponding brightness value, acquiring a brightness image, and regarding the brightness image as a reference image for house index extraction. The method comprises the following steps of utilizing a preset morphological linear operator to process an image by combining morphological top hat transformation, and extracting a region with a local highlight value in the image:
DMPTH(d,s)＝{|TH(d,s)-TH(d,s-Vs)|}
where TH represents the morphological top-hat transform, s and d refer to the scale and direction of the morphological linear operator, respectively, and Vs is the step size. On the basis, averaging the top hat transformation results obtained from different directions to further obtain the house index
Figure BDA0001735156590000091
Wherein N isdAnd NsThe number of directions and the number of scales of the linear operator, respectively. In the generated house index image, the area having a higher response value is more likely to belong to the house.
Shading index: the shadow index is a twin of the house index because the spectral characteristics of the shadow are completely opposite to those of the house, but the spatial characteristics of the shadow are similar, which is reflected in that the local contrast of the shadow area is high but the brightness is low. Thus, the construction of the shadow index can be done using morphological operations that are antithetical to the house index. The house operator highlights bright objects in the image with the top-hat transformed morphological contour sequence, while the shadow index is transformed with the black top-hat, labeling dark objects. The shadow operator is calculated in the following way:
Figure BDA0001735156590000092
wherein BTH stands for morphological black top hat transform, which is used to highlight local darker areas in the image.
Vegetation index: the vegetation index is used to represent the area of vegetation present in the area of interest, such as trees, grasses, and the like. The vegetation area has high reflectivity in the near infrared band and low reflectivity in the red light band. From this physical characteristic, a vegetation index can be calculated:
Figure BDA0001735156590000093
wherein NIR represents a response value in the near infrared band, and RED represents a response value in the RED light band.
And 3, removing the edge. The problem of OSM spatial accuracy is represented by the fact that part of the feature boundary has a certain deviation from the actual condition, and the overlapping degree of the central area is relatively high. Therefore, it is necessary to reserve the middle region of the OSM feature, remove the boundary region, and reduce the influence of the boundary noise on the classification result. The terrain mark raster image obtained by the OSM can be viewed as a two-class map. In order to remove the boundary region of the target, morphological erosion operation can be performed on each of the two types of marked graphs, and the formula is as follows:
Figure BDA0001735156590000094
wherein the content of the first and second substances,
Figure BDA0001735156590000095
is a corrosion operator, BxRepresenting the structural element B with the origin at position x on the image, the result of the erosion is a new image composed of all positions x where the structural element B can be completely contained in a. And selecting proper morphological operators and parameters by combining with the image spatial resolution, so that the sample of the target center can be reserved while the boundary error is removed.
And 4, clustering analysis. There may be errors in the OSM that the feature type does not match the actual situation, and step 3 cannot remove these erroneous samples. The false samples generally have two characteristics: (1) a small proportion of samples in the correct category; (2) the feature distribution is significantly different from the features of the correct class sample. Therefore, for each class of samples, using a k-means algorithm to perform cluster analysis, the position of the sub-cluster center in the feature space can be obtained, and the subclass attribution of each sample can be obtained.
The k-means algorithm is a mode based on a clustering center, and a clustering model is constructed by continuously optimizing the clustering center, so that the k-means algorithm has the greatest advantages of simple calculation and low time complexity. Suppose that it is necessary to include n sample data sets X ═ X1,x2,L,xnThe division into M subsets is such that the samples inside each subset have a higher similarity, while the samples belonging to different subsets have a lower similarity. Thus, when a sample is k-means clustered, it is assigned to the cluster center closest to it:
Figure BDA0001735156590000101
Figure BDA0001735156590000102
where t represents the t-th iteration, μiIs to belong to the class ciIs measured in the mean of the features of all samples.
The k-means algorithm calculation flow is as follows:
and 4.1, randomly selecting M seed points from the sample set as initial clustering centers.
And 4.2, calculating the distance from all samples to each clustering center in the feature space, and if the sample is closest to the mth clustering center, classifying the sample into a point group corresponding to the mth clustering center.
And 4.3, based on the obtained M point groups, recalculating the center of each point group, and using the recalculated center as the clustering center of the next iteration.
And 4.4, repeating the steps 4.3 and 4.4 until the cluster center is not changed or an iteration stop condition is reached.
And counting the number of samples belonging to each cluster subclass, and deleting the cluster subclass when the distance between the center of a certain cluster subclass in the feature space is too large and the distance between the centers of other subclasses is too small and the number of samples belonging to the subclass is too small. The method comprises the following specific steps: if the number of samples in the ith category is SiClustering into M subclasses through k-means; note the bookThe number of samples of the subclass is
Figure BDA0001735156590000111
The distance between the center of the subclass and the nearest center of the subclass is
Figure BDA0001735156590000112
