Asteroid multi-stage ground object identification and surface feature analysis method

Abstract

This invention provides a method for multi-stage ground feature identification and surface feature analysis of asteroids, comprising: predicting detection boxes for asteroids, craters, and artificial landmarks from distant asteroid images using natural language prompts; performing multi-target accurate detection using a hierarchical filtering optimization method based on the prediction results; drawing a distant-perception semantic map to identify regions of interest by fusing the segmentation method based on the detection box prompts and the segmentation method generated by automatic masking, based on the filtered results; quickly marking sandy areas in the detailed survey images of the identified regions of interest based on a sampling-first classification strategy; and identifying sampleable weathering layers in the marked sandy areas through granularity analysis, supporting the construction of a global sampleability map. This invention covers a complete technology chain from distant perception to detailed analysis in the detailed survey phase, improving the intelligent identification and analysis capabilities of ground features in asteroid exploration missions.

Description

Technical Field

[0001] This invention belongs to the field of asteroid surface feature identification, specifically involving a method for multi-stage asteroid feature identification and surface feature analysis. Background Technology

[0002] Asteroid surface feature identification is a remote sensing and computer vision technology designed to automatically identify and analyze surface features on asteroids through the processing and analysis of their images. This technology relies on high-resolution image data and can provide scientists with accurate information on surface features such as craters, rocks, sand, and other surface characteristics during asteroid exploration missions, thus providing fundamental data support for subsequent exploration, sampling, and landing missions.

[0003] In previous exploration missions, traditional ground feature identification methods relied mainly on manual or semi-automatic techniques, which were inefficient and unable to meet the real-time and high-accuracy requirements of deep space environments. With the advancement of deep learning technology, researchers have achieved accurate detection and segmentation of boulders on asteroid surfaces by combining convolutional neural networks with large amounts of labeled data. However, these methods are highly dependent on labeled data, especially in multi-object missions, where the generalization ability of the models is often limited due to differences in terrain features and image resolution.

[0004] Large-scale pre-trained models (such as Grounding DINO and SAM) leverage prior knowledge to reduce reliance on labeled data and significantly improve generalization capabilities through powerful detection and segmentation abilities. However, challenges remain in the unique environment of asteroid exploration, such as the confusion between different land cover categories, especially in craters and rocky areas where ambiguous semantic boundaries hinder detection. Furthermore, rocky areas, being densely populated and numerous land cover features, often cannot be fully covered by detection boxes generated from text prompts, resulting in insufficient recall. Therefore, improving and optimizing asteroid surface feature identification remains crucial.

[0005] The shortcomings of existing technologies for intelligent identification and analysis of asteroid surface features are mainly reflected in the following aspects: (1) There are many targets to be identified, with a wide variety of categories and significant differences in near and far scales. A single detection or segmentation method is difficult to meet all the needs; (2) Without relying on large-scale labeled data, different types of features are easily confused, and the recall of densely distributed rocks is insufficient, resulting in a decrease in identification accuracy; (3) The identification of sampleable weathering layers still relies on manual labor, and the intelligent detection process needs to be further improved. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a multi-stage method for asteroid surface feature identification and analysis. It designs a global perception technology for long-range feature identification and a detailed surface feature analysis technology for specific sections, enabling the step-by-step analysis of asteroid surface features and ultimately supporting the construction of a global sampleability map. Through the organic combination of long-range global perception and close-range detailed analysis, this framework can identify scientific targets and assist probes in actively sensing and selecting sampleable areas.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for multi-stage ground feature identification and surface feature analysis of asteroids includes the following steps:

[0009] Step 1: Using natural language prompts, predict the bounding boxes for asteroids, craters, and artificial landmarks in distant asteroid images;

[0010] Step 2: Based on the prediction results, a hierarchical filtering optimization method is used to perform accurate multi-target detection;

[0011] Step 3: Combine the segmentation method prompted by the detection box and the segmentation method generated by automatic masking to draw a long-distance perceptual semantic map to identify regions of interest;

[0012] Step 4: For the identified region of interest, based on a sampling-first classification strategy, quickly label the sandy area image in the detailed segment image of the region of interest;

[0013] Step 5: Based on the marked sandy land images, identify the sampleable weathered layers through grain size analysis and construct a global sampleability map.

