Spectroscopic identification system and method for targets of interest on planetary surfaces

Abstract

This invention discloses a system and method for spectral recognition of targets of interest on planetary surfaces, specifically including: data preprocessing, acquiring hyperspectral images, and performing meta-task partitioning on the hyperspectral images to obtain a multi-category sample set; receiving the sample set, extracting kernel attention features and kernel dispersion features through bi-branch feature extraction, respectively, and applying consistency constraints to obtain sample features; using the sample features to obtain prototype vectors for each category, constructing a graph structure with prototype vectors as nodes and sample features as associated nodes, and aggregating semantic and geometric information; using a gradient-aware metric function to calculate the gradient information between each sample feature and the corresponding category prototype vector, constructing a gradient force field, and adaptively adjusting the distance relationship between the sample and the prototype based on the gradient information; this invention can achieve highly accurate spectral intelligent recognition of targets of interest on planetary surfaces in environments lacking a large number of labeled samples and in harsh planetary surface environments.

Description

Technical Field

[0001] This invention relates to the field of target localization and identification technology, and more specifically to a spectral identification system for targets of interest on planetary surfaces.

[0002] This invention also relates to a method for spectral identification of targets of interest on planetary surfaces. Background Technology

[0003] In planetary surface remote sensing, hyperspectral imaging is an imaging technique capable of simultaneously acquiring the reflectance characteristics of targets across tens to hundreds of continuous spectral bands. The imaging results typically manifest as three-dimensional cubic data containing rich spectral and spatial dimensions. Each pixel carries a complete spectral reflectance curve, revealing information such as the mineral composition, chemical properties, structural state, and weathering degree of planetary surface materials. Therefore, hyperspectral image classification has become a crucial foundational technology in planetary exploration missions. By analyzing pixel spectra and local spatial structures, it is possible to automatically identify and label target areas of interest on the planetary surface, such as rock formations, mineral distribution, surface sediments, and frost-covered areas, providing critical support for geological analysis, resource exploration, environmental assessment, and mission decision-making.

[0004] However, planetary exploration is limited by complex observation environments, high-noise sensor conditions, and a lack of reliable spectral priors. Hyperspectral data often has characteristics such as high dimensionality, strong noise, severe spectral mixing, and scarce annotations. In addition, the surface composition and illumination conditions of different planets vary significantly, resulting in significant variations in spectral response between missions, locations, and even time.

[0005] Existing technologies in this field mainly fall into two categories: one is based on traditional spectral analysis and statistical learning methods, relying on manually designed features and classic classifiers for identification; the other is based on deep learning, which automatically learns joint spatial spectral features through end-to-end nonlinear modeling.

[0006] Traditional methods primarily rely on a pipelined design from feature engineering to dimensionality reduction to classifiers. They typically begin with the spectral dimension, characterizing the spectral differences of different ground features by analyzing the shape and statistical characteristics of reflectance curves. Typical methods such as spectral angle mapping are widely used due to their insensitivity to changes in illumination, while methods such as spectral information divergence provide a more sensitive measure of spectral differences when separability is insufficient.

[0007] Meanwhile, to enhance the modeling capability of scene spatial structure, spatial feature extraction strategies such as multi-scale filters, texture operators, and morphological structural elements were designed. These strategies capture the shape and structure of ground features by constructing dilation and erosion sequences at different scales, representing a significant technique in spatial modeling. Furthermore, models such as Markov random fields further enhance the spatial smoothness of classification results by imposing consistency constraints on neighboring pixels. With a deeper understanding of the joint relationship between spatial and spectral data, traditional methods have gradually developed various fusion schemes, including feature-level concatenation of spectral vectors and spatial textures, using 3D filters to capture local correlation structures in the spectral spatial dimension, and utilizing graph Laplacian matrices to construct manifold structures to enhance local geometric representation capabilities.

