CN115512159A - Object-oriented high-resolution remote sensing image earth surface coverage classification method and system - Google Patents

Abstract

The invention relates to the technical field of remote sensing image processing, and discloses a method and a system for classifying object-oriented high-resolution remote sensing image earth surface coverage, which comprises the following steps: obtaining remote sensing images of a high-resolution first satellite and a sentinel second satellite, and preprocessing the obtained sentinel second remote sensing image; performing image segmentation based on the obtained high-resolution first remote sensing image and the obtained sentinel second remote sensing image to obtain a high-resolution pixel and a sentinel pixel; constructing a multi-source characteristic space based on the obtained high-resolution pixels and the sentinel pixels, and ensuring that one high-resolution pixel at least comprises one sentinel pixel; and performing machine learning based on the constructed multi-source feature space, and performing earth surface coverage classification. The invention has low use cost and good distinguishing effect.

Description

Object-oriented high-resolution remote sensing image earth surface coverage classification method and system

Technical Field

The invention relates to the technical field of remote sensing image processing, in particular to a method and a system for classifying object-oriented high-resolution remote sensing image ground surface coverage.

Background

The ground surface coverings comprise organic substance coverings and inorganic coverings paved on the ground surface, and the ground surface covering types comprise cultivated land, forest, grassland, shrub land, wetland, water body, moss, artificial ground surface, bare land, glacier and permanent snow; the research on the surface coverage and the change thereof is essential basic information and key parameters for environmental change research, geographic national condition detection, sustainable development planning and the like.

The traditional earth surface coverage survey is time-consuming and labor-consuming, and the remote sensing technology provides a new technical means for rapidly acquiring earth surface coverage data by the characteristics of wide image acquisition range, large information amount, less limitation of the acquisition mode by ground conditions and the like. How to extract characteristic information from the image and classify the characteristic information so as to accurately obtain the earth surface coverage information is a hot point problem in the field of remote sensing. With the improvement of the image resolution, the high-resolution remote sensing image can clearly and accurately express the characteristic information of the spectrum, the shape, the texture and the like of the shot ground object. The traditional classification method based on the pixel only analyzes the single pixel from the spectral characteristics of the single pixel, cannot fully utilize other useful information, and is easy to generate the phenomenon of wrong classification and missing classification.

The existing research of the object-oriented high-resolution remote sensing image earth surface coverage classification method has some defects, so that the classification precision is not ideal. In the aspect of image segmentation, the segmentation accuracy is reduced due to the influence of artificial subjectivity when the segmentation scale is determined and the influence of complex earth surface morphology when the segmentation is performed; in the aspect of feature extraction, a single high-resolution image data source is adopted in most researches, so that the wave band data are less, and a comprehensive feature space with discrimination is difficult to construct; in the aspect of classification structure, the single-level classification structure is difficult to adjust different land categories on the segmentation scale and the feature space, and the information extraction and classification effects are influenced.

To solve the above problem, people usually purchase a high-precision input data set formed by data sources of a plurality of satellites through complicated processing from the acquired input data set. However, this method for improving the accuracy of classification results based on the high-accuracy input data set source does not have a potential for popularization due to its high investment cost. Because all the enterprises providing the high-precision input data set service at present basically obtain a specific high-precision input data set after one-to-one special processing is carried out on a plurality of different satellite data sources according to the use requirements of users, the processing mode is not universal and has no popularization premise; secondly, because the user needs are very different, there is a very large subjective adjustment space in selecting which satellite data source and how to achieve the specified precision, the uncertainty is large, and the data processing amount is huge, so that the overall processing cost is high, and the method is very not suitable for large-scale popularization.

Disclosure of Invention

The invention aims to provide a method for classifying the high-resolution remote sensing image earth surface coverage to an object so as to achieve the purposes of low cost and good distinguishing effect.

In order to achieve the purpose, the invention adopts the following technical scheme:

the object-oriented high-resolution remote sensing image earth surface coverage classification method comprises the following steps:

s1, acquiring an initial remote sensing image, namely acquiring a high-grade first remote sensing image from a high-grade first satellite, acquiring a sentinel second remote sensing image from a sentinel second satellite, and performing data preprocessing on the sentinel second remote sensing image; registering the high-grade first remote sensing image and the sentinel second remote sensing image to form a high-grade first registered image;

s2, image segmentation, namely performing image segmentation on the high-resolution first-order registration image to obtain block images, wherein each block image at least comprises one sentinel pixel;

s3, extracting object features and constructing a multi-source feature space;

s4, forming a multi-level classification structure based on object classification;

and S5, performing earth surface coverage classification based on the constructed multi-source feature space and multi-level classification structure.

The principle and the advantages of the scheme are as follows:

the scheme can directly utilize the remote sensing images of the high-grade first satellite and the sentry second satellite which can be obtained freely, and the cost is lower; through complementary use of the high-resolution first remote sensing image and the sentinel second remote sensing image, the defects of the high-resolution first remote sensing image and the sentinel second remote sensing image are eliminated as far as possible, when the image is segmented, the segmented image at least comprises one sentinel pixel, effective data volume is guaranteed, and a premise is provided for later classification accuracy; then, constructing a multi-source feature space; finally, based on object classification, a multi-level classification structure is formed, and ground surface coverage classification is realized;

compared with the existing ground surface coverage classification method, the scheme is different in that in inherent cognition, ground surface coverage classification needs more remote sensing image resources with high quality and high resolution to have a good classification effect; but the scheme adopts a data source which is almost unprocessed and can be obtained freely, and the cost is lower. Meanwhile, through combination with the sentinel II, the problem that the high-resolution I is lost in the near-infrared band is solved by utilizing the advantage of more bands, the spatial resolution advantage of the high-resolution I remote sensing data can be fully utilized, the spectral resolution advantage of the sentinel II remote sensing data can be fully utilized, and the better classification effect can be realized by using a low-cost database.

Compared with the short plate of the existing object-oriented analysis method on segmentation, although some methods for quantitatively calculating the optimal segmentation scale exist, the method is high in theoretical performance and scientificity, the determination process is complex, the intuitiveness is lacked, especially the optimal segmentation scale point exists high subjectivity, the judgment is difficult in scenes with many ground types and obvious differences, and the influence of complex ground surface morphology in high-resolution image data causes low segmentation accuracy. According to the scheme, the problem in the aspect is solved through the specially constructed multi-element feature space, even if the wave bands of general high-resolution remote sensing image data are few, the feature space with the comprehensive distinguishing degree is difficult to construct, the target accuracy can still be achieved on the premise that the distinguishing of similar ground types has limitation, and the problem that the precision of the existing general remote sensing data source is poor is effectively solved.

1. As an improvement, in step S2, quantitatively calculating a segmentation scale through a segmentation quality function formed by a Moran' S I index and an area weighted standard deviation, and an RMAS index to obtain a global optimal segmentation scale and optimal segmentation scales of different earth surface coverage types;

the segmentation quality function is:

GS＝V _norm +MI _norm

in the formula, GS is a segmentation quality function, vnorm is a normalized weighted standard deviation, and MINorm is normalized Moran's I

And (4) an index.

The effect is as follows: according to the scheme, the Moran's I index and a segmentation quality function formed by the area weighting standard deviation are adopted, so that the problem that the subjectivity of the existing segmentation scale is too high can be solved.

As an improvement, the surface covering types include cultivated land, forest land, construction land, transportation land, water area, and unused land.

The effect is as follows: with the above surface covering types, most surface covering classifications can be satisfied.

As an improvement, the global optimal segmentation scale is in the range of 10-150; the optimal division scale ranges of arable land, forest land, construction land, transportation land, water area and unused land are respectively 50-60, 75-85, 25-35, 55-65, 70-80 and 120-130.

