CN115272769A - Method and device for automatic extraction of lunar impact craters based on machine learning - Google Patents

Abstract

The invention discloses a moon impact pit automatic extraction method and device based on machine learning, and belongs to the technical field of geographical research. The automatic moon impact pit extraction method comprises the following steps: acquiring DEM data of an area to be identified on the moon; identifying the area to be identified according to a pre-constructed grid-level classifier and DEM data of the area to be identified to obtain candidate impact pits; obtaining a radial elevation profile of the candidate impact pit according to the candidate impact pit and DEM data of the area to be identified; and identifying the candidate meteorite crater according to a pre-constructed object-level classifier and a radial elevation profile of the candidate bumping crater to obtain the bumping crater. The device comprises: the device comprises an acquisition module, a first obtaining module, a second obtaining module and a third obtaining module. According to the technical scheme, the machine learning method is combined with the geomorphic elements, and a grid-oriented and object-oriented combined method is used in the machine learning, so that the identification accuracy of the impact pits is improved.

Description

Method and device for automatic extraction of lunar impact craters based on machine learning

technical field

The invention belongs to the technical field of geographical research, in particular to a method and device for automatically extracting lunar impact craters based on machine learning.

Background technique

As the most prominent landform feature on the lunar surface, impact craters show a relatively complete record of the impact history of the moon since its formation. Throughout the lunar history, characteristics such as the size, shape, and distribution of impact craters are crucial to the study of the evolutionary history of the moon. Therefore, the identification of impact craters is the most basic work in the study of impact craters, and many achievements have been made from manual interpretation to automatic extraction.

The extraction methods of impact craters can be divided into two categories: manual visual recognition and automatic extraction. Among them, the former method of identifying impact craters is relatively simple, and the results are relatively accurate. However, when studying the density of impact craters and other issues, a large number of impact craters need to be identified, and the manual extraction method is time-consuming and laborious. The principles of automatic recognition mainly include the following four categories: The first category is based on feature matching methods. Among these methods, some propose to automatically identify impact craters by using parameters such as the morphological characteristics of the annular structure of the lunar surface. The second category is based on image transformation and segmentation methods. In this kind of method, some proposed to use the object-oriented method to extract the impact craters on the lunar surface, and select the characteristic values "contrast of adjacent phase elements" and "aspect ratio" to identify the impact craters. The third category is the method based on geographic information fusion. Among these methods, some proposed three automatic extraction methods: fill, object-oriented classification, and fill-object-oriented classification, and carried out impact crater extraction tests on DEM (Digital Elevation Model, Digital Elevation Model). WA's object-oriented method has higher extraction accuracy. The fourth category is based on machine learning methods. In this type of method, some propose to establish a set of trainable algorithms to extract and identify impact craters of different sizes by referring to the theory of machine learning and computer vision. Compared with the first three types of automatic recognition methods of impact craters, the fourth type of machine learning method abandons the limitations based on two-dimensional image data, and uses traditional impact crater maps and spatial structure information knowledge obtained from digital terrain analysis to train machines. The learning classifier overcomes the shortcomings of terrain analysis and greatly improves the recognition efficiency.

Impact craters are not simple circular depressions due to the degradation and superposition of impact craters. Although the existing methods of detecting impact craters based on terrain analysis consider the terrain information of impact craters to a certain extent, they do not consider the spatial structure information of impact craters, which makes the extraction results low in accuracy.

Contents of the invention

In order to solve the above problems, the present invention proposes a method and device for automatic extraction of lunar impact craters based on machine learning, which involves a new grid-oriented and object-oriented automatic extraction method for impact craters. The spatial structure information of impact craters can be learned from digital terrain analysis through machine learning, and the automatic extraction of impact craters can be realized.