[0008] Due to the extremely high dimensionality of hyperspectral data and the scarcity of labeled samples, dimensionality reduction techniques play a central role in traditional hyperspectral classification algorithms. By compressing redundant dimensions and improving inter-class separation, these techniques enable backend classifiers such as Support Vector Machines and Random Forests to achieve more stable generalization capabilities under limited conditions. Overall, traditional hyperspectral image classification methods rely on carefully designed joint spectral spatial feature extraction and dimensionality reduction strategies, offering advantages such as well-defined structure and good interpretability. However, their complexity increases rapidly and they struggle to capture higher-level semantics and complex structural relationships.

[0009] With the development of deep learning, end-to-end learning methods such as convolutional neural networks have gradually become important technical routes for hyperspectral classification. Early methods typically used one-dimensional convolutional networks to extract spectral sequence features from the spectral dimension, or two-dimensional convolutions to obtain texture information from the spatial dimension. Subsequently, three-dimensional convolutions were used to jointly convolve in the spatial and spectral dimensions, which can simultaneously capture local spatial structures and continuous spectral curves, achieving richer joint spatial-spectral representation learning. Some hybrid network structures (such as one-dimensional and two-dimensional hybrid convolutions, convolutional networks combined with recurrent networks, spatial-spectral branching and fusion networks, etc.) are also used to further improve cross-dimensional feature modeling capabilities. Graph neural networks have also been introduced into hyperspectral classification tasks, building graph structures from pixels or superpixels and using graph convolutions to capture the connections between spatial regions. Attention mechanisms and Transformer structures have also been gradually introduced into hyperspectral classification to enhance long-range dependency modeling capabilities and improve generalization under unlabeled data conditions.

[0010] Because hyperspectral images typically have few annotations, recent research has gradually shifted towards small-sample studies. Prototype networks learn class centers from a small number of samples and classify unlabeled samples based on metric distance, or learn a generalizable spatial-spectral joint metric space suitable for cross-class transfer. Meta-learning, on the other hand, simulates the training and testing process through an episode mechanism to improve the model's adaptability to new classes. Furthermore, in self-supervised and contrastive learning frameworks, the problem of insufficient annotation is alleviated by constructing pre-training tasks using unlabeled pixels.

[0011] Existing hyperspectral classification techniques for planetary surfaces still have significant limitations when dealing with scenarios involving few samples:

[0012] There is a lack of dynamic modeling capabilities for fine-grained spectral variations. Most existing spectral feature extraction techniques rely on fixed convolution kernels, linear dimensionality reduction, or simple spectral attention mechanisms, which cannot adaptively reconstruct based on the spectral morphology of different samples. When faced with the weak spectral differences, mixed pixels, and multi-scale mineral composition commonly found on planetary surfaces, traditional methods often only obtain overly coarse or insufficiently informative spectral representations. They lack mechanisms for jointly modeling the geometric proximity between samples and category-level semantic priors, making it difficult to form stable and compact intra-class representations under conditions of scarce samples. This results in insufficient ability to identify fine spectral differences under limited sample conditions.

[0013] It is impossible to simultaneously utilize the geometric structure of samples and the semantic prior of categories. Existing spatial-spectral joint structure modeling methods generally rely on local neighborhoods (such as spatial convolution, K-nearest neighbor methods, superpixels, etc.) or edge weights in graph structures based solely on geometric proximity. However, very few methods can simultaneously introduce semantic guidance from class prototypes at the category level, nor can they establish a bidirectional relationship structure between samples and categories themselves in scenarios where samples are scarce. These methods often result in loose distribution of samples within a class and blurred boundaries between classes, which is especially unstable in scenarios such as the classification and recognition of targets of interest on planetary surfaces.

[0014] Current measurement methods rely solely on scalar distances, failing to handle directional differences and difficult-to-distinguish samples. Existing prototype networks and their improved methods generally employ scalar measures such as Euclidean and cosine distances, providing only the magnitude of the distance but lacking crucial directional information such as the rationality of the approach direction and the appropriate correction direction. Consequently, when faced with samples exhibiting similar spectral morphology, comparable surface composition, or cross-domain distribution variations, they struggle to effectively distinguish or adjust relationships. This is particularly problematic in planetary exploration, where unstable illumination, high noise levels, and severe component mixing result in numerous ambiguous samples, making traditional measurement methods highly prone to misjudgment. Summary of the Invention

[0015] The purpose of this invention is to provide a system and method for spectral identification of targets of interest on planetary surfaces, so as to improve the identification accuracy, stability and efficiency of the system under cross-domain and small sample conditions, and to achieve high-accuracy intelligent spectral identification of targets of interest on planetary surfaces in the absence of a large number of labeled samples and in harsh planetary surface environments.