The effect is as follows: the global optimal segmentation scale and the optimal segmentation scale range of different earth surface coverage types are respectively determined, so that the conditions of fragmentation of the earth class with a large area and missing classification of the earth class with a small area can be avoided by adopting different segmentation scales for different earth classes and the global situation.

As an improvement, the global optimal segmentation scale is 45; the optimal division scale of arable land, woodland, construction land, transportation land, water area and unused land is 55, 80, 30, 60, 75, 125 respectively.

The effect is as follows: the optimal segmentation scale is set according to the method, global segmentation and the most important earth surface coverage type classification can be considered simultaneously, and the data calculation amount is minimum on the premise of ensuring accurate segmentation.

As an improvement, after the segmentation is finished, a Canny edge detection algorithm is adopted for segmentation optimization, and the image edge information extracted according to the low threshold value of 150 and the high threshold value of 400 participates in the segmentation.

The effect is as follows: the method can well solve the problems that the homogeneity among the same land classes and the heterogeneity among different land classes in the high-resolution remote sensing image are weakened, the edges of part of the land classes are in a gradual change type, and obvious mutation is avoided, so that the edge outline of part of the segmented object is inaccurate after the image is segmented by adopting the optimal segmentation scale; compared with the existing scheme, the improvement in the aspect of segmentation obviously improves the recall degree of the land types with regular shapes such as buildings, and the overall precision is improved by 2.6 percent.

As an improvement, the multi-source feature space comprises a global multi-source feature space and various surface coverage type multi-source feature spaces, and each multi-source feature space comprises a spectral feature, a geometric feature, an exponential feature and a texture feature.

The effects are as follows: the established multi-source feature space can be consistent with the performance features of the actual region as much as possible.

As an improvement, in S3, an initial multi-source feature space is established first, and then the initial multi-source feature space is optimized, and a recursive elimination method is adopted to remove redundant features; and then selecting a random forest model as a feature optimization model, and respectively optimizing the global multi-source feature space and each earth surface coverage type multi-source feature space to obtain a global optimal multi-source feature space and each earth surface coverage type optimal multi-source feature space.

The effect is as follows: the obtained global optimal multi-source feature space and the optimal multi-source feature space of each earth surface coverage type are more fit with the actual area condition.

As an improvement, in S4, performing single-level classification on the constructed multi-source feature space, and performing separability calculation on different ground surface coverage types according to a confusion matrix obtained by the single-level classification to obtain a separability calculation result; according to the result of the calculation of the separability, an easy-to-first-difficult classification sequence is established, then a hierarchical clustering algorithm is utilized to decompose the multi-classification problem into two classification problems of extracting one earth surface coverage type from each layer, and a multi-layer classification structure is established.

The effect is as follows: calculating the separability between the terrains through single-layer classification, constructing a difficulty and easiness sequence of the terrains according to the separability between the terrains, and finally classifying by adopting a two-classification method according to the difficulty and easiness sequence; the classification calculation amount is small, and meanwhile, the problem of strong subjectivity in the existing multi-level classification can be avoided by taking the actual separability as the basis.

The application also provides an object-oriented high-resolution remote sensing image earth surface coverage classification system which adopts the classification method for classification. Different earth surface coverage types can be accurately and quickly distinguished.

Drawings

Fig. 1 is a flowchart of a method for classifying a surface coverage according to an embodiment of the present invention.

FIG. 2 is a graph comparing the results of the segmentation optimization according to the first embodiment of the present invention.

Fig. 3 is a schematic diagram of a multi-level classification structure according to an embodiment of the invention.

FIG. 4 is a graph showing the comparison between the overall accuracy and the Kappa coefficient of different classification experiments according to the first embodiment of the present invention.

FIG. 5 is a graph showing a comparison of production accuracy of different classification experiments according to a first embodiment of the present invention.

Fig. 6 is a user precision comparison diagram of different classification experiments according to a first embodiment of the present invention.

Detailed Description

The following is further detailed by way of specific embodiments:

example 1 is substantially as shown in figure 1: the object-oriented high-resolution remote sensing image earth surface coverage classification method in the embodiment comprises the following contents:

the method comprises the steps of firstly, obtaining an initial remote sensing image, obtaining a high-grade first remote sensing image from a high-grade first satellite, obtaining a sentinel second remote sensing image from a sentinel second satellite, and performing data preprocessing on the sentinel second remote sensing image; and registering the high-grade first remote sensing image and the sentinel second remote sensing image to form a high-grade first registered image.

Most remote sensing images of the high-score first satellite and the sentry second satellite can be obtained from a public free database, the scheme has no excessive requirement on an initial data source, and even if the initial data source is a free data source which can be obtained by the general public, ground surface coverage classification with certain precision can be completed through the method.

The high-resolution first-grade remote sensing image is fused to only have R, G and B three wave bands, the resolution can be 2 meters, 8 meters and 16 meters, and the cloud amount is less than 2% -5%. In this embodiment, the resolution of the high resolution first remote sensing image is 2 meters, and the cloud amount is less than 3%.

The high-resolution first satellite is the first high-resolution satellite independently developed in China, overcomes the key problem of incompatibility of space spectrum resolution, and adopts a wider breadth to carry out effective detection. During acquisition, the high-resolution first-grade remote sensing image is subjected to conventional preprocessing processes such as geometric correction, embedding, fusion, cutting, radiometric calibration, atmospheric correction and the like, and additional preprocessing is not needed.

The remote sensing image of the sentinel number two in this embodiment is an L1C-level image directly downloaded from the european space, and is an image that has undergone the above-mentioned atmosphere removal correction preprocessing step, and the cloud amount is less than 4.5%.

The euro defines the data after the atmospheric correction as L1A-level data, and provides a Sen2cor plug-in tool, so that L1C-level images can be subjected to atmospheric correction to form L1A-level images, but the L1A-level images cannot be directly downloaded. In this embodiment, the directly obtained L1C-level remote sensing image of the sentinel second is configured and called by using a CMD command through a Sen2cor plug-in, atmospheric correction is performed, all wave bands are resampled to 10 meters in SNAP software, data preprocessing is completed, and the processed remote sensing image of the sentinel second is obtained. Compared with the preprocessed image, the preprocessed image has higher saturation and darker color, and is more beneficial to registration and feature extraction of the next and high-resolution first remote sensing images.

And registering the sentinel second data with the high-resolution images, and resampling to the resolution which is the same as the resolution of the high-resolution first images by using a nearest neighbor method. Importing the processed sentinel second number data into ENVI for layerstack operation, and removing the Water vapours, the coast aerosol and the sentinel with lower resolution,

And (4) adding remote sensing data of 10 wave bands to the Cirrus three-wave band post-group.

In the embodiment, the image registration of the high-resolution first remote sensing image and the sentinel second remote sensing image only needs to be carried out in the conventional general image registration mode.

And secondly, image segmentation is carried out, the high-resolution first-order registration image is subjected to image segmentation to obtain block images, and each block image at least comprises one sentinel pixel, so that the high-spatial resolution advantage of the high-resolution first-order remote sensing data can be fully utilized, and the high-spectral resolution advantage of the sentinel second-order remote sensing data can be fully utilized.

First, a segmentation scale for an optimal segmentation is determined.

Quantitatively calculating segmentation indexes under different segmentation scales through a Moran's I index, a segmentation quality function of an area weighting standard deviation and an RMAS index, and respectively finding out a global optimal segmentation scale with the maximum internal homogeneity of an object and the maximum heterogeneity among the objects and different ground class optimal segmentation scales meeting the requirements of different layers of a multi-level classification structure; in this embodiment, the surface coverage type is abbreviated as ground class.

