Object-Oriented High-Resolution Remote Sensing Image Land Cover Classification Method and System

Abstract

The present invention relates to the technical field of remote sensing image processing, and discloses an object-oriented high-resolution remote sensing image land cover classification method and system, comprising the following steps: obtaining remote sensing images of the GF-1 satellite and the Sentinel-2 satellite, and preprocessing the obtained Sentinel-2 remote sensing images; performing image segmentation based on the obtained GF-1 remote sensing images and Sentinel-2 remote sensing images to obtain GF pixels and Sentinel pixels; constructing a multi-source feature space based on the obtained GF pixels and Sentinel pixels, and ensuring that at least one Sentinel pixel is included in one GF pixel; performing machine learning based on the constructed multi-source feature space and performing land cover classification. The present invention has low cost and good discrimination effect.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and particularly relates to an object-oriented high-resolution remote sensing image land cover classification method and system. Background Art

[0002] Land cover includes organic and inorganic coverings laid on the ground surface. Land cover types include cultivated land, forest, grassland, shrubland, wetland, water body, tundra, artificial land surface, bare land, glacier, and permanent snow cover. The study of land cover and its changes is essential basic information and key parameters for environmental change research, national geographic conditions monitoring, sustainable development planning, etc.

[0003] Traditional land cover surveys are time-consuming and laborious. Remote sensing technology, with its characteristics of wide image acquisition range, large amount of information, and less restriction by ground conditions in the acquisition method, provides a new technical means for the rapid acquisition of land cover data. How to extract feature information from images and classify them to accurately obtain land cover information is a hot issue in the remote sensing field. With the improvement of image resolution, high-resolution remote sensing images can clearly and accurately express the spectral, shape, texture and other feature information of the photographed ground objects. Traditional pixel-based classification methods only analyze single pixels from their spectral features and cannot make full use of other useful information, easily resulting in misclassification and missed classification. Object-oriented classification methods, on the other hand, comprehensively consider features such as spectrum, shape, and texture, and have a more powerful information extraction ability than pixel-based methods, which is conducive to improving the classification effect.

[0004] Currently, there are still some deficiencies in the research on object-oriented high-resolution remote sensing image land cover classification methods, resulting in less than ideal classification accuracy. For example, in image segmentation, the determination of the segmentation scale is affected by human subjectivity and the complex ground morphology during segmentation, leading to a reduction in segmentation accuracy; in feature extraction, most studies use a single high-resolution image data source with fewer band data, making it difficult to construct a comprehensive and discriminative feature space; in the classification structure, a single-level classification structure is difficult to adjust for different land classes in terms of segmentation scale and feature space, affecting information extraction and classification effects.

[0005] In order to solve the above problems, people usually start with the collected input data set and purchase a high-precision input data set formed by multiple satellite data sources after complex processing. However, this method of improving the accuracy of classification results based on the source of high-precision input data sets does not have promotion potential due to its high investment cost. Because now all companies that provide such high-precision input data set services basically obtain specific high-precision input data sets after one-to-one special processing of multiple different satellite data sources based on user needs. This processing method is not universal and does not have the premise for promotion. Secondly, because user needs vary greatly, there is a lot of subjective adjustment space in selecting which satellite data source to use and how to achieve the specified accuracy. The uncertainty is large, and the amount of data processed is huge, resulting in high overall processing costs, which is very unfavorable for large-scale promotion. Summary of the invention

[0006] The present invention aims to provide an object-oriented high-resolution remote sensing image surface coverage classification method to achieve the purpose of low cost and good distinction effect.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme:

[0008] The object-oriented high-resolution remote sensing image land cover classification method includes the following steps:

[0009] S1. Initial remote sensing image acquisition: acquiring Gaofen-1 remote sensing images from Gaofen-1 satellite, acquiring Sentinel-2 remote sensing images from Sentinel-2 satellite, and performing data preprocessing on Sentinel-2 remote sensing images; registering Gaofen-1 remote sensing images with Sentinel-2 remote sensing images to form Gaofen-1 registered images;

[0010] S2, image segmentation, performing image segmentation on the GF-1 registered image to obtain block images, each of which contains at least one sentinel pixel;

[0011] S3, object feature extraction, construction of multi-source feature space;

[0012] S4, based on object classification, forming a multi-level classification structure;

[0013] S5, land cover classification is performed based on the constructed multi-source feature space and multi-level classification structure.

[0014] The principles and advantages of this solution are:

[0015] This solution can directly utilize the remotely sensed images of GF-1 satellite and Sentinel-2 satellite which can be obtained for free, with relatively low cost. By using the GF-1 remotely sensed images and Sentinel-2 remotely sensed images complementarily, their respective disadvantages are eliminated as much as possible. During image segmentation, at least one Sentinel pixel is included in each segmented image to ensure the effective data volume, providing a prerequisite for the subsequent classification accuracy. Then, a multi-source feature space is constructed. Finally, based on object classification, a multi-level classification structure is formed to achieve land cover classification.

[0016] The difference between this solution and the existing land cover classification methods is that, in the inherent perception, land cover classification requires a large number of high-quality and high-resolution remotely sensed image resources to achieve good classification results. However, this solution uses almost unprocessed data sources that can be obtained for free, with relatively low cost. At the same time, by combining with Sentinel-2, the advantage of its multiple bands is used to make up for the lack of the near-infrared band of GF-1. It can not only make full use of the spatial resolution advantage of GF-1 remotely sensed data, but also make full use of the spectral resolution advantage of Sentinel-2 remotely sensed data, achieving better classification results with a low-cost database.

[0017] Compared with the shortcoming of the current object-oriented analysis method in segmentation, although there are some methods for quantitatively calculating the optimal segmentation scale, due to the strong theoretical and scientific nature of the methods, the determination process is relatively complex and lacks intuitiveness. Especially for the optimal segmentation scale point, there is a large degree of subjectivity, making it difficult to judge in scenarios with many and obvious differences in land cover types and being affected by the complex surface morphology in high-resolution remotely sensed image data, resulting in relatively low segmentation accuracy. This solution overcomes this problem through a specially constructed multi-source feature space. Even when there are few bands in general high-resolution remotely sensed image data and it is difficult to construct a comprehensive and discriminative feature space, and there are still limitations in distinguishing similar land cover types, it can still achieve the target accuracy, effectively solving the problem of poor accuracy of the current general remotely sensed data sources.

[0018] 1. As an improvement, in step S2, the segmentation quality function composed of Moran's I index and area-weighted standard deviation, and the RMAS index are used to quantitatively calculate the segmentation scale, obtaining the global optimal segmentation scale and the optimal segmentation scales of different land cover types.

[0019] The segmentation quality function is as follows:

[0020] GS = V norm + MI norm

[0021] In the formula, GS is the segmentation quality function, Vnorm is the normalized weighted standard deviation, and MInorm is the normalized Moran's I

[0022] index.

[0023] Effect: The proposed solution uses the Moran's I index plus the area-weighted standard deviation to form a segmentation quality function, which can avoid the problem of excessive subjectivity in the existing segmentation scales.

[0024] As an improvement, the land cover types include cultivated land, forest land, construction land, transportation land, water area, and unused land.

[0025] Effect: Using the above land cover types can meet the classification of most land cover objects.

[0026] As an improvement, the global optimal segmentation scale range is 10 - 150; the optimal segmentation scale ranges for cultivated land, forest land, construction land, transportation land, water area, and unused land are 50 - 60, 75 - 85, 25 - 35, 55 - 65, 70 - 80, and 120 - 130 respectively.

[0027] Effect: By establishing the global optimal segmentation scale and the optimal segmentation scale ranges for different land cover types respectively, different segmentation scales can be used for different land classes and globally, which can avoid the fragmentation of large-area land classes and the omission of small-area land classes.