[0016] The technical solution adopted in this invention is as follows: a spectral identification system for targets of interest on planetary surfaces, the system comprising:

[0017] The data preprocessing module is used to receive hyperspectral images and divide the hyperspectral images to obtain multi-category sample sets, which include support sets and query sets.

[0018] The spectral enhancement sensing module is used to receive the sample set, extract kernel attention class features and kernel dispersion features through dual-branch features and apply consistency constraints to obtain sample features;

[0019] The graph construction module uses the sample features to obtain the prototype vector of each category. With the prototype vector as the center, it constructs a bipartite graph structure between the sample and the prototype, and models the geometric distance information of the sample-level features and the non-geometric semantic information of the category level in a unified manner. The prototype vector is used as the category representation node, and the weight of each edge in the graph is calculated based on the spectral distance and probability correlation between the sample and the prototype.

[0020] The gradient-aware measurement module calculates the gradient information between each sample feature and the corresponding class prototype vector, constructs a gradient force field, and then applies adaptive constraints by setting loss functions for positive and negative samples respectively. Specifically, it applies corresponding positive and negative gradients to difficult-to-distinguish positive and negative samples respectively, adaptively adjusts the distance relationship between the sample and the prototype, and outputs the recognition result.

[0021] Furthermore, the dual-branch feature extraction specifically includes a central branch and peripheral branches;

[0022] Specifically, samples from the support set are used as input, and kernel attention class features are extracted through the central branch;

[0023] After spectral reconstruction and spectral sampling, the samples supporting the set are input into the peripheral branches to extract kernel dispersion features.

[0024] This invention also employs a method for spectral identification of targets of interest on planetary surfaces, using the aforementioned spectral identification system for targets of interest on planetary surfaces. The method includes:

[0025] Data preprocessing involves acquiring hyperspectral images and performing meta-task partitioning on the hyperspectral images to obtain multi-category sample sets, which include support sets and query sets.

[0026] Spectral enhancement sensing receives the sample set and extracts kernel attention class features and kernel dispersion features through bi-branch feature extraction, wherein the bi-branch feature extraction includes a central branch and a peripheral branch;

[0027] Specifically, samples from the support set are used as input, and kernel attention class features are extracted through the central branch;

[0028] After spectral reconstruction and spectral sampling, the samples supporting the set are input into the peripheral branches to extract kernel dispersion features;

[0029] Finally, the kernel attention features and kernel dispersion features are constrained by constructing a loss function to obtain the fused sample features;

[0030] The prototype vector-guided graph construction utilizes the sample features to obtain the prototype vector for each category. Centered on the prototype vector, a bipartite graph structure is constructed between the sample and the prototype, unifying the geometric distance information of the sample-level features with the non-geometric semantic information of the category level. The prototype vector is used as the category representation node, and the weights of each edge in the graph are calculated based on the spectral distance and probability correlation between the sample and the prototype.

[0031] The gradient-aware metric function calculates the gradient information between each sample feature and the corresponding class prototype vector, constructs a gradient force field, and then applies adaptive constraints to positive and negative samples by setting loss functions. Specifically, it applies corresponding positive and negative gradients to difficult-to-distinguish positive and negative samples respectively, adaptively adjusts the distance relationship between the sample and the prototype, and outputs the recognition result.

[0032] Furthermore, in the data preprocessing, the hyperspectral image is segmented to obtain several image blocks, and these image blocks are then divided into meta-tasks.

[0033] In each meta-task, N categories are selected, and a small number of K labeled samples are extracted from each category to form a support set for building category prototypes. Several samples are extracted from the remaining data of the same category to form a query set for model validation.