Specifically, firstly, determining a global optimal segmentation scale; and then establishing the optimal segmentation scale of different surface coverage types. By determining the global optimal segmentation scale, the heterogeneity inside the segmented object is as small as possible, the heterogeneity among objects is as large as possible, and the heterogeneity inside and among the objects of the whole image is balanced, so that a better distinguishing effect is achieved, and the scale is provided for single-level classification.

In the embodiment, an optimal segmentation scale calculation model (Johnson and Xie, 2011) based on a segmentation quality function is adopted, and a global optimal segmentation scale is calculated quantitatively by counting segmentation qualities under different segmentation scales. The method is different from other global optimal segmentation scale calculation, not only considers homogeneity information inside objects, but also adds Moran's I index to count correlation among global objects, and considers the principle of heterogeneity among objects. Compared with the traditional method for determining the global optimal segmentation scale through visual inspection and ESP software, the method is more objective and accurate, and can effectively reduce artificial subjective influence so as to improve the segmentation accuracy.

(1) Segmentation quality function calculation model theory

The function uses the area weighted standard deviation of the segmented objects to measure the internal spectral homogeneity of the objects, and uses the global Moran's I index in the spatial autocorrelation index to measure the heterogeneity among the objects.

1) The internal spectrum homogeneity of the object is expressed as

In the formula, wVar represents the area weighted spectral standard deviation, v _i Is the variance, a _i The area of the object i and n is the total number of the segmented objects in the image. The greater wVar, the greater the internal heterogeneity of the subject, and the less homogeneity.

2) The expression of heterogeneity between objects is

Where MI is the global Moran's I index,w _ij for a spatial adjacency between object i and object j, w is the case if object i is adjacent to object j (i.e., an area sharing a boundary) _ij =1, non-adjacent is 0,y _i And y _j Representing the spectral mean of object i and object j,

is the average value of the whole image spectrum. When the lower MI means the lower correlation between the objects, the higher heterogeneity, that is, the higher degree of separability.

3) Normalized expression

Since the two index metrics are computed for all bands during image segmentation, the normalization formula of the formula is used to re-adjust both to similar [0,1] ranges in order to allow the indices within and between the bands to be considered equally.

(X-X _min )/(X _max -X _min )

Where Xmin and Xmax are the minimum and maximum values of the spectral area weighted standard deviation or Moran's I index, respectively. Normalization brings the normalized value of each band with a low area weighted standard deviation or Moran's I index relatively close to zero.

4) The quality function of segmentation is expressed as

GS＝V _norm +MI _norm

In the formula, GS is a segmentation quality function, vnorm is normalized weighted standard deviation, and MINorm is normalized Moran's I index. And determining the global optimal segmentation scale by respectively calculating the GS value of each spectral band, averaging, and evaluating the segmentation quality by using the averaged GS value. The scale with the lowest average GS value is the optimal segmentation scale, because the weighted standard deviation of the spectrum and the spatial autocorrelation combination on the scale are the lowest, namely the homogeneity in the object is the highest, and the heterogeneity between the objects is the highest.

(2) Global optimal segmentation scale implementation

The range of the global optimal segmentation scale is set to be 10-150, multiple segmentation is carried out in 5 steps, and the segmentation quality function of each wave band obtained by each segmentation is counted, so that the segmentation times can be reduced as few as possible on the premise of meeting the requirement of accuracy, and the calculation amount is reduced. Considering the different scale segmentation effect of partial regions of the research region, when the segmentation scale in the research region is less than 10, over-segmentation occurs, and even the land class with the minimum area is divided into a plurality of objects, which is not beneficial to the integrity of subsequent feature extraction; when the segmentation scale exceeds 150, the segmentation is insufficient, and one object contains various types of places, so that the error rate is greatly increased, and the difficulty is caused in subsequent accurate classification. Therefore, the scale range of 10-150 is selected as the optimal selection.

When the method is implemented, firstly, importing high-score data of a high-score first registration image into eCoginization for multi-scale segmentation, then exporting standard deviation and area information of three bands of a segmented object to Excel, and calculating area weighted standard deviation of the three bands through the Excel; secondly, exporting the segmentation result vector to Arcgis, and calculating the Moran's I index of the three bands by using a spatial statistic tool in the Arcgis; and finally, statistically calculating the normalized Moran's I index and the weighted standard deviation.

And on the basis of determining that the global optimal segmentation scale range is 10-150, then establishing optimal segmentation scales of different surface coverage types.

Because the sizes of the land types in the image and the spectral differences are obvious, a single segmentation scale is difficult to ensure that the boundary of the segmentation object of each land type can better fit the real contour of the land type, so different segmentation scales need to be set for each land type, that is, the optimal segmentation scales of different land types are determined, and the segmentation scales are provided for the multi-level classification structure. In The embodiment, the Ratio Of The absolute Difference value between The object and The field Mean and The Standard Deviation Of The object, namely The RMAS index (The Ratio Of Mean differences To Neighbors To Standard development), is adopted, and The RMAS values Of different land types under different segmentation scales are calculated To obtain The optimal segmentation scale Of different land types, so that The artificial subjective influence is reduced, and The segmentation accuracy is improved. Under the condition of considering the internal homogeneity of the object, the heterogeneity between the objects is also considered by using the difference absolute value of the object and the field mean value, and the premise can be provided for accurately segmenting the high-resolution first-grade remote sensing image.

(1) RMAS index theory

1) The RMAS index is calculated according to the formula

In the formula,. DELTA.C _l The absolute value of the difference of the domain mean value is used for representing heterogeneity among objects, the difference of the mean value of the wave band l is calculated according to the side length between the adjacent objects, the difference can also be calculated by utilizing the area of the adjacent objects, and the embodiment adopts the domain side length; sigma _l And the standard deviation represents the internal homogeneity of the object, and the internal homogeneity is calculated according to the gray values of all n pixels of one object.

2) The standard deviation is expressed as

Where n is the number of pixels contained in the object, C _li The gray value of the ith pixel point of the ith wave band is represented,

mean value of gray scale, σ, representing image objects _l The closer to 0 the better the internal homogeneity.

3) The domain mean difference absolute value expression is

Where l is the length of the object side, lsj is the length of the common edge of the object adjacent to the jth object, Δ C _l Larger indicates larger differences between objects, and greater heterogeneity.

When the scale is smaller, the edge of the segmented object is smaller than the real ground object, the inside of the object is the same ground object, and the adjacent objects are the same ground object, at this time, the sigma is _l Smaller, Δ C _l Also smaller, so the RMAS value is smaller; with the increase of the scale, the edge of the segmented object is gradually attached to the real ground feature, at the moment, the same ground feature is still in the object, the adjacent objects are different ground features, and the sigma is _l Gradually increase, Δ C _l Also become large until their ratio RMAS index reaches a maximum; when the edge of the segmented object is larger than the real ground object, the inside of the object is different ground objects, sigma _l Gradually increase, Δ C _l The RMAS index decreases gradually and correspondingly. Therefore, RMAS graphs of different land classes can be established, and the corresponding scale is the optimal scale of the land class when the RMAS value reaches the maximum.

In this embodiment, the SSI index and the SDI index are used for evaluation of the optimal scales of different land types. When the optimal segmentation scale is reached and the edge goodness of fit between the segmented object and the actual ground object is high, the positions of the segmented object and the actual ground object are approximately the same, and the spectrums are similar.

1) The expression of the spectral similarity is

In the formula, DN1 is the average brightness value of the reference object, and DN2 is the average brightness value of the divided object.