[0014] The advantages of this invention compared to the prior art are as follows:

[0015] (1) This invention proposes a sampling-driven multi-stage asteroid surface feature identification and analysis technology framework, covering the complete technology chain from long-distance perception to close-range detailed analysis. The framework focuses on improving detection efficiency and sampling success rate, and realizes intelligent identification and analysis of asteroid surface features through the synergistic application of key technologies.

[0016] (2) This invention optimizes the application of large-scale training models in deep space environment. Based on the hierarchical filtering optimization method, it solves the problems of multi-target overlap and semantic ambiguity, significantly improves the adaptability of the model in deep space scene, and realizes accurate detection of scientific targets.

[0017] (3) This invention takes into account the computation and communication costs of the detector and proposes a sampling-first classification strategy to complete the rapid screening of regions locally on the detector and use the screening results as input for fine segmentation tasks, thus concentrating computational resources on high-value target regions.

[0018] In summary, the method employed in this invention is simple in principle and can achieve the goal of multi-stage ground feature identification and surface feature analysis of asteroids under long-distance and detailed image input. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for multi-stage ground feature identification and surface feature analysis of asteroids according to the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0021] like Figure 1 As shown, the present invention provides a method for multi-stage asteroid feature identification and surface characteristic analysis, comprising the following steps:

[0022] Step 1: Using natural language prompts, predict bounding boxes for asteroids, craters, and artificial landmarks in distant asteroid images, including;

[0023] The input for the long-range perception stage is a long-range image of the asteroid. The Grounding DINO model, based on a dual encoder-single decoder architecture, is used to detect specified targets such as asteroids, craters, and artificial landmarks based on natural language prompts.

[0024] First, a distant image of the asteroid is input into the model, and image and text features are extracted using the image backbone and text backbone, respectively.

[0025] Subsequently, the extracted image and text features are fused across modally through a feature enhancement module to align image and text features and obtain cross-modal features. These cross-modal features are then used by a language-guided query selection module to extract query information from the image features, enabling the decoder to further perform object detection and bounding box refinement. Through multimodal learning techniques, image and text features are mapped to a unified feature space. For distant images of asteroids, "Whiteball, Craterson asteroid." is selected as the text cue, and the model predicts a series of bounding boxes for the asteroid, craters, and artificial landmarks.

[0026] Step 2: Based on the prediction results, a hierarchical filtering optimization method is used to perform accurate multi-target detection, including:

[0027] In long-range imagery, asteroids, craters, and artificial landmarks exhibit significant uncertainties in semantic features and spatial distribution. To address the issues of multi-target overlap and semantic ambiguity, a triple filtering strategy based on semantic confidence, area constraint, and spatial overlap is proposed.

[0028] First, the detection bounding boxes with the highest confidence levels for both asteroids and artificial markers are selected as the recognition results:

[0029] ,

[0030] ,

[0031] in, and These are the index numbers for the asteroid detection frame and the artificial marker detection frame. and These represent the index numbers of the successfully detected asteroid detection frames and the artificial marker detection frames, respectively. and These represent the confidence levels of the bounding boxes for asteroids and artificial markers, respectively. and These represent the sets of detection box indices for asteroids and artificial markers, respectively.

[0032] Target category conflicts are reduced through area constraints and multi-level confidence filtering, and the area of ​​the asteroid is calculated based on the detection box. As a reference area, a set of targets with crater areas smaller than one-third of the asteroid frame area was selected. :

[0033] ,

[0034] in, This indicates the index number of the crater detection frame. This indicates the confidence level of the target bounding box in the crater. This represents the maximum confidence level of the target bounding box in the crater. This means that both conditions are met simultaneously.

[0035] Finally, tag overlap detection is performed to determine the spatial relationship between impact craters and artificial markers, and overlapping impact crater targets are removed.

[0036] ,

[0037] in, and This represents the boundary coordinates of the corresponding detection box. It is the crater detection frame index number that has passed area filtering. If it is true, then the target It will be removed.

[0038] Step 3: Combine the segmentation method provided by the detection box prompts and the segmentation method generated by automatic masking to draw a long-distance perceptual semantic map to identify regions of interest, including;

[0039] The segmentation method for the detection box prompts inputs the detection boxes generated in step 2 as prompt conditions into the SAM model. A visual transformer performs global feature extraction on the image, and the extracted features are mapped to a high-dimensional embedding space, forming a global feature representation of the entire image. The prompt boxes are embedded into the same high-dimensional embedding space, and an attention mechanism is used to locate the target region within the box in the global features.