A method for automatically extracting lunar impact craters based on machine learning, comprising: obtaining DEM data of an area to be identified on the moon; The area is identified to obtain candidate impact craters; according to the candidate impact craters and the DEM data of the area to be identified, the radial elevation profile of the candidate impact craters is obtained; according to the pre-built object-level classifier and the candidate The radial elevation profile of the impact crater is used to identify the candidate craters to obtain the impact craters.

In the method for automatically extracting lunar impact craters as described above, optionally, the area to be identified is identified according to the pre-built grid-level classifier and the DEM data of the area to be identified to obtain a candidate impact Pit, including: according to the DEM data of the area to be identified, determine the terrain element type of each grid in the area to be identified; Classify each grid of the region to be identified to obtain a candidate impact crater grid; obtain the candidate impact crater based on the candidate impact crater grid.

In the above-mentioned method for automatically extracting impact craters from the moon, optionally, the obtaining the candidate impact craters according to the grid of candidate impact craters includes: using a clustering algorithm to cluster the candidate grids of impact craters class to obtain the clustered objects; create a circumscribed circle based on the clustered objects to obtain the candidate impact craters.

In the method for automatically extracting lunar impact craters as described above, optionally, the clustering algorithm applies spatial clustering to density-based noise.

In the method for automatically extracting lunar impact craters as described above, optionally, the candidate impact craters are identified based on the pre-built object-level classifier and the radial elevation profile of the candidate impact craters to obtain The number of radial elevation profiles of the candidate impact craters used in the impact craters is 12, and the intervals are 30°.

Another aspect provides a device for automatically extracting lunar impact craters based on machine learning, which includes: an acquisition module for acquiring DEM data of areas to be identified on the moon; The classifier and the DEM data of the area to be identified identify the area to be identified to obtain candidate impact craters; the second obtaining module is used to obtain the candidate impact crater and the DEM data of the area to be identified. The radial elevation profile of the candidate impact crater; the third obtaining module is used to identify the candidate impact crater according to the pre-built object-level classifier and the radial elevation profile of the candidate impact crater, and obtain impact crater.

In the above-mentioned automatic extraction device for lunar impact craters, optionally, the first obtaining module includes: a determining unit, configured to determine each The topographic element type of the grid; the first obtaining unit is used to, based on the pre-built grid-level classifier and the topographic element type of each grid in the area to be identified, for each grid in the area to be identified performing classification to obtain a grid of candidate impact craters; a second obtaining unit configured to obtain the candidate impact craters according to the grid of candidate impact craters.

In the above-mentioned device for automatically extracting lunar impact craters, optionally, the second obtaining unit includes: a first obtaining subunit, configured to use a clustering algorithm to cluster the impact crater candidate grids, The clustered objects are obtained; the second obtaining subunit is configured to create a circumscribed circle based on the clustered objects to obtain the candidate impact craters.

In the aforementioned device for automatically extracting lunar impact craters, optionally, the clustering algorithm used by the first obtaining subunit is density-based noise application space clustering.

In the above-mentioned device for automatically extracting lunar impact craters, optionally, the number of radial elevation profiles of the candidate impact craters used by the third obtaining module is 12, and the intervals are 30°.

The beneficial effects brought by the technical solution provided by the embodiments of the present invention are:

By obtaining the DEM data of the area to be identified on the moon, according to the pre-built grid-level classifier and the DEM data of the area to be identified, identify the area to be identified, and obtain candidate impact craters. According to the DEM data of candidate impact craters and the area to be identified , get the radial elevation profile of the candidate impact crater, identify the candidate impact crater according to the pre-built object-level classifier and the radial elevation profile of the candidate impact crater, obtain the impact crater, and combine the machine learning method with the geomorphic elements , using a combination of grid-oriented and object-oriented methods in machine learning, thereby improving the recognition accuracy of impact craters.

