High-resolution remote sensing image bare soil extraction and fine classification method

Abstract

According to the high-resolution remote sensing image bare soil extraction and fine classification method, the high-resolution remote sensing image is obtained based on WorldView series satellites and resource No.3 satellites, bare soil information is extracted by an abstract normal form classification method and fine classification is carried out for bare soil extraction and fine classification of each region, and the accuracy of bare soil extraction and fine classification is improved. The key technology comprises image segmentation, feature entropy setting, classification system design and the like. According to the method, three bare soil indexes are designed for WorldView series images, two bare soil indexes are provided for resource No.3 images, mixed ground features are removed by using methods of supervised classification and index set construction, and the extraction precision can be greatly improved by adding a mixed ground feature removal rule. Analyzing object features including spectrum, geometry, spatial relation and the like, selecting features with good classification effects, and determining feature combinations extracted from various ground features such as bulldozing soil and non-farming farmland; a classification rule feature entropy set is established, a bare soil fine classification result is obtained, classification is accurate, and the precision is high.

Description

Technical Field

[0001] The present application relates to a method for classifying bare soil in high-resolution remote sensing images, and in particular to a method for extracting and finely classifying bare soil in high-resolution remote sensing images, belonging to the technical field of remote sensing image classification. Background Art

[0002] Currently, research on extracting data from high-resolution remote sensing imagery focuses on target features such as buildings, water bodies, and vegetation, while less research has focused on extracting bare soil. However, with the advancement of urbanization and new rural development, land use monitoring is urgently needed. Bare soil includes unused exposed surfaces, urban construction sites, uncultivated farmland, and backfill. Due to the lack of vegetation cover, bare soil seriously impacts the regional ecological environment and is detrimental to local soil and water conservation. Urban bare soil contributes to ground dust, and PM2.5 source analysis indicates that dust is a significant source of atmospheric particulate matter pollution. Extracting and studying the spatial distribution of urban bare soil is crucial for management decisions related to urban landscaping, sustainable land use, and atmospheric environmental protection.

[0003] Furthermore, cultivated land, as a primary consideration for urban expansion, also creates numerous problems for rural land use, primarily manifested in farmland destruction and extensive land use. Therefore, real-time monitoring of land dynamics during urbanization can help achieve a balance between urbanization and land use.

[0004] Among current monitoring methods, acquiring remote sensing images via satellite is an efficient and accurate approach. For specific target features, by assigning relevant image features based on their characteristics and forming effective feature combinations, they can be distinguished from background features. If necessary, a database of typical land use samples can be established to meet the needs of refined land use information extraction and interpretation.

[0005] However, there is still relatively little extraction of bare soil. This application uses the WorldView series and Ziyuan-3 satellite images to conduct experiments on bare soil and fine classification of bare soil, providing valuable information for my country's urban environmental construction and new rural construction.

[0006] The problems that need to be solved in the existing high-resolution remote sensing image bare soil classification and the key technical difficulties of this application include:

[0007] (1) Currently, the research on the extraction of high-resolution remote sensing images focuses on target objects such as buildings, water bodies and vegetation, while there is less research on the extraction of bare soil. With the advancement of urbanization and new rural construction, land use monitoring is also urgently needed. The existing technology lacks methods for extracting bare soil and fine classification of bare soil. There is a lack of bare soil index for WorldView series images and bare soil index for Resources-3 satellite images. Using only a single bare soil index cannot achieve satisfactory classification results. There is a lack of rules for removing confusing objects to improve extraction accuracy. There is no analysis of the spectral, geometric and spatial relationship object features of bare soil. There is no classification rule feature entropy set established. The overall accuracy of fine classification of bare soil is low.

[0008] (2) High-resolution remote sensing images provide richer details of land features and can provide more information for remote sensing land feature classification. The application of high-resolution remote sensing images to bare soil extraction is of great significance and significance. However, the existing technology lacks systematic bare soil extraction and fine classification research. It does not combine the spectral reflectance curves and high-resolution images of typical land features in the land feature spectral database to analyze the radiation characteristics of typical land features such as vegetation, bare land, and construction land. It does not establish various types of bare soil indexes. The accuracy is still far from enough for actual engineering applications. Since the spectral characteristics of construction land and bare soil are very similar, there is no further removal of confusing land features. There is a lack of spectral features, geometric features, etc. to establish a regular feature entropy set for fine classification of bare soil. The accuracy of bare soil extraction is low, the goal of bare soil classification is unclear, and the overall accuracy is low. Summary of the Invention

[0009] Based on high-resolution remote sensing imagery acquired by the WorldView series and the Ziyuan-3 satellite, this application investigates the extraction and fine-grained classification of bare soil in various regions. The application investigates the spectral characteristics of typical features, employing an abstract paradigm classification method to extract bare soil information and perform fine-grained classification. Key technologies involved include image segmentation, feature entropy configuration, and classification system design. Three bare soil indices were designed for WorldView imagery, and two were proposed for Ziyuan-3 imagery. These indices utilize supervised classification and constructed index sets, respectively, to remove confounding features. Compared to a single bare soil index, incorporating rules for confounding feature removal significantly improves extraction accuracy. For each target feature, in-depth analysis of its spectral, geometric, and spatial characteristics is performed to select features with optimal classification performance. Feature combinations for extracting various features, such as fill and uncultivated farmland, are then determined. Based on this, a feature entropy set of classification rules is developed, ultimately resulting in accurate and highly precise bare soil classification results for the region.

[0010] To achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0011] A method for extracting and finely classifying bare soil from high-resolution remote sensing images is developed. Based on the abstract paradigm classification method, three bare soil indices, TZ, LTZ, and LCT, are designed for WorldView series images. Two bare soil indices, ZYS and ZYA, are proposed for Resources-3 satellite images. The classification results of each indices are analyzed and their accuracy is evaluated. It is concluded that TZ>LCT>LTZ, and ZYS>ZYA are the best bare soil extraction indicators. Due to the influence of other land objects, a single bare soil index cannot achieve satisfactory classification results. The index with a high degree of conformity to actual land objects is optimized and used as the data basis for the next classification step. The classification results of the TZ and ZYS indices are used for further extraction of bare soil. Confusing land object removal rules are added to improve extraction accuracy. For WorldView series images, extraction is performed by establishing a rule feature entropy set. For Resources-3 images, extraction is performed by using the shortest distance classification method. The spectral, geometric, and spatial relationship object features of bare soil are analyzed, and features with good classification effects are selected. A classification rule feature entropy set is established to classify bare soil into two categories: non-cultivated farmland and pushback land.

[0012] 1) Using the eCongnition platform, various bare soil indices were established. For the WorldView series, TZ, LTZ, and LCT were used for preliminary classification. For the Ziyuan-3 satellite images, ZYS and ZYA indices were proposed for preliminary classification. The TZ index provided high classification accuracy, while the ZSY index provided high classification accuracy. Furthermore, the spectral characteristics of built-up land and bare soil are very similar, which further removed confusing features.

[0013] 2) The next step is to extract bare soil information based on the two indices TZ and ZYS with high classification accuracy. For the WorldView series images, the building index BI, vegetation index NDVI and other spectral features are used to remove confusing objects. For the Resources-3 satellite images, the shortest distance classification method is used to remove confusing objects. Subsequently, the spectral features and geometric features are used to establish a regular feature entropy set for fine classification of bare soil.