[0028] As an improvement, the global optimal segmentation scale is 45; the optimal segmentation scales for cultivated land, forest land, construction land, transportation land, water area, and unused land are 55, 80, 30, 60, 75, and 125 respectively.

[0029] Effect: Setting the optimal segmentation scale in this way can take into account both global segmentation and the classification of the main land cover types, and minimize the data calculation amount on the premise of ensuring accurate division.

[0030] As an improvement, after segmentation, the Canny edge detection algorithm is used for segmentation optimization, and the image edge information extracted according to the low threshold of 150 and the high threshold of 400 participates in the segmentation.

[0031] Effect: It can well solve the weakening of the homogeneity between the same land classes and the heterogeneity between different land classes in high-resolution remote sensing images. The edges of some land classes are gradually changing and there are no obvious mutations, resulting in inaccurate edge contours of some segmentation objects after the image is segmented using the optimal segmentation scale; at the same time, compared with the existing solutions, the improvement in segmentation significantly improves the recall rate of land classes with regular shapes such as buildings, and the overall accuracy is increased by 2.6%.

[0032] As an improvement, the multi-source feature space includes a global multi-source feature space and multi-source feature spaces for each land cover type, and each multi-source feature space includes spectral features, geometric features, index features, and texture features.

[0033] Effect: The established multi-source feature space can be made to be as consistent as possible with the characteristics of the actual area.

[0034] As an improvement, in S3, first establish an initial multi-source feature space, then optimize the initial multi-source feature space. Using the recursive elimination method, redundant features are eliminated; then a random forest model is selected as the feature optimization model to optimize the global multi-source feature space and the multi-source feature spaces of each land cover type respectively, so as to obtain the global optimal multi-source feature space and the optimal multi-source feature spaces of each land cover type.

[0035] Effect: Make the obtained global optimal multi-source feature space and the optimal multi-source feature spaces of each land cover type more conform to the actual area situation.

[0036] As an improvement, in S4, perform single-level classification on the constructed multi-source feature space, calculate the separability of different land cover types according to the confusion matrix obtained from the single-level classification to obtain the separability calculation result; according to the separability calculation result, construct an easy-first-and-difficult-second classification order, and then use the hierarchical clustering algorithm to decompose the multi-classification problem into a binary classification problem of extracting one land cover type at each level, and construct a multi-level classification structure.

[0037] Effect: Calculate the separability between land cover classes through single-level classification, then construct the order of difficulty of separating land cover classes according to the separability between land cover classes, and finally use the binary classification method for classification according to the order of difficulty; such classification has a small calculation amount, and at the same time, based on the actual separability, it can avoid the problem of strong subjectivity in the existing multi-level classification.

[0038] This application also provides an object-oriented high-resolution remote sensing image land cover classification system, which uses the above classification method for classification. It can accurately and quickly distinguish different land cover types. Brief Description of the Drawings

[0039] Figure 1 It is the flowchart of the land cover classification method in Embodiment 1 of the present invention.

[0040] Figure 2 It is the comparison diagram of the segmentation optimization results in Embodiment 1 of the present invention.

[0041] Figure 3 It is the schematic diagram of the multi-level classification structure in Embodiment 1 of the present invention.

[0042] Figure 4 It is the comparison diagram of the overall accuracy and Kappa coefficient of different classification experiments in Embodiment 1 of the present invention.

[0043] Figure 5This is a comparison chart of production accuracy for different classification experiments in Embodiment 1 of the present invention.

[0044] Figure 6 This is a comparison chart of user accuracy for different classification experiments in Embodiment 1 of the present invention. Detailed implementation manners

[0045] The following is a further detailed description through specific implementation manners:

[0046] Embodiment 1 is basically as shown in the appendix: Figure 1 The object-oriented high-resolution remote sensing image land cover classification method in this embodiment includes the following:

[0047] First step, initial remote sensing image acquisition. Obtain the GF-1 remote sensing image from the GF-1 satellite, obtain the Sentinel-2 remote sensing image from the Sentinel-2 satellite, and perform data preprocessing on the Sentinel-2 remote sensing image; register the GF-1 remote sensing image and the Sentinel-2 remote sensing image to form a GF-1 registered image.

[0048] Most of the remote sensing images of the GF-1 and Sentinel-2 satellites can be obtained from publicly available free databases. This solution has no excessive requirements for the initial data source. Even free data sources that can be obtained by the general public can complete land cover classification with a certain accuracy through this method.

[0049] Among them, the GF-1 remote sensing image, after fusion processing, has only three bands of R, G, and B, and the resolution can be 2 meters, 8 meters, or 16 meters, with a cloud cover of less than 2%-5%. The GF-1 remote sensing image in this embodiment has a resolution of 2 meters and a cloud cover of less than 3%.

[0050] The GF-1 satellite is the first high-resolution satellite independently developed in China, which has overcome the key problem of incompatible spatial and spectral resolutions and uses a relatively wide swath for effective detection. When acquired, the GF-1 remote sensing image has already undergone conventional preprocessing processes such as geometric correction, mosaicking, fusion, cropping, radiometric calibration, and atmospheric correction, and no additional preprocessing is required.

[0051] The Sentinel-2 remote sensing image in this embodiment is the L1C-level image directly downloaded from the European Space Agency, which is an image that has undergone the above preprocessing steps except for atmospheric correction, with a cloud cover of less than 4.5%.

[0052] The European Space Agency defines atmospherically corrected data as L1A data. It provides the Sen2cor plug-in tool, which can convert L1C-level images into L1A-level images after atmospheric correction, but L1A-level images cannot be downloaded directly. In this embodiment, the directly acquired L1C-level Sentinel-2 remote sensing image is called through the Sen2cor plug-in using the CMD command configuration, atmospheric correction is performed, and all bands are resampled to 10 meters in the SNAP software to complete data preprocessing to obtain the processed Sentinel-2 remote sensing image. The preprocessed image has a higher saturation and darker color than the pre-processed image, which is more conducive to the subsequent registration and feature extraction with the GF-1 remote sensing image.

[0053] The Sentinel-2 data were aligned with the high-resolution image and resampled to the same resolution as the high-resolution image using the nearest neighbor method. The processed Sentinel-2 data were imported into ENVI for layer stack operation to remove the low-resolution water vapour, coastal aerosol, The Cirrus three-band group adds 10 bands of remote sensing data.

[0054] In this embodiment, the image registration of the GF-1 remote sensing image and the Sentinel-2 remote sensing image can be performed using the existing general image registration method.

[0055] The second step is image segmentation. The GF-1 registered image is segmented to obtain block images. Each block image contains at least one sentinel pixel, so that it can fully utilize the high spatial resolution advantage of the GF-1 remote sensing data and the high spectral resolution advantage of the Sentinel-2 remote sensing data.

[0056] First, determine the optimal segmentation scale.

[0057] The segmentation indexes at different segmentation scales are quantitatively calculated through the segmentation quality function of Moran's I index, area-weighted standard deviation, and RMAS index, and the global optimal segmentation scale with the maximum homogeneity within the object and the maximum heterogeneity between objects and the optimal segmentation scale for different land types that meet the needs of different layers of the multi-level classification structure are found respectively; in this embodiment, the surface cover type is referred to as land type.

[0058] Specifically, the optimal global segmentation scale is first determined; then the optimal segmentation scale for different land cover types is established. By determining the global optimal segmentation scale, the internal heterogeneity of the segmented object is minimized as much as possible, and the heterogeneity between objects is maximized as much as possible, balancing the heterogeneity within and between objects in the entire image, thereby achieving a better distinction effect and providing a scale for single-level classification.

