A Feature Selection Method for Geo-Object Image Analysis

Abstract

The present invention discloses a feature selection method for object-oriented image analysis of geographical objects, which can optimize the feature space by simulating the process of manually constructing ground object extraction rules, use self-created evaluation indicators to achieve feature decorrelation and optimization of feature subsets, and give the classification order and soft threshold of features, which can be directly used as input to obtain a detailed dataset of regional land cover types. Compared with other feature selection algorithms, this method uses a feature combination evaluation indicator to construct a ground object classification rule instead of using a single feature evaluation indicator, and the constructed ground object extraction rule is more reasonable; it uses a membership function combined with a soft threshold to describe ground objects, rather than a "non-zero or one" classification criterion, making the classification result closer to the language and thinking mode of humans. The implementation process of the present invention is simple, greatly reducing the influence of manual intervention, with remarkable application effects, and can realize the automatic construction of ground object extraction rules in the process of object-oriented ground object classification.

Description

Technical Field

[0001] The present invention belongs to the field of feature selection for geographic object-based image analysis, and relates to a feature selection method for geographic object-based image analysis. This method takes the land cover classification of high-resolution remote sensing images as the actual application background and can be used in application directions such as land cover classification in complex environments, extraction of typical target objects, and change monitoring of land use types based on time series. Background Art

[0002] Geographic object-based image analysis (GEOBIA) technology is an emerging and rapidly developing research field in remote sensing information extraction. It makes full use of various features of remote sensing images, such as spectral, shape, texture, topology, semantics, and time phase, and comprehensively utilizes multi-source information, such as GIS data, digital elevation model (DEM), and landscape ecology and human geography thematic data. It integrates mainstream image analysis methods, such as supervised classification, fuzzy mathematics classification, and rule-based classification, and to a certain extent eliminates the phenomena of "same object with different spectra" and "same spectrum with different objects" that are prone to occur in the pixel-based classification process, effectively improving the accuracy of information extraction. Since the proposal of GEOBIA technology more than a decade ago, automation and intelligence have been the biggest development trends of this technology. However, feature selection and rule set construction have become the key factors restricting the automation development of GEOBIA. Currently, there are mainly two major problems in feature selection and rule set construction: one is that it is difficult to determine which features are effective for object extraction; the other is that the differences in data sources limit the portability of classification rule sets. Therefore, feature optimization and construction of classification rule sets are still time-consuming and challenging research tasks at present.

[0003] At present, feature selection methods are mainly divided into two categories: those based on image information content and those based on between-class separability. The former includes the optimal index factor (OIF), adaptive band selection (ABS), auto-subspace partition (ASP), etc.; the latter includes methods based on Euclidean distance (ED), density peaks cluster (DPC), scatter degree (SD), and separability and threshold (SEaTH), etc. Among them, the OIF method has a large computational amount. The ABS model is an improved OIF, which is convenient and fast, effectively shortening the running time of OIF. However, it is difficult to select truly representative bands, and the accuracy is not ideal for the classification of more complex image targets. The ASP method comprehensively considers band correlation and information content, but the selected bands are more likely to be distributed in the same subspace, making it difficult to achieve the best dimensionality reduction effect. The ED-based feature selection method only evaluates from the class mean, without considering the class distribution and between-class overlap. The SD model does not consider the mutual influence between features, and DPC also lacks the measurement of band information content.

[0004] The SEaTH algorithm involved in the present invention was first proposed by Nussbaum et al. This algorithm is based on the Gaussian mixture model, which can not only select classification features but also automatically calculate feature thresholds, significantly improving the accuracy and efficiency of classification. It is a relatively representative object-oriented feature optimization method and is widely used. However, this algorithm also has limitations. For example, it only uses the J-M distance to judge a single feature, there may be strong correlations between features and the classification order cannot be determined. In addition, the threshold given by this algorithm is only applicable to hard classification and cannot provide clear and adjustable feature descriptions. Relevant scholars at home and abroad have carried out optimization research on the above problems existing in the SEaTH algorithm, mainly focusing on the threshold calculation problem when the features do not conform to the normal distribution and comprehensively evaluating the feature bands using between-class distance and within-class distance, etc. However, the thresholds obtained by the existing improvement ideas are still not applicable to soft classification, and no scheme for finding the best feature combination is proposed, making the extraction of ground objects still unable to achieve the theoretically optimal effect, and the ground object classification order also depends too much on the professional skills and experience judgment of classification personnel to a certain extent. Summary of the Invention

