An automatic mapping method and system for lunar impact crater micro-topography

Abstract

This application provides an automatic mapping method and system for lunar impact crater micromorphology, belonging to the field of image data processing technology. The method involves acquiring a candidate raster dataset, including lunar digital terrain data, lunar image data, and terrain-derived data; the terrain-derived data includes multiple terrain-derived factors; filtering the candidate raster dataset to determine the optimal feature combination; combining the optimal feature combination with the labels of impact crater micromorphology subclasses to construct an impact crater micromorphology subclass dataset; inputting the impact crater micromorphology subclass dataset into a pre-built deep learning model for training to obtain a target model; and using the target model to automatically extract impact crater micromorphology subclasses across the entire lunar surface, thereby completing the mapping of the entire lunar landscape. This solution alleviates the current problem of insufficient efficiency in extracting impact crater micromorphology subclasses primarily through manual extraction, providing the possibility for high-precision and rapid mapping of the entire lunar landscape.

Description

Technical Field

[0001] This application relates to the field of image data processing technology, and in particular to an automatic mapping method and system for micromorphological landforms of lunar impact craters. Background Technology

[0002] The lunar surface features, especially impact craters, are complex and diverse. Extracting impact craters is a crucial step in lunar topographic mapping and research, and it is also a time-consuming and costly task.

[0003] With the development of artificial intelligence technologies such as machine learning and neural networks, related technologies employ supervised learning algorithms such as genetic algorithms, Adaboost, and support vector machines to train high-performance strong classifiers to achieve automatic identification and extraction of impact craters.

[0004] However, existing technologies for the automatic extraction of impact craters mostly focus on extracting the boundary and location information of impact craters, but there is little research on the automatic extraction and segmentation of impact crater micromorphological subclasses. The extraction accuracy is insufficient and it is difficult to meet the precision requirements of actual mapping. The efficiency of high-precision full-moon map of impact crater micromorphological subclasses is low.

[0005] Therefore, there is a need to provide an improved technical solution that addresses the shortcomings of the existing technology. Summary of the Invention

[0006] The purpose of this application is to provide an automatic mapping method and system for micromorphological topography of lunar impact craters, so as to solve or alleviate the problems existing in the prior art.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] This application provides an automatic mapping method for the micromorphological topography of lunar impact craters, including:

[0009] Obtain a candidate raster data set; wherein, the candidate raster data set includes lunar digital terrain data, lunar image data, and terrain-derived data; the terrain-derived data includes multiple terrain-derived factors;

[0010] The candidate raster data set is filtered to determine the optimal feature combination;

[0011] The optimal feature combination is combined with the labels of the impact crater micromorphology subclasses to construct an impact crater micromorphology subclass dataset.

[0012] The impact crater micromorphology subclass dataset is input into a pre-built deep learning model for training to obtain the target model;

[0013] The target model is used to automatically extract the micromorphological subclasses of impact craters across the entire lunar area, and the extraction results are obtained; then, based on the extraction results, the entire lunar topography is mapped.

[0014] The impact crater micromorphology subclasses are divided into four subclass unit types: central peak, crater bottom, crater wall, and crater edge.

[0015] In one possible implementation, the terrain-derived factor includes a macro-terrain factor, which includes lunar surface topographic relief.

[0016] The steps for generating the lunar surface topographic relief are as follows:

[0017] Based on the lunar digital terrain data, the mean variable point method is used to analyze the lunar surface relief and determine the optimal window for calculating the lunar surface relief; then, based on the optimal window for lunar surface relief, the window analysis method is used to calculate the terrain relief.

[0018] In one possible implementation, the macro-terrain factor further includes: roughness,

[0019] The roughness generation steps are as follows:

[0020] Based on the lunar digital terrain data, the surface roughness of the moon is calculated using the root mean square elevation method.

[0021] In one possible implementation, the terrain-derived factors include micro-terrain factors, which include slope, aspect, and curvature;

[0022] The slope, aspect, and curvature are all calculated based on the lunar digital terrain data.

[0023] In one possible implementation, the candidate raster data set is filtered to determine the optimal combination of features, including:

[0024] Normalization preprocessing is performed on each candidate raster data in the candidate raster data set, and feature value distribution maps of each candidate raster data in different impact crater micromorphology subclasses are plotted based on the results of normalization preprocessing.

[0025] Based on the feature value distribution map, the differences between different impact crater micromorphology subclasses on the same candidate raster data are analyzed to exclude candidate raster data with differences less than a preset threshold, and the first screening result is obtained.

[0026] The first screening result is used as the optimal feature combination.

[0027] In one possible implementation, after excluding candidate images with differences less than a preset threshold and obtaining the first screening result, the method further includes:

[0028] Correlation analysis was used to analyze the correlation between various terrain-derived factors in the first screening results;

[0029] Based on the results of the relevant analysis, the first screening results are subjected to a second screening to obtain the optimal feature combination.

[0030] In one possible implementation, the tags for the impact crater micromorphology subclasses are generated through the following steps:

[0031] Acquire the data of the first impact crater surface, the second impact crater surface, and the third impact crater surface;

[0032] The first impact crater surface data, the second impact crater surface data, and the third impact crater surface data are matched, and combined with the lunar digital terrain data, slope data, and mountain shadow data, the impact crater micromorphology subclasses are drawn and refined to obtain vector-form labels for the impact crater micromorphology subclasses.

[0033] Among them, the first impact crater surface data, the second impact crater surface data, and the third impact crater surface data are lunar geological maps compiled by different teams based on data from different sensors.

[0034] In one possible implementation, after obtaining the labels for the vector-based impact crater micromorphology subclasses, the method further includes:

[0035] The vector-form impact crater micromorphology subclasses are categorically encoded and rasterized. One-hot encoding is then used to convert the rasterized impact crater micromorphology subclasses into multi-channel data.

[0036] In one possible implementation, the deep learning model is constructed as follows: based on the deep convolutional network U-Net, the input channels are modified to be multi-channel, and the original U-Net's Upsampling module is modified to a MaxUnpooling module.

[0037] This embodiment also provides an automatic mapping system for the micromorphological topography of lunar impact craters, including:

[0038] The data acquisition unit is configured to acquire a set of candidate raster data; the set of candidate raster data includes lunar digital terrain data, lunar image data, and terrain-derived data.