[0059] This embodiment adopts an optimal segmentation scale calculation model based on the segmentation quality function (Johnson and Xie, 2011), and quantitatively calculates the global optimal segmentation scale by statistically analyzing the segmentation quality at different segmentation scales. Different from other global optimal segmentation scale calculations, this method not only considers the homogeneity information within the object, but also adds the Moran's I index to statistically analyze the correlation between global objects, taking into account the principle of heterogeneity between objects. It is more objective and accurate than the traditional visual and ESP software methods for determining the global optimal segmentation scale, and can effectively reduce the influence of human subjectivity to improve the accuracy of segmentation.

[0060] (1) Theory of the segmentation quality function calculation model

[0061] This function uses the area-weighted standard deviation of the objects after segmentation to measure the spectral homogeneity within the object, and uses the global Moran's I index in the spatial autocorrelation index to measure the heterogeneity between objects.

[0062] 1) The expression for spectral homogeneity within the object is

[0063]

[0064] where wVar represents the area-weighted spectral standard deviation, v i is the variance, a i is the area of object i, and n is the total number of objects after segmentation in the image. The larger the wVar, the greater the heterogeneity within the object and the smaller the homogeneity.

[0065] 2) The expression for heterogeneity between objects is

[0066]

[0067] where MI is the global Moran's I index, w ij is the spatial adjacency relationship between object i and object j. If object i is adjacent to object j (i.e., the regions sharing a boundary), then w ij = 1; otherwise, it is 0. y i and y j represent the spectral means of object i and object j, is the spectral mean of the entire image. The lower the MI, the lower the correlation between objects, the higher the heterogeneity, and thus the higher the separability.

[0068] 3) The normalization expression

[0069] Since both of the above two index metrics are calculated for all bands during the image segmentation process, in order to equally consider the indices within and between bands, the normalization formula in the formula is used to readjust both to a similar [0, 1] range.

[0070] (X-X min ) / (X max -X min )

[0071] In the formula, Xmin and Xmax are the minimum and maximum values of the standard deviation of the spectral area weighted or the Moran's I index respectively. Normalization makes the normalized values of the bands with low area weighted standard deviation or Moran's I index relatively close to zero.

[0072] 4) The expression of the segmentation quality function is

[0073] GS = V norm + MI norm

[0074] In the formula, GS is the segmentation quality function, Vnorm is the normalized weighted standard deviation, and MInorm is the normalized Moran's I index. The determination of the global optimal segmentation scale is achieved by calculating the GS values of each spectral band respectively, averaging them, and using the averaged GS value to evaluate the segmentation quality. The scale with the lowest average GS value is the optimal segmentation scale because at this scale, the combination of the weighted standard deviation of the spectrum and spatial autocorrelation is the lowest, that is, the within-object homogeneity is the highest and the between-object heterogeneity is the largest.

[0075] (2) Implementation of the global optimal segmentation scale

[0076] Set the range of the global optimal segmentation scale to 10 - 150, and perform multiple segmentations with a step size of 5, and count the segmentation quality function of each band obtained from each segmentation. This can minimize the number of segmentations and reduce the computational amount while meeting the accuracy requirements. Considering the segmentation effects at different scales in some areas of the study area, over-segmentation occurs when the segmentation scale in the study area is less than 10, and even the smallest land use type will be divided into multiple objects, which is not conducive to the integrity of subsequent feature extraction; when the segmentation scale exceeds 150, under-segmentation occurs, and multiple land use types are included in one object, greatly increasing the misclassification rate and causing difficulties for subsequent accurate classification. Therefore, the scale range of 10 - 150 is the optimal choice.

[0077] During implementation, first import the high-resolution data of the GF-1 registered image into eCognition for multi-scale segmentation, then export the standard deviation and area information of the three bands of the segmented objects to Excel, and calculate the area weighted standard deviation of the three bands through Excel; secondly, export the segmentation result vector to Arcgis, and use the spatial statistics tool in Arcgis to calculate the Moran's I index of the three bands; finally, statistically calculate the normalized Moran's I index and the weighted standard deviation.

[0078] On the basis of determining that the global optimal segmentation scale range is 10 - 150, the optimal segmentation scales for different land cover types are further determined.

[0079] Since the sizes and spectral differences of land classes in the image are obvious, a single segmentation scale is difficult to ensure that the boundaries of the segmentation objects of each land class can fit well with the true contours of the land classes. Therefore, different segmentation scales need to be set for each land class, that is, the optimal segmentation scales for different land classes are determined to provide segmentation scales for the multi-level classification structure. In this embodiment, the ratio of the absolute value of the difference between the object and the domain mean to the standard deviation of the object - the RMAS index (The Ratio Of Mean DifferenceTo Neighbors To Standard Deviation) is used. By calculating the RMAS values of different land classes at different segmentation scales, the optimal segmentation scales for different land classes are obtained to reduce the subjective influence of people and improve the segmentation accuracy. Considering the internal homogeneity of the object, the absolute value of the difference between the object and the domain mean is also used to consider the heterogeneity between objects, which can provide a premise for the accurate segmentation of GF-1 remote sensing images.

[0080] (1) RMAS index theory

[0081] 1) The calculation formula of the RMAS index is

[0082]

[0083] In the formula, ΔC l is the absolute value of the difference between the domain means, representing the heterogeneity between objects. The difference in the average value of band l is calculated according to the side lengths of adjacent objects, or the area of adjacent objects can also be used for calculation. In this embodiment, the domain side length is adopted; σ l is the standard deviation, representing the internal homogeneity of the object, which is obtained by operating on the gray values of all n pixels of an object.

[0084] 2) The expression of the standard deviation is

[0085]

[0086] In the formula, n is the number of pixels contained in the object, C li represents the gray value of the i-th pixel point in the l-th band, represents the gray mean value of the image object, and σ l the closer to 0, the better the internal homogeneity.

[0087] 3) The expression of the absolute value of the difference between the domain means is

[0088]

[0089] Where l is the side length of the object, lsj is the length of the common side of the object adjacent to the j-th object, and ΔC l The larger it is, the greater the difference and heterogeneity between objects.

[0090] At a small scale, the edge of the segmented object is smaller than the real ground object. The object interior is of the same type of ground object, and adjacent objects are also of the same type of ground object. At this time, σ l is small, and ΔC l is also small. Therefore, the RMAS value is small. As the scale increases, the edge of the segmented object gradually fits the real ground object. At this time, the object interior is still of the same type of ground object, while adjacent objects are of different types of ground objects. σ l gradually increases, and ΔC l also becomes larger until the ratio of them, the RMAS index, reaches the maximum. When the edge of the segmented object is larger than the real ground object, the object interior is of different types of ground objects, σ l gradually increases, and ΔC l gradually decreases, and the RMAS index decreases accordingly. Therefore, an RMAS curve graph for different land classes can be established. The scale corresponding to the maximum RMAS value is the optimal scale for this land class.

[0091] In this embodiment, the SSI index and SDI index are used to evaluate the optimal scale of different land classes. When the optimal segmentation scale is reached and the edge of the segmented object coincides well with the actual ground object, the position of the segmented object should be approximately the same as that of the actual ground object and the spectra should be similar.

[0092] 1) The spectral similarity expression is

[0093]

[0094] Where DN1 is the average brightness value of the reference object, and DN2 is the average brightness value of the segmented object.

[0095] 2) The position similarity expression is

[0096]

[0097] Where x1 is the abscissa of the centroid of the verification object, x2 is the abscissa of the centroid of the segmented object, y1 is the ordinate of the centroid of the verification object, and y2 is the ordinate of the centroid of the segmented object. When the spectral similarity SSI and the position similarity SDI are closer to 0, it indicates that the similarity between the segmented object and the verification object is higher and the segmentation effect is better.