[0005] The object of the present invention is to overcome the defects of the prior art and propose a feature selection method for geographical object image analysis - EXaST (extractability and softthreshold), which can optimize the feature space by simulating the process of manually constructing ground object extraction rules. EXaST first creates the extraction ability E of a single feature for a ground object a This evaluation index, combined with the Pearson correlation coefficient, decorrelates the feature library. Then, on the premise of retaining the extraction ability of the set for the target ground object, the feature subset is optimized, and all feature combinations with the same number of distinguishable classes as the subset are searched in the decorrelated subset, and the extraction ability C of the feature combination for the target ground object is created e This evaluation index measures the quality of the feature combination, and finally obtains the optimal feature combination. In addition, the SEaTH algorithm itself does not take into account the problem of classification order, and the calculated feature threshold is only applicable to hard classification, while the EXaST algorithm can determine the ground object classification order according to the number of distinguishable classes and the C value of the optimal feature combination required for extracting each ground object e value, then determines the soft threshold between classes according to the mean and variance of the distribution of ground object feature values, describes the ground object in combination with the fuzzy classification method, and then classifies the ground object.

[0006] The technical solution of the present invention is as follows:

[0007] A feature selection method for geographical object image analysis, comprising the following steps:

[0008] (1) Obtain image objects based on image segmentation technology, select typical samples according to the proportion of ground objects, and then output their features and feature values;

[0009] (2) Calculate the autocorrelation coefficient matrix of the ground object to be extracted and all other ground objects based on the Pearson correlation coefficient, and set an appropriate threshold to adjust the number of selected features;

[0010] (3) Perform feature decorrelation in the autocorrelation coefficient matrix in turn according to the extraction ability value of the feature for the target ground object;

[0011] (4) For the feature set obtained in (3), taking the extraction ability of the feature combination for the target ground object as the evaluation index, obtain the optimal feature combination of a certain type of ground object, and complete the optimization of the feature subset;

[0012] (5) Repeat (1)-(4), calculate the best feature combination for extracting each ground object, and determine which ground object to extract first based on the number of distinguishable classes and the extraction ability value of each extracted feature combination;

[0013] (6) Repeat steps (1)-(5) until only two types of ground objects remain unextracted, and obtain the optimal features that can separate them based on the separability between classes;

[0014] (7) Calculate the soft threshold for extracting the best features of each type of ground object, and then combine it with the membership function to construct the extraction rules for each type of ground object, thus realizing the extraction of ground objects.

[0015] Furthermore, the image segmentation method is multi-scale segmentation. The scale factor, spectral factor, and smoothness for segmentation are obtained through multiple experiments. After obtaining the ground object, according to the representativeness of the samples and the uniformity of their spatial distribution, approximately 2.5% of the typical samples are selected according to the proportion of the ground objects, and then their features and feature values are output.

[0016] Furthermore, the autocorrelation coefficient matrix r can be expressed as:

[0017]

[0018] In the above formula, r ij (i, j = 1,..., N) is the Pearson correlation coefficient, indicating the correlation degree between two features, and can be expressed as:

[0019]

[0020] Let the object category be C n =(c1, c2,..., c n ), and the original high-dimensional feature set be F N =(f1, f2,..., f N ). In the above formula, N represents the number of features in the feature set, represents the i-th feature value of the l-th sample in the feature set, represents the j-th feature value of the l-th sample in the feature set, represents the mean estimate of the i-th feature, represents the mean estimate of the j-th feature:

[0021]

[0022] Furthermore, the J value of classes c1 and c2 based on a certain feature is:

[0023] J = 2(1 - e -B )

[0024] where,

[0025]

[0026] In the formula, the value of J is the Jeffries-Matudita distance, abbreviated as the J-M distance, which can measure the between-class separability of a certain feature for two categories and is called the separation degree. m1 represents the mean of the distribution of category c1 based on this feature, m2 represents the mean of the distribution of category c2 based on this feature, σ1 represents the variance of the distribution of category c1 based on this feature, and σ2 represents the mean of the distribution of category c2 based on this feature;

[0027] The extraction ability E of the feature for the target ground object a The value is:

[0028]

[0029] In the formula, the separation degree between the i-th feature for c1 and c j (j = 2, 3,..., n) is s i represents the number of elements greater than or equal to 1.25 in represents the values greater than 1.25 in

[0030] The rule for decorrelation is: successively examine the correlation coefficient rij of every two features in the autocorrelation coefficient matrix r of the ground object to be extracted and all other ground objects. When the correlation coefficient of a certain two features exceeds the threshold, then remove the feature with a smaller E ij value. a The smaller feature.