[0039] The data filtering unit is configured to filter the candidate raster data set to determine the optimal feature combination;

[0040] The training set construction unit is configured to combine the optimal feature combination with the labels of the impact crater micromorphology subclasses to construct an impact crater micromorphology subclass dataset.

[0041] The model training unit is configured to input the impact crater micromorphology subclass dataset into a pre-built deep learning model for training to obtain the target model;

[0042] The automatic extraction unit is configured to use the target model to automatically extract the micromorphological subclasses of impact craters across the entire lunar range, and obtain the extraction results; then, based on the extraction results, to complete the mapping of the impact crater topography across the entire lunar range.

[0043] The impact crater micromorphology subclasses are divided into four subclass unit types: central peak, crater bottom, crater wall, and crater edge.

[0044] The technical solution of this application embodiment has the following beneficial effects:

[0045] The technical solution provided in this application selects the best feature combination from lunar digital terrain data, lunar image data and terrain-derived data, combines it with the labels of impact crater micromorphology subclasses to construct a high-quality impact crater micromorphology subclass dataset, then trains a deep learning model, and then uses the trained deep learning model (target model) to automatically extract impact crater micromorphology subclasses, so as to achieve high-precision rapid mapping of the entire lunar landform. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating an automatic mapping method for lunar impact crater micromorphology according to some embodiments of this application.

[0047] Figure 2 A schematic diagram of the micromorphological subclasses of lunar impact crater topography.

[0048] Figure 3 This is a schematic diagram of the process for automatically extracting micromorphological subclasses of lunar impact craters according to some embodiments of this application.

[0049] Figure 4 This is a schematic diagram of terrain-derived data.

[0050] Figure 5 The eigenvalue distribution of candidate raster data in different subclasses of impact crater micromorphology.

[0051] Figure 6 A schematic diagram showing the results of drawing and finishing the micromorphological subclasses of impact craters.

[0052] Figure 7 A schematic diagram of the deep learning model structure provided for some embodiments of this application.

[0053] Figure 8 The image shows the sub-class segmentation results of Al-Biruni impact crater micromorphology based on the Crater3-Net model for some embodiments of this application.

[0054] Figure 9 The image shows the Jackson impact crater micromorphology subclass segmentation results based on the Crater3-Net model, provided for some embodiments of this application.

[0055] Explanation of reference numerals in the attached figures:

[0056] 1-Pit rim, 2-Pit wall, 3-Pit bottom, 4-Central peak, 5-Lunar crust. Detailed Implementation

[0057] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0058] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0059] It should be noted that, due to the limited research on the automatic extraction and segmentation of impact crater micromorphological subclasses in existing technologies, there is currently a technological gap in the industry regarding which feature combinations to use in deep learning models to achieve the automatic extraction of different impact crater micromorphological subclasses.

[0060] Some existing technologies also involve mapping lunar impact crater types such as ejecta, crater rims, crater walls, crater floors, and central peaks in lunar geological maps. However, this mapping process does not use deep learning technology. Instead, it converts DEM data into grayscale classification maps and matches them with high-precision lunar images. The boundaries of different impact crater micromorphological subclasses are delineated by searching for the maximum value of the DEM or the maximum value of the grayscale.

[0061] However, the inventors discovered that the above-mentioned process of classifying impact crater micromorphological subclasses ignores the complexity of lunar surface topography, which may lead to inaccurate boundary delineation and thus affect the accuracy of lunar topographic mapping. Specifically, the crater rim often exhibits irregular topographic features in multiple directions. Delineating feature points solely based on the maximum slope or the highest DEM value may not accurately reflect the actual morphology of the crater rim. This is especially true for more complex impact craters, which may contain local highlands or depressions. These topographic features may not be captured by a single maximum slope or elevation. Furthermore, in reality, the edges of different impact crater micromorphological subclasses are not regular linear structures but may contain irregular shapes. Using feature points to connect and define the boundaries of each impact crater micromorphological subclass may not fully capture these irregular shapes, leading to inaccurate classification.

[0062] The following reference Figure 2 To elaborate on the complexity and diversity of lunar impact crater micromorphological subclasses.

[0063] like Figure 2 As shown, lunar impact craters, located on the lunar crust 5, are the most common and prominent geomorphic units and geological markers on the lunar surface. They are numerous, varied in shape, and exhibit a ring-shaped crater structure of varying sizes and uneven clustering. Based on their formation mechanism, impact crater landforms can be divided into four sub-types: central peak 4, crater floor 3, crater wall 2, and crater rim 1. The spatial boundary and topographic features between central peak 4 and crater floor 3 are significantly different. In well-preserved impact craters, the spatial boundaries between crater floor 3 and crater wall 2, as well as between crater wall 2 and crater rim 3, also exhibit distinct characteristics. However, in the micromorphic sub-types of degraded and altered impact craters, the topographic features and spatial boundaries between crater floor 3 and crater wall 2 are not significantly distinct. Given the complexity and diversity of impact crater micromorphic sub-types, how to effectively extract these sub-types using deep learning models, what the candidate feature dataset should include, and which features can effectively distinguish between different micromorphic sub-types are worthy of further research.

[0064] In addition, although deep learning models have been applied to lunar mapping, these technologies typically only involve extracting the overall boundary and location information of impact craters based on lunar DEM data and image data, without considering how to extract the boundaries of various micromorphological subclasses of impact craters. Based on this, this embodiment attempts to select features for impact crater micromorphological subclasses. It not only introduces lunar DEM data and lunar image data as alternative input data for deep learning, but also adds terrain-derived data, in order to enable the deep learning model to learn the most effective features of impact crater micromorphological subclasses, thereby achieving high-precision automatic identification and extraction of micromorphological subclass boundaries.

[0065] The embodiments of this application will now be described with reference to the accompanying drawings.

[0066] This embodiment provides an automatic mapping method for the micromorphological topography of lunar impact craters, such as... Figures 1-9 As shown, the method includes:

[0067] Step S101: Obtain the set of candidate raster data.

[0068] In this embodiment, the candidate raster dataset is used as the candidate feature set for the deep learning model, which uses this data for feature extraction and learning.

[0069] Raster data is spatial data stored in a grid format. Each raster unit (pixel) corresponds to a specific value, representing a certain spatial attribute or observation value, also known as a cell value.