[0014] Preferably, the classification system and classification rule feature entropy set: specific rules are established based on the following three levels:

[0015] (1) Establishing rules for each level of classification: defining the classification rules based on the spectral, geometric, and topological characteristics of the object;

[0016] (2) Inheritance of parent types by subtypes within a layer: If a subtype exists, it first inherits the judgment rules of its parent type, and then adds its unique spectral features, geometric features, and topological features as judgment rules;

[0017] (3) Merge and transfer the classification results of each layer to form the final classification judgment rules;

[0018] The establishment of each rule does not necessarily have to include the above three levels. If the type of land feature can be judged well, only one level can be used to form a rule. The feature formation rules can also be flexibly set at each level, and it is not required to include all features.

[0019] Preferably, classification based on chaotic logic abstraction: classification based on chaotic logic abstraction paradigm is adopted, the membership of the sample is obtained through the membership function, and the membership function converts any feature value range into a unified range [0, 1] to describe the membership of a type, combining two classification methods:

[0020] 1) Shortest distance classification: For each image object, find the nearest sample object in the feature space. If the nearest sample object of an image object belongs to class A, the object will be classified as class A. This is done through a membership function. The closer the distance between the image object and the sample object belonging to class A in the feature space, the greater the membership of class A.

[0021] 2) Membership function classification: Establish a semantic hierarchy and classify images based on various features. For each feature, calculate the feature value, set the membership function, and assign it to a membership degree of [0-1]. When there are different features, combine them through logical operations. The features described by the class include three categories: object features, inter-class features, and global features.

[0022] Preferably, the bare soil index design based on spectral characteristics: the index design is based on the spectral characteristics of the target object, quantitatively describing the attributes of the object. The calculation model is based on the bands of the multispectral image, and the weakest and strongest response bands are screened from each band for ratio calculation. The weak response is placed in the denominator and the strong response is placed in the numerator. The band division method is used to widen the gap between the two bands, reduce the brightness of the background object, and increase the brightness of the target area, so as to make the target area more obvious. After the index calculation, the information of the target area is highlighted. At the same time, some other objects with similar spectra to the target area are also amplified in brightness, forming interference between objects. When using the index feature, a critical value is set to achieve the extraction effect.

[0023] Corresponding bare soil indices were constructed for different regions for preliminary extraction of bare soil. Combining the spectral curves of regional bare soil, three bare soil indices, TZ, LTZ, and LCT, were constructed for the WorldView series images, and two bare soil indices, ZYS and ZYA, were constructed for the Resources-3 images, for preliminary extraction of bare soil.

[0024] Preferably, the classification joint model of TZ, LTZ, and LCT indexes: first, perform multi-scale segmentation on the region, and set the segmentation parameters as follows: scale parameter SP = 50, shape factor Shape = 0.1, compactness factor Ct = 0.5, set the weights of the three layers R, G, and B to 1, and the weights of the remaining layers to 0 to obtain the segmentation result;

[0025] (1) TZ index bare soil extraction

[0026] Based on the difference between bare soil and vegetation status, the surface areas including high vegetation cover and large amounts of bare soil are distinguished using the formula:

[0027]

[0028] Among them, Red and Blue represent the red and blue bands in visible light respectively, SWIR3 represents the third short-wave infrared band, and Nir1 represents the first near-infrared band. The index critical value is adjusted to achieve the best extraction effect. When TZ ≥ -0.06, the extraction effect is the best;

[0029] (2) LTZ index bare soil extraction

[0030] To distinguish the highlighted buildings from the bare soil, based on the difference in the spectral curves of bare soil and buildings, the formula is:

[0031]

[0032] Among them, Red, Blue and Green represent the red, blue and green bands in visible light, respectively; SWIR3 and SWIR8 represent the third and eighth bands of shortwave infrared, respectively. The bare soil is extracted according to the LTZ index, and the critical value of the index is adjusted to achieve the best extraction effect. When -0.02≤LTZ≤-0.1, the extraction effect is the best;

[0033] (3) LCT index bare soil extraction

[0034] The SWIR and NIR bands are used to represent the difference in soil reflectance values. The soil response values between the green and yellow bands in the visible light range have unique differences. This feature is used to construct the LCT index to extract soil. Bare soil extraction is performed based on the LTZ index. The critical value of the index is adjusted to achieve the best extraction effect. When -0.06≤LCT≤0, the extraction effect is best.

[0035] Preferably, the classification joint model of the ZYS and ZYA indexes: multi-scale segmentation is performed on the region, the segmentation parameters are set as follows: scale parameter SP=40, shape factor Shape=0.1, compactness factor Ct=0.5, and the layer weight is set to R:G:B:Nir=1:1:1:0 to obtain the segmentation result;

[0036] (1) ZYS index bare soil extraction

[0037] The spectral brightness value of vegetation in Band 2-Band 3 of the Resource-3 image shows a downward trend, while the spectral curve of bare soil increases. Based on this characteristic of the spectral curve, the bare soil index ZYS is set, and the formula is:

[0038] ZYS=Red-Green Formula 3

[0039] Among them, Red and Green represent the red and green bands in visible light, respectively. The bare soil is extracted according to the ZYS index, and the critical value of the index is adjusted to achieve the best extraction effect. When ZYS ≤ 16, the extraction effect is the best;

[0040] (2) Extraction of ZYA index for bare soil

[0041] The spectral curve of the image shows that the brightness value of the bare soil spectrum is between that of vegetation and buildings. Since the image contains only three types of objects, a suitable critical value is set to distinguish the bare soil from the other two types of objects. Based on this feature, the ZYA index model is established. The formula is:

[0042] ZYA=(Blue+Green+Red)×2 Formula 4

[0043] In order to widen the gap between the two types of land features, the sum of the three in the index is doubled. Blue, Green and Red represent the blue, green and red bands in visible light respectively. The bare soil is extracted according to the ZYA index, and the critical value of the index is adjusted to achieve the best extraction effect. When 2336≤ZYA≤3200, the extraction effect is best.

[0044] Preferably, feature entropy is set: for a specific target object, relevant image features are set according to its object characteristics and an effective feature combination is formed to distinguish it from background objects. The feature entropy setting is based on the following features:

[0045] (1) Spectral characteristics: reflect the radiation characteristics of the ground object. The magnitude of the ground object's radiation energy corresponds to the magnitude of the pixel grayscale value on the image. The spectral characteristics used for classification include: object brightness, mean value of different bands, standard deviation, difference between different bands, and custom features constructed from different bands;

[0046] (2) Shape features: These are the geometric shapes formed by the boundaries of objects. Rivers and roads appear as long strips, houses appear as regular rectangles or combinations of rectangles, and farmland appears as regular blocks. Shape features include area, aspect ratio, length, width, perimeter, shape index, density, and asymmetry.

[0047] (3) Position feature: It is the position of an image object relative to the entire image, measured by two scales: distance and coordinates.

[0048] Preferably, the classification system is constructed: first, the land use types are divided according to the application purpose, and the features are classified into structural categories based on the image separability, the land cover type of the region, and the specific conditions of different types of features in the construction land;

[0049] Secondly, a rule feature entropy set is constructed based on the feature knowledge base: the establishment of a classification system is a process of feature entropy setting, feature combination, and feature space optimization. The classification rule construction is targeted at a certain land feature. The obtained rules are converted into a feature knowledge base, which is then described in mathematical language and presented as a set of feature values within a specific interval. This is used to determine the subordinate relationship between the object and the land feature type.