[0031] Furthermore, the extraction ability C of the feature combination e The value is:

[0032]

[0033] Let the feature subset after decorrelation be F M = (f1, f2,..., f M ), represents the separation degree between the i-th feature for c1 and c j (j = 2, 3,..., n), C i represents the set of all j of i Let the union of C a (i = 1, 2,..., N) be C M Use the feature set F k (k ∈ C a ) to separate the categories c1 and c e and take the feature subset with the largest C value as the best feature combination for extracting the target ground object;

[0034] In the above formula, A is the number of elements in C a J iIndicates the maximum J value between the feature combination and the classes corresponding to the i-th element in the target feature class and C a where n i represents the number of elements within the feature combination, and σ j represents the variance after normalizing all the maximum J values.

[0035] Furthermore, the definition rule of the classification order is as follows: Repeat steps (1)-(4) to calculate and extract the optimal feature combination required for each type of feature and the C e value of this combination, and place the target feature corresponding to the feature combination that can distinguish the most classes in the front for extraction; if the number of distinguishable classes is equal, first extract the one with a larger C e value; if the C e values are equal, first extract the one with fewer feature objects, because this approach facilitates the classifier to evaluate the classification effect during the classification process. After selection, delete this type of feature from the feature class library.

[0036] Furthermore, the selection method of the membership function is as follows: Adopt the assignment method, and apply the existing fuzzy distribution according to the threshold form given by the SEaTH (separability and threshold) algorithm. Among them, the threshold form given by the SEaTH algorithm is obtained through the optimal separation threshold of a certain feature that can distinguish the target feature from all other features. This threshold is calculated by the following formula:

[0037]

[0038] where

[0039]

[0040] In the formula, n1 and n2 represent the number of samples of two feature classes, m1 represents the mean of class c1 based on this feature distribution, m2 represents the mean of class c2 based on this feature distribution, σ1 represents the variance of class c1 based on this feature distribution, and σ2 represents the mean of class c2 based on this feature distribution.

[0041] The calculation method of the soft threshold is as follows: Let the lower limit of the soft threshold be T a , and the upper limit be T b . The calculation method of the upper and lower limits of the soft threshold defined in the present invention is as follows:

[0042]

[0043]

[0044] In the above formula, T′ is the hard threshold parameter, which is expressed as:

[0045]

[0046] Let \(x\sim N(m,\sigma 2 ^2)\). According to the characteristics of the normal distribution, 95% of the eigenvalues are distributed in the interval \([m - 1.96\sigma, m + 1.96\sigma]\), and more than 95% are distributed in the interval \([m - 1.96\sigma, +\infty)\) or \((-\infty, m + 1.96\sigma]\). If \(m\pm1.96\sigma\) is used as the boundary of the soft threshold, theoretically, more than 95% of the ground objects can be successfully classified, meeting the requirements for extracting ground objects.

[0047] The present invention has the following advantages compared with the prior art:

[0048] The present invention proposes a feature selection method - EXaST for object - oriented image analysis of geographical objects. The features and thresholds calculated by this method can be directly used as the input for classification. Based on the original SEaTH algorithm, EXaST first performs decorrelation according to the Pearson correlation coefficient and the feature evaluation index E a , and then uses the feature combination evaluation index C e to find the optimal feature combination and determine the classification order. Finally, the soft threshold between classes is determined based on the mean and variance of the distribution of ground object feature values, and the ground objects are described by combining the fuzzy classification method, and then the ground object objects are classified. The new method reduces the dimension of the original feature set while maximizing the retention of the classification ability of the original set, and improves the effectiveness of features and classification accuracy while reducing the influencing factors of manual intervention in the classification process. The implementation process of this method is simple, the application effect is remarkable, and it can realize the automatic construction of ground object extraction in the object - oriented ground object classification process. Brief Description of the Drawings

[0049] Figure 1 is the flow chart of the present invention;

[0050] Figure 2 is the schematic diagram of two fuzzy distributions assigned by the present invention;

[0051] Figure 3 is the schematic diagram of calculating and extracting the soft threshold of the features of class c1 by the present invention;

[0052] Figure 4 is the schematic diagram of the experimental data used as an example in the present invention;

[0053] Figure 5 is the classification map of ground objects in the study area obtained by using the EXaST algorithm for feature selection and ground object extraction;

[0054] Figure 6 is the result map of classifying the study area by using the SEaTH algorithm and the object - oriented nearest - neighbor classification method;

[0055] Figure 7 is Figure 5 and Figure 6Three-dimensional schematic diagram of the landslide extraction result at the corresponding position in the white frame;

[0056] Figure 8 Yes Figure 5 And Figure 6 Schematic diagram of the extraction result of the residential area at the corresponding position within the black frame in the figure. Detailed implementation manners

[0057] The present invention will be further described below with reference to the accompanying drawings.