Number of samples in subclass of this class
Figure BDA0001735156590000113
Mean value of SuiStandard deviation is SsiOf variable quantity
Figure BDA0001735156590000114
Has a mean value of DuiStandard deviation is DsiIf a subclass satisfies
Figure BDA0001735156590000115
And is
Figure BDA0001735156590000116
The subclass is considered invalid and the samples in the subclass are deleted.
And 5, classifying the images. The images are classified using a Support Vector Machine (SVM). The SVM is a classifier based on a structural risk minimization criterion, and the basic idea is to convert a sample from a low-dimensional feature space to a high-dimensional feature space through a mapping function and find an optimal hyperplane in the high-dimensional space to finish the differentiation of different classes. Assume that there is a mapping function Φ (g) and a training sample set comprising n labeled samples { (x)1,y1),(x2,y2),....(xn,yn) Y ∈ (-1, +1) }, the classification model may be expressed as:
f(x)＝w·Φ(x)+b
where w represents the weight vector and b represents the bias. In order to maximize the classification boundary and minimize the classification error, resulting in an optimal classification hyperplane, the following function needs to be solved:
Figure BDA0001735156590000117
Figure BDA0001735156590000118
wherein, ξiIs a relaxation variable and C is a penalty coefficient. The above function can be solved using optimization of transforming it into a dual problem using lagrange multiplier method. The function can also be expressed as:
Figure BDA0001735156590000119
Figure BDA00017351565900001110
wherein, αiFor the lagrange multiplier vector, K (,) represents the kernel function. The present invention uses a Radial Basis Function (RBF), expressed as:
K(xi,xj)＝exp(-γ||xi-xj||2)
where γ is the bandwidth of the RBF core. The decision function of the SVM can be expressed as:
Figure BDA0001735156590000121
for the multi-class classification problem in the remote sensing image, a one-to-many strategy is combined with a plurality of two-class SVM to construct a multi-class SVM, a certain class of sample is selected as a positive sample, all other training samples are taken as negative samples, and K classifiers are constructed. Through K SVM classification models, K classification hyperplanes of a high-dimensional space can be obtained, and for each pixel, the distance between each pixel and each classification hyperplane can be obtained. If d is usedk(x) Representing its distance from the k-th class hyperplane, the probability that the pixel belongs to class k can be expressed as:
Figure BDA0001735156590000122
wherein A iskControlling the rate of decay of probability with distance, BkControl dk(x) In this embodiment, the two parameters are specifically calculated in the Probabistic Outputs for Support vectors and generalized L Ikelihood Methods, which are not described herein.
And substituting different characteristics of the samples into the K SVM classifiers for training, thereby obtaining the probability that each pixel belongs to different categories under different characteristics.
And 6, classifying and fusing. Each pixel of the image presents different characteristics under different characteristics, and the probability of belonging to each category is different, so that a plurality of classification results need to be organically combined.
Firstly, combining the obtained class probability to obtain the reliability of the classification result corresponding to different characteristics:
Figure BDA0001735156590000123
wherein K represents the number of categories defined in the image,
Figure BDA0001735156590000131
and outputting corresponding to the probability after descending order. R (x) ranges from 0 to 1, with larger values of R (x) indicating higher classification reliability. And (3) using the classification certainty as the weight of different classification results, fusing class probability output results corresponding to different features, and acquiring weighted probability output:
Figure BDA0001735156590000132
wherein the content of the first and second substances,
Figure BDA0001735156590000133
represents the weighted probability that the pixel x belongs to the class k, and satisfies
Figure BDA0001735156590000134
And Rf(x) Respectively representing the probability that the pixel x belongs to the class k and the corresponding classification reliability in case of using the spatial feature f (including the gray level co-occurrence texture, the morphological contour and the multi-exponential feature described in step 2). For each pixel, the class with the highest probability of correspondence is selected as its label
Figure BDA0001735156590000135
Where C (x) is the class label for pixel x.
The precision of the method of the invention is compared with the traditional method:
according to the method, the surface feature type information provided by the OSM is utilized, manpower and material resources consumed by manual marking are avoided, and the automatic interpretation degree of the remote sensing image can be improved. On a remote sensing image of a certain area acquired by a resource No. 3 satellite, the overall accuracy obtained by the traditional OSM classification method is 83.54%. On the same data, the overall accuracy obtained by the method of the invention is 91.22%, which is significantly better than that obtained by the conventional method. The reason is that the traditional method only utilizes the spectral information of the image, and does not consider the influence of error markers in the OSM data on the classification result. The method identifies the ground objects in a multi-feature combined classification mode, enhances the separability of the spectrum-similar ground objects by using the spatial features, improves the reliability and accuracy of the sample by combining morphological filtering and cluster analysis, and is favorable for constructing a classification model.
The specific embodiments described herein are to be considered in all respects as illustrative and not restrictive. Those skilled in the art to which the invention relates may make modifications and additions to the described embodiments without departing from the spirit of the invention or exceeding the scope thereof as defined by the appended claims.