[0098] (2) Realization of the optimal segmentation scale for different land classes

[0099] 1) Determination of the optimal segmentation scale for different land classes

[0100] Using Arcgis, pixel points were selected as training sample points for 6 land types set in the study area on high-resolution remote sensing images, and segmentation at different scales was carried out in the eCognition software. The segmentation scale range was from 10 to 150, with a step size of 5. After segmentation, the RMAS values of the objects where the training sample points were located at different scales were calculated. Then, referring to the data, the true contours of different land types were outlined as verification sample surfaces, and verification sample points were selected in the verification sample surfaces to determine the segmentation objects where the verification sample points were located after segmentation at different scales. Furthermore, the degree of coincidence between the segmentation objects and the verification sample surfaces in terms of area, spectrum, and spatial distance was calculated to evaluate the segmentation effect. The quantitative evaluation calculation formulas used were SDI and SSI.

[0101] Secondly, the segmentation quality evaluation indices SSI and SDI were calculated from a quantitative perspective. The smaller the SSI, the higher the spectral similarity between the segmentation object and the verification sample surface. The smaller the SDI, the smaller the difference in the centroid between the segmentation object and the verification sample surface. Therefore, the smaller both of them are, the closer the segmentation object is to the verification sample surface of the true contour, and the better the segmentation effect. In eCognition, the average brightness and centroid coordinates of the objects where the verification points were located were exported, and the segmentation quality evaluation indices SSI and SDI were calculated with the corresponding verification surfaces. In the calculation result table, for each land type except cultivated land and unused land, when the RMSA value reached the peak, SSI and SDI obtained the minimum values and the sum of the two was the smallest, which was consistent with the change of RMAS. After calculation, the optimal segmentation scale ranges for cultivated land, forest land, construction land, transportation land, water area, and unused land among different land cover types were 50 - 60, 75 - 85, 25 - 35, 55 - 65, 70 - 80, and 120 - 130 respectively. When segmenting different land types, on the basis of ensuring a limited number of segmentation times, the segmentation accuracy was higher and it was more convenient for identification.

[0102] In this embodiment, the global optimal segmentation scale was established as 45; among the optimal segmentation scales of different land cover types, the values for cultivated land, forest land, construction land, transportation land, water area, and unused land were 55, 80, 30, 60, 75, and 125 respectively; establishing the global segmentation scale as 45 can avoid omitting small land types during land cover classification. The optimal segmentation scales of the above main land types can ensure the smallest calculation amount on the basis of accurate identification of the corresponding land cover categories.

[0103] Secondly, segmentation optimization.

[0104] After determining the global optimal segmentation scale and the main land-cover type segmentation scales in this embodiment, segmentation optimization is carried out. The Canny edge detection algorithm is introduced to suppress the maxima of the relatively thick and blurred edges in the high-resolution image, fit the accurate edge information of the land-cover types, and extract an appropriate amount of the fitted edge information by setting appropriate parameters. Then, combined with the high-resolution image, segmentation is performed using the optimal segmentation scale calculated above, further making the segmentation more accurate.

[0105] An edge refers to the pixels between the end of one land-cover type and the start of another adjacent different land-cover type, which is the part where the pixels mutate between different land-cover types in the image. Edge detection mainly determines whether a point is an edge point by the rate of change of direction and amplitude. The gray value of a pixel point in the image changes at a relatively small rate in the edge direction and at a relatively large rate in the direction perpendicular to the edge. According to this change characteristic, the derivative method is usually used to extract edge information. Common edge detection algorithms include the Sobel operator, the Laplacian operator, etc. However, these operators more or less have problems such as inaccurate edge localization, thick edge lines, missing edge information, and are easily affected by noise, misextracting noise as edges, resulting in certain errors. The Canny edge detection operator was proposed by John F Canny and is widely recognized and widely cited. He optimized on the basis of general edge detection, added non-maximum suppression to sparsify the edges and improve the accuracy of edge extraction. Its steps are mainly divided into four steps, namely image denoising, calculating gradients and directions, non-maximum suppression, and edge detection and connection.

[0106] In this embodiment, an appropriate amount of edge information is obtained by setting appropriate parameters, and the edge information is added as a separate band to the segmentation to optimize the segmentation effect. The edge information in this embodiment has a low threshold equal to 150 and a high threshold equal to 400. When the high threshold is equal to 400, the edge contour is obvious and does not contain too much or too little unnecessary contour information. However, there are discontinuous edges, so an appropriate low threshold needs to be set to increase its connectivity. When the low threshold is less than 150, the boundary information is too rich, resulting in too much unnecessary information such as cultivated land ridges or detailed house contours; when the low threshold is greater than 150, the boundary information is greatly reduced, showing a defect phenomenon; when the low threshold is equal to 150, the boundary information is relatively complete, which can effectively extract the ground object contour and will not be too rich to cause difficulties in subsequent processing.

[0107] The image edge information extracted with a low threshold of 150 and a high threshold of 400 is added as a separate layer to eCognition for segmentation. The segmentation scale is selected for segmentation optimization under the global optimal segmentation scale and the optimal segmentation scales of different land-cover types calculated above. The specific process is the same as the above general process and will not be elaborated here.

[0108] The segmentation optimization effect at the global optimal segmentation scale is shown as follows. The band weights are set as High-Resolution Blue Band: High-Resolution Green Band: High-Resolution Red Band: Edge Information = 1:1:1:1. The comparison of some segmentation results is as Figure 2 shown. It can be clearly seen that after adding edge information, the problems of inaccurate edge contours of segmentation objects such as the adhesion of adjacent similar ground objects, fragmented building outlines, and discontinuous roads can be optimized.

[0109] In this embodiment, not only by determining the optimal global segmentation scale and the optimal segmentation scale for different land classes, the segmentation is made more accurate, providing a premise for high-resolution land cover classification. At the same time, after the optimal segmentation scale has been established, the Canny edge detection algorithm is also adopted to fit the land class edge information by using maximum suppression, and the extracted edge information is given weights and added to the segmentation process, which can greatly improve the accuracy of land class segmentation and effectively solve the problem that the accuracy of land class segmentation is affected due to the homogeneity between the same land classes and the weakening of the heterogeneity between different land classes in high-resolution remote sensing images.

[0110] Compared with the current object-oriented high-resolution classification method, the determination of the segmentation scale generally adopts the trial-and-error method, which requires certain experience and is subjective. It is difficult to judge for scenes with many and obvious differences in land objects, which has an impact on the segmentation accuracy. In this embodiment, by quantitatively calculating the optimal segmentation scale, the optimal segmentation scale is determined for single-level classification and multi-level classification. At the same time, the edge detection algorithm is introduced to optimize the problem that the edge contours of segmentation objects are still inaccurate due to the influence of the complex surface morphology of high-resolution remote sensing images under the optimal scale segmentation, further improving the accuracy of segmentation, thereby enhancing the integrity of subsequent feature extraction and the accuracy of classification.

[0111] The third step is object feature extraction.

[0112] On the basis of ensuring that at least one sentinel pixel is included in a segmented block image, a multi-source feature space is constructed for feature extraction.

[0113] First, the construction of the multi-source feature space.

[0114] In this paper, some Sentinel-2 band data are added to the data of GF-1 to form multi-source data. By extracting the band features of the same area of the Sentinel image and the high-resolution remote sensing image, a comprehensive multi-source feature space is constructed for collaborative classification.

[0115] In this paper, 10 Sentinel band data are added to the high-resolution data obtained from the GF-1 registered image to form multi-source data. The band information of the multi-source data is shown in Table 1:

[0116] Table 1 Correspondence table of multi-source data bands

[0117]

[0118] Table 2 Multi-source Feature Space

[0119]

[0120] Establish a multi-source feature space as shown in Table 2. There are 28 spectral features, 32 geometric features, 12 texture features, and 4 index features in the multi-source feature space, for a total of 76 features.