[0070] A feature set refers to the input data used for model training. For deep learning, a feature set is usually a multi-dimensional data matrix (such as pixel values ​​of an image, elevation values ​​of raster data, etc.). The raster dataset serves as the input features of the deep learning model, helping the model learn the complex patterns of the micromorphology of lunar impact craters.

[0071] In this embodiment, the candidate raster data set includes Lunar Digital Elevation Model (DEM), lunar imagery data, and terrain-derived data.

[0072] In other words, lunar DEM data, lunar image data, and terrain-derived data are all stored in the form of raster data. Together, they serve as candidate features for deep learning models to help the models identify impact craters of different micromorphological subclasses.

[0073] Specifically, lunar digital terrain data is digital data acquired through satellite remote sensing, probes, or other space technologies. It is used to represent the elevation of the lunar surface and is an important basic data for studying the landforms, geological features, and other related scientific research of the moon.

[0074] Terrain-derived data refers to raster data derived from basic terrain data such as lunar DEM data, used to describe impact craters of different micromorphological subtypes on the Moon. These raster data can characterize the spatial distribution features of lunar landforms.

[0075] Furthermore, the terrain-derived data includes multiple terrain-derived factors, each of which is used to describe a certain attribute of the lunar landform. Different terrain-derived factors can describe the morphological characteristics of the lunar landform from different perspectives.

[0076] Lunar imagery data refers to imagery data acquired through space probes and satellite equipment, typically used to study the geological features, landforms, and lunar environment of the lunar surface. In this embodiment, lunar imagery data can specifically be optical imagery data. In some embodiments, lunar imagery data can also be multispectral images or thermal infrared images. This embodiment does not limit this.

[0077] In this embodiment, lunar digital terrain data and lunar imagery data can be acquired through various means, such as different space missions, satellite observations, and public data platforms. Alternatively, lunar digital terrain data can be constructed based on technologies such as lidar, radar imaging, and stereo imaging from lunar probes and satellites to create a three-dimensional dataset reflecting the lunar surface elevation. Lunar imagery data can also be acquired through various satellite missions, probes, and space telescopes. This embodiment does not limit the acquisition method of the above data.

[0078] As an example, the lunar digital terrain data and lunar imagery data utilize the latest acquired multi-source remote sensing data of the lunar surface. The lunar imagery data includes data acquired by the US Lunar Reconnaissance Orbiter (LRO) and data acquired by the Japanese Lunar Science and Engineering Explorer (SELENE). The Lunar Reconnaissance Orbiter Camera (LROC) onboard the US LRO includes a Wide Angle Camera (WAC) and a Narrow Angle Camera (NAC), with the Wide Angle Camera imagery (LROC WAC) having a resolution of approximately 100 meters per pixel and the Narrow Angle Camera imagery (LROC NAC) having a resolution of approximately 0.5 meters per pixel. The imagery data acquired by the Terrain Camera (TC) onboard the Japanese Kaguya (SELENE) satellite has a resolution of approximately 10 meters per pixel. The lunar digital terrain data includes: LOLA elevation data from the Lunar Orbiter Laser Altimeter (LOLA), a multi-beam laser altimeter aboard the Lunar Reconnaissance Orbiter (LRO). LOLA emits light at a wavelength of 1064.4 nm at a frequency of 28 Hz. LOLA provides data from over 650 million laser altimeter points across the entire lunar surface, with an elevation accuracy of approximately 10 cm and a precision of approximately 1 m. It offers high-precision global coverage, and the resulting digital elevation model (DEM) data has a resolution of approximately 118 m / pixel, with a vertical accuracy better than 1 m. The lunar digital terrain data also includes SLDEM data, a higher-resolution DEM (SLDEM2015) generated by researchers combining LOLA and SELENETC (Terrain Camera) data. This data covers the region between 60 degrees north and south latitude on the Moon, with a resolution of approximately 59 m / pixel and a vertical accuracy of approximately 3–4 m.

[0079] WAC data with a resolution of 100m / pixel, LOLA digital elevation data with a resolution of 118m / pixel, SLDEM data with a resolution of 59m / pixel, mountain shadow data generated based on SLDEM with a resolution of 59m / pixel, SELENETC imagery data and DTM data with a resolution of 10m / pixel, etc.

[0080] Step S102: Filter the candidate raster data set to determine the best feature combination.

[0081] Furthermore, in step S102, the candidate raster data set is filtered to determine the optimal feature combination. Filtering and optimizing the feature combination can help the model focus on the most important features, thereby improving the model's accuracy and predictive ability, while accelerating the training process and reducing the consumption of computing resources.

[0082] Step S103: Combine the best feature combination with the labels of the impact crater micromorphology subclasses to construct an impact crater micromorphology subclass dataset. This effectively utilizes digital terrain data (DEM), image data, terrain-derived data, and labeled impact crater micromorphology subclasses on the lunar surface to train an accurate model to identify and classify different impact crater micromorphological features.

[0083] The impact crater micromorphology subclass is divided into four subclass units: central peak, crater bottom, crater wall, and crater rim. In other words, the label includes four classification types: central peak, crater bottom, crater wall, and crater rim.

[0084] Specifically, the optimal feature combination is combined with the label of the impact crater micromorphology subclass to form a training dataset (i.e., impact crater micromorphology subclass dataset). Each sample in the training dataset corresponds to a region covered by an impact crater micromorphology subclass in spatial location. Furthermore, each sample consists of a feature vector composed of the optimal feature combination and its corresponding label of the impact crater micromorphology subclass.

[0085] Step S104: Input the impact crater micromorphology subclass dataset into the pre-built deep learning model for training to obtain the target model.

[0086] The purpose of step S104 is to use a deep learning model to train a classification model for impact crater micromorphology subclasses. The model is trained based on the feature vectors and labels in the impact crater micromorphology subclass dataset, so that the model can automatically learn and distinguish different impact crater micromorphological features, and thus accurately classify and identify them in the lunar data to be predicted.

[0087] During training, a certain proportion of samples from the impact crater micromorphology subclass dataset can be randomly selected as the training set, and then the validation set and test set can be divided.

[0088] Furthermore, to prevent gradient vanishing and gradient diffusion during model training, the input data of each class in the impact crater micromorphology subclass dataset is normalized before model training.

[0089] The deep learning model can be a fully convolutional neural network (FCN), U-Net, UNet++, SegNet (based on an encoder-decode architecture), ResNet (a residual network), or Google's DeepLab series models. This embodiment does not limit the specific selection of the deep learning model.