[0034] Furthermore, the method also includes parameter updating, as detailed below:

[0035] Pseudo-labels are generated based on the query set, and these pseudo-labels are used as monitoring signals to update the parameters of the entire method using the cross-entropy function.

[0036] Furthermore, in the prototype vector-guided graph construction, the prototype vector for each category is calculated through metric learning based on sample features.

[0037] In summary, due to the adoption of the above technical solution, the beneficial effects of this application are:

[0038] (1) Reduce dependence on a large amount of labeled data: Based on a small sample framework with a dual structure of sample level and category level, the model can be trained in the case of small samples, which can reduce the strong correlation with labeled data, while ensuring the efficiency of training speed and the accuracy of classification of targets of interest on planetary surfaces.

[0039] (2) Enhance spectral representation capabilities and maintain stable recognition performance on data with significant spectral shifts in different domains: Existing methods often use fixed convolution kernels or simple attention mechanisms, which are difficult to capture local details and global spectral changes at the same time. This invention uses a dual-branch spectral recombination mechanism to dynamically perform fine-grained and diffusion-based modeling of spectral dimensions, amplifying subtle class-discriminative spectral differences and improving the spectral robustness of the model's feature extraction. Even under small sample conditions, it can still form more discriminative sample-level features.

[0040] (3) Effectively reduce inter-class confusion, improve the common class drift in cross-domain tasks, and make the feature distribution structure more stable and generalizable: Traditional methods rely on the local neighborhood geometric relationship between samples and lack class-level semantic structure. This invention constructs a bipartite graph of sample categories guided by prototype vectors, unifying geometric neighborhood information and class semantic priors into the same graph space, making intra-class features more compact, inter-class boundaries more explicit, and the graph structure more in line with the segmented and implicit semantic structure of hyperspectral data, significantly making up for the shortcomings of traditional global class-level feature modeling.

[0041] (4) Significantly improves the recognition accuracy of difficult-to-distinguish samples and enhances the discriminativeness of the overall classification boundary, so that the model can still maintain high reliability under small sample conditions: Most existing metric learning uses scalar distance, which cannot express the directional information between the sample and the prototype. The gradient-aware metric function proposed in this invention regards the gradient as a push-pull force vector, which can adaptively refine the relationship, apply stronger pull to easily confused positive samples to promote intra-class convergence, apply push to negative samples close to the wrong class to increase inter-class interval, and is particularly effective for blurry samples in the scene. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the structure of the planetary surface target of interest spectral identification system of the present invention;

[0043] Figure 2 This is a flowchart illustrating the training phase of the planetary surface target of interest spectral identification method of the present invention;

[0044] Figure 3 This is a flowchart illustrating the testing phase of the method for spectral identification of targets of interest on planetary surfaces in this invention. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings.

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] Example 1

[0048] This embodiment provides a spectral identification system for targets of interest on a planetary surface, such as... Figure 1 As shown, it specifically includes:

[0049] Data preprocessing module: The training data is divided into a series of independent meta-tasks. Each meta-task simulates the few-shot recognition scenario in real deployment. In each meta-task, N categories are randomly selected, and only a very small number of K labeled samples are extracted from each category to form a support set to simulate the few-shot supervision condition. A number of samples are extracted from the remaining data of the same category to form a query set to evaluate the learning effect under the meta-task.

[0050] The spectral enhancement perception module employs a dual-branch structure, comprising a central branch and peripheral branches. The central branch focuses on key spectral regions, enhancing bands with significant differences to form stable intra-class aggregation features and acquire kernel attention features. The peripheral branches, through perturbation, resampling, and fragmentation of the spectral dimensions, break local sequential dependencies, enabling the model to perceive weak and dispersed spectral differences and acquire kernel dispersion features. The two branches complement each other, allowing the system to capture both globally stable differences and be sensitive to subtle changes, thereby significantly improving the granularity and robustness of sample-level spectral representation and providing high-quality input features for subsequent category-level graph construction.