[0058] Referring to Figure 1 , the specific implementation steps of the present invention are as follows:

[0059] Step 1. Obtain image objects based on image segmentation technology, select typical samples according to the proportion of ground objects, and then output their features and feature values;

[0060] Among them, the image segmentation method is multi-scale segmentation, and the scale factor, spectral factor, and smoothness of the segmentation are obtained through multiple experiments. After obtaining the ground object objects, according to the representativeness of the samples and the uniformity of their spatial distribution, about 2.5% of the typical samples are selected according to the proportion of ground objects, and then their features and feature values are output.

[0061] Step 2. Calculate the autocorrelation coefficient matrix of the ground object to be extracted and all other ground objects based on the Pearson correlation coefficient, and set an appropriate threshold to adjust the number of selected features;

[0062] Among them, the autocorrelation coefficient matrix r can be expressed as:

[0063]

[0064] In the above formula, r ij (i, j = 1,..., N) is the Pearson correlation coefficient, indicating the degree of correlation between two features, and can be expressed as:

[0065]

[0066] Let the object category be C n =(c1, c2,..., c n ), and the original high-dimensional feature set be F N =(f1, f2,..., f N ). In the above formula, N represents the number of features in the feature set, represents the i-th feature value of the l-th sample in the feature set, represents the j-th feature value of the l-th sample in the feature set, represents the mean estimation of the i-th feature:

[0067]

[0068] Similarly, represents the mean estimate of the j-th feature.

[0069] Step 3. According to the extraction ability E of the feature for the target ground object a values are successively decorrelated in r;

[0070] Among them, the J value of categories c1 and c2 based on a certain feature is:

[0071] J = 2(1 - e -B )

[0072] Among them,

[0073]

[0074] In the formula, the J value is the Jeffries-Matudita distance, abbreviated as the J-M distance, which can measure the between-class separability of a certain feature for two categories and is called the separation degree. m1 represents the mean of the distribution of category c1 based on this feature, m2 represents the mean of the distribution of category c2 based on this feature, σ1 represents the variance of the distribution of category c1 based on this feature, and σ2 represents the mean of the distribution of category c2 based on this feature;

[0075] The extraction ability E of the feature for the target ground object a value is:

[0076]

[0077] In the formula, the separation degree between the i-th feature for c1 and c j (j = 2, 3,..., n) is s i represents the number of those greater than or equal to 1.25 in represents the values greater than 1.25 in

[0078] The method of decorrelation is: successively examine the correlation coefficient r ij of every two features in the autocorrelation coefficient matrix r of the ground object to be extracted and all other ground objects. When the correlation coefficient of a certain two features exceeds the threshold, then remove the feature with a smaller E a value.

[0079] Step 4. For the feature set obtained in Step 3, using the extraction ability of the feature combination for the target ground object as the evaluation index, obtain the optimal feature combination for a certain type of ground object to complete the optimization of the feature subset;

[0080] Among them, the extraction ability C e value of the feature combination is:

[0081]

[0082] Let the feature subset after decorrelation be F M =(f1, f2, …, f M ). Denote the separability between the i-th feature for c1 and c j (j = 2, 3, …, n), C i denotes the set of all j, let the union of C i (i = 1, 2, …, N) be C a . Use the feature set F M to separate the class c1 and c k (k ∈ C a ), and take the feature subset with the largest C e value as the best feature combination for extracting the target ground object;

[0083] In the above formula, A is the number of elements in C a , J i represents the maximum J value between the target ground object category of the feature combination and the category corresponding to the i-th element of C a , n i represents the number of elements in the feature combination, σ j represents the variance after normalizing all the maximum J values.

[0084] Step 5. Repeat Steps 1 - 4 to calculate the best feature combination for extracting each type of ground object, and determine which type of ground object to extract first based on the number of distinguishable categories and the extraction ability value of each extracted feature combination;

[0085] Among them, the definition rule of the classification order is: Repeat Steps 1 - 4 to calculate the best feature combination required for extracting each type of ground object and the C e value of this combination, and place the target ground object corresponding to the feature combination that can distinguish the most categories in the front for extraction; if the number of distinguishable categories is equal, then extract the one with a larger C e value first; if the C e values are equal, then extract the one with fewer ground object objects first, because this approach is convenient for the classifier to evaluate the classification effect during the classification process. After selection, delete this type of ground object from the ground object category library;

[0086] Step 6. Repeat Steps 1 - 5 until only two types of ground objects remain unextracted, and obtain the optimal features that can separate them based on the separability between classes;

[0087] Step 7. Calculate the soft threshold for extracting the best features of each type of ground object and then construct the extraction rule for each type of ground object in combination with the membership function.