[0090] Step S105: Use the target model to automatically extract the micromorphological subclasses of impact craters across the entire lunar area to obtain the extraction results; then, based on the extraction results, complete the mapping of the entire lunar topography.

[0091] By using a trained target model to automatically process data across the entire lunar area, the micromorphological subclasses of impact craters can be extracted quickly and efficiently. Based on the extraction results, a detailed lunar surface topographic map is constructed, showing impact craters of different micromorphological subclasses and their characteristic regions, which greatly improves the efficiency and accuracy of impact crater micromorphological subclass identification and topographic mapping.

[0092] Preferably, the terrain-derived factors include macro-terrain factors, which include lunar surface topographic relief.

[0093] In this embodiment, the terrain relief is also referred to as topographic relief or relative height.

[0094] The steps for generating lunar surface topographic relief are as follows:

[0095] Based on lunar digital terrain data, the mean variable point method is used to analyze the lunar surface relief and determine the optimal window for calculating the lunar surface relief. Then, based on the optimal window for lunar surface relief, the window analysis method is used to calculate the terrain relief.

[0096] Topographic relief is the difference in elevation within a given regional unit (analysis window). The relief value changes within a certain range as the analysis window changes, exhibiting a certain scale dependence.

[0097] The analysis windows mainly include rectangular windows, circular windows, ring windows, and fan-shaped windows.

[0098] For example, the optimal window could be a circular window. Using a circular window, rather than a rectangular window, to determine the optimal window for lunar surface relief can better match the shape of impact craters, which helps to improve the accuracy of relief calculation.

[0099] In this embodiment, the mean-point method is used to analyze the lunar surface undulation and determine the optimal window size for calculating the lunar surface undulation.

[0100] Due to the significant undulations on the lunar surface and the large amount of data required for the entire month, this embodiment employs a two-step mean-variable-point method to accurately determine the optimal window for undulation. The first step uses the mean-variable-point method to estimate the possible window for optimal undulation. For example, the initial window is a circular window with a radius of 5 pixels and a step size of 5. The average undulation and its corresponding statistics are calculated up to 500 pixels to obtain the number of pixels corresponding to the possible window for optimal undulation. The second step uses the mean-variable-point method for precise analysis of the optimal window. For example, the initial window is 1 pixel with a step size of 1. The calculation is repeated up to twice the number of pixels corresponding to the possible window for optimal undulation obtained from the pre-analysis, yielding the precise number of pixels for the optimal window.

[0101] Preferably, the macroscopic terrain factor further includes: roughness, and the roughness generation steps are as follows: based on lunar digital terrain data, the roughness of the lunar surface is calculated using the root mean square elevation method.

[0102] Specifically, roughness is used to describe how the elevation of the lunar surface changes with a horizontal scale. In other words, the roughness of the lunar surface quantitatively describes the topographic relief on a horizontal scale, and this factor can reveal, to some extent, the spatial differentiation characteristics of lunar landforms.

[0103] The root mean square elevation (RMS) method is used to calculate lunar surface roughness. By calculating the deviation and square mean of elevation values, the influence of accidental noise or local irregularities can be reduced. Even if there are some outliers or small noise in the data, the RMS value can remain relatively stable. At the same time, the RMS value directly reflects the roughness of the region. It is an intuitive and simple calculation method that is suitable for the calculation of large-scale raster data on the lunar surface.

[0104] Preferably, the terrain-derived factors include micro-terrain factors, which include slope, aspect, and curvature; slope, aspect, and curvature are all calculated based on lunar digital terrain data.

[0105] The slope of any point on the lunar surface refers to the angle between the tangent plane passing through that point and the horizontal ground, indicating the degree of inclination of the local surface slope. The slope directly affects the scale and intensity of surface material flow and energy conversion.

[0106] Furthermore, the maximum average method was used to calculate the slope data for the entire month based on lunar digital terrain data.

[0107] Curvature includes sectional curvature and planar curvature. Sectional curvature is the rate of change of slope, while planar curvature is the rate of change of aspect. Curvature can reveal the bending characteristics of the lunar surface and helps to understand the processes of change in the lunar surface morphology.

[0108] In this embodiment, the slope aspect and curvature are calculated based on lunar digital terrain data.

[0109] Figure 4 The diagram illustrates terrain-derived data, where (a) represents relief amplitude (RA), (b) slope, (c) profile curvature, (d) aspect, (e) plan curvature, and (f) root mean square height. As can be seen from the diagram, different terrain-derived factors can describe impact crater micromorphological subclasses from different perspectives and can serve as alternative input data for deep learning models.

[0110] Preferably, the candidate raster data set is screened to determine the optimal feature combination, including:

[0111] Normalization preprocessing is performed on each candidate raster data in the candidate raster dataset, and feature value distribution maps of each candidate raster data in different impact crater micromorphology subclasses are plotted based on the results of normalization preprocessing.

[0112] Based on the eigenvalue distribution map, the differences between different impact crater micromorphological subclasses on the same candidate raster data are analyzed to exclude candidate raster data with differences less than a preset threshold, and the first screening result is obtained; the first screening result is used as the optimal feature combination.

[0113] In the above steps, firstly, multiple typical impact craters of different ages and morphological characteristics are selected. After normalization preprocessing of each candidate raster data in the candidate raster data set, the distribution characteristics of each candidate raster data in different impact crater micromorphological subclasses are plotted to obtain the feature value distribution map. Based on the feature value distribution map, features with little difference are excluded to improve the model's recognition effect, reduce computational overhead, and improve the model's interpretability.

[0114] For example, such as Figure 5 As shown in the eigenvalue distribution map, LDEM represents lunar DEM data, Aspect is slope aspect, Plan_Cur is planar curvature, Profile_Cur is profile curvature, Slope is slope gradient, RA is topographic relief, MAS is roughness represented by root mean square height, Shade is mountain shadow, and WAC is lunar image data. Figure 5It can be seen that the differences in the distribution of the various micromorphological subclasses of impact craters are not the same across different candidate raster data. Among the four candidate raster data of planar curvature, slope, undulation, and terrain roughness, the four micromorphological subclasses of impact craters have the best distinguishability. The distinguishability is second best in lunar DEM data, mountain shadow, lunar image data, and profile curvature data. However, in the aspect data, the micromorphological subclasses of impact craters show almost no difference. Based on this, the aspect can be excluded from the candidate raster data set, resulting in the first screening result.

