Intelligent sampling method and system for scientific interest targets on planetary surfaces

Abstract

The application relates to the field of deep space exploration and planetary surface scientific research, in particular to a planetary surface scientific interest target intelligent sampling method and system. The method comprises the following steps: step 1: a rover end acquires a planetary surface visible light image, autonomously identifies a potential scientific interest target through a target detection model and determines a spatial position, inputs spectral data of a target region into a spectral classification model for identification, executes step 2 when a confidence degree is lower than a preset threshold, otherwise, outputs a category label and executes step 3; step 2: a low-confidence-degree target is marked as a potential new category target, image, spectral and spatial position information of the target are transmitted to a ground center, a classification result is obtained through manual marking, an incremental learning training is conducted on the spectral classification model, adaptive optimization and new and old category compatible identification are realized, and the updated model is redeployed to the rover end; and step 3: according to the target classification result and the spatial position information, the rover is controlled to complete scientific interest target sampling operation.

Description

Technical Field

[0001] This application relates to the fields of deep space exploration and planetary surface science research, and in particular to an intelligent sampling method and system for planetary surface science targets of interest. Background Technology

[0002] Current deep space exploration missions, such as lunar and Martian surface exploration, typically employ a "space-ground coordinated" scientific research model. Specifically, after acquiring planetary surface image data, the rover transmits it back to the ground control center via a communication link. Ground scientists then interpret and select data based on their mission experience and expertise, and finally control the rover to conduct in-situ analysis or sample return missions via downlink commands. While this model is widely used in existing exploration missions, it heavily relies on ground-based human intervention and tracking and control resources, making it difficult to meet the demands of future long-term, highly autonomous deep space exploration missions.

[0003] The main reason for the above problems is that existing scientific exploration payloads do not have the ability to autonomously and intelligently select scientific interests, and cannot cope with the challenges of identifying new types of scientific targets on planetary surfaces, thus limiting the exploration efficiency and scientific output of deep space exploration missions.

[0004] The main difficulties encountered in solving the above problems are: first, the scientific payloads configured on existing mission platforms lack the ability to autonomously identify and screen scientific interests in complex environments; second, existing systems lack mechanisms for continuous learning and self-updating, making it difficult to achieve compatible identification and continuous understanding of new and old categories of scientific goals. Summary of the Invention

[0005] The purpose of this application is to overcome the above-mentioned deficiencies of the prior art, thereby providing an intelligent sampling method and system for planetary surface scientific interest targets.

[0006] To address the aforementioned technical problems, the present application provides an intelligent sampling method for planetary surface scientific interest targets, comprising: Step 1: Acquire visible light images of the planetary surface on the rover, autonomously identify potential targets of scientific interest using a target detection model and determine their spatial locations, collect spectral data of the target area, and input the spectral data into a spectral classification model for classification and identification; when the confidence level of the classification and identification result is lower than a preset threshold, proceed to Step 2; when the confidence level of the classification and identification result is higher than or equal to the preset threshold, output the category result label and proceed to Step 3; Step 2: Targets with confidence scores below a preset threshold are marked as potential new categories of scientific interest targets. The corresponding image, spectrum, and spatial location information of these targets are transmitted to the ground center. The new category is manually labeled to obtain the classification result label. Incremental learning training is performed on the spectral classification model to achieve adaptive optimization of the spectral classification model and compatible recognition of new and known categories of scientific interest targets. The spectral classification model updated by incremental learning is then redeployed to the rover. Step 3: Based on the classification results and spatial location information of the targets, control the rover to perform sampling operations for targets of scientific interest.

[0007] As an improvement to the above method, during the incremental learning training process of the spectral classification model, the feature replay strategy is used to maintain the spectral classification model's memory of features of known categories.

[0008] As an improvement to the above method, the feature replay strategy is implemented by constructing an adaptive component Gaussian mixture model for each known category during the basic training phase of the spectral classification model, thereby representing the feature distribution of each known category in a parameterized manner.