[0088] Among them, the method for selecting the membership function is as follows: Using the assignment method, apply the existing fuzzy distributions according to the threshold forms given by the SEaTH (separability and threshold) algorithm. Taking the extraction of class c1 as an example, the specific assignment method is as follows: For a certain feature, if the ground objects of class c1 can be separated from the rest of the ground objects with a single feature threshold, then give the form of a cubic function; if the ground objects of class c1 require a threshold interval to ensure the separation effect from the rest of the ground objects, then give the form of a quadratic function (parabola).

[0089] Among them, the threshold form given by the SEaTH algorithm is obtained through the optimal separation threshold of a certain feature that can distinguish the target ground objects from all other ground objects, and this threshold is calculated by the following formula:

[0090]

[0091] Among them,

[0092]

[0093] In the formula, n1 and n2 represent the sample numbers of two ground object classes, m1 represents the mean value of class c1 based on the distribution of this feature, m2 represents the mean value of class c2 based on the distribution of this feature, σ1 represents the variance of class c1 based on the distribution of this feature, and σ2 represents the mean value of class c2 based on the distribution of this feature.

[0094] Regarding the setting of the upper and lower limits of the soft threshold, assume x ∼ N(m, σ 2 ). According to the characteristics of the normal distribution, 95% of the eigenvalue is distributed in the section of [m - 1.96σ, m + 1.96σ], and more than 95% is distributed in the section of [m - 1.96σ, +∞] or [-∞, m + 1.96σ]. If m ± 1.96σ is used as the boundary of the soft threshold, theoretically it can ensure that more than 95% of the ground objects are successfully classified, meeting the requirements for extracting ground objects. Assume the lower limit of the soft threshold is T a , and the upper limit is T b . The calculation method for the upper and lower limits of the soft threshold defined in the present invention is as follows:

[0095]

[0096]

[0097] In the above formula, T′ is the hard threshold parameter, which is expressed as:

[0098]

[0099] For a certain feature, if the membership function given by assignment is in the form of a cubic function, then the soft threshold interval of this feature for extracting class c1 is [T a , T b, denoted as "T > [T a , T b ". Among them, T a is the lower limit of the soft threshold, and T b is the upper limit of the soft threshold. It means that objects with feature values less than T a are classified into other categories, objects with feature values greater than T b are classified into class c1, and objects with feature values between T a and T b calculate their membership degrees according to the fuzzy distribution in the form of a cubic function, and then determine their categories (such as Figure 3 a).

[0100] For a certain feature, if the membership function given by assignment is in the form of a quadratic function, then the soft threshold interval of this feature for extracting class c1 is [T a , T' b , denoted as "T ~ [T a , T' b ". Among them, T a is the lower limit of the soft threshold interval between the category where the feature value distribution is on the left side of class c1, that is, the feature mean is less than class c1, and the category closest to class c1 and class c1, and T' b is the upper limit of the soft threshold interval between the category where the feature value distribution is on the right side of class c1, that is, the feature mean is greater than class c1, and the category closest to class c1 and class c1. It means that objects with feature values less than T a are classified into class c2, objects with feature values greater than T' b are classified into class c3, and objects with feature values between T a and T' b calculate their membership degrees according to the fuzzy distribution in the form of a parabola, and then determine their categories (such as Figure 3 b).

[0101] Compared with other feature selection algorithms, this method uses a feature combination evaluation index to construct the ground object classification rule instead of using a single feature evaluation index, and the constructed ground object extraction rule is more reasonable; it uses the membership function combined with the soft threshold to describe the ground object, rather than the "either zero or one" classification criterion, making the classification result closer to the human language and thinking mode. The implementation process of the present invention is simple, greatly reducing the influence of manual intervention, with significant application effects, and can realize the automatic construction of ground object extraction rules in the object-oriented ground object classification process.

[0102] Using the extraction rules, feature soft thresholds, and classification order of each type of ground object given by the EXaST algorithm as input, a detailed dataset of regional land surface cover types can be obtained. Further combined with manual visual interpretation and news materials, literature data, etc. for verification, it can be used for ground object classification in complex environments, extraction of typical target ground objects, and monitoring of land use type changes based on time series, providing a scientific basis for economic development, urban and rural construction, land use planning, and prevention and control of geological disasters.

[0103] The effects of the present invention can be illustrated by the following numerical examples:

[0104] 1. Numerical example setting

[0105] Here, to prove the accuracy and effectiveness of the new algorithm proposed by the present invention, an actual numerical example is set.