[0121] Among them, each feature in the multi-source feature space is characterized and calculated using the following expressions:

[0122] (1) Spectral features: Record the spectral reflection information of the object, which is the most effective feature for distinguishing land cover types and mainly includes the following categories.

[0123] 1) The mean value expression is

[0124]

[0125] In the formula is the spectral mean of band l, n is the number of objects, and C li is the spectral value of the i-th object within band l.

[0126] 2) The brightness mean value expression is

[0127]

[0128] In the formula, b is the brightness mean value, n is the number of objects, is the spectral mean of the i-th band.

[0129] 3) The standard deviation expression is

[0130]

[0131] In the formula, σ l is the standard deviation of the l-th band, n is the number of objects, and C li represents the spectral value of the i-th object in the l-th band, represents the spectral mean of the l-th band.

[0132] 4) The maximum variance expression is

[0133]

[0134] In the formula, CMAXD is the maximum variance, is the average brightness of object v, is the average brightness of object v in the i-th band, and K B is the brightness weight of the image layer.

[0135] (2) Shape features (geometric features): Used to describe the geometric information of objects. Different objects have different shape manifestations. For example, roads and rivers are strip-shaped; houses and farmlands are rectangular and planar, etc. It mainly includes the number of pixels, area, length, skeleton branching degree, width, radius of the largest enclosed ellipse, number of line segments, relative boundary, maximum branch length, roundness, shape index, asymmetry, volume, radius of the smallest enclosed ellipse, density, rectangle fitting, boundary length, compactness, boundary index, ellipse fitting, number of polygon sides, standard deviation of polygon side lengths, length of the longest polygon side, polygon compactness, number of internal objects in the polygon, average side length of the polygon, polygon perimeter, area including internal polygons, area excluding internal polygons, polygon self-intersection, average area represented by line segments, standard deviation of the area represented by line segments.

[0136] (3) Texture features: Used to describe the surface attributes of ground objects in the image, such as the density, fineness, and uniformity of the ground object texture. The texture features use the first-order statistics of the gray-level co-occurrence matrix and the second-order statistics of the gray-level difference. The gray-level co-occurrence matrix feature quantities and gray-level difference vectors are obtained respectively. The categories and descriptions can be defined using the general texture feature definition.

[0137] (4) Custom index features: According to the usage requirements, band operations can be performed on each band of the image to highlight the index features with obvious differences between certain types of ground objects and other types of ground objects, such as vegetation index, soil index, water body index, building index, etc.

[0138] Secondly, optimize the multi-source feature space.

[0139] Feature space optimization is mainly divided into feature extraction and feature selection.

[0140] Feature extraction is to use transformation means to calculate features and combine them into new features to achieve dimensionality reduction. Feature selection is to delete some secondary features. In current feature space optimization, in order to explore the influence of the original bands on classification, the method of feature selection without transformation is often used for feature space optimization.

[0141] Feature selection mainly includes filter selection (Filter) and wrapper selection (Wrapper). Filter selection first selects features according to the divergence and correlation of the analyzed features, and then inputs the selected features into the classifier for training. The screening process is independent of the classifier. Its advantages are fast and intuitive, but it does not consider the correlation between features, it is difficult to estimate the correlation criteria, and the relevant subsets may not be optimal for the classifier model. Wrapper selection uses the final classification accuracy of the classifier as the criterion for evaluating the quality of feature selection and optimizes features according to the classifier. Therefore, from the perspective of the performance of the final classifier, wrapper selection is better than filter selection.

[0142] Recursive Feature Elimination (RFE) is a typical wrapper method. After constructing a comprehensive multi-source feature space in combination with sentinel data in this embodiment, redundant features are first effectively eliminated by the recursive feature elimination method, and then the random forest model is selected as the model for feature optimization in this article. The global multi-source feature space and the multi-source feature space of each land class are optimized respectively to obtain the global optimal multi-source feature space and the optimal multi-source feature space of each land class.

[0143] Because the random forest model has lower computational complexity and higher interpretability compared with other models, it has strong ability to learn complex classification functions, is easy to use, and does not require too many parameters. More importantly, both the pixel-based random forest classification research and the coupling research with object-oriented are specially designed in this solution. By combining with the random forest model, this solution can not only use the importance of the random forest to calculate feature variables to optimize the feature space, but also effectively reduce the computational amount, and make the consumed computational cost as little as possible without affecting the accuracy, so as to improve the overall classification speed.

[0144] (1) Global optimal multi-source feature space

[0145] First, the image composed of training sample vector points and multi-source data is imported into eCognition. The training sample vector points are:

[0146] Random points are selected in Arcgis as training samples and test samples, and artificial interpretation is carried out by comparing with high-resolution remote sensing images in Google Earth. The selected sample points are manually interpreted and marked with land class categories. After randomly selecting training sample points, typical sample points of each land class are appropriately added to make a vector data set. The statistics of the number of sample points are shown in Table 3. There are 1080 in the training set and 1658 in the test set.

[0147] Table 3 Statistics of the number of sample points

[0148]

[0149] In this embodiment, multi-scale segmentation is used. Due to scale differences, the objects after segmentation at each scale are different. The actual sample attributes are the attributes of the objects at the positions of the sample points. For the convenience of data acquisition, the samples are selected in the form of points. After the segmentation is completed, the attributes of the objects where the sample points are located are extracted as the sample data set for training and testing. When multiple sample points fall within the same object, they are repeatedly counted.

[0150] Then, multi-scale segmentation is carried out in eCognition, and 76-dimensional features of the objects where the sample points are located after segmentation are extracted into Excel to form a feature value table as shown in Table 4, and then it is input into the random forest classifier for training.

[0151] Table 4 Object Feature Value Attributes

[0152]

[0153] After training, an importance ranking was performed on the 76-dimensional initial feature space. Among the initial multi-source feature spaces, the index features such as buildings, water bodies, and vegetation ranked high in importance, followed by spectral features. The texture features and geometric features ranked relatively low, which also confirmed that the spectral contribution was relatively high and the texture and geometric feature contributions were relatively small in land cover classification.

[0154] Then, according to the importance ranking of the above initial feature space, the step size of each recursive elimination was set to 1. The classifier was trained using the new feature space after removing the feature with the lowest rank each time, and the accuracy of the classified results was predicted. A curve graph of the relationship between the prediction accuracy and the number of features was established. As the number of features decreased, the prediction accuracy fluctuated continuously. The accuracy reached the highest when the number of features was 30, started to gradually decline when it was less than 30, and decreased significantly when the number of features was less than a certain value. Therefore, 30-dimensional features were selected as the globally optimal multi-source features for training and classification. The final globally optimal multi-source feature space is shown in Table 5, including 18 spectral features, 5 geometric features, 4 texture features, and 3 index features, for a total of 30 features.

[0155] Table 5 Globally Optimal Multi-Source Feature Space

[0156]

[0157] Through the above globally optimal multi-source feature space, on the premise of limited computational effort, the extraction of object features can be made more in line with the true contributions of each land type to the image, making the classification and recognition of land types in the image more accurate.

[0158] In this embodiment, through the optimization of the multi-source feature space, the data of the 76-dimensional feature space in the initial multi-source feature space was reduced to 30 dimensions. On the premise of not overly affecting the accuracy, the data volume was effectively reduced, redundant features were removed, and then the single-level classification and multi-level classification feature spaces were obtained. In the past, the optimization of the feature space mostly determined the optimal classification features manually through the accuracy of multiple classification experiments with different feature combinations. This method was not only time-consuming and laborious, required certain experience and was subjective, while this solution effectively solved these problems and could complete the feature space optimization more objectively, accurately and with limited computational effort.