2) The expression of position similarity is

In the formula, x1 is an abscissa of the center of gravity of the verification object, x2 is an abscissa of the center of gravity of the division object, y1 is an ordinate of the center of gravity of the verification object, and y2 is an ordinate of the center of gravity of the division object. When the spectral similarity SSI and the position similarity SDI are closer to 0, the similarity between the segmentation object and the verification object is higher, and the segmentation effect is better.

(2) Optimal segmentation scale implementation for different land classes

1) Optimal segmentation scale determination for different land classes

The method comprises the steps of selecting pixel points of 6 types of places set in a research area on a high-resolution remote sensing image by Arcgis as training sample points, and carrying out segmentation of different scales in eCoginization software, wherein the segmentation scale range is 10-150, and the step length is 5. After segmentation is completed, RMAS values of the training sample points under different scales of the object are calculated, real outlines of different land types are drawn by reference data to serve as verification sample faces, the verification sample points are selected from the verification sample faces and used for determining the segmentation objects where the verification sample points are located after segmentation of the different scales, the matching degree of the segmentation objects and the verification sample faces in the aspects of area, spectrum, space distance and the like is further calculated to evaluate the segmentation effect, and SDI and SSI are adopted as quantitative evaluation calculation formulas.

And secondly, calculating the segmentation quality evaluation indexes SSI and SDI from the quantitative point of view. When the SSI is smaller, the spectral similarity between the segmentation object and the verification sample surface is higher, and when the SDI is smaller, the gravity center difference between the segmentation object and the verification sample surface is smaller. Therefore, the smaller the two are, the closer the segmentation object is to the verification sample surface of the real contour, and the better the segmentation effect is. And (3) deriving the average brightness and barycentric coordinates of an object where the verification point is located in the eCooginion, calculating segmentation quality evaluation indexes SSI and SDI with a corresponding verification surface, and obtaining the minimum value of SSI and SDI when the RMSA value reaches the peak value and the minimum value of the sum of SSI and SDI when the division scale of all land types except cultivated land and unused land in a calculation result table, wherein the minimum value of the SSI and the SDI is consistent with the RMAS change. Through calculation, the optimal distribution scale ranges of arable land, woodland, construction land, transportation land, water area and unused land in the optimal division scales of different earth surface coverage types are respectively 50-60, 75-85, 25-35, 55-65, 70-80 and 120-130, and the division accuracy is higher and the identification is more convenient on the basis of ensuring the limited division times during the division of different land types.

In this embodiment, the global optimal segmentation scale is established to be 45; the cultivated land, the forest land, the construction land, the transportation land, the water area and the unutilized land in the optimal division scale of the different surface coverage types are respectively 55, 80, 30, 60, 75 and 125; the global segmentation scale is established to be 45, so that the omission of land types with small areas during the land surface coverage classification can be avoided, and the optimal segmentation scale of each main land type can ensure the minimum calculated amount on the basis of accurate identification of the corresponding land surface coverage type.

Second, segmentation optimization.

After the global optimal segmentation scale and the main geo-type segmentation scale are determined, segmentation optimization is performed, a Canny edge detection algorithm is introduced, gradual change type is thick in a high-resolution image, fuzzy edges are restrained through a maximum value, accurate edge information of the geo-type is fitted, a proper amount of fitted edge information is extracted through setting proper parameters, and the optimal segmentation scale obtained through calculation is combined with the high-resolution image to perform segmentation, so that the segmentation is further accurate.

An edge refers to a pixel between the end of a land class and the beginning of another adjacent different land class, and is a part of the image where the pixel is abruptly changed between different land classes. The edge detection mainly judges whether the edge point is an edge point according to the change rate of the direction and the amplitude. The gray value of the pixel point in the image has a small change rate in the edge direction and a large change rate in the edge vertical direction. According to the variation characteristics, the derivation method is usually adopted to extract the edge information. Common edge detection algorithms include Sobel operators, laplacian operators and the like, but the operators have the problems of inaccurate edge positioning, thick edge lines, missing edge information and the like more or less, are easily interfered by noise, extract noise into edges by mistake and have certain errors. The Canny edge detection operator is proposed by John F Canny and widely accepted and cited in a large amount, and is optimized on the basis of general edge detection, non-maximum suppression is added, edges are thinned, and the edge extraction precision is improved. The method mainly comprises four steps of image noise reduction, gradient and direction calculation, non-maximum value inhibition, edge detection and connection.

In the embodiment, a proper amount of edge information is obtained by setting proper parameters, and the edge information is added into segmentation as a single waveband, so that the optimization of the segmentation effect is realized. The edge information in this embodiment is that the low threshold is equal to 150 and the high threshold is equal to 400. When the high threshold is equal to 400, the edge contour is obvious, excessive or too little unnecessary contour information is not contained, but the edge is not communicated, so that a proper low threshold needs to be set to increase the connectivity of the edge, and when the low threshold is less than 150, the boundary information is too rich, and excessive unnecessary information such as detailed contours of farmland ridges or houses can be generated; when the low threshold is larger than 150, the boundary information is greatly reduced, and the defect phenomenon occurs; when the low threshold value is equal to 150, the boundary information is perfect, the ground feature outline can be effectively extracted, and the subsequent processing difficulty is not large due to excessive enrichment.

The image edge information extracted with the low threshold of 150 and the high threshold of 400 is divided as a single layer by adding the eCoginization. The segmentation scale is selected as the global optimal segmentation scale obtained by the calculation and the segmentation optimization is carried out under different terrain optimal segmentation scales, the specific process is the general process, and the detailed description is omitted.

The segmentation optimization effect under the global optimal segmentation scale is used as a display, and the wave band weight is set to be a high-score blue wave band: high-resolution green band: high red band: edge information =1:1:1:1, as shown in fig. 2, it is obvious that the problems of inaccurate edge contour of the segmented object, such as adhesion of adjacent similar objects, trivial building contour, road disconnection and the like can be optimized after the edge information is added.

According to the method, the optimal global segmentation scale and the optimal segmentation scale of different land types are determined, so that the segmentation is more accurate, the premise is provided for high-resolution earth surface coverage classification, meanwhile, after the optimal segmentation scale is established, a Canny edge detection algorithm is adopted, fitted land type edge information is restrained by using a maximum value, the extracted edge information is given with weight and added into the segmentation process, the accuracy of land type segmentation can be greatly improved, and the problem that the accuracy of land type segmentation is influenced due to the fact that homogeneity among the same land types and heterogeneity among different land types are weakened in the high-resolution remote sensing image is effectively solved.

Compared with the existing object-oriented high-resolution classification method, the segmentation scale is generally determined by adopting a trial-and-error method, certain experience is required, subjectivity exists, the judgment is difficult for scenes with more ground species and obvious differences, and the segmentation accuracy is influenced; in the embodiment, the optimal segmentation scale is determined for single-level classification and multi-level classification by a method for quantitatively calculating the optimal segmentation scale, and meanwhile, an edge detection algorithm is introduced to optimize the problem that the edge contour of a segmented object is still inaccurate due to the influence of the complex earth surface form of a high-resolution remote sensing image under the optimal scale segmentation, so that the segmentation accuracy is further improved, and the integrity of subsequent feature extraction and the classification accuracy are improved.

And thirdly, extracting object features.

On the basis of ensuring that at least one sentinel pixel is included in a segmented image formed by segmentation, a multi-source feature space is constructed for feature extraction.

First, a multi-source feature space is constructed.

The method is characterized in that multi-source data are formed by adding sentinel second part waveband data on the basis of high-grade first part data, a comprehensive multi-source characteristic space is constructed by extracting waveband characteristics of sentinel images and high-grade remote sensing images in the same area, and the sentinel images and the high-grade remote sensing images are classified in a coordinated mode.