[0106] As Figure 4 shown, the Heifangtai area of Yangguoxia Town, Yongjing County, Linxia Hui Autonomous Prefecture, Gansu Province, China is selected, and the high-resolution satellite image No. 2 obtained on November 9, 2020 and the 12.5m resolution DEM data obtained in 2020 in this area are used to test the new feature selection method for object-based image analysis proposed by the present invention. Among them, the coverage range of the high-resolution satellite image No. 2 data is from 103°17′30″ to 103°19′15″ east longitude and from 36°5′00″ to 36°6′20″ north latitude, and the spatial resolution is 0.8m. Its false color composite image (the NIR band is used to enhance the G band to highlight vegetation) is as Figure 4 shown in a. First, the DEM is resampled to the same resolution as the image, and based on this, terrain information such as elevation ( Figure 4 c), slope ( Figure 4 d), and terrain undulation ( Figure 4 e) can be obtained. Then, multi-scale segmentation is performed on the image. The segmentation scale finally adopted in this numerical example after multiple experiments is 200, the spectral factor is 0.7, the shape factor is 0.3, and the smoothness and roughness are both 0.5. In addition, in this numerical example, the slope map and the terrain undulation map are also added as the sixth and seventh bands respectively to participate in feature extraction, and the DEM is added as the fifth band to participate in the segmentation with the multi-spectral bands with equal weights of 1:1:1:1:1 to improve the segmentation effect. The segmented image is as Figure 4 shown in b.

[0107] Finally, the experimental data are divided into 5 major categories: landslides, bedrock, residential areas, wasteland, and vegetation. According to the representativeness of the samples and the uniformity of their spatial distribution, about 2.5% of the typical samples are selected according to the proportion of ground objects, and their features and feature values are output. Feature selection experiments and ground object classification are respectively carried out using the SEaTH algorithm and the EXaST algorithm, and the classification results of the three methods are evaluated using the classification accuracy evaluation methods based on the confusion matrix and error analysis.

[0108] 2. Example Results

[0109] To compare the effects of different methods on feature selection and final classification, three methods, namely EXaST, SEaTH, and object-oriented nearest neighbor classification method, were used to classify the ground objects in the study area. Considering the comparability of the classification results, the same multi-scale segmentation parameters were used when obtaining ground objects. As the input of the algorithm, the same samples and feature values were also used.

[0110] Figure 5 The classification map of the ground objects in the study area, where feature selection was carried out using the EXaST algorithm and the ground objects were extracted, is given.

[0111] Figure 6 The result maps of classifying the study area using the SEaTH algorithm ( Figure 6 a) and the object-oriented nearest neighbor classification method ( Figure 6 b) are given.

[0112] Figure 7 The result of overlaying the landslide extraction result at the corresponding position in the white box in Figure 5 with Figure 6 and displaying it with the image, and obtaining the elevation of ground pixels from the high-precision DSM for three-dimensional stereo display is given. Based on the landslide information released by the Ministry of Natural Resources of the country and combined with the remote sensing identification marks of landslides, visual interpretation is carried out, and the manually interpreted landslide boundary is given for reference ( Figure 7 a). By comparing the extraction effects of the SEaTH algorithm and the EXaST algorithm, it can be seen that the SEaTH algorithm misjudged some objects with brightness values higher than the surrounding bedrock ground objects on the slope as landslides, and some loess-bedrock landslides on the right side of the Dangchuan loess landslide group were not identified as landslides because the brightness mean value was lower than the threshold. And because some landslides occurred a long time ago, the traces of some landslides were re-covered by the surface soil. The extraction order and feature combination given by the SEaTH algorithm did not identify these objects as landslides, resulting in discontinuity inside the landslides in the final classification result ( Figure 7 b). The EXaST algorithm first extracts relatively easy-to-extract ground objects such as vegetation, and extracts the bedrock and landslides at the end, which can more accurately describe the difference between the two and better extract the landslides. The extraction result is basically matched with the manually visual interpretation result ( Figure 7 c). It can be seen from this that the EXaST algorithm can plan a more reasonable classification order.

[0113] Figure 8 The result of Figure 5 and Figure 6Schematic diagram of the extraction result of the residential area corresponding to the position within the black frame. The vegetation and residential areas are densely distributed within the region, and the objects in the segmentation result inevitably contain both types of ground features at the same time. Figure 8 a). Combining Table 1, it can be seen that the SEaTH algorithm selects the 90° directional gray difference vector entropy and the standard deviation feature of the R-band value to extract the residential area. Although the J values of both these two features are relatively large and suitable for extracting the residential area, the composed feature combinations only ensure a good separation effect between the residential area and the surrounding wasteland, and some objects with a relatively large proportion of vegetation will be recognized as residential areas. Figure 8 b). At this time, if the threshold is adjusted, the residential area cannot be completely extracted. The EXaST algorithm uses the 0° directional gray difference vector contrast to extract the residential area, and this feature ensures good separability between the target ground feature, that is, the residential area and all other ground features. Figure 8 c). Therefore, the EXaST algorithm can preferably select more reasonable features for extracting ground features and reduce the possibility of misclassification and missed classification.