[0050] The following rule feature entropy set is established for the removal of confusing objects and fine classification of bare soil in WorldView series images:

[0051] 1) Land feature type: bare soil; regular feature entropy set: TZ ≥ -0.06, BI ≥ 0.58, NDVI > 0.39, Brightness > 2828

[0052] 2) Landform type: Bulldozer fill; brightness ≥ 2241, Y distance to scene bottom border > 595, Length / Width ≤ 4.9, Area ≥ 50pxl Mean Red ≤ 2100

[0053] 3) Landform type: farmland; Brightness < 241;

[0054] To remove confusing objects in the Ziyuan-3 image, the shortest distance classification method was used. In the fine classification, the rule feature entropy set established is as follows:

[0055] 1) Land feature type: push fill; regular feature entropy set: Mean Green ≥ 588, Length / Width ≤ 2.2, Area ≥ 196pxl, Asymmetry ≤ 0.7, Mean Red ≤ 627, Rectangular Fit ≤ 0.864

[0056] 2) Land feature type: farmland; regular feature entropy set: Mean Green > 588.

[0057] Preferably, bare soil extraction based on multiple feature entropy: bare soil confusing features removal is based on the TZ index extraction result, and confusing features are removed according to the image characteristics;

[0058] (1) Removing buildings: In the blue band, the spectral curve of the building has a small peak, while the bare soil has a small trough in this band, and then a small peak appears in the near-red band. Based on this feature, the building index BI is constructed;

[0059] (2) Removing vegetation: Bare soil contains less vegetation, and the NDVI index is used to extract vegetation in the soil class;

[0060] (3) Removal of white roofs: removal based on the difference in brightness characteristics between white roofs and bare soil;

[0061] The rule feature entropy set constructed by removing confusing objects based on the initial classification results of the TZ index is as follows:

[0062] Object characteristics: TZ ≥ -0.06 Ground features: Soil index within a fixed range

[0063] Object characteristics: BI ≥ 0.58 Landform characteristics: Building index is higher than bare soil and vegetation features

[0064] Object characteristics: NDVI>0.39 Ground features: The NDVI index of vegetation mistakenly classified as bare soil is higher than that of bare soil

[0065] Object characteristics: Brightness>2828 Ground features: The brightness value of buildings with white roofs is higher than that of bare soil.

[0066] The building index is used to extract most of the buildings in the image, and the aspect ratio feature is used to extract the roads.

[0067] Preferably, bare soil extraction from Resource No. 3 image:

[0068] 1) Bare soil extraction based on the shortest distance: First, the shortest distance classification method is used to classify the samples and add the features required for classification;

[0069] Buildings are distributed in clusters and have irregular shapes, so texture information is used to distinguish them. The color of the soil is yellowish, and the color of the buildings is reddish, so spectral features are added to distinguish them. The shortest distance algorithm is used to separate the building land from the bare soil.

[0070] 2) Bare soil fine classification of Ziyuan-3 image

[0071] Bare soil is divided into two categories: farmland and fill soil. The fill soil is distributed near buildings and has the following construction characteristics:

[0072] Object characteristics: Mean Green ≥ 588. Features: Strong reflection in the green band of fill soil, with high grayscale value.

[0073] Object characteristics: Length / Width≤2.2 Land feature: Compared with farmland, the length and width of the fill are smaller

[0074] Object characteristics: Area ≥ 196pxl Land feature: Filled soil is distributed in sheets, and the area is larger than farmland

[0075] Object characteristics: Asymmetry≤0.7. Features: The shape of the farmland is long and its asymmetry index is high.

[0076] Compared with the existing technology, the innovation and advantages of this application are:

[0077] (1) Based on high-resolution remote sensing images obtained by the WorldView series satellites and the Resource-3 satellite, this application studies the spectral characteristics of typical landforms for the extraction and fine classification of bare soil in various regions. The abstract paradigm classification method is used to extract bare soil information and perform fine classification. The key technologies involved include image segmentation, feature entropy setting, and classification system design. Three bare soil indices are designed for the WorldView series images, and two bare soil indices are proposed for the Resource-3 images. The supervised classification and index set construction methods are used to remove mixed landforms. Compared with a single bare soil index, the addition of rules for removing mixed landforms can greatly improve the extraction accuracy. For the target landforms, the object features including spectrum, geometry, and spatial relationships are deeply analyzed, and features with good classification effects are selected. The feature combination for extracting various types of landforms such as fill and non-cultivated farmland is determined. On this basis, a classification rule feature entropy set is developed and established, and finally the fine classification results of bare soil in the region are obtained. The classification is accurate and the accuracy is high.

[0078] (2) The current research on the extraction of high-resolution remote sensing images focuses on target objects such as buildings, water bodies and vegetation, while there is less research on the extraction of bare soil. This application uses the WorldView series (1.2m resolution) and Resources-3 (5.8m resolution) images as data sources, and based on the abstract paradigm classification method, extracts and finely classifies bare soil; designs three bare soil indices, TZ, LTZ and LCT, for the WorldView series images, and proposes two bare soil indices, ZYS and ZYA, for the Resources-3 satellite images. Using only a single bare soil index cannot achieve a satisfactory classification effect, and optimizes the index with a high degree of conformity to the actual objects as the data basis for the next classification; further extracts bare soil from the classification results of the TZ and ZYS indices, and adds confusing object removal rules to improve the extraction accuracy. For the WorldView series images, the extraction is carried out by establishing a rule feature entropy set, and the overall extraction accuracy reaches 94%; for the Resources-3 images, the extraction is carried out by the shortest distance classification method, and the overall extraction accuracy reaches 93%;

[0079] By analyzing the spectral, geometric, and spatial relationship characteristics of bare soil objects, the team selected features with the best classification results and established a feature entropy set for classification rules. The team then classified bare soil into two categories: uncultivated farmland and landfill. The accuracy of the classification results was then evaluated. Experiments showed that the overall accuracy of fine-grained bare soil classification for WorldView images reached 93%, and for Resource-3 images reached 92%.

[0080] (3) High-resolution remote sensing images provide richer details of ground objects and can provide more information for remote sensing ground object classification. Therefore, the application of high-resolution remote sensing images to bare soil extraction is of great significance. This application systematically carried out the extraction and fine classification of bare soil using WorldView series and Resource-3 images. Combining the spectral reflectance curves and high-resolution images of typical ground objects in the ground object spectral database, using eCongnition as the platform and various bare soil indices as the entry point, the WorldView series was preliminarily classified using TZ, LTZ, and LCT. For Resource-3 satellite images, the ZYS and ZYA indices were proposed for preliminarily classification. The conclusion was that the former had a higher classification accuracy using the TZ index, while the latter had a higher classification accuracy using the ZSY index. The next step of bare soil information extraction was carried out based on the two indices (TZ and ZYS) with higher classification accuracy. For the WorldView series images, the building index, vegetation index NDVI and other spectral features were used to remove the confusing ground objects. For the Resource-3 satellite images, the shortest distance classification method was used to remove the confusing ground objects. Then, the spectral features, geometric features, etc. were used to establish a regular feature entropy set to finely classify the bare soil. The results show that further confusing the extraction of features based on the bare soil index can effectively improve the accuracy of bare soil extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is a schematic diagram of the WorldView series bare soil index segmentation results.