[0159] (2) Globally Optimal Multi-Source Feature Space for Different Land Types

[0160] According to the multi-level classification order, multiple eigenvalue attribute tables are constructed in sequence. Each attribute table calculates the optimal feature space for the land cover samples to be stripped and the remaining land cover samples at this level. For example, if a certain land cover has been stripped, the information of the samples of this land cover will be deleted in the eigenvalue attribute table of the next level, and the "categories" of the classes to be stripped and the remaining classes in each attribute table are set to two different values, as shown in Table 6. Then, they are sequentially input into the random forest classifier for training.

[0161] Table 6 Object Eigenvalue Attributes

[0162]

[0163] After the training is completed, the importance of the features at each level is ranked. Using the recursive feature elimination method, the step size of each recursive elimination is set to 1, and the new feature space after removing the last ranked feature each time is used to train the classifier and predict the accuracy. The optimal multi-source feature spaces for different land covers after recursive elimination are shown in Table 7.

[0164] Table 7 Optimal Multi-Source Feature Spaces for Different Land Covers

[0165]

[0166] The features with the best discrimination effect between water areas and the other five categories are 44-dimensional in total, the features with the best discrimination effect between forest land and the other four categories are 17-dimensional in total, the features with the best discrimination effect between cultivated land and the other three categories are 28-dimensional in total, the features with the best discrimination effect between unused land and the other two categories are 5-dimensional in total, and finally, the features with the best discrimination effect between construction land and transportation land are 72-dimensional in total.

[0167] On the basis of establishing the global optimal multi-source feature space, the optimal multi-source feature space for each land cover is established, so as to ensure the accuracy of the classification of each land cover while reducing the dimension and streamlining.

[0168] Object-oriented classification is based on feature differences. Therefore, the extraction of object features is a key step. This solution takes advantage of the significant differences in useful features among different land classes to enhance the discriminative ability during the model classification process and improve the classification accuracy. In contrast, redundant features have the opposite effect and affect the classification accuracy. In current research in the field of land cover classification, most studies use a single data source. However, most high-resolution remote sensing image data have relatively few bands, making it difficult to construct a comprehensive and discriminative feature space, which has limitations in differentiating similar land classes. Moreover, in the research on combining medium- and low-resolution data and pixel-based classification, although there are methods that use multi-source data classification, most of them fuse the spectral information of multiple data sources. However, the fusion requires manual selection of fusion data, fusion methods, etc. according to specific objects and purposes, which is rather cumbersome. This solution uses special image segmentation and object feature extraction, and conducts collaborative classification using multi-source data to extract the features of different remote sensing images to construct a multi-source feature space, enhancing the differentiability of land class features. Redundant features are eliminated through feature optimization methods to provide the optimal feature space for single-level and multi-level classification.

[0169] Fourth, object-based classification.

[0170] First, set the multi-level classification structure.

[0171] In this embodiment, single-level classification is performed using the optimized segmentation combined with edge detection information under the global optimal segmentation scale and the global optimal multi-source feature space, aiming to improve the classification accuracy and obtain an accurate confusion matrix to provide a data basis for the multi-level structure. Taking the improved single-level classification mentioned above as an example, the process of constructing the multi-level classification structure through the confusion matrix is as follows:

[0172] The confusion matrix M is obtained from the improved single-level classification result, as shown in Table 8. The element in the i-th row and j-th column of the confusion matrix is Mij, and the value of Mij represents the number of samples classified as class j from class i.

[0173] (1) First, construct the normalized confusion matrix shown in Table 9 through the following expression for each element

[0174]

[0175] (2) Second, construct each element Oij of the overlap matrix O shown in Table 10 through the following expression.

[0176]

[0177] (3) Then, construct each element Dij of the inter-class distance matrix D shown in Table 11 through the following expression.

[0178] D ij = 1 - O ij

[0179] Table 8 Confusion Matrix M

[0180]

[0181] Table 9 Normalized Confusion Matrix

[0182]

[0183] Table 10 Overlap Matrix O

[0184]

[0185] Table 11 Inter-class Distance Matrix D

[0186]

[0187] After obtaining the inter-class distance matrix, the average distance between each pair of land classes is calculated as the separability, and the calculation formula is as follows.

[0188]

[0189] Through calculation, the average distances of cultivated land, forest land, construction land, transportation land, water area and unused land classes are 0.965, 0.9686, 0.9384, 0.9638, 0.9796, 0.9786 respectively. Therefore, the water area is the most easily distinguishable land class. So, after extracting the water area as the first layer, the average distance between the remaining 5 classes is calculated for the second time, and this cycle continues until only two land classes remain. Finally, according to the magnitude of the average distance, the classification order is obtained as follows: water area (average distance from other 5 classes is 0.9796), unused land (average distance from other 4 classes is 0.9733), cultivated land (average distance from other 3 classes is 0.9572), forest land (average distance from other 2 classes is 0.9600), transportation land and construction land. Therefore, through the hierarchical clustering method, different classes are generated into a multi-level classification structure from large to small according to the magnitude of the average distance as Figure 3 shown. In this classification structure, X represents the classifier, and at each layer of the multi-level structure, the optimal segmentation scale of different land classes obtained from the above calculation that conforms to the current layer can be used, and edge information is added for optimized segmentation, and then different land classes' optimal feature spaces are used for training classification.

[0190] In view of the problem that the single-level classification structure is difficult to adjust land classes with obvious area differences, resulting in fragmented classification of land classes with large areas and missed classification of land classes with small areas, this solution constructs a multi-level classification structure and adopts segmentation scales and feature spaces suitable for different land classes at different levels. Calculate the mean distance between a certain class and other classes through the confusion matrix of a single-level classification once, obtain the separability between classes, classify the difficulty of separation between classes according to the separability, and determine the classification order according to the principle of "easier first and then more difficult". The hierarchical clustering algorithm is used in the structure, and one land class is peeled off at each level. The classification order of "easier first and then more difficult" ensures that the class with the least confusion is classified first each time, ensuring the accuracy of subsequent classifications; the binary classification structure simplifies the multi-classification problem into multiple binary classification problems, reduces the complexity of each level of classification, and improves the classification performance.

[0191] In addition, since the last layer of the binary classification structure is different from peeling off one land class at each previous layer and needs to peel off the last two types of ground objects, there are two optimal segmentation scales. In this paper, the smaller segmentation scale is selected as the segmentation scale of the last layer to avoid under-segmentation caused by too large a segmentation scale.

[0192] Compared with most of the existing multi-level classification structures that are constructed based on experience and are subjective, this solution reasonably sets the classification order of land classes according to the differences between land classes, simplifies the multi-classification problem into multiple binary classification problems by adopting the idea of binary classification, and has a reasonable multi-level classification structure, making the multi-level classification structure not rely on experience but directly construct according to the distinctiveness of each land class, making the whole structure more objective and more in line with the actual situation of each land class.

[0193] Secondly, classifier training and classification settings.

[0194] Substitute the sample points set as described above into the multi-level classification structure for training calculation, and classify the surface cover of the GF-1 registered image corresponding to the sample points to obtain the classification result.

[0195] The fifth step is accuracy evaluation.

[0196] For the classification result, according to the existing technology, comprehensively evaluate the classification result from the overall accuracy, Kappa coefficient, user accuracy, and cartographic accuracy.

[0197] Setting of comparative experimental examples

[0198] To ensure scientificity and effectiveness, the random forest classifier (Ma et al, 2017), which is the most commonly used, has the best classification performance, and is the most stable in the current land cover classification, is selected as the classifier for both the method and the comparative method in this paper. On this classifier, a traditional object-oriented single-level classification is constructed as a comparative experiment. By controlling variables, on the basis of this classifier, the image segmentation, object feature extraction, and classification structure are successively replaced with the methods used in the classification system constructed in this paper for classification experiments. A total of five comparative experiment schemes are constructed as shown in Table 12.