[0115] To further improve the sensitivity of features, in some embodiments, after excluding candidate images with differences less than a preset threshold and obtaining the first screening result, the method further includes: using correlation analysis to analyze the correlation between various terrain-derived factors in the first screening result.

[0116] Based on the results of the correlation analysis, the first screening results are further screened to obtain the optimal feature combination.

[0117] Correlation analysis of different topographically derived data within the micromorphological subclasses of impact craters revealed that while elevation showed a high correlation with most topographic factors, elevation data was retained as input channel data due to the spatial structure it represents. Furthermore, slope exhibited a high correlation with topographic relief, mountain shadow, and profile curvature; therefore, slope was selected as input channel data, while topographic relief, mountain shadow, and profile curvature, which showed high correlations, were excluded. Planar curvature showed a low correlation with slope, topographic relief, lunar surface roughness, and profile curvature; therefore, planar curvature was considered as an alternative input data.

[0118] By utilizing correlation analysis, and comprehensively considering the feature representation capabilities of different candidate feature data and the strength of correlation between different terrain-derived data, terrain-derived data with high overlap are eliminated, while data with strong feature representation capabilities are retained, thus reducing computational load. Furthermore, correlation analysis is used as a supplement to eigenvalue distribution map analysis. This two-stage screening of the candidate feature set ensures that the feature vectors input to the model can establish an effective mapping relationship with the model output, guaranteeing that the deep learning model can correctly identify different micromorphological subclasses of impact craters, thereby obtaining accurate identification results.

[0119] In one alternative implementation, the optimal feature combination could be: lunar DEM data, slope, and planar curvature. These three features will serve as input channel data for the deep learning model, participating in subsequent model training and the automatic extraction of impact crater micromorphological subclasses.

[0120] Considering that deep learning algorithms require a sufficient number of samples to summarize, learn, and identify different features of impact crater micromorphology subclasses for network training, and that the trained model is used to segment impact crater micromorphology subclasses, it is necessary to obtain label data from the training sample set. Preferably, the labels for impact crater micromorphology subclasses are generated through the following steps:

[0121] Acquire the data of the first impact crater surface, the second impact crater surface, and the third impact crater surface;

[0122] The data of the first, second, and third impact crater surfaces were matched, and combined with lunar digital terrain data, slope data, and mountain shadow data, the micromorphological subclasses of impact craters were drawn and refined to obtain vector-form labels for the micromorphological subclasses of impact craters.

[0123] Among them, the data on the first impact crater, the second impact crater, and the third impact crater are lunar geological maps compiled by different teams based on data from different sensors.

[0124] In this embodiment, the actual boundary mapping of impact crater micromorphological subclasses is mainly based on expert knowledge through manual visual interpretation. The base map is derived from lunar geological maps compiled by different teams using data from different sensors.

[0125] Due to differences in data acquired by different sensors and the resulting images, the boundaries of impact craters drawn based on different data exhibit certain positional deviations. Furthermore, the details of impact craters depicted in lunar geological maps compiled by different teams vary. For example, some lunar geological maps only depict the actual boundaries of impact craters—the lines connecting points where the slope changes at the crater rim—without distinguishing between the micromorphological subclasses of the impact craters. In contrast, other lunar geological maps identify the crater bottom subclass for larger-diameter impact craters, without distinguishing between the crater walls and rims. In some datasets, impact craters from different eras are superimposed without being segmented. Therefore, it is necessary to match impact crater surface data from different sources and combine them with lunar digital terrain data, slope data, and mountain shadow data to refine and enhance the micromorphological subclasses of impact craters, thereby improving the accuracy of the labeled data and ultimately the quality of the training dataset.

[0126] Figure 6 The results of the mapping and refinement of impact crater micromorphological subclasses are shown. Among them, (a) is the impact crater micromorphological subclass region mapped on a certain lunar geological map, and (b), (c), and (d) are the mapping and refinement results of impact crater micromorphological subclasses under different base maps.

[0127] In addition, it should be noted that impact craters that are older and have degraded due to spatial weathering, whose micromorphological subclasses cannot be identified and classified, should be excluded when preparing samples to further improve the quality of training samples.

[0128] Preferably, after obtaining the vector-form labels for the impact crater micromorphology subclasses, the method further includes:

[0129] The vector-form impact crater micromorphology subclasses are categorically encoded and rasterized. One-hot encoding is then used to convert the rasterized impact crater micromorphology subclasses into multi-channel data.

[0130] For the label data Y of the impact crater micromorphology subclass, it corresponds to the category value. In this embodiment, there are four target categories: central peak, crater bottom, crater wall, and crater edge. Including the background pixels, there are a total of five categories. In order to input the label data into the deep learning model, it is necessary to process the vector-form label data by category encoding, rasterization, and one-hot encoding.

[0131] For example, category encoding includes: converting the central peak, pit bottom, pit wall, and pit edge into numerical values ​​and encoding them as 1, 2, 3, and 4 respectively, while encoding the remaining pixels as 0; rasterization includes: rasterizing the vector data at the same resolution scale as the corresponding raster data to obtain the corresponding label raster, with the raster value ranging from 0 to 4; one-hot encoding includes: semantic segmentation is essentially a pixel-level classification task, while the research in this embodiment corresponds to a multi-classification task, so the data after rasterization in the previous step is one-hot encoded to convert it from single-channel data to 5-channel data, where the corresponding data in each channel has only 0 and 1 values, where 1 represents a pixel of this category and 0 represents a pixel of a different category.

[0132] Preferably, the deep learning model is constructed as follows: based on the deep convolutional network U-Net, the input channels are modified to be multi-channel, and the original U-Net's Upsampling module is modified to a Max Unpooling module.

[0133] Traditional U-Net structures have only one input channel. However, in lunar terrain research, lunar terrain features are a complex combination, and data based on a single channel cannot accurately represent the characteristics of terrain types. For example, image data can reflect the texture features of terrain but cannot represent the topographic features related to elevation data. Furthermore, automatic classification of lunar terrain requires input from different datasets to represent the multiple features of terrain types. Therefore, in this embodiment, the traditional U-Net structure needs to be improved to construct a U-Net network with multiple input channels. For ease of description, the improved U-Net network is referred to as Crater3-Net.