[0082] Figure 2 This is a schematic diagram of the TZ index bare soil extraction of the WorldView series.

[0083] Figure 3 This is a comparison chart of the original image TZ index and LTZ index white roof extraction.

[0084] Figure 4 This is the comparison chart of the TZ index and LTZ index of the original image and the gray cement ground extraction.

[0085] Figure 5 This is a comparison chart of the accuracy evaluation of the WorldView series bare soil index classification results.

[0086] Figure 6 This is a schematic diagram of the bare soil extraction results of the ZYS index.

[0087] Figure 7 This is a schematic diagram of the ZYA index extraction results for bare soil.

[0088] Figure 8 This is an analysis diagram of the efficiency of image relationship construction.

[0089] Figure 9 It is the result map of land use type classification based on application purpose.

[0090] Figure 10 It is a land feature classification structure category map that combines image separability, land cover type, and construction land.

[0091] Figure 11 This is the result of fine classification of bare soil from the WorldView series of images.

[0092] Figure 12 This is the feature selection and training sample map for the shortest distance classification of ZY-3 images.

[0093] Figure 13 This is the confusion matrix and accuracy evaluation diagram of the bare soil classification results of the ZY-3 image.

[0094] Figure 14 This is the result of fine classification of bare soil based on the rule feature entropy set.

[0095] Figure 15 It is a sample diagram for accuracy evaluation of fine classification of bare soil.

[0096] Figure 16 The output error matrix and classification accuracy are used to draw the classification result graph. DETAILED DESCRIPTION

[0097] The following, in conjunction with the accompanying drawings, further describes the technical solution of the high-resolution remote sensing image bare soil extraction and fine classification method provided by this application, so that those skilled in the art can better understand this application and implement it.

[0098] Currently, research on extracting features from high-resolution remote sensing images focuses on target objects such as buildings, water bodies, and vegetation, while research on extracting bare soil is relatively limited. However, with the advancement of urbanization and new rural construction, land use monitoring is also urgently needed. This application uses the WorldView series (1.2m resolution) and Resource-3 (5.8m resolution) images as data sources, and based on the abstract paradigm classification method, conducts the following research on the extraction and fine classification of bare soil:

[0099] (1) Three bare soil indices, TZ, LTZ, and LCT, were designed for WorldView series images, and two bare soil indices, ZYS and ZYA, were proposed for Ziyuan-3 satellite images. The classification results of each index were analyzed and the accuracy of the results was evaluated. It was found that in terms of bare soil extraction ability, TZ>LCT>LTZ, and ZYS>ZYA. Due to the influence of other land features, using only a single bare soil index could not achieve satisfactory classification results. The index with the highest degree of conformity to the actual land features was optimized and set as the data basis for the next step of classification.

[0100] (2) The classification results of the TZ and ZYS indices were used to further extract bare soil, and the rules for removing confusing objects were added to improve the extraction accuracy. For the WorldView series images, the extraction was performed by establishing a rule feature entropy set, and the overall extraction accuracy reached 94%. The Resource No. 3 images were extracted using the shortest distance classification method, and the overall extraction accuracy reached 93%.

[0101] (3) Analyze the spectral, geometric, and spatial relationship characteristics of bare soil objects, select features with good classification results, establish a classification rule feature entropy set, classify bare soil into two categories: uncultivated farmland and backfill, and evaluate the accuracy of the classification results. Experiments show that the overall accuracy of fine classification of bare soil in the WorldView series imagery reaches 93%, and the overall accuracy of fine classification of bare soil in the Resource-3 imagery reaches 92%.

[0102] 1. Abstract paradigm classification of high-resolution remote sensing images

[0103] (1) Multi-scale segmentation

[0104] Abstracting features into symbols or attributes that are easy to recognize includes two aspects: one is the scope represented by the abstract feature symbols, and the other is the scale and resolution of the features. High-resolution remote sensing images themselves contain rich information, and the attributes of target features at different scales are considered, including spectral information and texture information.

[0105] The bottom layer is the initial image layer, i.e. the pixel layer. In the last layer, the segmented image has the largest area and the least number of polygons. The new layer is placed above, in the middle, or below the existing layer, or covered on the original layer. The sub-objects of each layer are merged from the sub-objects of the next layer, and are merged in order from bottom to top. Different feature categories are extracted in different segmentation layers during feature classification.

[0106] The ground object target information is extracted in layers of different scales. Some larger spatial ground object categories are extracted in the large-scale segmentation object layer. Some ground object categories with many and complex categories are extracted in the small-scale segmentation object layer. After extracting the ground object information of each layer, the layers are superimposed to obtain a new layer. In addition, in the image segmentation process, vector data parameters are added for segmentation to increase the accuracy of image information extraction.

[0107] 1. Multi-scale segmentation parameter setting

[0108] According to the spectral characteristics and shape characteristics of each target object in the image, the image segmentation result is optimized. The segmentation parameters include:

[0109] (1) Setting of layer weights: weights are set based on the impact of each band on the image segmentation quality and results. If the use of certain layers has little or no impact on the segmentation results, the weights are set to 0, indicating that the layer does not participate in the image segmentation. Including spatial data in the segmentation improves the image classification accuracy.

[0110] (2) Setting of homogeneity factors: including shape (smoothness and compactness) and color factors to improve classification accuracy and reliability.

[0111] (2) Classification system and classification rule feature entropy set

[0112] The specific rules are established based on the following three levels:

[0113] (1) Establishing rules for each level of classification: defining the classification rules based on the spectral, geometric, and topological characteristics of the object;

[0114] (2) Inheritance of parent types by subtypes within a layer: If a subtype exists, it first inherits the judgment rules of its parent type, and then adds its unique spectral features, geometric features, and topological features as judgment rules;

[0115] (3) Merge and transfer the classification results of each layer to form the final classification judgment rules;

[0116] The establishment of each rule does not necessarily have to include the above three levels. If the type of land feature can be judged well, only one level can be used to form a rule. The feature formation rules can also be flexibly set at each level, and it is not required to include all features.

[0117] (3) Abstract classification based on chaotic logic

[0118] The classification is based on the chaotic logic abstract paradigm. The membership of the sample is obtained through the membership function. The membership function converts any feature value range into a unified range [0, 1] to describe the membership of a type. It combines two classification methods:

[0119] 1) Shortest distance classification: For each image object, find the nearest sample object in the feature space. If the nearest sample object of an image object belongs to class A, the object will be classified as class A. This is done through a membership function. The closer the distance between the image object and the sample object belonging to class A in the feature space, the greater the membership of class A.

[0120] 2) Membership function classification: Establish a semantic hierarchy and classify images based on various features. For each feature, calculate the feature value, set the membership function, and assign it to a membership degree of [0-1]. When there are different features, combine them through logical operations. The features described by the class include three categories: object features, inter-class features, and global features.

[0121] 3.4 Chapter Summary

[0122] First, image segmentation is performed to obtain homogeneous objects. Then, according to the classification requirements, the spectrum, shape, texture, and position information contained in the image objects are used to set and combine feature entropy, and a classification system is constructed based on this. Finally, remote sensing image classification is performed based on chaotic logic abstraction.

[0123] Bare soil has no distinct spectral characteristics, making it difficult to distinguish it from other land features based on the spectral information of each pixel. This is especially true for land features that are easily confused with built-up land. However, built-up land has a distinct geometric structure and rich spatial features such as shape and texture. This information can be fully utilized during the extraction and fine-grained classification of bare soil to distinguish the two types of land features. Therefore, this application adopts a classification method based on the chaotic logic abstract paradigm.