Description of drawings

Fig. 1 is a schematic flow chart of a method for automatically extracting lunar impact craters based on machine learning provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram of an impact crater database with a diameter greater than 20 km provided in the prior art;

FIG. 3 is a schematic structural diagram of a training area based on FIG. 2 provided by an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of a test area based on FIG. 2 provided by an embodiment of the present invention;

Fig. 5 is a schematic diagram of distribution of geomorphic elements at different analysis scales based on Fig. 3 provided by an embodiment of the present invention;

FIG. 6 is a schematic diagram of a training process of a raster-level classifier provided by an embodiment of the present invention;

Fig. 7 is a schematic diagram of a radial elevation section provided by an embodiment of the present invention;

Fig. 8 is a schematic diagram of a normalized radial elevation section provided by an embodiment of the present invention;

FIG. 9 is a schematic structural diagram of positive samples and negative samples of a training region provided by an embodiment of the present invention;

Fig. 10 is another schematic flow chart of automatic extraction of lunar impact craters based on machine learning provided by an embodiment of the present invention;

Fig. 11 is a schematic structural diagram of a device for automatically extracting lunar impact craters based on machine learning provided by an embodiment of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and examples.

Referring to Fig. 1, the embodiment of the present invention provides a kind of automatic extraction method of lunar impact crater based on machine learning, and it comprises the following steps:

Step 101, acquiring DEM data of the region to be identified on the moon.

This embodiment does not limit the production method of the acquired lunar DEM data. It can be produced by using the laser altimeter height data carried by the Chang'e-1 satellite, or by using the image data returned by the Chang'e satellite CCD three-line array camera and the satellite-borne laser. Altimeter combined to make. In the application, the digital elevation data can be selected from the DEM data of Chang'e-1 with a resolution of 500m, and its vertical accuracy is about 60m. After the DEM data is produced, select the area to be identified.

Step 102, according to the pre-built raster-level classifier and the DEM data of the area to be identified, the area to be identified is identified to obtain candidate impact craters.

Specifically, the geomorphic element type is extracted from the EDM data of the area to be identified, and the type of geomorphic element in the area to be identified is obtained. The completed raster classifier identifies the area to be identified according to the type of geomorphic element in the area to be identified. The geomorphic element type of the region is used as the input of the constructed raster-level classifier to obtain the impact crater candidate grid and non-impact crater grid, perform spatial clustering on the impact crater candidate grid, and create the minimum circumscribed circle to obtain the candidate impact Crater (or candidate impact crater object). The algorithm of spatial clustering can select DBSCAN algorithm (Density-BasedSpatial Clustering of Applications with Noise, has the clustering method based on the density of noise), in other embodiments, can also select other spatial clustering algorithms for use, the present embodiment is to This is not limited. Both the impact crater candidate grid and the non-impact crater grid are image grids. Since there are many impact craters on the lunar surface and the spatial structure is complex, the obtained binary image has more noise. Therefore, before spatial clustering, the impact crater candidate The raster is processed by opening operation to remove noise.

The following describes the construction process of the classifier:

To construct the data base of the classifier, specifically, the existing impact crater database can be selected, for example, the existing impact crater database with a diameter greater than 20 km, as shown in Figure 2. The impact crater database adopts manual visual interpretation method, which has high accuracy. The digital elevation data used to extract the spatial structure information of the impact crater can use the DEM data of Chang'e-1 with a resolution of 500m, and the vertical accuracy is about 60m. Select the training area and the test area in the existing impact crater database. The two areas are located in the mid-latitude region on the back of the moon, which is covered by many impact craters. Among them, there are 18 training areas and 92 test areas. The impact crater is shown in Figure 3. In Figure 3, the left figure shows the training area, and the right figure shows the test area.

The training method of the classifier is RF (Random Forest, Random Forest), which is a widely used machine learning classification method and has been successfully applied in many fields such as geomorphology and geology. During training, each tree in RF is based on randomly selected input features and randomly selects samples from the training set. For unlabeled data, the RF-based classification result is an aggregation of the classification results of each tree, weighted by their probability estimates. Avoids overfitting when learning complex relationships and works well with relatively sparse, unbalanced data and noise. Therefore, the raster-level classifier can be called a raster-level random forest classifier, and the following object-level classifier can be called an object-level random forest classifier. In other embodiments, other classification training methods can also be used , which is not limited in this embodiment.