Claims (6)

1. A remote sensing image multi-feature classification method based on OpenStreetMap comprises the following steps:
1) registering OpenStreetMap (OSM) with the remote sensing image to be classified, selecting a vector layer corresponding to the interested category from the OSM according to the actual classification requirement, and converting the vector layer into a raster image;
2) extracting a plurality of different spatial features of the remote sensing image to be classified, and describing the distribution characteristics of the ground features of the remote sensing image to be classified from a plurality of aspects;
3) performing morphological processing on the OSM grid images corresponding to the categories to obtain a sample of removing target edge pixels and reserving center pixels;
4) performing clustering analysis on the samples of each category by using a clustering algorithm according to the characteristics of pixels, and rejecting the samples marked by errors according to the distribution of clustering centers of each subclass and the number of the samples in each subclass;
in the step 4), a k-means algorithm is adopted to perform cluster analysis on the samples of each category, and the calculation flow of the k-means algorithm is as follows:
step 4.1, randomly selecting M seed points from the sample set as initial clustering centers;
step 4.2, calculating the distance from all samples to each clustering center in the feature space, and if the sample is closest to the mth clustering center, classifying the sample into a point group corresponding to the mth clustering center;
step 4.3, based on the obtained M point groups, recalculating the center of each point group, and using the recalculated center as a clustering center of the next iteration;
4.4, repeating the steps 4.3 and 4.4 until the clustering center is not changed or the iteration stop condition is reached, and clustering the sample set into M subclasses;
in step 4), the elimination of the incorrectly labeled samples is realized as follows,
counting the number of samples belonging to each cluster subclass, and deleting a cluster subclass when the distance between the center of the cluster subclass in the feature space and the centers of other subclasses is too large and the number of samples belonging to the subclass is too small, specifically as follows: if the number of samples in the ith category is SiClustering into M subclasses through k-means; the number of samples of the token class is
Figure FDA0002480468420000021
The distance between the center of the subclass and the nearest center of the subclass is
Figure FDA0002480468420000022
Number of samples in subclass of this class
Figure FDA0002480468420000023
Mean value of SuiStandard deviation is SsiOf variable quantity
Figure FDA0002480468420000024
Has a mean value of DuiStandard deviation is DsiIf a subclass satisfies
Figure FDA0002480468420000025
And is
Figure FDA0002480468420000026
The subclass is considered invalid and the samples in the subclass are deleted;
5) classifying the remote sensing images to be classified by using the acquired samples and using various different spatial features respectively in combination with a support vector machine, and acquiring the probability that each pixel belongs to different classes under different spatial features;
6) and calculating the reliability of each classification result according to the output class probability, and performing weighted fusion on the output results of the plurality of classification results on the basis of the reliability, thereby obtaining the final classification result.
2. The OpenStreetMap-based remote sensing image multi-feature classification method according to claim 1, characterized by: in the step 1), the spatial resolution of the obtained raster image is consistent with the spatial resolution of the remote sensing image to be classified.
3. The OpenStreetMap-based remote sensing image Multi-Feature classification method as claimed in claim 1, wherein in the step 2), Gray-level symbiotic texture (Gray-L eveco-occurrence Matrix, G L CM) is used for extracting diversified spatial texture information of the remote sensing image, Morphological Profiles (MPs) are used for extracting spatial structure information of the remote sensing image, Multi-Index features (Multi-Index features, MIF) are used for extracting complex scenes of the remote sensing image, including house indexes, shadow indexes and vegetation indexes, and the problem that the spatial information of the image is difficult to be fully expressed by a single spatial Feature is solved by combining the complementary information of the 3 different spatial features.
4. The OpenStreetMap-based remote sensing image multi-feature classification method according to claim 1, characterized by: in step 3), the implementation manner of obtaining the sample of removing the target edge pixel and retaining the central pixel is as follows,
each raster image can be regarded as a two-class label map, labels belonging to the class are 1, labels not belonging to the class are 0; in order to remove the boundary region of the target, morphological corrosion operation is performed on each of the two types of marked graphs, and the formula is as follows:
Figure FDA0002480468420000031
wherein the content of the first and second substances,
Figure FDA0002480468420000032
is a corrosion operator, BxRepresenting the structural element B with the origin at position x on the image, the result of the erosion is a new image composed of all positions x where the structural element B can be completely contained in a.
5. The OpenStreetMap-based remote sensing image multi-feature classification method according to claim 1, characterized by: the implementation of step 5) is as follows,
supposing that the remote sensing image to be classified needs to be classified into K ground object types, selecting a certain type of sample from the samples obtained in the step 4) as a positive sample, taking all other training samples as negative samples, constructing K classifiers, training through the K SVM classifiers and a certain characteristic of the sample to obtain K classified hyperplanes in high-dimensional space, and solving the pixel x of the hyperplane in high-dimensional spaceDistance from each classification hyperplane and using dk(x) Representing its distance from the k-th class hyperplane, to calculate the probability that the pixel belongs to class k,
Figure FDA0002480468420000033
wherein A iskControlling the rate of decay of probability with distance, BkControl dk(x) Probability intercept at the origin, two parameters obtained by minimizing the mean square error between the original mark of the sample and the class output;
and substituting different characteristics of the samples into the K SVM classifiers for training, thereby obtaining the probability that each pixel belongs to different categories under different characteristics.
6. The OpenStreetMap-based remote sensing image multi-feature classification method according to claim 5, characterized by: the implementation of step 6) is as follows,
(a) the formula for calculating the reliability of the classification result according to the class probability output result obtained by different spatial features is as follows:
Figure FDA0002480468420000041
wherein K represents the number of categories defined in the image,
Figure FDA0002480468420000042
outputting corresponding to the probability after descending order; r (x) ranges from 0 to 1, with larger values of R (x) indicating higher classification reliability;
(b) and (3) using the classification certainty as the weight of different classification results, fusing class probability data results corresponding to different spatial features, and acquiring weighted probability output:
Figure FDA0002480468420000043
wherein the content of the first and second substances,
Figure FDA0002480468420000044
represents the weighted probability that the pixel x belongs to the class k, and satisfies
Figure FDA0002480468420000045
Figure FDA0002480468420000046
And Rf(x) Respectively representing the probability of the pixel x belonging to the class k and the corresponding classification reliability in the case of using the spatial feature f;
(c) for each pixel, the class with the highest corresponding probability is selected as its label:
Figure FDA0002480468420000047
where C (x) is the class label for pixel x.
CN201810792087.9A 2018-07-18 2018-07-18 Remote sensing image multi-feature joint classification method based on OpenStreetMap Expired - Fee Related CN109063754B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810792087.9A CN109063754B (en) 2018-07-18 2018-07-18 Remote sensing image multi-feature joint classification method based on OpenStreetMap