[0124] 2. Bare soil index design and bare soil extraction

[0125] (1) Bare soil index design based on spectral characteristics

[0126] The index design is based on the spectral characteristics of the target object, and the attributes of the object are quantitatively described. The calculation model is based on the bands of the multispectral image, and the bands with the weakest response and the bands with the strongest response are screened out from each band for ratio calculation. The weak response is placed in the denominator and the strong response is placed in the numerator. The band division method is used to widen the gap between the two bands, reduce the brightness of the background objects, and increase the brightness of the target area, so as to make the target area more obvious. After the index calculation, the information of the target area is highlighted. At the same time, the brightness of some other objects with similar spectra to the target area is also amplified, forming interference between objects. When using the index feature, a critical value is set to achieve the extraction effect.

[0127] For easily confused ground features, a new setting is made based on a single index. Since the spectral characteristics of bare soil and built-up land are very similar, a single bare soil index cannot accurately extract bare soil. This application constructs corresponding bare soil indices for different regions to perform preliminary extraction of bare soil.

[0128] Combined with the spectral curve of regional bare soil, three bare soil indices, TZ, LTZ, and LCT, were constructed for the WorldView series images, and two bare soil indices, ZYS and ZYA, were constructed for the Resources-3 image. Preliminary bare soil extraction experiments were carried out and the accuracy of the classification results were analyzed.

[0129] (2) WorldView series bare soil index image extraction

[0130] 1) Classification joint model of TZ, LTZ, and LCT indices

[0131] First, perform multi-scale segmentation on the region. The segmentation parameters are set as follows: scale parameter SP = 50, shape factor Shape = 0.1, compactness factor Ct = 0.5, and set the weights of the three layers R, G, and B to 1, and the weights of the remaining layers to 0. The segmentation results are as follows: Figure 1 As shown, there are 8866 segmented objects in total.

[0132] (1) TZ index bare soil extraction

[0133] Based on the difference between bare soil and vegetation status, the surface areas including high vegetation cover and large amounts of bare soil are distinguished using the formula:

[0134]

[0135] Among them, Red and Blue represent the red and blue bands in visible light respectively, SWIR3 represents the third short-wave infrared band, and Nir1 represents the first near-infrared band. The index critical value is adjusted to achieve the best extraction effect. When TZ ≥ -0.06, the extraction effect is the best. The extraction results are as follows Figure 2 As shown:

[0136] Compared with the initial image, the bare soil extraction is relatively complete. Large areas of vegetation and a few buildings with black roofs have been distinguished. However, the original intention of the design of this index was to distinguish only vegetation from bare soil, while most buildings, asphalt roads, concrete floors and other construction sites were not separated from the bare soil.

[0137] (2) LTZ index bare soil extraction

[0138] To distinguish the highlighted buildings from the bare soil, based on the difference in the spectral curves of bare soil and buildings, the formula is:

[0139]

[0140] Among them, Red, Blue and Green represent the red, blue and green bands in visible light respectively, SWIR3 and SWIR8 represent the third and eighth bands of shortwave infrared respectively. Bare soil is extracted according to the LTZ index, and the critical value of the index is adjusted to achieve the best extraction effect. When -0.02≤LTZ≤-0.1, the extraction effect is best.

[0141] From the extraction results, taking the initial image as a reference, compared with the results of the TZ index, some buildings with white and black roofs are separated in the results of the LTZ index extraction ( Figure 3 ), in addition, part of the gray cement was also separated ( Figure 4 ).

[0142] (3) LCT index bare soil extraction

[0143] The SWIR and NIR bands are used to represent the difference in soil reflectance values. The soil response values between the green and yellow bands in the visible light range have unique differences. This characteristic is used to construct the LCT index to extract soil. The LTZ index is used for bare soil extraction. The critical value of the index is adjusted to achieve the best extraction effect. When -0.06≤LCT≤0, the extraction effect is the best.

[0144] Compared with the initial image, the vegetation part is clearly separated from the extraction results of the LCT index. The blue houses are also separated. In addition, the milky white roofs are separated, which are features not reflected in the first two indices.

[0145] 2) Classification accuracy analysis of TZ, LTZ, and LCT indices

[0146] The three index extraction results are analyzed with the same samples. Based on the above samples, the accuracy is evaluated in eCognition, and the error matrix and classification accuracy are output as follows: Figure 5 The three indices are ranked as TZ > LCT > LTZ, but the overall extraction accuracy of all three indices is relatively low. Furthermore, the classification results show that a large portion of building land in the upper half of the image is misclassified as bare soil. Bare soil lacks distinctive features, making it difficult to fully extract it using a single index, as is the case with vegetation and water bodies. Therefore, in addition to using the bare soil index, other rules should be incorporated to eliminate confusing features when extracting bare soil.

[0147] (3) Bare soil extraction from Resource No. 3 image based on bare soil index

[0148] 1) Classification joint model of ZYS and ZYA index

[0149] The region is segmented at multiple scales, and the segmentation parameters are set as follows: scale parameter SP = 40, shape factor Shape = 0.1, compactness factor Ct = 0.5, and the layer weight is set to R:G:B:Nir = 1:1:1:0. The segmentation results are 9149 segmented objects in total.

[0150] (1) ZYS index bare soil extraction

[0151] The spectral brightness value of vegetation in Band 2-Band 3 of the Resource-3 image shows a downward trend, while the spectral curve of bare soil increases. Based on this characteristic of the spectral curve, the bare soil index ZYS is set, and the formula is:

[0152] ZYS=Red-Green Formula 3

[0153] Among them, Red and Green represent the red and green bands in visible light respectively. The bare soil is extracted according to the ZYS index, and the critical value of the index is adjusted to achieve the best extraction effect. When ZYS≤16, the extraction effect is the best. Figure 6 As shown;

[0154] The ZYS index effectively distinguishes the asphalt road from bare soil in the through-area. Furthermore, in the eastern portion of the image, white and blue-roofed buildings are effectively separated from the bare soil. In the central portion of the image, the less abundant vegetation in this area is separated from the bare soil. However, as shown in the spectral curves, the spectral curves of bare soil and building sites are similar, which can lead to confusion between the two when extracting features. This is particularly true for buildings with red brick roofs, which cannot be effectively separated using this index.

[0155] (2) Extraction of ZYA index for bare soil

[0156] The spectral curve of the image shows that the brightness value of the bare soil spectrum is between that of vegetation and buildings. Since the image contains only three types of objects, a suitable critical value is set to distinguish the bare soil from the other two types of objects. Based on this feature, the ZYA index model is established. The formula is:

[0157] ZYA=(Blue+Green+Red)×2 Formula 4

[0158] To widen the gap between the two types of land features, the sum of the three is doubled in the index. Blue, Green, and Red represent the blue, green, and red bands in visible light, respectively. The bare soil is extracted according to the ZYA index, and the critical value of the index is adjusted to achieve the best extraction effect. When 2336≤ZYA≤3200, the extraction effect is the best. Figure 7 As shown;

[0159] Comparing the initial image with the image classification results, the resulting index effectively distinguishes vegetation, some buildings, and bare soil, including white-roofed and blue-roofed buildings. Compared to the ZYS index, the ZYA index also extracts some brick-red houses. A closer look reveals a drawback to this index: only one of the five asphalt roads running through the entire image is clearly distinguished. A large area of bare soil, the main subject of this application, is also omitted, as can be seen on the eastern side of the image.