[0114] Table 1 presents the optimal extraction feature results of the ground features in the study area obtained by the EXaST algorithm, and the order of ground feature extraction is from top to bottom.

[0115] Table 1 Optimal Extraction Feature Table of Ground Features by EXaST Algorithm

[0116]

[0117] Table 2 presents the optimal extraction feature results of the ground features in the study area obtained by the SEaTH algorithm, and the order of ground feature extraction is from top to bottom. From Table 1 and Table 2, it can be seen that the SEaTH algorithm only judges the effectiveness of a single feature based on the J-M distance, optimizes the original feature space (188 dimensions), and obtains a 7-dimensional feature subset: NDVI, NDWI, 90° directional gray difference vector entropy (GLDV Entropy(90°)), standard deviation of the R-band (StandarddeviationLayer3), Mean Layer6, 0° directional gray level co-occurrence matrix value (GLCM Mean(0°)), and average brightness value (MeanBrightness). The EXaST algorithm first decorrelates the features according to the Pearson correlation coefficient and E a value, then optimizes the remaining feature subset, obtains several feature combinations with the same ground feature extraction ability as the original subset, and finally measures the ability of the feature combination to extract the target ground feature according to the evaluation index C e Through this optimized feature space, a 4-dimensional feature subset is obtained: NDVI, GLDV Contrast(0°), MeanLayer6, and Max different. After preliminary visual interpretation of the classification results of the EXaST algorithm Figure 5f), it can be seen that its classification results roughly conform to the actual ground object distribution. From this, it can be concluded that the EXaST algorithm can better remove bad features and better reduce the feature dimension, and it is an effective feature dimension reduction method.

[0118] Table 2 Optimal Feature Extraction Table of Ground Objects by the SEaTH Algorithm

[0119]

[0120] Table 3 gives the comparison of the landslide extraction accuracy evaluation between the SEaTH algorithm and the EXaST algorithm based on error analysis. From this, it can be known that the landslide extraction accuracy rate of the EXaST algorithm is 69.5% and the landslide recognition rate is 75.1%, both of which can meet the basic requirements of landslide extraction and are better than the SEaTH algorithm.

[0121] Table 3 Landslide Extraction Accuracy Evaluation Table of the SEaTH Algorithm and the EXaST Algorithm

[0122]

[0123] Tables 4, 5, and 6 respectively give the classification accuracy evaluation tables of the EXaST algorithm, the SEaTH algorithm, and the object-oriented nearest neighbor classification method based on the confusion matrix. The results show that the overall accuracy of the EXaST algorithm classification is 94% and the Kappa coefficient is 0.919, which are 10% and 0.135 higher than those of the SEaTH algorithm respectively, and 20% and 0.268 higher than those of the nearest neighbor classification method respectively. In addition, the producer accuracy and user accuracy of each ground object in the classification results of the EXaST algorithm are generally higher than those of the SEaTH algorithm and the nearest neighbor classification algorithm. From this, it can be known that the EXaST algorithm is a better object-oriented feature selection method.

[0124] Table 4 Confusion Matrix for Verifying the Classification Accuracy of the EXaST Algorithm

[0125]

[0126] Table 5 Confusion Matrix for Verifying the Classification Accuracy of the SEaTH Algorithm

[0127]

[0128] Table 6 Confusion Matrix for Verifying the Accuracy of the Nearest Neighbor Classification Method

[0129]