[0134] like Figure 7 As shown, Crater3-Net's improvements over the traditional U-Net network are:

[0135] (1) Modify the input channel to multi-channel (e.g., 3 channels) so that the network can make full use of the characteristics of the lunar DEM and its derived data;

[0136] (2) Modify the original Upsampling module in the traditional U-Net network to the Max Unpooling module to adapt to the segmentation features of the micro-morphological landforms of lunar impact craters, while keeping the original main feature positions unchanged.

[0137] For example, the execution flow of the Crater3-Net network is as follows:

[0138] (1) The input data (i.e. the best combination of features) is processed and enhanced by image processing and is converted into a multi-channel tensor, such as a tensor with 3 channels and a height and width of 512.

[0139] (2) Perform convolution and pooling operations on the generated multi-channel tensor. Max pooling is used with a stride of 2, meaning that the feature map size is halved after each pooling operation. Meanwhile, to maintain the same amount of information, the number of channels is doubled after the convolution operation. Add BatchNorm and PReLU activation function layers after each convolutional layer.

[0140] (3) Repeat step (2) above until the size of the feature map becomes 32*32, at which point the number of channels is 256.

[0141] It should be noted that during the max pooling operation in steps (2) and (3), the position of the maximum value should be recorded.

[0142] (4) Perform maximum unpooling on the results obtained above to gradually restore the size of the feature map of the left half of the network, and then perform channel splicing on the feature maps of the corresponding sizes.

[0143] (5) Obtain the final output result with 5 channels through convolution, which has a size of 512*512. During training, use this output result and the label data to calculate the loss, and then perform backpropagation of the gradient to update the network parameters; during inference, perform the maximum value index operation in the channel dimension to obtain the final segmentation result.

[0144] Considering the complexity and diversity of impact crater morphology, some impact craters may lack a central peak, or some may lack a bottom, and one-hot encoding may result in sample imbalance (the proportion of target pixels is less than that of background pixels). Therefore, DiceLoss is chosen as the loss function during deep learning model training. The specific calculation formula for DiceLoss can be found using existing techniques and will not be elaborated here.

[0145] To better train the model and enhance its range capabilities, the following image processing and enhancement methods are used during model training:

[0146] (1) Size scaling: To facilitate network training, U-Net-like networks generally require a fixed image size during training. Considering the data of the impact craters selected in this study, the raster data within the corresponding range of each impact crater is interpolated to a size of 512*512. To avoid introducing new noise, the nearest neighbor interpolation method is used in the interpolation process.

[0147] (2) Horizontal flip: During training, the input data and the corresponding label data are horizontally flipped with a probability of 0.5.

[0148] (3) Vertical Flip: During training, the input data and corresponding label data are horizontally flipped with a probability of 0.5. In particular, when using the model for inference, only the image size is scaled, and the size and coordinate information of the original data are recorded before the data is input into the model. Then, this information is used to restore the output data of the model to ensure that it has the same size and geographical location as the original data, so as to perform subsequent vectorization and mapping.

[0149] Furthermore, the Crater3-Net network employs a dynamic learning rate adjustment strategy during training. Specifically, the initial learning rate is set to 0.001, and a Reduce LR On Plateau learning rate adjustment strategy is used. This strategy dynamically adjusts the model's learning rate based on the changes in the validation set loss during training. When the loss on the validation set no longer changes for N consecutive epochs, the learning rate is reduced by a factor of S. During model training, N is set to 3 and S is set to 0.5.

[0150] To verify the performance of the Crater3-Net network provided in this embodiment, four models were trained using the same model parameter settings, and the results were evaluated and compared:

[0151] (1) Crater3-Net: The model is obtained based on the network architecture above. The input channels include three data channels: elevation data, plane curvature and slope. The upsampling module adopts Max Unpooling.

[0152] (2) Crater1-Net: The model is obtained based on the network architecture described above, but the input channel only retains elevation data, and the upsampling module uses Max Unpooling.

[0153] (3) U3-Net: The original U-Net architecture is modified to have 3 input channels, including three data channels: elevation data, plane curvature and slope data. Upsampling uses the nearest neighbor interpolation method.

[0154] (4) U1-Net: The original U-Net architecture has three input channels: elevation data, plane curvature and slope data. Upsampling uses the nearest neighbor interpolation method.

[0155] The mean Intersection over Union (mIoU) and mean pixel accuracy (MPA) of the four models were calculated as evaluation metrics. During training, the learning rate change curves, loss change curves on different training sets and different validation sets, and change curves of each evaluation metric were plotted for the four models.

[0156] The results show that the timing of the first learning rate decay varies among the different models. Crater1-Net decays the fastest, followed by U1-Net, while Crater3-Net decays the slowest. The different timing of learning rate decay under the same random seed indicates differences in learning rates among the models. After training begins, the loss function values ​​of all four models decrease rapidly with the increase in the number of data iterations. After approximately 50 epochs, all four models have essentially converged. The evaluation metrics mIOU and MPA show differences in the final performance of the four models: after convergence, the mIOU on the validation set is close to 0.8, with Crater3-Net performing best, followed by U3-Net, then U1-Net, and finally Crater1-Net. The MPA metric also shows similar results, with the MPA values ​​of each model around 0.95 after convergence. Crater3-Net performs best, followed by U3-Net, then U1-Net and Crater1-Net. Looking at the changes in the mIOU MPA metric on the validation set, using elevation data, plane curvature, and slope data simultaneously as input data yielded better results than using only elevation data. From another perspective, when the input data has 3 channels, in the decoder part, using Max Unpooling for upsampling is better than using nearest neighbor interpolation. However, when the input data has a single channel, nearest neighbor interpolation is slightly better than Max Unpooling. Overall, however, the Crater3-Net model shows the best segmentation performance in terms of both mIOU and MPA.

[0157] Furthermore, the four trained models were used to segment the micromorphological subclasses of impact craters on the test dataset, and their evaluation metrics were statistically analyzed, resulting in the following table:

[0158] Table 1 Comparison of segmentation accuracy

[0159]

[0160] As can be seen from the table above, the method provided in this embodiment has the highest segmentation accuracy on the test dataset, which is close to 0.8.

[0161] In addition, to further verify the practical effect of this embodiment, impact craters of different morphologies (such as Cauchy impact craters and Al-Biruni impact craters) were selected for comparison of segmentation effects. The segmentation results were evaluated from both qualitative and quantitative perspectives, as shown in Table 2.