[0009] As an improvement to the above method, the spectral classification model includes a feature projection module for learning the mapping relationship between feature extractors of known categories and feature extractors of new categories. In step 2, the incremental learning training process of the spectral classification model includes: during incremental learning, known category feature data sampled from the Gaussian mixture model and manually labeled new category feature data are used together for training; during training, the feature projection module learns the mapping relationship between feature extractors of known categories and feature extractors of new categories, so that the feature space of known categories and the feature space of new categories are semantically aligned; based on the feature replay strategy and feature alignment results, the network parameters of the spectral classification model are updated, thereby enabling the spectral classification model to retain the memory of known category features while learning new category features.

[0010] As an improvement to the above method, the target detection model is trained based on a multi-source joint dataset, which includes images of scientific interest acquired from historical exploration missions and scientific interest target data generated by simulating planetary surface environments.

[0011] As an improvement to the above method, the multi-source joint dataset also includes hyperspectral data corresponding to images of scientific interest acquired in historical exploration missions; the spectral classification model is trained based on the hyperspectral data in the multi-source joint dataset during the basic training phase to identify scientific interest targets of known categories.

[0012] As an improvement to the above method, the visible light image includes an RGB image.

[0013] To achieve another objective of the present invention, the present invention also provides an intelligent sampling system for planetary surface scientific targets of interest, used to implement the above-described intelligent sampling method for planetary surface scientific targets of interest, comprising: The acquisition module is used to acquire visible light images of the planetary surface at the rover end, autonomously identify potential targets of scientific interest using a target detection model and determine the spatial location of the targets, acquire spectral data of the target area, and input the spectral data into a spectral classification model for classification and identification; A spectral classification model is used to identify the category of a target and output the classification result and the corresponding confidence score. The comparison module is used to compare the confidence level of the classification and identification results with a preset threshold. The labeling module is used to output classification result labels for targets with confidence levels higher than or equal to a preset threshold; it is used to mark targets with confidence levels lower than a preset threshold as potential new categories of scientific interest targets, and to transmit the image, spectral and spatial location information corresponding to the potential new categories of scientific interest targets to the ground center. The potential new categories of scientific interest targets are manually labeled to obtain the classification result labels of the new categories. The incremental learning module is used to incrementally train the spectral classification model, enabling adaptive optimization of the model and compatible recognition of new and known categories of scientific interest targets; it is also used to redeploy the incrementally updated spectral classification model to the rovers; and The sampling operation module controls the rover to perform sampling operations for scientifically interested targets based on the classification results and spatial location information of the targets.

[0014] The intelligent sampling method and system for scientific targets of interest on planetary surfaces provided in this application are based on visible light images captured by the panoramic camera of a rover. Using an intelligent target detection method, scientific targets on the planetary surface are autonomously detected and located. After acquiring the spectral information of the targets through an infrared hyperspectral instrument, multiple categories of scientific targets are classified and identified. When a new category of scientific target appears, the spectral classification model is incrementally learned to guide the rover to independently complete the intelligent sampling of points of interest. This solves the problems of heavy reliance on ground support, insufficient exploration autonomy, and insufficient environmental adaptability in existing technologies.

[0015] Compared with the prior art, the advantages of this application are: 1. It can autonomously identify and screen scientific points of interest at the rover end, and can select targets without relying on ground control commands, which significantly improves the automation and intelligence level of the detection process; 2. It can adaptively update the spectral classification model during the detection process, improving the system's adaptability and scientific discovery capabilities in unknown environments; 3. It is clearly more suitable for the application needs of my country's subsequent long-term, highly autonomous deep space exploration missions for international lunar research stations. Attached Figure Description

[0016] Figure 1 Flowchart of an intelligent sampling method for targets of scientific interest on planetary surfaces; Figure 2 This is a schematic diagram of a spectral classification model. Detailed Implementation

[0017] The technical solutions provided in this application are further illustrated below with reference to the embodiments.

[0018] Example 1 This embodiment provides an intelligent sampling method for scientific points of interest on planetary surfaces. The method first acquires RGB images of the planetary surface and automatically identifies and locates potential scientific points of interest based on a deep learning-based target detection method. Then, spectral data is collected from the target region to complete classification and identification. The system innovatively introduces an incremental learning mechanism, enabling adaptive updates to the spectral classification model after the discovery of new categories of scientific targets, thus possessing the ability to recognize both new and old categories of scientific targets. This embodiment not only has technological innovations but also achieves breakthroughs in detection modes, enabling autonomous sampling and dynamic recognition of scientific points of interest on planetary surfaces.