[0040] ,

[0041] in, The segmentation mask representing the target. The bounding box indicates the target, img represents the distant image of the asteroid, t corresponds to the target index of the asteroid, crater, and artificial marker, and the semantic target mask set is represented as:

[0042] ,

[0043] The automatic mask generation segmentation method utilizes the automatic_mask_generator function of SAM, which does not require detection box prompts and directly generates all possible ground feature segmentation masks through global analysis of image content. :

[0044] ,

[0045] The automatically generated mask contains all possible target regions, which are then filtered in conjunction with the mask of the key target. For each automatically generated mask... Calculate its relationship with the key target mask set. For areas with high mask overlap that do not belong to independent rock regions, their index k is added to the set to be removed. The mask after removal is the segmentation result of the rock region, represented as:

[0046] ,

[0047] After segmenting various land features, a multi-level semantic result overlay method is applied to construct a long-distance perceptual semantic map and identify regions of interest.

[0048] Step 4: For the identified region of interest, based on a sampling-first classification strategy, quickly label the sandy area image in the detailed segment image of the region of interest, including:

[0049] After standardizing the detailed segment image of the region of interest, it is fed into the ResNet network to perform an image-level classification task and output a category score. The categories include hard surface, rock accumulation area and sandy land. The softmax function is used to transform it into a probability distribution, and the category with the highest probability is selected as the final prediction result.

[0050] Step 5: For the marked sandy land images, identify sampleable weathering layers through graininess analysis and construct a global sampleability map, including:

[0051] The SAM (Sand Image Processing) method is used to perform global segmentation on the marked sandy land image, generating initial masks for all land features. These initial masks are then merged to form a single global mask, denoted as [Mask Name]. ;

[0052] Utilizing image spatial resolution Unit: mm / pixel. This calculates the pixel scale of the particles; approximately 5cm corresponds to a pixel size of:

[0053] ,

[0054] Morphological etching was used to remove particles smaller than 5 cm, with a core size of [missing value]. This forms the etched mask. :

[0055] ,

[0056] in, This indicates a morphological etching operation;

[0057] The etched mask is processed using a dilation operation, with the size of the dilation kernel set according to the camera resolution and grain spacing. The expanded mask is represented as :

[0058] ,

[0059] in, Indicates morphological dilation operation;

[0060] A morphological erosion operation is performed again, and then the mask is inverted to obtain the distribution area of ​​the sampleable weathering layer. By combining the orbital information with the long-distance sensing semantic map for registration, a global sampleability map is constructed.

[0061] Therefore, this invention can effectively improve the accuracy of ground feature identification and surface feature analysis in both long-range and detailed survey stages. In the long-range perception stage, the hierarchical filtering optimization method solves the problems of semantic ambiguity and multi-label overlap, achieving efficient screening and contour extraction of ground features. In the detailed survey stage, the combination of the sampling area-first classification strategy and the sampleable weathering layer area analysis method not only optimizes the resource allocation of the detector but also provides high-confidence regional support for the sampling task, achieving high-precision multi-stage ground feature identification and surface feature analysis for asteroids.

[0062] Contents not described in detail in this specification are common knowledge to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications will be readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the inventive concept are protected.