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810792087.9A CN109063754B (en) 2018-07-18 2018-07-18 Remote sensing image multi-feature joint classification method based on OpenStreetMap

Publications (2)

Publication Number Publication Date
CN109063754A CN109063754A (en) 2018-12-21
CN109063754B true CN109063754B (en) 2020-08-07

Family

ID=64817156

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810792087.9A Expired - Fee Related CN109063754B (en) 2018-07-18 2018-07-18 Remote sensing image multi-feature joint classification method based on OpenStreetMap

Country Status (1)

Country Link
CN (1) CN109063754B (en)

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948697B (en) * 2019-03-19 2022-08-26 中南大学 Method for extracting urban built-up area by using multi-source data to assist remote sensing image classification
CN110188613A (en) * 2019-04-28 2019-08-30 上海鹰瞳医疗科技有限公司 Image classification method and equipment
CN110598564B (en) * 2019-08-16 2022-02-11 浙江工业大学 OpenStreetMap-based high-spatial-resolution remote sensing image transfer learning classification method
CN110517575B (en) * 2019-08-21 2021-03-02 华北水利水电大学 Method and device for mapping surface water body
CN110782904A (en) * 2019-11-07 2020-02-11 四川长虹电器股份有限公司 User account switching method of intelligent voice equipment
CN110991335A (en) * 2019-11-29 2020-04-10 福州大学 Visible light unmanned aerial vehicle remote sensing image forest tree species classification method based on multi-feature optimization
CN111079807B (en) * 2019-12-05 2023-07-07 二十一世纪空间技术应用股份有限公司 Ground object classification method and device
CN111784606B (en) * 2020-06-30 2023-11-24 长江大学 Remote sensing image classification post-processing method, storage medium and system
CN112069981A (en) * 2020-09-03 2020-12-11 Oppo广东移动通信有限公司 Image classification method and device, electronic equipment and storage medium
CN113535861B (en) * 2021-07-16 2023-08-11 子亥科技(成都)有限公司 Track prediction method for multi-scale feature fusion and self-adaptive clustering