[0019] This method is as follows Figure 1 The process is divided into four stages, with the specific steps as follows.

[0020] Phase 1: This phase of work was completed during the ground-based development process. It collected RGB (Red-Green-Blue) images and their corresponding hyperspectral data obtained from previous lunar and Mars exploration missions, and incorporated scientific interest target (SPI) data generated in simulated planetary surface environments as a valuable supplement, thus constructing a multi-source dataset specifically for planetary exploration missions. Using this dataset, a YOLOv8s deep learning network (a type of object detection model) was trained to create a dedicated RGB target detection model for planetary surfaces, used for automatically locating SPI regions. After acquiring the SPI regions, their corresponding spectral data was extracted, and based on this, a spectral classification model was trained in the basic training phase to identify SPI categories known to be of interest.

[0021] Phase Two: After landing on the planetary surface, the rover acquires RGB images using its navigation camera, and then uses a target detection model to automatically identify potential points of scientific interest (POIs) in the images and output their location information. Subsequently, a hyperspectral camera acquires target spectral data, which is then fed into a classification model for identification. When the confidence level of the identification result is higher than a preset threshold, the system directly outputs a category label; when the confidence level is lower than the threshold, the system marks the target as a potential new category of scientific target and caches its image, spectrum, and location information. The cached data is packaged and transmitted to Earth after meeting certain conditions to support subsequent incremental model training.

[0022] Phase Three: After receiving data on potential new categories of scientific targets, the ground center assigns categories by experts and initiates an incremental learning optimization process for the spectral classification model built in the first phase of training. For example... Figure 2 As shown, during incremental learning, this embodiment does not require access to the original data of the old categories. Instead, it maintains the model's memory of the old categories through a feature replay strategy. Specifically, in the basic training phase, this embodiment constructs an adaptive component Gaussian mixture model (GMM) for each old category to represent the feature distribution of the old categories in a parameterized manner. Simultaneously, this embodiment designs a feature projection module to learn the mapping relationship between the old and new feature extractors. Features sampled from the old category GMM are semantically aligned with the feature space of the new feature extractor via the feature projection module. These replayed old category features participate in training together with newly labeled new category samples, enabling the model to effectively learn new category features while continuously consolidating its ability to distinguish the old categories. The updated model is uploaded to the rovers and deployed in orbit, achieving continuous evolution of the detection and cognitive level.

[0023] Phase Four: Based on the classification and identification results of scientific points of interest and their spatial location information, the rover, either autonomously or after confirmation by the ground, is controlled to plan its path to approach the target and perform pre-set scientific operations such as high-precision imaging, elemental composition analysis, or robotic arm sampling.

[0024] In the first stage, this embodiment constructs a dedicated RGB and hyperspectral joint dataset for scientific points of interest on the lunar and Martian surfaces, and designs a dedicated deep learning target detection model and spectral classification model for special application scenarios on planetary surfaces. In the second stage, a dedicated target detection method for RGB images is designed to achieve autonomous identification and spatial positioning of scientific points of interest. Furthermore, in the spectral classification model in the third stage, an incremental learning strategy designed specifically for deep space exploration missions is introduced to achieve adaptive model updates and compatible identification of scientific targets of new and old categories.

[0025] Example 2 This embodiment provides an intelligent sampling system for planetary surface scientific targets of interest, used in the intelligent sampling method for planetary surface scientific targets of interest in Embodiment 1, including: The acquisition module is used to acquire visible light images of the planetary surface at the rover end, autonomously identify potential targets of scientific interest using a target detection model and determine the spatial location of the targets, acquire spectral data of the target area, and input the spectral data into a spectral classification model for classification and identification; A spectral classification model is used to identify the category of a target and output the classification result and the corresponding confidence score. The comparison module is used to compare the confidence level of the classification and identification results with a preset threshold. The labeling module is used to output classification result labels for targets with confidence levels higher than or equal to a preset threshold; it is used to mark targets with confidence levels lower than a preset threshold as potential new categories of scientific interest targets, and to transmit the image, spectral and spatial location information corresponding to the potential new categories of scientific interest targets to the ground center. The potential new categories of scientific interest targets are manually labeled to obtain the classification result labels of the new categories. The incremental learning module is used to incrementally train the spectral classification model, enabling adaptive optimization of the model and compatible recognition of new and known categories of scientific interest targets; it is also used to redeploy the incrementally updated spectral classification model to the rovers; and The sampling operation module controls the rover to perform sampling operations for scientifically interested targets based on the classification results and spatial location information of the targets.