[0162] Table 2 Comparison of segmentation accuracy for micromorphological subclasses of impact craters of different morphological types

[0163]

[0164] Figure 8 The results of the Al-Biruni impact crater micromorphological subclass segmentation are shown. (Combined with Table 2 and...) Figure 8 The results show that the Crater3-Net model achieves good overall segmentation of micromorphological subclasses for both the Cauchy and Al-Biruni impact craters. The segmentation accuracy of the Cauchy impact crater's bottom and walls is high, with an MPA of up to 0.9, while the segmentation accuracy of the crater edge is relatively low, at about 0.84. The segmentation results of the Al-Biruni impact crater include the bottom, walls, and edge, with cPAs of 0.99, 0.91, and 0.96, respectively, and an MPA close to 0.96.

[0165] The transitional impact crater Hoffmeister N, located in the highlands of the far side of the Moon, formed during the Nectarian period. A Nectarian impact crater, Hoffmeister, exists to its northern part, and the two craters partially overlap. When segmenting using the Crater3-Net model, the model can effectively segment the micromorphological subclasses of impact craters Hoffmeister N and Hoffmeister, and it also cuts out the overlap between them. The segmentation accuracy for the crater floor, walls, and rim is high, but some missegmentation issues exist. For example, when an impact crater overlaps with Hoffmeister, the model segments it as a central peak.

[0166] Jackson crater is a typical complex impact crater in highland areas. Figure 9 The image shows the segmentation results for the Jackson impact crater micromorphology subclass. It is clear from the image that the segmentation effect of the crater wall and rim is better than that of the crater bottom and central peak in the micromorphology subclass. Combined with Table 2, it can be seen that for the segmentation of complex impact craters, the segmentation accuracy of the crater wall and rim is higher, with a cPA around 0.9, while the segmentation accuracy of the crater wall and central peak is lower, with cPAs of 0.70 and 0.63, respectively. Furthermore, this paper converts the segmentation results into vector data and compares them with manually drawn sample data. Based on WAC and slope data, the vector data of the Jackson impact crater segmentation results more accurately fits the boundaries of the crater bottom and central peak. From another perspective, when the manually labeled samples have large errors, the Crater3-Net model constructed in this embodiment can, to some extent, correct the errors of manual drawing, thereby obtaining better boundary results for the impact crater micromorphology subclass.

[0167] The trained Crater3-Net model (target model) was used to automatically extract the micromorphological subclasses of impact craters across the entire month, and the extraction results were obtained.

[0168] In this embodiment, the Crater3-Net model is used to obtain lunar impact craters with a direct length greater than 5km, and these craters are segmented. Based on the segmentation results, a micromorphological subclass database of impact craters is obtained, and the landform of lunar impact craters is mapped.

[0169] The following reference Figure 3 A further illustrative description is provided of the process for automatically extracting micromorphological subclasses of lunar impact craters.

[0170] Figure 3 This is a schematic diagram illustrating the process of automatically extracting micromorphological subclasses of lunar impact craters according to some embodiments of this application. Figure 3 As shown, the automatic extraction of micromorphological subclasses of lunar impact craters can be performed according to the following steps:

[0171] Dataset construction includes selecting input data, extracting terrain factors to select raster datasets, screening images to determine the best feature combination, labeling samples, and then constructing a subclass dataset of impact crater micromorphology.

[0172] Classification data preparation: The impact crater micromorphology subclass dataset is divided into training data, test data, and validation data.

[0173] Data preprocessing includes data resampling, image cropping, and image enhancement. Here, "image" refers to raster data within the impact crater micromorphology subclass data.

[0174] Model construction: A multi-channel Crater3-Net network is constructed based on the U-Net network.

[0175] Training process: Input the training data into the Crater3-Net semantic segmentation network, calculate the loss function, and adjust the learning rate until the model converges to obtain the optimal parameters of the Crater3-Net model.

[0176] Model evaluation: Based on test and validation data, the converged Crater3-Net model is compared and its accuracy is evaluated with Crater1-Net, U3-Net, and U1-Net models. Evaluation metrics include mIOU and MPA.

[0177] Through the above steps, the automatic extraction and segmentation of micromorphological sub-types of lunar impact craters are achieved. Based on this, a map of the entire lunar landscape is completed.

[0178] Taking the impact landform mapping of the Chang'e 5 landing area as an example, there are 197 types of impact crater landforms classified into micromorphological subclasses, including the bottom of impact craters with small undulations at low altitudes in lunar maria, the walls of impact craters with large undulations at extremely high altitudes in highlands, and the edge of impact craters with small undulations at extremely low altitudes in the Aitken region of Antarctica.

[0179] In summary, impact craters are a typical lunar landform. Due to the diverse and complex shapes of impact craters, boundary extraction requires expert knowledge. Currently, automatic extraction and identification of impact craters often only determine their location and diameter, with limited research on the actual boundary extraction of impact crater micromorphological subclasses. To address this technological gap, this embodiment uses various terrain-derived factors as candidate features in the training dataset construction phase to determine the optimal feature combination for the model. This provides high-quality training data, ensuring the model can learn the features of impact crater micromorphological subclasses. Based on U-Net, a deep learning network segmentation model Crater3-Net for impact crater micromorphological subclasses is constructed. This model automatically segments and extracts impact craters with diameters greater than 5 km across the entire lunar surface, forming a database of impact crater micromorphological subclasses with diameters greater than 5 km. This alleviates the current problem of insufficient efficiency due to reliance on manual extraction of impact crater micromorphological subclasses, providing the possibility for high-precision rapid mapping of lunar impact landforms across the entire moon, and laying the foundation for further research on lunar topography.

[0180] Based on the same inventive concept, this embodiment also provides an automatic mapping system for the micromorphological topography of lunar impact craters, the system comprising:

[0181] The data acquisition unit is configured to acquire a candidate raster data set; the candidate raster data set includes lunar digital terrain data, lunar image data, and terrain-derived data.

[0182] The data filtering unit is configured to filter the candidate raster data set to determine the optimal combination of features;

[0183] The training set building unit is configured to combine the best feature combination with the labels of the impact crater micromorphology subclasses to build the impact crater micromorphology subclass dataset.