[0051] The prototype vector-guided graph construction module aims to address the problem that traditional graph modeling cannot simultaneously utilize sample geometric proximity relationships and category semantic priors. Centered on category prototype vectors, this module constructs a bidirectional bipartite graph structure between samples and prototypes, unifying the modeling of geometric distance information at the sample level and non-geometric semantic information at the category level. Based on features extracted by the spectral enhancement perception module, prototype vectors for each category are generated and used as category representation nodes. Subsequently, edge weights in the graph are calculated based on the spectral distance and probabilistic correlation between samples and prototypes, thus explicitly establishing the dependency relationship between samples and categories. This not only enhances intra-class feature aggregation and strengthens inter-class separability but also improves the stability and robustness of category representation by integrating non-geometric semantic priors under conditions of sample scarcity and cross-domain scenarios.

[0052] The gradient-aware measurement module introduces gradient information—both gradient direction and gradient intensity—in the modeling of prototype sample relationships. It treats the gradient as a vector signal, similar to an attractive or repulsive force, reflecting the direction and strength in which samples should move closer to or away from the prototype class. Through analysis of the gradient field, this module can adaptively adjust the relationship between samples and prototypes during model updates, resulting in stronger aggregation trends for positive samples, clearer discrimination corrections for difficult-to-distinguish samples, and effective removal of negative samples from the error class boundary. This significantly enhances class separability and classification reliability under cross-domain and weakly discriminative conditions.

[0053] Example 2

[0054] This embodiment provides a method for spectral identification of targets of interest on planetary surfaces, the method specifically including:

[0055] During the training phase, such as Figure 2 As shown, the system first cuts the input hyperspectral image into fixed-size image patches, and then performs meta-task partitioning to ensure that small sample episodes can work properly. A small sample episode includes a support set and a query set.

[0056] The samples from the support set are input into the central branch of the spectral enhancement perception module to extract the kernel attention class features. Then, the samples from the support set are input into the peripheral branches through spectral reconstruction and spectral sampling to extract the kernel dispersion features. Consistency constraints are applied to the two branches to ensure that the two branches make consistent judgments on the sample categories. The fused sample features are then output.

[0057] Based on the dual-branch fusion features of support set samples, a prototype network is constructed using metric learning methods to obtain the prototype vector of each category. The prototype vector is then input into the graph construction module guided by the prototype vector. The graph structure is constructed with the prototype vector as the node and the sample feature as the associated node, which aggregates semantic and geometric information, so that the features have both category discrimination and spatial correlation.

[0058] Then, using a gradient-aware metric function, the gradient information between the sample features and the prototype vector is calculated to establish a gradient force field. Loss functions are set for positive and negative samples respectively to ensure that there are large positive and negative gradients for difficult-to-distinguish positive and negative samples respectively. Specifically, a large positive gradient is applied to difficult-to-distinguish positive samples (i.e., samples of the same class that are far from the prototype), pulling the samples closer to the prototype center; a large negative gradient is applied to difficult-to-distinguish negative samples (i.e., samples of different classes that are close to the prototype), pushing the samples away from the prototype center; so that the class center, i.e., the prototype vector, has a certain ability to accept and reject positive and negative samples.

[0059] The method in this embodiment also includes parameter updating, specifically:

[0060] Generate overall predicted pseudo-labels: Combining the results of bi-branch features and prototype measurement, specifically, after bi-branch feature extraction and consistency constraints, a preliminary prediction confidence of a sample belonging to each category is output; the prototype measurement is the matching degree between the sample features and the prototypes of each category, outputting the category confidence based on prototype matching. Weights are assigned to the preliminary confidence of the bi-branch features and the confidence of the prototype measurement, respectively. The weights can be dynamically learned. The final category confidence is obtained by weighted summation, generating pseudo-labels for the query set samples. This is suitable for semi-supervised / self-supervised small sample scenarios.

[0061] Calculate cross-entropy loss: Using pseudo-labels as supervision signals, backpropagate to optimize the parameters of the entire network, achieving end-to-end training;

[0062] Finally, the network parameters are updated using the overall predicted pseudo-labels and the cross-entropy function.