On the basis of high-score data acquired by a high-score first registration image, 10 sentinel wave band data are added to construct multi-source data, and the multi-source data wave band information is shown in table 1:

TABLE 1 Multi-Source data band correspondence Table

TABLE 2 Multi-Source feature space

A multi-source feature space as shown in table 2 is established, wherein the multi-source feature space has 28 spectral features, 32 geometric features, 12 texture features, 4 exponential features and 76 features.

Each feature in the multivariate feature space is represented and calculated by the following expression:

(1) Spectral characteristics: the recorded spectral reflection information of an object is the most effective characteristic for distinguishing the ground class, and mainly comprises the following classes.

1) The mean value expression is

In the formula

Is the spectral mean of the band l, n is the number of objects, C _li Is the ith object spectral value in the band l.

2) The brightness mean value is expressed as

Wherein b is the mean value of brightness, n is the number of objects,

is the spectral average of the ith band.

3) The standard deviation is expressed as

In the formula σ _l Is the standard deviation of the l-th band, n is the number of objects, C _li A spectral value of the ith object representing the l-th band,

the spectral mean of the l-th band is shown.

4) The maximum variance is expressed as

Wherein CMAXD is the maximum variance,

is the average luminance of the subject v,

is the average brightness, K, of the ith band object v _B Is the image layer luminance weight.

(2) Shape characteristics (geometric characteristics): geometric information used for describing objects, and the shapes of different objects are different, such as roads and rivers which are in strip shapes; a house and a farmland are in a rectangular surface shape, etc. The method mainly comprises the following steps of pixel number, area, length, framework branching degree, width, maximum closed ellipse radius, line segment number, relative boundary, maximum branching length, roundness, shape index, asymmetry, volume, minimum closed ellipse radius, density, rectangle fitting, boundary length, compactness, boundary index, ellipse fitting, polygon edge number, polygon edge length standard deviation, polygon longest edge length, polygon compactness, polygon internal object number, polygon average edge length, polygon perimeter, area including internal polygon, area not including internal polygon, polygon self-intersection, line segment representation average area and line segment representation area standard deviation.

(3) Texture characteristics: the method is used for describing the surface attributes of the ground objects in the image, such as the density, the thickness and the uniformity of the texture of the ground objects. The texture features adopt gray level co-occurrence matrix first-order statistics and gray level difference second-order statistics to respectively obtain gray level co-occurrence matrix feature quantity and gray level difference value vectors, and the category and the description of the gray level co-occurrence matrix feature quantity and the gray level difference value vectors are defined by general texture features.

(4) Self-defining index characteristics: the wave band operation can be carried out on each wave band of the image according to the use requirement so as to highlight index characteristics such as vegetation indexes, soil indexes, water body indexes, building indexes and the like which have obvious difference between a certain class of land features and other classes of land features.

Second, multi-source feature space optimization.

The feature space optimization is mainly divided into feature extraction and feature selection.

The feature extraction is to use a transformation means to carry out feature calculation and combine new features, thereby realizing dimension reduction. Feature selection is the deletion of some of the secondary features. In the existing feature space optimization, in order to explore the influence of original wave bands on classification, a feature space optimization method without transformation is usually adopted.

Filtering selection (Filter) and wrapping selection (Wrapper) are mainly included in feature selection. The filtering type firstly selects the characteristics according to the divergence, the correlation and the like of the analysis characteristics, secondly inputs the screened characteristics into a classifier for training, and the screening process is irrelevant to the classifier; the advantage is that it is fast and intuitive, but does not take into account the inter-feature correlation effects, it is difficult to estimate the criteria of correlation and the correlated subset may not be optimal for the classifier model. And the final classification precision of the classifier is used as a standard for judging the quality of the feature selection in a wrapping mode, and feature optimization is carried out according to the classifier. So wrapped is superior to filtered from a final classifier performance point of view.

The Recursive Feature Elimination (RFE) method is a typical wrapping method, and after a comprehensive multi-source Feature space is constructed by combining sentinel data, firstly, redundant features are effectively eliminated by the Recursive Feature Elimination method, then, a random forest model is selected as a model for text Feature optimization, and the global multi-source Feature space and each land multi-source Feature space are optimized respectively to obtain a global optimal multi-source Feature space and each land optimal multi-source Feature space.

Because the random forest model has lower computational complexity and higher interpretability compared with other models, the capability of learning a complex classification function is strong, the random forest model is easy to use, and too many parameters are not needed; more importantly, both the random forest classification research based on the pixels and the object-oriented coupling research are specially designed in the scheme, and the scheme is combined with a random forest model, so that the feature space can be optimized by using the importance of the random forest to calculate the feature variables, the calculated amount can be effectively reduced, the consumed calculation cost is reduced as much as possible on the premise of not influencing the precision, and the overall classification speed is improved.

(1) Globally optimal multi-source feature space

Firstly, an image composed of training sample vector points and multi-source data is imported into eCogniation. The training sample vector points are:

randomly selecting points in Arcgis to be used as training samples and testing samples, manually interpreting in Google Earth in contrast with high-resolution remote sensing images, manually interpreting the selected sample points and marking the categories of the land. After training sample points are randomly selected, typical sample points of various types are properly added to prepare vector data sets. The statistics of the number of sample points are shown in table 3, with a total of 1080 training sets and 1658 test sets.

TABLE 3 sample number points statistics

In the embodiment, multi-scale segmentation is used, objects are different after segmentation in each scale due to scale difference, and the actual sample attribute is the attribute of the object at the position of the sample point. To facilitate data acquisition, samples are selected in the form of dots. After the segmentation is finished, the attributes of the object where the sample points are located are extracted to be used as a sample data set for training and testing, and when a plurality of sample points fall in the same object, repeated statistics is carried out.

Then, multi-scale segmentation is carried out in eCoginization, 76-dimensional features of the object where the segmented sample points are located are extracted to Excel to form a feature value table shown in a table 4, and the feature value table is input into a random forest classifier to be trained.

TABLE 4 object eigenvalue attributes

After training is completed, importance ranking is carried out on 76-dimensional initial feature space, index features such as buildings, water bodies and vegetation are arranged in front of the initial multi-source feature space in the importance ranking, spectral features are arranged behind the initial multi-source feature space, most of the initial multi-source feature space in the ranking are texture features and geometric features, and therefore the fact that spectral contribution is high and texture and geometric feature contribution is small in ground surface coverage classification is also confirmed.

And then, according to the importance ranking of the initial feature space, setting the recursive elimination step length at each time to be 1, training a classifier by using a new feature space after removing 1 most-ranked feature at each time, and carrying out precision prediction on a classified result. Establishing a relation curve graph of the prediction precision and the feature quantity, wherein the prediction precision fluctuates continuously along with the reduction of the feature quantity, the precision reaches the highest when the feature quantity is 30, the precision starts to decline gradually when the feature quantity is less than 30, and the precision is obviously declined when the feature quantity is less than a certain value, so that the 30-dimensional features are selected as the global optimal multi-source features for training and classification. The final global optimal multi-source feature space is shown in table 5, and has 18 spectral features, 5 geometric features, 4 texture features, 3 exponential features, and 30 features in total.

TABLE 5 Global optimal Multi-Source feature space

By the global optimal multi-source feature space, the object feature extraction can better accord with the real contribution of each land class to the image on the premise of limited calculation amount, and the classification and identification of the land classes in the image are more accurate.