[0184] The model training unit is configured to input the impact crater micromorphology subclass dataset into a pre-built deep learning model for training to obtain the target model;

[0185] The automatic extraction unit is configured to automatically extract the micromorphological subclasses of impact craters across the entire lunar range using the target model, and then complete the mapping of the impact crater topography across the entire lunar range based on the extraction results.

[0186] The micromorphological subclasses of impact craters are divided into four subclass units: central peak, crater bottom, crater wall, and crater edge.

[0187] The automatic mapping system for lunar impact crater micromorphology provided in this embodiment can realize the steps and processes of the automatic mapping method for lunar impact crater micromorphology provided in any of the above embodiments, and achieve the same technical effect, which will not be described in detail here.

[0188] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An automatic mapping method for the micromorphological topography of lunar impact craters, characterized in that, include: Obtain a candidate raster data set; wherein, the candidate raster data set includes lunar digital terrain data, lunar image data, and terrain-derived data; the terrain-derived data includes multiple terrain-derived factors; The candidate raster data set is filtered to determine the optimal feature combination; The optimal feature combination is combined with the labels of the impact crater micromorphology subclasses to construct an impact crater micromorphology subclass dataset. The impact crater micromorphology subclass dataset is input into a pre-built deep learning model for training to obtain the target model; The target model is used to automatically extract the micromorphological subclasses of impact craters across the entire lunar area, and the extraction results are obtained; then, based on the extraction results, the entire lunar topography is mapped. The impact crater micromorphology subclasses are divided into four subclass unit types: central peak, crater bottom, crater wall, and crater edge. The candidate raster data set is filtered to determine the optimal feature combination, including: Normalization preprocessing is performed on each candidate raster data in the candidate raster data set, and feature value distribution maps of each candidate raster data in different impact crater micromorphology subclasses are plotted based on the results of normalization preprocessing. Based on the feature value distribution map, the differences between different impact crater micromorphology subclasses on the same candidate raster data are analyzed to exclude candidate raster data with differences less than a preset threshold, and the first screening result is obtained. Correlation analysis was used to analyze the correlation between various terrain-derived factors in the first screening results; Based on the results of the correlation analysis, a second screening is performed on the first screening results; including: using correlation analysis, comprehensively considering the feature representation ability of different candidate feature data and the correlation strength between different terrain-derived data, thereby eliminating terrain-derived data with high overlap and retaining data with strong feature expression ability. The optimal feature combination is the feature combination obtained after secondary screening; the optimal feature combination is: lunar DEM data, slope, and plane curvature.

2. The method according to claim 1, characterized in that, The terrain-derived factors include macro-terrain factors, which include lunar surface topographic relief. The steps for generating the lunar surface topographic relief are as follows: Based on the lunar digital terrain data, the mean variable point method is used to analyze the lunar surface relief and determine the optimal window for calculating the lunar surface relief; then, based on the optimal window for lunar surface relief, the window analysis method is used to calculate the terrain relief.

3. The method according to claim 2, characterized in that, The macroscopic terrain factors also include: roughness, The roughness generation steps are as follows: Based on the lunar digital terrain data, the surface roughness of the moon is calculated using the root mean square elevation method.

4. The method according to claim 1, characterized in that, The topographic derivative factors include micro-topographic factors, which include slope, aspect, and curvature; The slope, aspect, and curvature are all calculated based on the lunar digital terrain data.

5. The method according to claim 1, characterized in that, The tags for the impact crater micromorphology subclasses are generated through the following steps: Acquire the data of the first impact crater surface, the second impact crater surface, and the third impact crater surface; The first impact crater surface data, the second impact crater surface data, and the third impact crater surface data are matched, and combined with the lunar digital terrain data, slope data, and mountain shadow data, the impact crater micromorphology subclasses are drawn and refined to obtain vector-form labels for the impact crater micromorphology subclasses. Among them, the first impact crater surface data, the second impact crater surface data, and the third impact crater surface data are lunar geological maps compiled by different teams based on data from different sensors.

6. The method according to claim 5, characterized in that, After obtaining the labels for the vector-form impact crater micromorphology subclasses, the following is also included: The vector-form impact crater micromorphology subclasses are categorically encoded and rasterized. One-hot encoding is then used to convert the rasterized impact crater micromorphology subclasses into multi-channel data.

7. The method according to claim 1, characterized in that, The deep learning model is constructed as follows: based on the deep convolutional network U-Net, the input channels are modified to be multi-channel, and the original U-Net's Upsampling module is modified to a Max Unpooling module.

8. An automatic mapping system for micromorphological topography of lunar impact craters, characterized in that, include: The data acquisition unit is configured to acquire a candidate raster data set; the candidate raster data set includes lunar digital terrain data, lunar image data, and terrain-derived data; the terrain-derived data includes multiple terrain-derived factors. The data filtering unit is configured to filter the candidate raster data set to determine the optimal feature combination; The training set construction unit is configured to combine the optimal feature combination with the labels of the impact crater micromorphology subclasses to construct an impact crater micromorphology subclass dataset. The model training unit is configured to input the impact crater micromorphology subclass dataset into a pre-built deep learning model for training to obtain the target model; The automatic extraction unit is configured to use the target model to automatically extract the micromorphological subclasses of impact craters across the entire lunar range, and obtain the extraction results; then, based on the extraction results, to complete the mapping of the impact crater topography across the entire lunar range. The impact crater micromorphology subclasses are divided into four subclass unit types: central peak, crater bottom, crater wall, and crater edge. The candidate raster data set is filtered to determine the optimal feature combination, including: Normalization preprocessing is performed on each candidate raster data in the candidate raster data set, and feature value distribution maps of each candidate raster data in different impact crater micromorphology subclasses are plotted based on the results of normalization preprocessing. Based on the feature value distribution map, the differences between different impact crater micromorphology subclasses on the same candidate raster data are analyzed to exclude candidate raster data with differences less than a preset threshold, and the first screening result is obtained. Correlation analysis was used to analyze the correlation between various terrain-derived factors in the first screening results; Based on the results of the correlation analysis, a second screening is performed on the first screening results; including: using correlation analysis, comprehensively considering the feature representation ability of different candidate feature data and the correlation strength between different terrain-derived data, thereby eliminating terrain-derived data with high overlap and retaining data with strong feature expression ability. The optimal feature combination is the feature combination obtained after secondary screening; the optimal feature combination is: lunar DEM data, slope, and plane curvature.

Images