Impact craters are characterized by lower, wider, and flatter interiors, and higher, narrower, and steeper edges, as shown in Figure 4, but some craters have a central bulge inside. Compared with terrain attributes (ie, slope, curvature, etc.), the type of geomorphic elements (geomorphological elements or topographic elements or geomorphic elements) at a location can fully reflect the spatial structure of the surrounding terrain at a certain analysis scale. The geomorphic element type is the spatial structure information, which can include: mountain peaks, ridges, flat areas, valleys, slope shoulders, slope feet, straight back slopes, convex back slopes, concave back slopes, and depressions. The extraction method of geomorphic element types is mainly to use raster DEM data to space-divide the surface morphology and specify the terrain element category to which it belongs. The extraction method can use existing technologies, such as multi-scale extraction method, which can be a classification method of comprehensive discrimination of terrain elements at multiple analysis scales. Under this method, the extracted geomorphic element types are called multi-scale geomorphic element types.

In this embodiment, multi-scale geomorphic elements are used to describe the spatial structure information of the grid in the impact crater, and these features (ie, multi-scale geomorphic elements) are input into the classifier for training. Since the geomorphic elements are not stable on a smaller analysis scale, the method of determining the types of geomorphic elements on the preset analysis scale is adopted. A feature point is determined on each preset analysis scale, which refers to one of the eight field directions, from the end point of the analysis scale to the point where the elevation deviation between the elevation profile of the unit of interest and the straight elevation line is the largest pixel.

Since 96.2% of the impact craters on the moon have a diameter less than 60 km, for the grid in the impact crater, the maximum analysis scale of 60 km can effectively cover the impact crater, and can avoid including more information about the impact craters that are far away. The selected analysis scale starts from 3km, and the step length is 1km, that is, two pixels, and the analysis reaches 60km. While reducing the amount of calculation, it can also reduce the loss of the spatial structure of the impact crater to a certain extent. Figure 5 is the distribution map of geomorphic elements on the preset analysis scale of the training area. The analysis scale of Figure 5A is 20km, the analysis scale of Figure 5B is 40km, and the analysis scale of Figure 5C is 60km. It can be seen from the figure that as the analysis scale increases, more grids in the impact crater are identified as landform elements with negative topography, while the edge of the crater is identified as landform elements with positive topography. This feature is consistent with the macroscopic structure of the impact crater. unanimous. That is to say, multi-scale analysis can quantify the spatial structure of impact craters, so that impact craters can be extracted more effectively.

At the grid level, the topography of grids inside impact craters and their surrounding areas often show variations in convexity and concavity at different analysis scales (i.e., different window sizes). Therefore, the input features of each sample of the raster-level classifier (or called raster-level classifier) are the multi-scale geomorphic elements of the sample unit on a series of continuous analysis scales. According to the existing impact crater database, determine the grid-level positive and negative samples to identify the impact crater, and use the DEM data with the spatial structure information of the impact crater (that is, multi-scale landform elements) as the input features of the classifier, and train the grid level classifier. A positive sample indicates that there are impact craters in the sample, and a negative sample indicates that there are no impact craters in the sample.

After training the raster-level classifier in the training area, apply the classifier to the test area to verify the accuracy of the classifier. When the accuracy requirement is met, the training of the raster-level classifier is completed; otherwise, the parameters of the raster-level classifier are adjusted until the accuracy requirement is met in the test area.