[0199] Table 12 Comparative Experiment Examples

[0200]

[0201] Experiment A: This is the basic comparative experiment of this paper. In determining the optimal segmentation scale in image segmentation, the ESP software combined with visual discrimination is used. In object feature extraction, a single high-resolution feature space is used, and the feature recursive elimination method is used for optimization. In the classification structure, a single-level classification structure is used.

[0202] Experiment B: On the basis of Experiment A, the image segmentation is replaced. The optimal segmentation scale is calculated quantitatively and combined with edge information to optimize the segmentation. By comparing with Experiment A, it is explored whether the optimization in image segmentation can improve the segmentation accuracy and thus enhance the classification effect.

[0203] Experiment C: The same method as Experiment B is used for a separate classification experiment on the Sentinel feature space combined in this paper. It provides evidence for subsequent collaborative classification experiments with multi-source feature data.

[0204] Experiment D: On the basis of Experiment B, the object feature extraction is replaced. A multi-source feature space and the globally optimal multi-source feature space after feature optimization are used. This experiment provides the classification results for constructing a multi-level classification structure and is also used as a comparative test. By comparing with Experiment B and Experiment C, it is explored whether the optimization in object feature extraction can make up for the features of missing bands, improve the discrimination between land classes, and thus enhance the classification effect.

[0205] Experiment E: On the basis of Experiment D, the classification structure is replaced. A multi-level classification structure is adopted, and the optimal segmentation scale suitable for different land classes combined with edge information to optimize the segmentation and the optimal multi-source feature space suitable for different land classes are used in different layers. This experiment is a multi-level classification method constructed through the complete classification system of this paper. By comparing with Experiment D, it is explored whether the optimization in the classification structure of this paper can effectively solve the problem that the single-level classification method is difficult to adjust for land classes with obvious area differences, resulting in fragmented classification of larger land classes and missed classification of smaller land classes. By comparing with Experiment A, the advantages and disadvantages of the multi-level classification method constructed through the complete classification system of this paper compared with the traditional single-level method are summarized.

[0206] The specific parameter settings for the above experiments are as follows:

[0207] (1) Data

[0208] The band data of Experiments A, B, C, D, and E are shown in Tables 13, 14, and 15.

[0209] Table 13 Corresponding Table of Band Data for Experiment A and Experiment B

[0210]

[0211] Table 14 Corresponding Table of Band Data for Experiment C

[0212]

[0213] Table 15 Corresponding Table of Band Data for Experiment D and Experiment E

[0214]

[0215] (2) Segmentation Parameters

[0216] 1) Band weights: For Experiments A, B, and C, the band weights are set as Band 1: Band 2: Band 3 = 1:1:1; for Experiment D, since the Sentinel data resolution is lower than that of the high-resolution data and the segmentation fineness is much lower than that of the high-resolution data, only the high-resolution data with higher resolution is used for segmentation, and the Sentinel data is only used for subsequent feature extraction, and the band weights are set as Band 1: Band 2: Band 3: Band 14 = 1:1:1:1; for Experiment E, which is a multi-level classification experiment, each layer is set as Band 1: Band 2: Band 3: Band 14 = 1:1:1:1. Bands not mentioned in the above experiments are all set to 0 (the bands of different experiments correspond to the bands in the corresponding tables of different experimental data).

[0217] 2) Homogeneity factor: In this paper, all combinations of homogeneity factors are experimented through the trial-and-error method, and finally, the homogeneity factor weights of Experiments A, B, C, D, and E are determined as Color factor: Shape factor = 0.7:0.3. Moreover, the ground objects in the study area have different shapes, including both regular-shaped and messy-shaped objects. Therefore, the compactness factor: smoothness factor in the shape factor is set as 0.5:0.5.

[0218] 3) Segmentation scale: For Experiment A, through the ESP software combined with visual discrimination, it is finally determined that when the segmentation scale of Experiment A is 40, the global land cover segmentation effect is the best; for Experiments B, C, and E, the globally optimal segmentation scale of 45 obtained from the previous quantitative calculation is adopted; for Experiment F, the optimal segmentation scales for each ground object obtained from the previous quantitative calculation are used, and they are set as 75, 125, 55, 80, and 60 respectively for each layer.

[0219] (3) Feature Space

[0220] For Experiments A and B, the feature spaces are shown in Table 16, with 8 spectral features, 23 geometric features, and 12 texture features, for a total of 52 features; for Experiment C, the feature space is shown in Table 17, with 22 spectral features, 32 geometric features, 12 texture features, and 4 index features, for a total of 70 features; Experiment D uses the globally optimal feature space, as shown in Table 5; Experiment E uses the optimal feature space for different land classes, as shown in Table 7.

[0221] Table 16 Feature Spaces of Experiments A and B

[0222]

[0223] Table 17 Feature Space of Experiment C

[0224]

[0225] Import the vector data of the previously randomly selected verification points into eCognition, use the ErrorMatrix based on Samples function in eCognition to calculate the confusion matrix, and calculate accuracy evaluation indicators such as user accuracy, mapping accuracy, overall accuracy, and Kappa coefficient through the confusion matrix. The classification accuracies of Experiments A, B, C, D, and E are shown in Tables 18, 19, 20, 21, and 22.

[0226] Table 18 Classification Accuracy of Experiment A

[0227]

[0228] Table 19 Classification Accuracy of Experiment B

[0229]

[0230] Table 20 Classification Accuracy of Experiment C

[0231]

[0232] Table 21 Classification Accuracy of Experiment D

[0233]

[0234] Table 22 Classification Accuracy of Experiment E

[0235]

[0236] First, comprehensively compare and analyze the five experimental schemes in terms of overall accuracy and Kappa coefficient, as Figure 4As shown in the figure. The multi-level classification experiment E that incorporates the optimizations in three aspects of this paper achieved the highest overall accuracy and Kappa coefficient. Except for the verification experiment C, the optimization experiments of this paper in image segmentation, feature information extraction, and classification structure, compared with the basic single-level classification experiment A, increased the overall accuracy by 2.6%, 4.22%, and 1.99% respectively, and also improved in terms of the Kappa coefficient. Therefore, it can be preliminarily proven that this solution has obvious advantages in terms of overall accuracy and Kappa coefficient.

[0237] The comparison of different classification experiments in terms of production accuracy and user accuracy is as Figure 5 、 Figure 6 shown. Production accuracy represents the recall rate, and the lower the accuracy, the more serious the phenomenon of missed classification. User accuracy represents the precision rate, and the lower the accuracy, the more serious the phenomenon of misclassification.

[0238] From the above charts, it can be seen that experiment E after optimizing the classification structure further narrowed the difference between the production accuracy and user accuracy of different land types, and reduced the phenomenon of misclassification and missed classification. Finally, the multi-level classification method optimized by this paper, except that the production accuracy of forest land and transportation land is slightly lower than one of the experiments, is greater than other methods for the rest of the land types, achieving the best classification effect. In terms of the user accuracy of the unused land rate, it is slightly lower than other methods, and for the rest of the land types, it is also greater than or similar to other methods. In terms of forest land, the production accuracy of the method in this paper is slightly lower than other methods because when determining the optimal segmentation scale for different land types, forest land is mostly distributed in large clusters, so the scale is set to a relatively large 80, resulting in small forest land that cannot be segmented out, causing missed classification. However, in terms of user accuracy, the method in this paper is much greater than other methods, indicating that the selection of the optimal scale and optimal features in this paper for large forest land is relatively correct, and the selected scale and features can accurately distinguish large forest land.