Claims

1. A feature selection method for geographic object-oriented image analysis, characterized in that Including the following steps: (1) Obtain image objects based on image segmentation technology, select typical samples according to the proportion of ground objects, and then output their features and feature values; (2) Calculate the autocorrelation coefficient matrix of the ground object to be extracted and all other ground objects based on the Pearson correlation coefficient, and set an appropriate threshold to adjust the number of selected features; (3) Perform feature decorrelation in the autocorrelation coefficient matrix in sequence according to the extraction ability value of the feature for the target ground object; Among them, the extraction ability E of the feature for the target ground object a The value is: where the separation degree of the i-th feature between c1 and c j is s i represents the number greater than or equal to 1.25 in represents the value greater than 1.25 in, where j = 2, 3,..., n; n is the number of sample categories; (4) For the feature set obtained in step (3), take the extraction ability of the feature combination for the target ground object as the evaluation index, obtain the optimal feature combination of a certain type of ground object, and complete the optimization of the feature subset; Extraction ability C of feature combination e The value is: Let the feature subset after correlation be F M =(f1, f2, …, f M ), denotes the separation degree of the i-th feature between c1 and c j , C i denotes the set of all j; Let the union of C i be C a , where i = 1, 2, …, N, and N represents the number of features in the feature set; Use the feature set F M to separate the classes c1 and c k , and take the feature subset with the largest C e value as the best feature combination for extracting the target ground object; k ∈ C a ; In the above formula, A is the number of elements in C a The number of elements inside, J i Indicates the maximum J value between the feature combination in the target ground object category and the category corresponding to the i-th element of C a The maximum J value, n i Indicates the number of elements in the feature combination, σ j Indicates the variance after normalizing all the maximum J values; (5) Repeat steps (1)-(4), calculate the best feature combination for extracting each type of ground object, and determine which ground object to extract first based on the number of distinguishable categories and the extraction ability value of each extracted feature combination; The definition rule of the classification order is as follows: repeat steps (1)-(4) to calculate and extract the optimal feature combination required for each type of ground object and the C value of this combination, and place the target ground object corresponding to the feature combination that can distinguish the most categories in the front for extraction; e The target ground object corresponding to the feature combination that can distinguish the most categories is placed in the front for extraction; If the number of distinguishable categories is equal, extract C first e with a larger value; if the C e values are equal, extract the one with fewer ground object numbers first. After selection, delete this type of ground object from the ground object category library; (6) Repeat steps (1)-(5) until only two types of ground objects remain unextracted, and obtain the optimal features that can separate them based on the separability between classes; (7) Calculate the soft threshold for extracting the best features of each type of ground object, and then combine it with the membership function to construct the extraction rule for each type of ground object to achieve ground object extraction.

2. The new method for feature selection for geographic object image analysis according to claim 1, characterized in that: The image segmentation method used in step (1) is multi-scale segmentation. The scale factor, spectral factor, and smoothness of the segmentation are obtained through multiple experiments. After obtaining the ground object, according to the representativeness of the samples and the uniformity of their spatial distribution, about 2.5% of the typical samples are selected according to the proportion of ground objects, and then their features and feature values are output.

3. The new feature selection method for geographical object-oriented image analysis according to claim 1, wherein: In step (2), the autocorrelation coefficient matrix r can be expressed as: In the above formula, r ij is the Pearson correlation coefficient, indicating the degree of correlation between two features; Let the object category be C n =(c1, c2, …, c n ), and the original high-dimensional feature set be F N =(f1, f2, …, f N ). In the above formula, where i = 1, …, N; j = 1, …, N; N represents the number of features in the feature set, represents the i-th eigenvalue of the l-th sample in the feature set, represents the j-th eigenvalue of the l-th sample in the feature set, represents the mean estimate of the i-th feature, represents the mean estimate of the j-th feature.

4. The new feature selection method for geographic object-oriented image analysis according to claim 3, characterized in that: In step (3), the J value of categories c1 and c2 based on a certain feature is: J = 2(1 - e -B ) Where, In the formula, the J value is the separation degree; m1 represents the mean of the distribution of category c1 based on this feature, m2 represents the mean of the distribution of category c2 based on this feature, σ1 represents the variance of the distribution of category c1 based on this feature, and σ2 represents the mean of the distribution of category c2 based on this feature.

5. The new feature selection method for geographic object-oriented image analysis according to claim 4, characterized in that: In step (3), the method for removing correlation is as follows: successively examine the correlation coefficient r between each pair of features in the autocorrelation coefficient matrix r of the feature to be extracted and all other features ij , and when the correlation coefficient between a certain pair of features exceeds the threshold, then remove E a the feature with a smaller value.

6. The new feature selection method for geographical object-oriented image analysis according to claim 1, characterized in that: In step (7), the formula for calculating the soft threshold of the membership function is Where n1 and n2 represent the number of samples of two ground object categories, m1 represents the mean of the distribution of category c1 based on this feature, m2 represents the mean of the distribution of category c2 based on this feature, σ1 represents the variance of the distribution of category c1 based on this feature, and σ2 represents the mean of the distribution of category c2 based on this feature.

7. The new feature selection method for geographical object image analysis according to claim 6, characterized in that: In step (7), set the lower limit of the soft threshold as T a , and the upper limit as T b . The calculation method of the upper and lower limits of the soft threshold is as follows: T ′ is the hard threshold parameter.

Images