After the construction of the raster-level classifier is completed, the types of geomorphic elements are extracted from the EDM data of the area to be identified, and the types of geomorphic elements in the area to be identified are obtained. According to the geomorphic feature type of the raster, it can be classified as a crater candidate raster or a non-crater raster. The division is done with a raster-level classifier. Use the completed raster-level classifier to identify the area to be identified according to the type of geomorphic elements in the area to be identified, that is, the type of geomorphic element in the area to be identified is used as the input of the constructed raster-level classifier, and the candidate raster of the impact crater and non-crater candidate rasters, and then convert the crater candidate rasters to candidate craters. The conversion process includes: removing outliers (for example, regions smaller than the first preset threshold, or regions whose distance from other candidate grids is greater than the second preset threshold), and then: using the morphological opening operation in mathematics to remove noise, Then use the DBSCAN algorithm to create the candidate grid of impact craters, that is, gather the grids based on the density of the grids to obtain grid clusters of arbitrary shape, and then create the minimum circumscribed circle to obtain candidate impact craters, that is, candidate impact craters . Referring to FIG. 6 , the figure schematically shows the training process of the raster-level classifier.

It should be noted that the impact craters formed by the impact are basically nearly circular. Although the superposition and degradation of the later impact craters will change the shape of the impact craters, they are all circular in theory, so the circumscribed circle is used for creation.

Step 103, according to the candidate impact crater and the DEM data of the area to be identified, obtain the radial elevation profile of the candidate impact crater.

In step 102, the obtained candidate impact craters are only possible aggregates of grid cells of impact craters, and how to judge whether they are real impact craters needs to consider the radial elevation profile. Elevation profiles can reflect the spatial structure of impact craters to a certain extent. Although the theoretical impact crater is a circular concave area, in reality, the spatial structure of the impact crater is more complex due to the degradation and superposition of the impact crater. Therefore, the elevation profiles of impact craters in different directions can effectively characterize their detailed spatial structure characteristics. For a single candidate impact crater, multiple radial elevation profiles, such as 12, can be selected.

Step 104, according to the pre-built object-level classifier and the radial elevation profile of the candidate impact crater, identify the candidate impact crater to obtain the impact crater.

Figure 8 is a radial elevation profile of an impact crater. If the input feature is the radial elevation profile, it can be determined whether the candidate impact crater is an impact crater. In order to improve the recognition accuracy, for candidate impact craters, the input features are multiple Radial elevation profiles, for example, 12 radial elevation profiles, with an interval of 30 degrees between adjacent radial elevation profiles, to determine whether each radial elevation profile is an impact crater, and if it is judged to be the number of impact craters If it is greater than or equal to the preset number threshold, the candidate impact crater is an impact crater, otherwise it is a non-impact crater. The preset number threshold is a positive number, which may be half of the multiple, or 0.75 times the multiple, etc. This embodiment does not limit the specific value of the preset threshold.

The following describes the construction process of the object-level classifier:

This process mainly includes two parts: collecting object-level training samples and input features with spatial structure information. The radial elevation profile of the impact crater is shown in Figure 7. The elevation profile of the impact crater is used as a training sample of the impact crater. The elevation profiles in different directions can effectively indicate the detailed spatial structure characteristics of the impact crater.

In general, the elevation profiles of different impact craters usually have different lengths and buffers, so each elevation profile is normalized, as shown in Figure 8, and the average elevation of each part of the radial elevation profile is converted to It is between 0 and 1, and 0 and 1 correspond to the minimum and maximum value of the profile elevation respectively. The input features for each sample to the object-level classifier are the normalized elevation values [0,1] from the middle of the sample profile.

In order to train the object-level random forest classifier, 12 radial elevation profiles of a single impact crater in the training area are collected as positive samples, and all angles between adjacent profiles are 30°, as shown in Fig. 9, Fig. The abscissa in 9 indicates that the training area is divided into 9 sections in the horizontal direction, and each section is a unit of length. Each crater in the training area provides many elevation profiles. Even if the training area has few impact craters, a large number of positive samples can be collected at the object level to train an object-level random forest classifier. For each crater in the training region, randomly generate a circle of similar size in the non-cratered region of the training data. Subsequently, 12 radial elevation profiles were collected from these circles as negative samples for training a random forest classifier construction at the object level.