In this embodiment, through multi-source feature space optimization, 76-dimensional feature space data in an initial multi-source feature space is reduced to 30 dimensions, on the premise that accuracy is not affected too much, data volume is effectively reduced, redundant features are removed, and single-level classified and multi-level classified feature spaces are obtained. In the past, the optimal classification characteristic is manually determined mostly through the precision of classification experiments of multiple different characteristic combinations, the method is time-consuming and labor-consuming, needs certain experience and is strong in subjectivity, and the problems are effectively solved by the scheme, so that the characteristic space optimization can be completed more objectively and accurately under the limited calculation amount.

(2) Optimal multisource feature space for different land classes

According to a multi-level classification sequence, a plurality of characteristic value attribute tables are sequentially constructed, each attribute table calculates the optimal characteristic space of the local layer of land samples needing to be peeled and other residual land samples, if a certain land sample is peeled, the information of the land sample is deleted in the next layer of characteristic value attribute table, and the 'category' of the land sample needing to be peeled and other residual land samples in each attribute table is set to be two different values, as shown in table 6. And then inputting the sequence into a random forest classifier for training.

TABLE 6 object eigenvalue attributes

After training is finished, the importance of each layer of features is sorted, a special recursive feature elimination method is used, the step length of each recursive elimination is set to be 1, and a classifier is trained and precision prediction is carried out by using a new feature space after 1 most-ranked feature is removed each time. The optimal multi-source feature space of different land types after recursive elimination is shown in table 7.

TABLE 7 optimal multisource feature space for different terrain classes

The water area and other five types of features with the best distinguishing effect are 44-dimensional, the forest land and other four types of features with the best distinguishing effect are 17-dimensional, the farmland and other three types of features with the best distinguishing effect are 28-dimensional, the unused land and other two types of features with the best distinguishing effect are 5-dimensional, and finally the building land and the traffic land are 72-dimensional.

On the basis of establishing the global optimal multi-source feature space, the optimal multi-source feature space of each land type is established for different land types, so that the accuracy of classification of each land type can be ensured while the dimensionality reduction and simplification are realized.

Object-oriented classification is performed according to feature differences, and therefore, extraction of object features is a key step. The scheme utilizes the great difference of useful characteristics on different land types, enhances the discrimination capability in the model classification process, improves the classification precision, and the redundant characteristics are opposite to the redundant characteristics and influence the classification precision. At present, in the field of ground surface coverage classification research, a single data source is mostly adopted, but most high-resolution remote sensing image data have few wave bands, a comprehensive and distinguishing characteristic space is difficult to construct, limitation exists in distinguishing similar ground types, a multi-source data classification method is adopted in combination with medium-low resolution and pixel-based classification research, spectral information of multiple data sources is mostly fused, but fusion needs to artificially select fusion data, a fusion method and the like according to specific objects and purposes, and the method is complicated. According to the scheme, through special image segmentation and object feature extraction, the features of different remote sensing images are extracted by utilizing multi-source data collaborative classification to construct a multi-source feature space, the distinguishing degree of the land features is improved, redundant features are eliminated through a feature optimization method, and an optimal feature space is provided for single-level and multi-level classification.

And fourthly, classifying based on the object.

Firstly, a multi-level classification structure is arranged.

The embodiment adopts the global optimal segmentation scale to combine with the edge detection information optimal segmentation and adopts the global optimal multi-source characteristic space to perform single-level classification, and aims to improve the classification precision, obtain an accurate confusion matrix and provide a data basis for a multi-level structure. Taking the improved single-level classification as an example, the process of constructing the multi-level classification structure by the confusion matrix is as follows:

the confusion matrix M is obtained by improving the single-layer classification result, as shown in table 8. The elements in the ith row and jth column in the confusion matrix are Mij, and the value of Mij represents the number of samples for dividing the category i into the category j.

(1) First, the normalized confusion matrix shown in Table 9 is constructed by the following expression

Each element in (1)

(2) Next, each element Oij in the overlap matrix O shown in table 10 is constructed by the following expression.

(3) Then, each element Dij in the inter-class distance matrix D shown in table 11 is constructed by the following expression.

D _ij ＝1-O _ij

TABLE 8 confusion matrix M

TABLE 9 normalized confusion matrix

TABLE 10 overlap matrix O

TABLE 11 distance between classes matrix D

After the inter-class distance matrix is obtained, the average distance between the classes is calculated as separable, and the calculation formula is shown as follows.

The average distances of cultivated land, woodland, construction land, transportation land, water area and unused land are respectively 0.965, 0.9686, 0.9384, 0.9638, 0.9796 and 0.9786 through calculation, so that the water area is the most easily distinguished land, the average distance between the remaining 5 types is calculated for the second time after the water area is extracted as the first layer, and the process is circulated until only two types of land remain. The water area (with the average distance of 0.9796 from other 5 types), the unused land (with the average distance of 0.9733 from other 4 types), the farmland (with the average distance of 0.9572 from other 3 types), the forest land (with the average distance of 0.9600 from other 2 types), the transportation land and the construction land are sequentially classified according to the average distance. Therefore, a hierarchical clustering method is used to generate a multi-level classification structure from different classes according to the average distance from large to small, as shown in fig. 3, X in the classification structure is represented as a classifier, each layer in the multi-level structure can use the optimal segmentation scale of different classes of land which is obtained by the calculation in the above way and accords with the current layer, and the optimal segmentation scale is optimized by adding edge information, and then training classification is performed by adopting the optimal feature space of different classes of land.

Aiming at the problems that the single-layer classification structure is difficult to adjust the land types with obvious area difference, the land types with large area are classified and broken, and the land types with small area are missed, the scheme adopts the segmentation scale and the feature space which are suitable for different land types on different layers by constructing a multi-layer classification structure. Calculating the distance mean value of a certain class and other classes through a confusion matrix of one-time single-layer classification to obtain the separability between the classes, dividing the difficulty of distinguishing between the classes according to the separability, and determining the classification sequence according to the first difficulty and the second difficulty. Structurally, a hierarchical clustering algorithm is adopted, and one land class is stripped from each layer. The easy-to-difficult classification sequence ensures that the classification with the smallest confusion is classified firstly in each classification, and the accuracy of subsequent classification is ensured; the two-classification structure simplifies the multi-classification problem into a plurality of two-classification problems, reduces the complexity of classification of each layer and improves the classification performance.

In addition, because the last layer of the binary classification structure is different from each layer of the binary classification structure in the prior art in the aspect of stripping a land type, the last two types of land objects need to be stripped, so that two optimal segmentation scales exist, the segmentation scale with the smaller segmentation scale is selected as the segmentation scale of the last layer, and under-segmentation caused by overlarge segmentation scale is avoided.

Compare in current multilayer classified structure and adopt experience to construct mostly, there is subjectivity, this scheme is through the difference between each land, rationally sets up land classification order, and adopt two categorised thoughts to simplify many classification problems into a plurality of two classification problems, reasonable multilayer classified structure makes multilayer classified structure not rely on experience and directly constructs according to the distinctiveness of each land, makes whole structure more objective, more accords with each land actual conditions.

Secondly, training and setting the classifier.

And substituting the set sample points into a multi-level classification structure for training calculation, and performing surface covering classification on the high-grade first-grade registration image corresponding to the sample points to obtain a classification result.

And fifthly, evaluating the precision.

For the classification result, according to the prior art, the classification result is comprehensively evaluated from the overall accuracy, kappa coefficient, user accuracy and drawing accuracy.

Comparative Experimental example setup

In order to ensure the scientificity and effectiveness, the most common random forest classifier (Ma et al, 2017) with the best classification performance and the most stable in the current earth surface coverage classification is selected as the classifier of the text method and the contrast method. Traditional object-oriented single-level classification is constructed on the classifier to serve as a comparison experiment, and image segmentation, object feature extraction and classification structures are sequentially replaced by the method used in the classification system constructed in the text on the basis of the classifier through controlling variables to carry out the classification experiment, so that five comparison experiment schemes are constructed in total and shown in table 12.