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778547A (en) * 2016-11-29 2017-05-31 中南大学 The construction method and device of remote sensing image terrain classification pattern library
CN107451614A (en) * 2017-08-01 2017-12-08 西安电子科技大学 The hyperspectral classification method merged based on space coordinates with empty spectrum signature
CN107516103A (en) * 2016-06-17 2017-12-26 北京市商汤科技开发有限公司 A kind of image classification method and system
CN105787937B (en) * 2016-02-25 2019-03-01 武汉大学 A kind of high resolution remote sensing image road change detecting method based on OSM

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105787937B (en) * 2016-02-25 2019-03-01 武汉大学 A kind of high resolution remote sensing image road change detecting method based on OSM
CN107516103A (en) * 2016-06-17 2017-12-26 北京市商汤科技开发有限公司 A kind of image classification method and system
CN106778547A (en) * 2016-11-29 2017-05-31 中南大学 The construction method and device of remote sensing image terrain classification pattern library
CN107451614A (en) * 2017-08-01 2017-12-08 西安电子科技大学 The hyperspectral classification method merged based on space coordinates with empty spectrum signature

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"A Note on Platt’s Probabilistic Outputs for Support Vector Machines";Hsuan-Tien Lin等;《DOI 10.1007/s10994-007-5018-6》;20070625;期刊第1节 *

Also Published As

Publication number Publication date
CN109063754A (en) 2018-12-21

Similar Documents

Publication Publication Date Title
CN109063754B (en) Remote sensing image multi-feature joint classification method based on OpenStreetMap
Zhang et al. Object-oriented method for urban vegetation mapping using IKONOS imagery
Huang et al. Morphological building/shadow index for building extraction from high-resolution imagery over urban areas
CN106909902B (en) Remote sensing target detection method based on improved hierarchical significant model
CN103679675B (en) Remote sensing image fusion method oriented to water quality quantitative remote sensing application
CN110309781B (en) House damage remote sensing identification method based on multi-scale spectrum texture self-adaptive fusion
CN104331698B (en) Remote sensing type urban image extracting method
CN107330875B (en) Water body surrounding environment change detection method based on forward and reverse heterogeneity of remote sensing image
CN113449594B (en) Multilayer network combined remote sensing image ground semantic segmentation and area calculation method
CN109034233B (en) High-resolution remote sensing image multi-classifier joint classification method combined with OpenStreetMap
CN104182767B (en) The hyperspectral image classification method that Active Learning and neighborhood information are combined
Peng et al. Object-based change detection from satellite imagery by segmentation optimization and multi-features fusion
Deng et al. Cloud detection in satellite images based on natural scene statistics and gabor features
CN108428220A (en) Satellite sequence remote sensing image sea island reef region automatic geometric correction method
Shaoqing et al. The comparative study of three methods of remote sensing image change detection
Nurmasari et al. Oil palm plantation detection in Indonesia using Sentinel-2 and Landsat-8 optical satellite imagery (case study: Rokan Hulu regency, Riau Province)
Ziaei et al. A rule-based parameter aided with object-based classification approach for extraction of building and roads from WorldView-2 images
CN110598564A (en) OpenStreetMap-based high-spatial-resolution remote sensing image transfer learning classification method
Balcik et al. Greenhouse mapping using object based classification and sentinel-2 satellite imagery
Poterek et al. Deep learning for automatic colorization of legacy grayscale aerial photographs
CN109447111A (en) A kind of remote sensing supervised classification method based on subclass training sample
Genitha et al. Classification of satellite images using new fuzzy cluster centroid for unsupervised classification algorithm
CN109886146A (en) Flood information remote-sensing intelligent acquisition method and equipment based on Machine Vision Detection
Hasanlou et al. A sub-pixel multiple change detection approach for hyperspectral imagery
Tang et al. A novel sample selection method for impervious surface area mapping using JL1-3B nighttime light and Sentinel-2 imagery

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20200807

Termination date: 20210718

CF01 Termination of patent right due to non-payment of annual fee