Claims

1. A method for multi-stage ground feature identification and surface feature analysis of asteroids, characterized in that, Includes the following steps: Step 1: Using natural language prompts, predict the bounding boxes for asteroids, craters, and artificial landmarks in distant asteroid images; Step 2: Based on the prediction results, a hierarchical filtering optimization method is used for accurate multi-target detection; this includes a triple filtering strategy based on semantic confidence, area constraint, and spatial overlap. First, the detection bounding boxes with the highest confidence levels for both asteroids and artificial markers are selected as the recognition results: ， ， in, and These are the index numbers for the asteroid detection frame and the artificial marker detection frame. and These represent the index numbers of the successfully detected asteroid detection frames and the artificial marker detection frames, respectively. and These represent the confidence levels of the bounding boxes for asteroids and artificial markers, respectively. and These represent the sets of detection box indices for asteroids and artificial markers, respectively. Calculate the area of ​​the asteroid based on the detection frame. As a reference area, a set of targets with crater areas smaller than one-third of the asteroid frame area was selected. : ， in, This indicates the index number of the crater detection frame. This indicates the confidence level of the target bounding box in the crater. This represents the maximum confidence level of the target bounding box in the crater. This means that both conditions are met simultaneously; Finally, tag overlap detection is performed to determine the spatial relationship between impact craters and artificial markers, and overlapping impact crater targets are removed. ， in, and This represents the boundary coordinates of the corresponding detection box. It is the index number of the crater detection frame that has passed area filtering. If it is true, then the target It will be removed; Step 3: Combine the segmentation method prompted by the detection box and the segmentation method generated by automatic masking to draw a long-distance perceptual semantic map to identify regions of interest; Step 4: For the identified region of interest, based on a sampling-first classification strategy, quickly label the sandy area image in the detailed segment image of the region of interest; Step 5: For the marked sandy land images, identify the sampleable weathering layer through grain size analysis and construct a global sampleability map.

2. The method for multi-stage ground feature identification and surface feature analysis of asteroids according to claim 1, characterized in that, Step 1 includes: The distant image of the asteroid is input into the Grounding DINO model based on a dual encoder-single decoder architecture, which extracts image and text features using the image backbone and text backbone, respectively. The extracted image and text features are fused across modally through a feature enhancement module to align the image and text features and obtain cross-modal features. The cross-modal features are obtained by a language-guided query selection module that extracts query information from image features and then decodes it to obtain a series of detection boxes for asteroids, craters and artificial landmarks.

3. The method for multi-stage ground feature identification and surface feature analysis of asteroids according to claim 1, characterized in that, Step 3 includes: The segmentation method for the detection box prompts inputs the detection boxes generated in step 2 as prompt conditions into the SAM model. A visual transformer performs global feature extraction on the image, and the extracted features are mapped to a high-dimensional embedding space, forming a global feature representation of the entire image. The prompt boxes are embedded into the same high-dimensional embedding space, and an attention mechanism is used to locate the target region within the box in the global features. ， in, The segmentation mask representing the target. The bounding box indicates the target, img represents the distant image of the asteroid, t corresponds to the target index of the asteroid, crater, and artificial marker, and the semantic target mask set is represented as: ， The automatic mask generation segmentation method utilizes the automatic_mask_generator function of SAM, which does not require detection box prompts and directly generates all possible ground feature segmentation masks through global analysis of image content. : ， For each automatically generated mask Calculate its relationship with the key target mask set. For areas with high mask overlap that do not belong to independent rock regions, their index k is added to the set to be removed. The mask after removal is the segmentation result of the rock region, represented as: ， After segmenting various land features, a multi-level semantic result overlay method is applied to construct a long-distance perceptual semantic map and identify regions of interest.

4. The method for multi-stage ground feature identification and surface feature analysis of asteroids according to claim 1, characterized in that, Step 4 includes: After standardizing the detailed segment image of the region of interest, it is fed into the ResNet network to perform an image-level classification task and output a category score. The categories include hard surface, rock accumulation area and sandy land. The softmax function is used to transform it into a probability distribution, and the category with the highest probability is selected as the final prediction result.

5. The method for multi-stage ground feature identification and surface feature analysis of asteroids according to claim 1, characterized in that, Step 5 includes: The SAM (Sand Image Processing) method is used to perform global segmentation on the marked sandy land image, generating initial masks for all land features. These initial masks are then merged to form a single global mask, denoted as [Mask Name]. ; Utilizing image spatial resolution Unit: mm / pixel. Calculates the pixel scale of the particles; the pixel size corresponding to 5cm is: ， Morphological etching was used to remove particles smaller than 5 cm, with a core size of [missing value]. This forms the etched mask. : ， in, This indicates a morphological etching operation; The etched mask is processed using a dilation operation, with the size of the dilation kernel set according to the camera resolution and grain spacing. The expanded mask is represented as : ， in, Indicates morphological dilation operation; A morphological erosion operation is performed again, and then the mask is inverted to obtain the distribution area of ​​the sampleable weathering layer. By combining the orbital information with the long-distance sensing semantic map for registration, a global sampleability map is constructed.

Images