TABLE 12 comparative examples

Experiment A: for the basic contrast experiment, ESP software is adopted to combine visual discrimination in the aspect of determining the optimal segmentation scale in the aspect of image segmentation, a single high-resolution feature space is adopted in the aspect of object feature extraction, a feature recursive elimination method is used for optimization, and a single-level classification structure is adopted in the aspect of a classification structure.

Experiment B: the image segmentation aspect is replaced on the basis of the experiment A, the optimal segmentation scale is calculated quantitatively, edge information is combined for optimal segmentation, and whether the optimization in the image segmentation aspect improves the segmentation accuracy or not is researched through comparison with the experiment A, so that the classification effect is improved.

Experiment C: a separate classification experiment was performed using the same method as experiment B on the sentinel feature space incorporated herein. And the method provides evidences for subsequent collaborative classification experiments by combining multi-source characteristic data.

Experiment D: and replacing the object feature extraction aspect on the basis of the experiment B, adopting a multi-source feature space and a globally optimal multi-source feature space after feature optimization, wherein the experiment provides a classification result for constructing a multi-level classification structure and is also used as a comparison experiment. Compared with the experiment B and the experiment C, whether the optimization in the aspect of object feature extraction makes up the features of the missing wave bands or not is researched, the discrimination between the land categories is improved, and then the classification effect is improved.

Experiment E: and (3) replacing the classification structure on the basis of the experiment D, adopting a multi-level classification structure, adopting optimal segmentation scales suitable for different land types in different layers, and combining edge information optimization segmentation and optimal multi-source feature space suitable for different land types. The experiment is a multi-level classification method constructed by a text complete classification system, and compared with an experiment D, the multi-level classification method is used for researching whether the optimization of the text in the aspect of classification structure effectively solves the problems that the single-level classification method is difficult to adjust the land types with obvious area difference, so that the land types with large areas are classified and crushed, and the land types with small areas are missed, and compared with an experiment A, the multi-level classification method constructed by the text complete classification system is summarized compared with the traditional single-level method.

The specific parameters set for the above experiment were as follows:

(1) Data of

The band data of experiments a, B, C, D, E are shown in tables 13, 14, 15.

TABLE 13 wave band data correspondence table for experiment A and experiment B

TABLE 14 band data correspondence table for experiment C

TABLE 15 wave band data correspondence table for experiment D and experiment E

(2) Segmentation parameters

1) The wave band weight is as follows: experiments a, B, C were set to band 1: band 2: band 3=1:1:1; experiment D because sentinel data resolution is lower than high-score data, the segmentation fineness degree is far less than that of the high-score data, therefore only the high-score data with higher resolution is used for segmentation, the sentinel data is only used for subsequent feature extraction, and the wave band weight is set as wave band 1: band 2: band 3: band 14=1:1:1:1; experiment E is a multi-level classification experiment, so each layer is set to band 1: band 2: band 3: band 14=1:1:1:1. the bands not mentioned in the above experiments are all set to 0 (the bands of different experiments correspond to the bands in different experiment data correspondence tables, respectively).

2) Homogeneity factor: the test is carried out on all the homogenization factors through a trial and error method, and the homogenization factor weights of the experiments A, B, C, D and E are finally determined as color factors: form factor =0.7:0.3, the shapes of the ground objects in the research area are different, and the ground objects have regular shapes and messy shapes, so that the compactness factor in the shape factors is set as follows: smoothness factor =0.5:0.5.

3) And (3) segmentation scale: the experiment A is judged by combining ESP software with visual observation, and finally when the segmentation scale of the experiment A is determined to be 40, the global classification segmentation effect is optimal; the experiments B, C and E adopt the global optimal segmentation scale 45 obtained by the quantitative calculation; in experiment F, the optimal segmentation scale of each feature obtained by the above quantitative calculation was used, and the hierarchical levels were set to 75, 125, 55, 80, and 60, respectively.

(3) Characteristic space

The characteristic space of the experiment A and the experiment B are shown in table 16, the number of the spectral characteristics is 8, the number of the geometric characteristics is 23, the number of the texture characteristics is 12, and the total number of the characteristics is 52; the characteristic space of experiment C is shown in table 17, and there are 22 spectral features, 32 geometric features, 12 texture features, 4 exponential features, and 70 features in total; experiment D uses a global optimal feature space, as shown in table 5; experiment E used a different ground class optimal feature space as shown in table 7.

TABLE 16 feature spaces for experiment A and experiment B

TABLE 17 characteristic space of experiment C

The vector data of the randomly selected verification points are led into an eCoginization, an Error Matrix based on Samples function in the eCoginization is used for counting a confusion Matrix, precision evaluation indexes such as user precision, drawing precision, overall precision, kappa coefficient and the like are calculated through the confusion Matrix, and the classification precision of experiments A, B, C, D and E is shown in tables 18, 19, 20, 21 and 22.

TABLE 18 Classification accuracy of experiment A

TABLE 19 Classification accuracy of experiment B

TABLE 20 Classification accuracy of experiment C

TABLE 21 Classification accuracy of experiment D

TABLE 22 Classification accuracy of experiment E

First, a comprehensive comparative analysis was performed on 5 experimental protocols in terms of overall accuracy and Kappa coefficient, as shown in fig. 4. The multi-level classification experiment E which combines the optimization in three aspects achieves the highest overall precision and Kappa coefficient. In addition to the evidence experiment C, the optimization experiment in the aspects of image segmentation, characteristic information extraction and classification structure is respectively improved by 2.6%, 4.22% and 1.99% in the overall precision compared with the basic single-level classification experiment A, and the Kappa coefficient is also improved. It can thus be preliminarily demonstrated that this solution is clearly superior in terms of overall accuracy and Kappa coefficient.

The comparison of different classification experiments in terms of production accuracy and user accuracy is shown in fig. 5 and 6. The production precision represents the recall ratio, the lower the precision, the more serious the missing score phenomenon, the lower the precision, the user precision represents the precision ratio, and the lower the precision, the more serious the wrong score current situation.

From the above graphs, the experiment E after the optimization of the classification structure further reduces the difference between the production precision and the user precision of different types of land, and reduces the phenomena of misclassification and omission. Finally, the production precision of the multi-level classification method optimized by the method is slightly lower than that of one experiment except for the forest land and the traffic land, and the other land types are all larger than that of the other methods, so that the optimal classification effect is obtained. The accuracy of the user is slightly lower in the unused land, and the other land types are larger or similar to the other methods. The production precision of the method is slightly lower than that of other methods in the aspect of the forest land because the forest land is mostly distributed in a large scale cluster when the optimal division scales of different land types are determined, so that the scale is set to be 80 large, fine forest lands cannot be divided, and a division missing situation is generated.

The analysis shows that the method has obvious effects on image segmentation, object feature extraction and classification structure optimization, the overall precision is improved by 8.81 percent compared with a single-level classification method based on a random forest classifier, the overall precision reaches 87.21 percent, the overall precision is improved by 2.6 percent in segmentation, the overall precision is improved by 4.22 percent in object extraction, and the overall precision is improved by 1.99 percent in hierarchy.

According to the scheme, partial remote sensing images can be obtained from a free open database to carry out earth surface coverage classification, and compared with the existing earth surface coverage classification, the cost of data sources is low; and the advantages and the disadvantages of the remote sensing images of the first high-resolution and the second sentinel are fully utilized, and the extracted features can effectively represent the land type distribution of the actual region corresponding to the remote sensing images.