[0026] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent sampling of scientific targets of interest on planetary surfaces, characterized in that, include: Step 1: Acquire visible light images of the planetary surface on the rover, autonomously identify potential targets of scientific interest using a target detection model and determine their spatial locations, collect spectral data of the target area, and input the spectral data into a spectral classification model for classification and identification; when the confidence level of the classification and identification result is lower than a preset threshold, proceed to Step 2; when the confidence level of the classification and identification result is higher than or equal to the preset threshold, output the category result label and proceed to Step 3; Step 2: Targets with confidence scores below a preset threshold are marked as potential new categories of scientific interest targets. The corresponding image, spectrum, and spatial location information of these targets are transmitted to the ground center. The new category is manually labeled to obtain the classification result label. Incremental learning training is performed on the spectral classification model to achieve adaptive optimization of the spectral classification model and compatible recognition of new and known categories of scientific interest targets. The spectral classification model updated by incremental learning is then redeployed to the rover. Step 3: Based on the classification results and spatial location information of the targets, control the rover to perform sampling operations for targets of scientific interest; During the incremental learning training process, the spectral classification model retains its memory of features of known categories through a feature replay strategy. The feature replay strategy is achieved by constructing an adaptive component Gaussian mixture model for each known category during the basic training phase of the spectral classification model, and representing the feature distribution of each known category in a parameterized manner. The spectral classification model includes a feature projection module, which is used to learn the mapping relationship between feature extractors of known categories and feature extractors of new categories; In step 2, the incremental learning training process of the spectral classification model includes: during incremental learning, known category feature data sampled from the Gaussian mixture model and manually labeled new category feature data are used together for training; during training, the feature projection module learns the mapping relationship between the feature extractors of known categories and the feature extractors of new categories, so that the feature space of known categories and the feature space of new categories are semantically aligned; based on the feature replay strategy and the feature alignment results, the network parameters of the spectral classification model are updated, so that the spectral classification model can maintain the memory of known category features while learning new category features.

2. The intelligent sampling method for planetary surface scientific interest targets according to claim 1, characterized in that, The target detection model is trained on a multi-source joint dataset, which includes images of scientific interest acquired from historical exploration missions and scientific interest target data generated by simulating planetary surface environments.

3. The intelligent sampling method for planetary surface scientific interest targets according to claim 2, characterized in that, The multi-source joint dataset also includes hyperspectral data corresponding to images of scientific interest acquired in historical exploration missions; the spectral classification model is trained on the hyperspectral data in the multi-source joint dataset during the basic training phase to identify scientific interest targets of known categories.

4. The intelligent sampling method for planetary surface scientific interest targets according to claim 1, characterized in that, The visible light image includes: RGB image.

5. An intelligent sampling system for planetary surface scientific interest targets, used to implement the intelligent sampling method for planetary surface scientific interest targets as described in any one of claims 1-4, comprising: The acquisition module is used to acquire visible light images of the planetary surface at the rover end, autonomously identify potential targets of scientific interest using a target detection model and determine the spatial location of the targets, acquire spectral data of the target area, and input the spectral data into a spectral classification model for classification and identification; A spectral classification model is used to identify the category of a target and output the classification result and the corresponding confidence score. The comparison module is used to compare the confidence level of the classification and identification results with a preset threshold. The label module is used to output the classification result labels for targets with a confidence level higher than or equal to a preset threshold; This tool is used to mark targets with confidence levels below a preset threshold as potential new categories of scientific interest targets, and to transmit the image, spectral, and spatial location information corresponding to the potential new categories of scientific interest targets to the ground center. The potential new categories of scientific interest targets are manually labeled to obtain the classification result labels of the new categories. The incremental learning module is used to incrementally train the spectral classification model, enabling adaptive optimization of the model and compatible recognition of new and known categories of scientific interest targets; it is also used to redeploy the incrementally updated spectral classification model to the rovers; and The sampling operation module controls the rover to perform sampling operations for scientifically interested targets based on the classification results and spatial location information of the targets.

Images