[0239] Through analysis, it is clear that the optimizations of this paper in image segmentation, object feature extraction, and classification structure all have obvious effects. The overall accuracy is improved by 8.81% compared with the single-level classification method based on the random forest classifier, reaching 87.21%. It is improved by 2.6% in segmentation, 4.22% in object extraction, and 1.99% in the hierarchy.

[0240] This solution can obtain some remote sensing images from publicly available databases for land cover classification, with a lower data source cost compared to existing land cover classification; and it fully utilizes the advantages and disadvantages of GF-1 and Sentinel-2 remote sensing images, effectively ensuring that the extracted features can represent the actual regional land type distribution corresponding to the remote sensing images.

[0241] It should be noted that when classifying surface cover at present, although single-level classification has a small computational load and a simple structure, the single-level classification method is usually not adopted because there are major problems in single-level classification for surface cover classification, such as it is difficult to adjust land classes with obvious area differences, which will lead to fragmented classification of land classes with larger areas and missed classification of land classes with smaller areas; while this solution cleverly uses the results of single-level classification to calculate the separability between land classes, and then adopts the multi-level classification method and uses the separability for surface cover classification; through the combination of the single-level classification method and the multi-level classification method, this solution avoids the problem of strong subjectivity existing in the existing multi-level classification, and at the same time has the advantage that the computational load is smaller compared with the existing technology when performing specific classification.

[0242] The multi-level classification method provided by this solution has higher classification accuracy and better classification effect than the single-level classification method based on the random forest classifier that is currently used more. The overall accuracy reaches 87.21%, and the Kappa coefficient reaches 0.8258. It can achieve relatively accurate object-oriented automatic classification of surface cover of high-resolution remote sensing images in the case of the lack of GF-1 bands.

[0243] An existing computer system that runs using an object-oriented high-resolution remote sensing image surface cover classification method with low cost, accurate classification and small computational load is an object-oriented high-resolution remote sensing image surface cover classification system. According to actual needs, this system can be a system connected to a cloud platform.

[0244] Embodiment 2

[0245] Different from Embodiment 1, in this embodiment, when selecting GF-1 remote sensing images and Sentinel-2 remote sensing images, it is obtained by a time-division method according to weather conditions. Within the same time period, when the cloud amount difference is not significant, it is preferably to obtain images under rainless conditions, and compare the images within the specified time period, and select the image with better conditions as the initial image for subsequent classification operations. For example, within one hour, when the cloud amount difference is less than one percent, it is preferably to obtain images under rainless conditions, and compare whether there are better remote sensing images at intervals of every half hour after obtaining the images. If so, replace the latest remote sensing image with the initial remote sensing image. Such a setting can make the initial image as the classification object clearer, more real, with a high scene restoration degree, and does not affect the subsequent classification accuracy without significantly increasing the computational load.

[0246] Embodiment 3

[0247] Different from Embodiment 1, in this embodiment, when performing surface cover classification, land type blocks are established based on geographical coordinates, and edge labeling is performed for different land type blocks. Combining the meteorological information crawled from the Internet, the edge positions of each land type block are dynamically adjusted. For the convenience of distinction, the same color is used for labeling the same land type; and after obtaining a new remote sensing image for classification, it is compared with the previous land type blocks, specifically including the comparison of the edges of the land type blocks and the comparison of the internal features. The internal features include the colors between the land type blocks. For example, during the rainy season, after the edge of the water area land type block is determined in Embodiment 1, the meteorological information obtained is combined to dynamically predict whether the edge of the water area will expand or contract in the next time period. According to the prediction result, a predicted labeling line is formed outward or inward from the determined edge of the water area land type block; the change in the distance between the predicted labeling line and the edge line of the land type block is proportional to the change in the prediction result; the predicted labeling line is only labeled when the precipitation exceeds one-half of the average precipitation or is less than one-third of the average precipitation. After comparing two adjacent remote sensing images, when the distance by which the predicted labeling line expands outward exceeds one-tenth of the width of the water area land type block at this position, it is determined that a flood may occur and an alarm is issued.

[0248] Based on the basic classification, this embodiment combines meteorological information to predict the area change of the land type blocks and reasonably sets an alarm, which is convenient for the wider application of the classification results.

[0249] The above are only embodiments of the present invention. Specific technical solutions and / or common knowledge such as characteristics known in the art are not described in detail herein. It should be noted that for those skilled in the art, without departing from the technical solution of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, which will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be subject to the content of its claims, and the specific implementation manners described in the specification can be used to interpret the content of the claims.

Claims

1. Object-oriented high-resolution remote sensing image land cover classification method, characterized in that, it includes the following steps: S1. Initial remote sensing image acquisition: Obtain the GF-1 remote sensing image from the GF-1 satellite, obtain the Sentinel-2 remote sensing image from the Sentinel-2 satellite, and perform data preprocessing on the Sentinel-2 remote sensing image; register the GF-1 remote sensing image and the Sentinel-2 remote sensing image to form a GF-1 registered image; S2. Image segmentation: Segment the GF-1 registered image to obtain segmented images, and each segmented image contains at least one Sentinel pixel; S3. Object feature extraction: Construct a multi-source feature space; S4. Object-based classification: Form a multi-level classification structure; S5. Based on the constructed multi-source feature space and multi-level classification structure, perform land cover classification; In step S2, the segmentation quality function composed of Moran's I index and area-weighted standard deviation, and the RMAS index are used to quantitatively calculate the segmentation scale to obtain the global optimal segmentation scale and the optimal segmentation scales of different land cover types; The segmentation quality function is: where GS is the segmentation quality function, Vnorm is the normalized weighted standard deviation, and MInorm is the normalized Moran's I index; The multi-source feature space includes a global multi-source feature space and multi-source feature spaces for each land cover type, and each multi-source feature space includes spectral features, geometric features, index features, and texture features; In S3, first establish an initial multi-source feature space, and then optimize the initial multi-source feature space. The recursive elimination method is used to eliminate redundant features; Then, a random forest model is selected as the feature optimization model to optimize the global multi-source feature space and multi-source feature spaces for each land cover type respectively, to obtain the global optimal multi-source feature space and the optimal multi-source feature spaces for each land cover type; In S4, perform single-level classification on the constructed multi-source feature space, calculate the separability of different land cover types according to the confusion matrix obtained from the single-level classification to obtain the separability calculation result; according to the separability calculation result, construct a classification order from easy to difficult, and then use the hierarchical clustering algorithm to decompose the multi-classification problem into a binary classification problem of extracting one land cover type for each layer, and construct a multi-level classification structure.

2. The object-oriented high-resolution remote sensing image land cover classification method according to claim 1, characterized in that, the land cover types include cultivated land, forest land, construction land, transportation land, water area, and unused land.

3. The object-oriented high-resolution remote sensing image land cover classification method according to claim 2, characterized in that, the range of the global optimal segmentation scale is 10-150; the ranges of the optimal segmentation scales of cultivated land, forest land, construction land, transportation land, water area, and unused land are 50-60, 75-85, 25-35, 55-65, 70-80, and 120-130 respectively.

4. The object-oriented high-resolution remote sensing image land cover classification method according to claim 2, characterized in that, The global optimal segmentation scale is 45; the optimal segmentation scales of cultivated land, forest land, construction land, transportation land, water area and unused land are 55, 80, 30, 60, 75 and 125 respectively.

5. The object-oriented high-resolution remote sensing image surface cover classification method according to claim 1, characterized in that, after the segmentation is completed, the Canny edge detection algorithm is used for segmentation optimization, and the image edge information extracted according to the low threshold of 150 and the high threshold of 400 participates in the segmentation.

6. An object-oriented high-resolution remote sensing image surface cover classification system, characterized in that, the object-oriented high-resolution remote sensing image surface cover classification method described in any one of claims 1-5 is used for classification.

Images