After the object-level random forest classifier is trained in the training area, the candidate impact craters obtained in step 102 are detected. For each crater candidate object, its radial elevation profile is extracted in the same way as collecting positive samples at the object level. Each radial elevation profile was converted into input features for an object-level random forest classifier to assess whether the profile is an impact crater. If a candidate for an impact crater is characterized by a radial elevation profile and is judged to be an impact crater, then the candidate for an impact crater is a confirmed impact crater. Therefore, the proposed method can effectively detect those degenerate and superimposed impact craters.

Referring to Fig. 10, that is to say: this method trains a classifier (such as a random forest classifier) based on the existing impact craters and combined with the spatial structure information of the impact craters obtained by digital terrain analysis, so as to realize the automatic extraction of impact craters. In the first stage, a grid-level random forest classifier is trained to identify candidate grids of impact craters in the target area, taking the grid of impact craters as a sample and using the multi-scale terrain element information obtained from terrain analysis as a feature , to obtain candidate impact crater objects; in the second stage, an object-level random forest classifier is trained with the radial elevation profile of impact craters in the impact crater distribution map to determine whether each candidate object obtained in the first stage is impact crater.

Overall, the present method combines machine learning methods with geomorphic elements. In this method, the random forest effectively utilizes the multi-scale geomorphic elements obtained by digital terrain analysis, not only quantifies the morphological information, but also quantifies the spatial structure inside a single impact crater, combining the spatial structure information obtained by terrain analysis with machine learning together, improving the accuracy of the machine learning model and extending the method to extract landform types.

Referring to FIG. 11 , an embodiment of the present invention provides a device for automatically extracting lunar impact craters based on machine learning, which includes: an acquisition module 201 , a first acquisition module 202 , a second acquisition module 203 and a third acquisition module 204 .

Specifically, the acquisition module 201 is used to acquire the DEM data of the region to be identified on the moon. The first obtaining module 202 is configured to identify the area to be identified according to the pre-built grid-level classifier and the DEM data of the area to be identified, and obtain candidate impact craters. The second obtaining module 203 is used to obtain the radial elevation profile of the candidate impact crater according to the DEM data of the candidate impact crater and the area to be identified. The third obtaining module 204 is configured to identify the candidate craters according to the pre-built object-level classifier and the radial elevation profiles of the candidate craters, and obtain the impact craters.

Optionally, the first obtaining module 202 includes: a determining unit, a first obtaining unit and a second obtaining unit. The determination unit is used to determine the geomorphic element type of each grid in the area to be identified according to the DEM data of the area to be identified. The first obtaining unit is configured to classify each grid in the area to be identified based on the pre-built grid-level classifier and the geomorphic element type of each grid in the area to be identified, so as to obtain candidate grids of impact craters. The second obtaining unit is used to obtain candidate impact craters according to the grid of impact crater candidates.

Optionally, the second obtaining unit includes: a first obtaining subunit and a second obtaining subunit. The first obtaining subunit is used to cluster the impact crater candidate grids using a clustering algorithm to obtain clustered objects. The second obtaining subunit is used to create a circumscribed circle based on the clustered objects to obtain candidate impact craters.

Optionally, the clustering algorithm used by the first deriving subunit is density-based noise application spatial clustering.

Optionally, the number of radial elevation profiles of candidate impact craters used by the third obtaining module 204 is 12, and the intervals are 30°.

It should be noted that when the automatic extraction device for lunar impact craters provided by the above-mentioned embodiments is extracted, only the division of the above-mentioned functional modules is used as an example for illustration. In practical applications, the above-mentioned function distribution can be completed by different functional modules according to needs. , which divides the internal structure of the device into different functional modules to complete all or part of the functions described above. In addition, the device for automatically extracting lunar impact craters provided in the above-mentioned embodiments belongs to the same concept as the embodiment of the method for automatically extracting lunar impact craters, and its specific implementation process is detailed in the method embodiments, and will not be repeated here.