[0160] 2) Accuracy analysis of ZYS and ZYA index classification

[0161] The two indices were analyzed for accuracy using the same sample. The results show that the overall accuracy of the two indices, ZYS, exceeds ZYA. Specifically, the ZYS index shows a clear advantage over the ZYA index for the soil classification, a fact also observed in the classification results. While some buildings in the image are extracted, the majority of residential areas are misclassified as bare soil, a category dominated by buildings with red roofs. This misclassification results in a low Kappa coefficient for both indices. This also demonstrates that using only the bare soil index alone is insufficient for fully extracting the target ground objects, requiring further removal of confounding features.

[0162] 3. Bare soil extraction based on rule feature entropy set

[0163] (1) Feature entropy setting and classification system construction

[0164] 1) Feature entropy setting

[0165] For a specific target object, relevant image features are set according to its object characteristics and an effective feature combination is formed to distinguish it from the background objects. The feature entropy setting is based on the following features:

[0166] (1) Spectral characteristics: reflect the radiation characteristics of the ground object. The magnitude of the ground object's radiation energy corresponds to the magnitude of the pixel grayscale value on the image. The spectral characteristics used for classification include: object brightness, mean value of different bands, standard deviation, difference between different bands, and custom features constructed from different bands;

[0167] (2) Shape features: These are geometric figures formed by the boundaries of objects. Rivers and roads appear as long strips, houses appear as regular rectangles or combinations of rectangles, and farmland appears as regular blocks. Shape features include area, aspect ratio, length, width, perimeter, shape index, density, and asymmetry. The shape features used in the classification process and their calculation formulas are as follows: Figure 8 As shown;

[0168] (3) Position feature: It is the position of an image object relative to the entire image, measured by two scales: distance and coordinates.

[0169] 2) Construction of classification system

[0170] First, the land use types are divided into Figure 9 ,Combined with the image separability, the land cover type of the region, the specific conditions of different types of construction land features, and the land feature classification structure categories, a Figure 10 ;

[0171] Secondly, a rule feature entropy set is constructed based on the feature knowledge base: the establishment of a classification system is a process of feature entropy setting, feature combination, and feature space optimization. The classification rule construction is targeted at a certain land feature. The obtained rules are converted into a feature knowledge base, which is then described in mathematical language and presented as a set of feature values within a specific interval. This is used to determine the subordinate relationship between the object and the land feature type.

[0172] The following rule feature entropy set is established for the removal of confusing objects and fine classification of bare soil in WorldView series images:

[0173] 1) Land feature type: bare soil; regular feature entropy set: TZ ≥ -0.06, BI ≥ 0.58, NDVI > 0.39, Brightness > 2828

[0174] 2) Landform type: Bulldozer fill; brightness ≥ 2241, Y distance to scene bottom border > 595, Length / Width ≤ 4.9, Area ≥ 50pxl Mean Red ≤ 2100

[0175] 3) Land feature type: farmland; Brightness < 241.

[0176] To remove confusing objects in the Ziyuan-3 image, the shortest distance classification method was used. In the fine classification, the rule feature entropy set established is as follows:

[0177] 1) Land feature type: push fill; regular feature entropy set: Mean Green ≥ 588, Length / Width ≤ 2.2, Area ≥ 196pxl, Asymmetry ≤ 0.7, Mean Red ≤ 627, Rectangular Fit ≤ 0.864

[0178] 2) Landform type: farmland; Rule feature entropy set: Mean Green > 588

[0179] (2) Bare soil extraction from WorldView series images

[0180] 1) Bare soil extraction based on multi-surface entropy

[0181] Bare soil confusing objects are removed based on the TZ index extraction results. Large areas of buildings are mistakenly extracted, including concrete floors, blue roofs, white roofs, and some asphalt roads. Some vegetation is also included in the extraction results. The confusing objects are removed based on the characteristics of the image.

[0182] (1) Removing buildings: In the blue band, the spectral curve of the building has a small peak, while the bare soil has a small trough in this band, and then a small peak appears in the near-red band. Based on this feature, the building index BI is constructed;

[0183] (2) Removing vegetation: Bare soil contains less vegetation, and the NDVI index is used to extract vegetation in the soil class;

[0184] (3) Removal of white roofs: They are removed based on the difference in brightness characteristics between them and bare soil.

[0185] The rule feature entropy set constructed by removing confusing objects based on the initial classification results of the TZ index is as follows:

[0186] Object characteristics: TZ ≥ -0.06 Ground features: Soil index within a fixed range

[0187] Object characteristics: BI ≥ 0.58 Landform characteristics: Building index is higher than bare soil and vegetation features

[0188] Object characteristics: NDVI>0.39 Ground features: The NDVI index of vegetation mistakenly classified as bare soil is higher than that of bare soil

[0189] Object characteristics: Brightness>2828 Ground features: The brightness value of buildings with white roofs is higher than that of bare soil.

[0190] The building index is used to extract most of the buildings in the image, and the aspect ratio feature is used to extract the roads. However, since the farmland in the image is mostly rectangular due to its own structural segmentation shape, the aspect ratio feature does not achieve a better separation effect.

[0191] 2) Fine classification of bare soil

[0192] Bare soil in remote sensing images is divided into two categories: farmland and fill. Fill is mostly concentrated around buildings. Based on the characteristics of the two types of bare soil, a regular feature entropy set is constructed to finely classify the bare soil. The constructed regular feature entropy set is as follows:

[0193] Object characteristics: brightness ≥ 2241 Ground feature: The brightness of the fill is higher than that of the farmland

[0194] Object characteristics: Length / Width≤4.9 Land feature: Compared with farmland, the length and width of the fill are relatively small.

[0195] Object characteristics: Area ≥ 50pxl Feature characteristics: Land objects appear in patches. Areas that are too small do not belong to this category. Object characteristics: MeanRed ≤ 2100 Feature characteristics: Farmland has a reddish hue and a high grayscale value in the red band.

[0196] The feature entropy set of farmland extraction rules from WorldView series images is as follows:

[0197] Object characteristics: Brightness < 2241. Land feature: The brightness of farmland is lower than that of fill.

[0198] 3) Classification results and accuracy evaluation

[0199] (1) Bare soil extraction results and accuracy evaluation

[0200] The final bare soil extraction results of the WorldView series images are obtained by classifying the confusing objects as non-bare soil based on the TZ index classification results. The accuracy of the samples is evaluated in eCognition, and the error matrix and classification accuracy are output to plot the classification results.

[0201] According to the classification results, after removing confusing features, the overall classification accuracy reached 94.5%, and the Kappa coefficient reached 87.4%. Using the same accuracy evaluation sample, compared with the classification results using only the bare soil index (TZ), the overall classification results improved by 27%, and the Kappa coefficient increased by 43%. This shows that using more features can optimize extraction results.

[0202] (2) Bare soil fine classification results and accuracy assessment

[0203] According to the rule feature entropy set, the result of bare soil fine classification is Figure 11 , select the accuracy evaluation samples for bare soil fine classification;

[0204] Based on the above samples, perform accuracy evaluation in eCognition, output the error matrix and classification accuracy to plot the classification results;

[0205] This experiment conducted a precision analysis of the fine classification of three categories: non-bare soil, push-fill, and farmland. The overall accuracy of the confusion matrix for these three categories was 93%, with a kappa coefficient of 85.4%, indicating good classification results. For the two types of bare soil to be distinguished, the push-fill user accuracy reached 93.3%, indicating that this method is generally feasible.