It is worth explaining that, when the ground surface coverage is classified at present, although the single-level classification calculation amount is small and the structure is simple, the single-level classification method cannot be adopted generally, and because the single-level classification has the problems that the adjustment of the land types with obvious area difference is difficult to carry out in the ground surface coverage classification, the classification of the land types with larger area is broken, and the land types with smaller area are missed; according to the scheme, the separability between the land types is calculated by skillfully utilizing the single-level classification result, and then a multi-level classification method is adopted and the separability is utilized to carry out the land surface coverage classification; the scheme has the advantages that through the combination of a single-level classification method and a multi-level classification method, the problem of strong subjectivity in the existing multi-level classification is solved, and meanwhile, the calculated amount during specific classification is smaller than that in the prior art.

Compared with the existing single-layer classification method based on random forest classifiers, the multi-layer classification method provided by the scheme has the advantages that the classification precision is higher, the classification effect is better, the overall precision reaches 87.21%, the Kappa coefficient reaches 0.8258, and the accurate object-oriented high-resolution remote sensing image ground surface coverage automatic classification can be realized under the condition that a high-resolution one wave band is lost.

The existing computer system which is operated by adopting the object-oriented high-resolution remote sensing image earth surface coverage classification method with low cost, accurate classification and small calculated amount is the object-oriented high-resolution remote sensing image earth surface coverage classification system. In combination with the actual need, the system may be a system connected to a cloud platform.

Example 2

Different from the embodiment 1, in the embodiment, when the high-score first remote sensing image and the sentinel second remote sensing image are selected, the high-score first remote sensing image and the sentinel second remote sensing image are acquired in a time-interval mode according to the weather condition. In the same time period, under the condition that the cloud amount is not large, the image acquisition is preferably carried out under the rain-free condition, the image comparison is carried out in a specified time period, and the image with better condition is selected as the initial image to carry out the subsequent classification operation. For example, within one hour, when the cloud amount difference is less than one percent, preferably no rain, the image acquisition is performed, and after the image acquisition, every half hour interval, whether the comparison is better remote sensing image or not is determined, and if yes, the latest remote sensing image is replaced by the initial remote sensing image. By the arrangement, the initial image serving as the classification object can be clearer and truer without obviously increasing the calculation amount, the scene reduction degree is high, and the classification accuracy of the rear part is not influenced.

Example 3

Different from the embodiment 1, in the embodiment, when performing the ground surface coverage classification, the land type blocks are established for the geographic coordinates, the edge labeling is performed for different land type blocks, and the edge positions of the land type blocks are dynamically adjusted in combination with the meteorological information crawled on the network. For convenience of distinguishing, the same color is adopted when the same land type is marked; and after obtaining the new remote sensing image for classification, comparing the new remote sensing image with the previous land type blocks, specifically comprising land type block edge comparison and internal feature comparison, wherein the internal features comprise colors among the land type blocks. In rainy season, after the edge of the water area land type block is determined in the embodiment 1, combining the acquired meteorological information to dynamically predict whether the edge of the water area in the next time period is expanded or contracted outwards, and forming a prediction marking line outwards or inwards on the determined edge of the water area land type block according to a prediction result; the distance change between the prediction marking line and the edge line of the land type block is in direct proportion to the change of the prediction result; the predicted annotation line is annotated only after the precipitation exceeds one-half of the average precipitation or is less than one-third of the average precipitation. And comparing the remote sensing images of two adjacent times, and judging that the flooding is likely to occur when the distance of the outward expansion of the marked line is predicted to exceed one tenth of the width of the position of the water area land type block so as to give an alarm.

The embodiment combines weather information to predict the area change of the land parcel on the basis of basic classification, and reasonably sets an alarm, thereby facilitating the wider application of classification results.

The above description is only an example of the present invention, and the general knowledge of the known specific technical solutions and/or characteristics and the like in the solutions is not described herein too much. It should be noted that, for those skilled in the art, without departing from the technical solution of the present invention, several variations and modifications can be made, and these should also be considered as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicability of the patent. The scope of the claims of the present application shall be determined by the contents of the claims, and the description of the embodiments and the like in the specification shall be used to explain the contents of the claims.

Claims (10)

1. The object-oriented high-resolution remote sensing image earth surface coverage classification method is characterized by comprising the following steps of:

s1, obtaining an initial remote sensing image, namely obtaining a first high-grade remote sensing image from a first high-grade satellite, obtaining a second sentinel remote sensing image from a second sentinel satellite, and preprocessing data of the second sentinel remote sensing image; registering the high-grade first remote sensing image and the sentinel second remote sensing image to form a high-grade first registered image;

s2, image segmentation, namely performing image segmentation on the high-resolution first-order registration image to obtain block images, wherein each block image at least comprises one sentinel pixel;

s3, extracting object features and constructing a multi-source feature space;

s4, forming a multi-level classification structure based on object classification;

and S5, performing earth surface coverage classification based on the constructed multi-source feature space and multi-level classification structure.

2. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to claim 1, characterized in that in step S2, the segmentation scale is quantitatively calculated through a segmentation quality function composed of a Moran' S I index and an area weighted standard deviation, and an RMAS index, so as to obtain a global optimal segmentation scale and optimal segmentation scales of different earth surface coverage types;

the segmentation quality function is:

GS＝V _norm +MI _norm

in the formula, GS is a segmentation quality function, vnorm is a normalized weighted standard deviation, and MINorm is a normalized Moran's I index.

3. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to claim 2, wherein the earth surface coverage types comprise cultivated land, forest land, construction land, transportation land, water area and unused land.

4. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to claim 3, characterized in that the global optimal segmentation scale range is 10-150; the optimal division scale ranges of arable land, forest land, construction land, transportation land, water area and unused land are respectively 50-60, 75-85, 25-35, 55-65, 70-80 and 120-130.

5. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to the claim 3, characterized in that the global optimal segmentation scale is 45; the optimal division scales of arable land, woodland, construction land, transportation land, water area and unused land are 55, 80, 30, 60, 75 and 125 respectively.

6. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to claim 2, characterized in that after the segmentation is completed, a Canny edge detection algorithm is adopted for segmentation optimization, and the segmentation is participated in according to image edge information extracted according to the conditions that a low threshold value is 150 and a high threshold value is 400.

7. The object-oriented high-resolution remote sensing image surface coverage classification method according to claim 1, characterized in that the multi-source feature space comprises a global multi-source feature space and multi-source feature spaces of various surface coverage types, and each multi-source feature space comprises a spectral feature, a geometric feature, an exponential feature and a texture feature.

8. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to claim 7, characterized in that in S3, an initial multi-source feature space is established, then the initial multi-source feature space is optimized, and redundant features are removed by a recursive elimination method; and then selecting a random forest model as a feature optimization model, and respectively optimizing the global multisource feature space and each earth surface coverage type multisource feature space to obtain a global optimal multisource feature space and each earth surface coverage type optimal multisource feature space.

9. The object-oriented high-resolution remote sensing image earth surface coverage classification method according to claim 1, characterized in that in S4, single-level classification is performed on the constructed multi-source feature space, and separable degree calculation results are obtained by performing separable degree calculation on different earth surface coverage types according to a confusion matrix obtained by the single-level classification; according to the result of the calculation of the separability, an easy-to-first-difficult classification sequence is established, then a hierarchical clustering algorithm is utilized to decompose the multi-classification problem into two classification problems of extracting one earth surface coverage type from each layer, and a multi-layer classification structure is established.

10. The object-oriented high-resolution remote sensing image earth surface coverage classification system is characterized in that the object-oriented high-resolution remote sensing image earth surface coverage classification method of claims 1-9 is adopted for classification.

Images