It can be known from common technical knowledge that the present invention can be realized through other embodiments without departing from its spirit or essential features. Accordingly, the above-disclosed embodiments are, in all respects, illustrative and not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are embraced by the present invention.

Claims (10)

1. An automatic moon impact pit extraction method based on machine learning is characterized by comprising the following steps:

acquiring DEM data of an area to be identified on the moon;

identifying the area to be identified according to a pre-constructed grid-level classifier and DEM data of the area to be identified to obtain candidate impact pits;

obtaining a radial elevation profile of the candidate impact pit according to the candidate impact pit and the DEM data of the area to be identified;

and identifying the candidate meteorite crater according to a pre-constructed object-level classifier and the radial elevation profile of the candidate impact crater to obtain the impact crater.

2. The automatic moon impact pit extraction method according to claim 1, wherein the identifying the area to be identified according to a pre-constructed grid-level classifier and DEM data of the area to be identified to obtain candidate impact pits comprises:

according to the DEM data of the area to be identified, determining the type of the landform element of each grid in the area to be identified;

classifying each grid of the area to be identified based on a pre-constructed grid-level classifier and the landform element type of each grid in the area to be identified to obtain a collision pit candidate grid;

and obtaining the candidate impact pit according to the impact pit candidate grid.

3. The automatic moon impact pit extraction method according to claim 2, wherein the obtaining of the candidate impact pits from the impact pit candidate grid comprises:

clustering the impact pit candidate grids by using a clustering algorithm to obtain clustered objects;

and creating a minimum circumcircle based on the clustered objects to obtain the candidate impact pits.

4. The method for automatically extracting moon impact pits according to claim 3, wherein the clustering algorithm is a density-based noise application spatial clustering algorithm.

5. The method for automatically extracting moon impact craters according to claim 1, wherein the candidate merle crates are identified based on a pre-constructed object-level classifier and the radial elevation profiles of the candidate impact crates, and the number of the radial elevation profiles of the candidate impact crates used in the impact crates is 12, and the profiles are spaced by 30 °.

6. An automatic moon impact pit extraction device based on machine learning, characterized in that the automatic moon impact pit extraction device comprises:

the acquisition module is used for acquiring DEM data of an area to be identified on the moon;

the first obtaining module is used for identifying the area to be identified according to a pre-constructed grid-level classifier and DEM data of the area to be identified to obtain candidate impact pits;

the second obtaining module is used for obtaining a radial elevation profile of the candidate impact pit according to the candidate impact pit and DEM data of the area to be identified;

and the third obtaining module is used for identifying the candidate meteorite crater according to a pre-constructed object-level classifier and the radial elevation profile of the candidate crater to obtain the crater.

7. The automatic moon impact pit extraction device according to claim 6, wherein the first obtaining module comprises:

the determining unit is used for determining the type of the landform element of each grid in the area to be identified according to the DEM data of the area to be identified;

the first obtaining unit is used for classifying each grid of the area to be identified based on a pre-constructed grid-level classifier and the landform element type of each grid in the area to be identified to obtain a collision pit candidate grid;

and the second obtaining unit is used for obtaining the candidate impact pit according to the impact pit candidate grid.

8. The automatic moon pit extraction apparatus according to claim 7, wherein the second obtaining unit includes:

the first obtaining subunit is used for clustering the impact pit candidate grids by using a clustering algorithm to obtain clustered objects;

and the second obtaining subunit is used for creating a circumscribed circle based on the clustered object to obtain the candidate impact pit.

9. The automatic moon impact pit extraction device according to claim 8, wherein the clustering algorithm used by the first deriving subunit applies spatial clustering for density-based noise.

10. The automatic moon impact pit extraction device according to claim 8, wherein the number of radial elevation profile of the candidate impact pits used by the third obtaining module is 12, and the candidate impact pits are spaced 30 ° apart from each other.

Images