[0206] (3) Bare soil extraction from Resource No. 3 image

[0207] 1) Bare soil extraction based on the shortest distance

[0208] In the ZYS index extraction results for Resource 3, the confusing features in the ZYS extraction results are buildings. Among these buildings, those with red brick roofs occupy a large area, so removing these buildings is crucial. The spectral curves of buildings and bare soil in the image are very similar, making it difficult to distinguish the two types of features using only spectral features. Therefore, this application first uses the shortest distance classification method for classification, selecting samples and adding the required features.

[0209] Buildings are distributed in clusters and have irregular shapes, so texture information is used to distinguish them. The color of the soil is yellowish and the color of the buildings is reddish, so spectral features are added to distinguish them. The selected features and training samples are as follows: Figure 12 . The shortest distance algorithm is used to separate the building land from the bare soil.

[0210] 2) Bare soil fine classification of Ziyuan-3 image

[0211] Bare soil is divided into two categories: farmland and fill soil. The fill soil is distributed near buildings and has the following construction characteristics:

[0212] Object characteristics: Mean Green ≥ 588 Ground features: Strong reflection in the green band of fill soil, high grayscale value

[0213] Object characteristics: Length / Width≤2.2 Land feature: Compared with farmland, the length and width of the fill are smaller

[0214] Object characteristics: Area ≥ 196pxl Land feature: Filled soil is distributed in sheets, and the area is larger than farmland

[0215] Object characteristics: Asymmetry≤0.7 Landform characteristics: The shape of the farmland is long and its asymmetry index is high.

[0216] 3) Classification results and accuracy evaluation

[0217] (1) Bare soil classification results and accuracy evaluation

[0218] The non-bare soil objects in the image are merged into one category, and the classification results of bare soil are obtained after removing the confusing objects from the Resource No. 3 image; the accuracy evaluation is performed in eCognition based on the samples, and the error matrix and classification accuracy classification results are output as follows Figure 13 ;

[0219] The classification results show that the shortest distance classification method effectively improves classification accuracy, increasing it by 20% compared to the TZ index. This algorithm separates most of the building sites from the bare soil in the image, and the kappa coefficient also increases significantly, demonstrating the feasibility of this method. Comparing the initial image with the classification results reveals that voids appear in the building complex, misclassifying these buildings as bare soil, resulting in lower classification accuracy in non-bare soil areas.

[0220] (2) Bare soil fine classification results and accuracy evaluation

[0221] According to the rule feature entropy set, the results of bare soil fine classification are as follows Figure 14 ;

[0222] Select the bare soil fine classification accuracy evaluation samples, such as Figure 15 As shown;

[0223] The accuracy of the three categories of push fill (soil1), farmland (soil2) and non-bare soil was evaluated in the eCognition software, and the error matrix and classification accuracy were output to draw the classification results. Figure 16 As shown;

[0224] In terms of evaluation accuracy, the overall classification accuracy reached 92.1%, and the classification accuracy of backfill reached 89.3%.

Claims

1. A method for extracting and finely classifying bare soil from high-resolution remote sensing images, characterized in that: Based on the abstract paradigm classification method, three bare soil indices, TZ, LTZ, and LCT, were designed for the WorldView series images, and two bare soil indices, ZYS and ZYA, were proposed for the Resources-3 satellite images. The classification results of each indices were analyzed and the accuracy of the results was evaluated. It was concluded that TZ>LCT>LTZ, and ZYS>ZYA were the best bare soil extraction methods. Due to the influence of other land objects, a single bare soil index could not achieve satisfactory classification results. The index with a high degree of conformity to actual land objects was optimized and used as the data basis for the next classification. The classification results of the TZ and ZYS indices were used for further bare soil extraction, and confusing land object removal rules were added to improve the extraction accuracy. For the WorldView series images, the extraction was performed by establishing a rule feature entropy set; for the Resources-3 images, the extraction was performed by the shortest distance classification method. The spectral, geometric, and spatial relationship object features of the bare soil were analyzed, and the features with the best classification effect were selected. A classification rule feature entropy set was established to classify the bare soil into two categories: non-cultivated farmland and pushback land. 1) Using the eCongnition platform, various bare soil indices were established. For the WorldView series, TZ, LTZ, and LCT were used for preliminary classification. For the Ziyuan-3 satellite images, ZYS and ZYA indices were proposed for preliminary classification. The TZ index provided high classification accuracy, while the ZSY index provided high classification accuracy. Furthermore, the spectral characteristics of built-up land and bare soil are very similar, which further removed confusing features. 2) The next step is to extract bare soil information based on the two indices with high classification accuracy, TZ and ZYS. For WorldView series images, confusing objects are removed using the building index BI, vegetation index NDVI, and other spectral features. For Ziyuan-3 satellite images, confusing objects are removed using the shortest distance classification method. Then, the spectral features and geometric features are used to establish a regular feature entropy set for fine classification of bare soil.

2. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1, characterized in that: Classification system and classification rule feature entropy set: Specific rules are established based on the following three levels: (1) Establishing rules for each level of classification: defining the classification rules based on the spectral, geometric, and topological characteristics of the object; (2) Inheritance of parent types by subtypes within a layer: If a subtype exists, it first inherits the judgment rules of its parent type, and then adds its unique spectral features, geometric features, and topological features as judgment rules; (3) Merge and transfer the classification results of each layer to form the final classification judgment rules; The establishment of each rule does not necessarily have to include the above three levels. If the type of land feature can be judged well, only one level can be used to form a rule. The feature formation rules can also be flexibly set at each level, and it is not required to include all features.

3. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1 is characterized in that: Classification based on chaotic logic abstraction: Adopting classification based on chaotic logic abstraction paradigm, the sample's membership is obtained through membership function. The membership function converts any feature value range into a unified range [0, 1] to describe the membership of a type. It combines two classification methods: 1) Shortest distance classification: For each image object, find the nearest sample object in the feature space. If the nearest sample object of an image object belongs to class A, the object will be classified as class A. This is done through a membership function. The closer the distance between the image object and the sample object belonging to class A in the feature space, the greater the membership of class A. 2) Membership function classification: Establish a semantic hierarchy and classify images based on various features. For each feature, calculate the feature value, set the membership function, and assign it to a membership degree of [0-1]. When there are different features, combine them through logical operations. The features described by the class include three categories: object features, inter-class features, and global features.

4. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1 is characterized in that: Bare soil index design based on spectral characteristics: The index design is based on the spectral characteristics of the target object, quantitatively describing the attributes of the object. The calculation model is based on the bands of the multispectral image, and the weakest and strongest response bands are screened from each band for ratio calculation. The weak response is placed in the denominator and the strong response is placed in the numerator. The band division method is used to widen the gap between the two bands, reducing the brightness of the background objects and increasing the brightness of the target area, making the target area more obvious. After the index calculation, the information of the target area is highlighted. At the same time, some other objects with similar spectra to the target area are also amplified in brightness, forming interference between objects. When using the index feature, a critical value is set to achieve the extraction effect; Corresponding bare soil indices were constructed for different regions for preliminary extraction of bare soil. Combining the spectral curves of regional bare soil, three bare soil indices, TZ, LTZ, and LCT, were constructed for the WorldView series images, and two bare soil indices, ZYS and ZYA, were constructed for the Resources-3 images, for preliminary extraction of bare soil.

5. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1 is characterized in that: Classification joint model of TZ, LTZ, and LCT indices: First, perform multi-scale segmentation on the region. The segmentation parameters are set as follows: scale parameter SP = 50, shape factor Shape = 0.1, and compactness factor Ct = 0.

5. The weights of the R, G, and B layers are set to 1, and the weights of the remaining layers are set to 0 to obtain the segmentation result. (1) TZ index bare soil extraction Based on the difference between bare soil and vegetation status, the surface areas including high vegetation cover and large amounts of bare soil are distinguished using the formula: Among them, Red and Blue represent the red and blue bands in visible light respectively, SWIR3 represents the third short-wave infrared band, and Nir1 represents the first near-infrared band. The index critical value is adjusted to achieve the best extraction effect. When TZ ≥ -0.06, the extraction effect is the best; (2) LTZ index bare soil extraction To distinguish the highlighted buildings from the bare soil, based on the difference in the spectral curves of bare soil and buildings, the formula is: Among them, Red, Blue and Green represent the red, blue and green bands in visible light, respectively; SWIR3 and SWIR8 represent the third and eighth bands of shortwave infrared, respectively. The bare soil is extracted according to the LTZ index, and the critical value of the index is adjusted to achieve the best extraction effect. When -0.02≤LTZ≤-0.1, the extraction effect is the best; (3) LCT index bare soil extraction The SWIR and NIR bands are used to represent the difference in soil reflectance values. The soil response values between the green and yellow bands in the visible light range have unique differences. This feature is used to construct the LCT index to extract soil. Bare soil extraction is performed based on the LTZ index. The critical value of the index is adjusted to achieve the best extraction effect. When -0.06≤LCT≤0, the extraction effect is best.

6. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1, characterized in that: Classification joint model of ZYS and ZYA index: Perform multi-scale segmentation on the region with the following parameters: scale parameter SP = 40, shape factor Shape = 0.1, compactness factor Ct = 0.5, and set the layer weights to R:G:B:Nir = 1:1:1:0 to obtain the segmentation result; (1) ZYS index bare soil extraction The spectral brightness value of vegetation in Band 2-Band 3 of the Resource-3 image shows a downward trend, while the spectral curve of bare soil increases. Based on this characteristic of the spectral curve, the bare soil index ZYS is set, and the formula is: ZYS=Red-Green Formula 3 Among them, Red and Green represent the red and green bands in visible light, respectively. The bare soil is extracted according to the ZYS index, and the critical value of the index is adjusted to achieve the best extraction effect. When ZYS ≤ 16, the extraction effect is the best; (2) Extraction of ZYA index for bare soil The spectral curve of the image shows that the brightness value of the bare soil spectrum is between that of vegetation and buildings. Since the image contains only three types of objects, a suitable critical value is set to distinguish the bare soil from the other two types of objects. Based on this feature, the ZYA index model is established. The formula is: ZYA=(Blue+Green+Red)×2 Formula 4 In order to widen the gap between the two types of land features, the sum of the three in the index is doubled. Blue, Green and Red represent the blue, green and red bands in visible light respectively. The bare soil is extracted according to the ZYA index, and the critical value of the index is adjusted to achieve the best extraction effect. When 2336≤ZYA≤3200, the extraction effect is best.

7. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1, characterized in that: Feature entropy setting: For a specific target object, relevant image features are set according to its object characteristics and an effective feature combination is formed to distinguish it from background objects. Feature entropy setting is based on the following features: (1) Spectral characteristics: reflect the radiation characteristics of the ground object. The magnitude of the ground object's radiation energy corresponds to the magnitude of the pixel grayscale value on the image. The spectral characteristics used for classification include: object brightness, mean value of different bands, standard deviation, difference between different bands, and custom features constructed from different bands; (2) Shape features: These are the geometric shapes formed by the boundaries of objects. Rivers and roads appear as long strips, houses appear as regular rectangles or combinations of rectangles, and farmland appears as regular blocks. Shape features include area, aspect ratio, length, width, perimeter, shape index, density, and asymmetry. (3) Position feature: It is the position of an image object relative to the entire image, measured by two scales: distance and coordinates.

8. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1, characterized in that: Classification system construction: First, the land use types are divided according to the application purpose, and the land feature classification structure categories are classified based on the image separability, the land cover type of the region, and the specific conditions of different types of land features in construction land; Secondly, a rule feature entropy set is constructed based on the feature knowledge base: the establishment of a classification system is a process of feature entropy setting, feature combination, and feature space optimization. The classification rule construction is targeted at a certain land feature. The obtained rules are converted into a feature knowledge base, which is then described in mathematical language and presented as a set of feature values within a specific interval. This is used to determine the subordinate relationship between the object and the land feature type. The following rule feature entropy set is established for the removal of confusing objects and fine classification of bare soil in WorldView series images: 1) Ground feature type: bare soil; Rule feature entropy set: TZ ≥ -0.06, BI ≥ 0.58, NDVI > 0.39, Brightness > 2828 2) Landform type: Bulldozer fill; brightness ≥ 2241, Y distance to scene bottom border > 595, Length / Width ≤ 4.9, Area ≥ 50pxl Mean Red ≤ 2100 3) Landform type: farmland; Brightness < 241; To remove confusing objects in the Ziyuan-3 image, the shortest distance classification method was used. In the fine classification, the rule feature entropy set established is as follows: 1) Land feature type: push fill; regular feature entropy set: Mean Green ≥ 588, Length / Width ≤ 2.2, Area ≥ 196pxl, Asymmetry ≤ 0.7, Mean Red ≤ 627, Rectangular Fit ≤ 0.864 2) Land feature type: farmland; regular feature entropy set: Mean Green > 588.

9. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1, characterized in that: Bare soil extraction based on multi-object entropy: Bare soil confusing objects are removed based on the TZ index extraction results and the confusing objects are removed according to the image characteristics; (1) Removing buildings: In the blue band, the spectral curve of the building has a small peak, while the bare soil has a small trough in this band, and then a small peak appears in the near-red band. Based on this feature, the building index BI is constructed; (2) Removing vegetation: Bare soil contains less vegetation, and the NDVI index is used to extract vegetation in the soil class; (3) Removal of white roofs: removal based on the difference in brightness characteristics between them and bare soil; The rule feature entropy set constructed by removing confusing objects based on the initial classification results of the TZ index is as follows: The building index is used to extract most of the buildings in the image, and the aspect ratio feature is used to extract the roads.

10. The method for extracting and finely classifying bare soil from high-resolution remote sensing images according to claim 1, characterized in that: Bare soil extraction from Resource 3 image: 1) Bare soil extraction based on the shortest distance: First, the shortest distance classification method is used to classify the samples and add the features required for classification; Buildings are distributed in clusters and have irregular shapes, so texture information is used to distinguish them. The color of the soil is yellowish, and the color of the buildings is reddish, so spectral features are added to distinguish them. The shortest distance algorithm is used to separate the building land from the bare soil. 2) Bare soil fine classification of Ziyuan-3 image Bare soil is divided into two categories: farmland and fill soil. The fill soil is distributed near buildings and has the following construction characteristics:

Images