[0063] During the testing phase, such as Figure 3 As shown, the KNN classifier is first trained using features extracted from a small number of labeled samples. Then, the KNN classifier is used to classify the obtained unlabeled samples, and the classification accuracy is calculated.

[0064] Through the above-mentioned modules and training and testing process design, this invention can achieve high-accuracy intelligent spectral recognition of targets of interest on planetary surfaces in the absence of a large number of labeled samples and in harsh planetary surface environments.

[0065] In summary, the advantages of this invention are:

[0066] A small sample framework based on a two-tiered structure of sample-level and category-level architecture:

[0067] This invention proposes a hierarchical architecture from sample-level spectral perception to category-level graph structure aggregation to achieve spectral classification and identification of targets of interest on planetary surfaces. This architecture first reconstructs the spectra of the original samples, and then constructs a graph structure based on prototype vectors at the category level. This allows the model to simultaneously utilize local geometric structures and global semantic topological relationships, thereby improving the problem of insufficient global structure caused by traditional methods relying solely on local geometry.

[0068] The spectral enhancement sensing module for dynamic spectral reconstruction and its dual-branch mechanism:

[0069] This invention designs a sample-level module for dynamic spectral reconstruction. This module includes two complementary branches: a central branch for extracting fine-grained details of the region of interest, and a peripheral branch for diffusely modeling large-scale spectral variations. The combination of these two branches can capture subtle spectral differences and improve the discriminative power of single-sample representations, providing a foundation for robust classification and recognition in small-sample and planetary surface scenarios.

[0070] Construction of a bipartite graph of sample categories guided by prototype vectors:

[0071] This invention introduces a bipartite graph of samples and prototype vectors at the category level to fuse local geometric neighborhood relationships in the metric network with non-geometric category priors at the semantic level. By connecting each sample to the prototype vector of the corresponding category, this invention achieves a joint aggregation of category semantics and geometric structure, effectively improving intra-class compactness and inter-class separability.

[0072] Gradient-aware metric function for refining the relationship between prototypes and samples:

[0073] Traditional metric functions rely solely on scalar distance and cannot describe the directional relationship between samples and prototypes. This invention proposes a novel gradient-aware metric function that treats the back gradient as a force vector. It dynamically guides the application of a pulling force to hard-to-distinguish positive samples and a pushing force to easily confused negative samples through the magnitude and direction of the gradient. This achieves finer intra-class compactness and inter-class separation, and improves the robustness of spectral classification and recognition of targets of interest in planetary surface scenes.

[0074] The proposed intelligent spectral recognition system for targets of interest on planetary surfaces exhibits significant recognition accuracy, stability, and generalization ability under limited sample conditions. Its structural design balances fine-grained spectral modeling, category-level structural aggregation, and direction-sensitive relational correction, thus possessing broad application prospects and significant engineering value. Based on a meta-task partitioning mechanism, this invention can rapidly adapt to new planetary surface observation scenarios under conditions of extremely scarce labeled samples. A spectral enhancement sensing module strengthens subtle spectral differences and weak signal features, maintaining reliable performance even in planetary exploration scenarios with complex terrain, varying illumination, and significant mineral mixing. A prototype vector-guided graph construction module effectively integrates geometric proximity relationships and semantic priors, forming a stable and compact category representation. A gradient-aware metric function further enhances the discrimination ability of difficult-to-distinguish samples, making it suitable for conditions with noise interference, weak texture regions, and uncertainties in cross-platform data.

[0075] With the aforementioned characteristics, this invention can provide technical support for various mission scenarios, including planetary remote sensing, deep space science missions, surface material identification of unmanned rover probes, risk assessment of planetary landing areas, and resource exploration and planning. Furthermore, the technical framework of this invention has good scalability and can be transferred to applications such as airborne remote sensing that require spectral identification under limited sample conditions, demonstrating significant economic and social value. With the increasing number of deep space exploration missions and the growing demand for automated planetary surface sensing, this invention has a broad application market and is expected to become an important technology for next-generation hyperspectral intelligent identification systems.

[0076] This article uses specific embodiments to illustrate the principles and implementation methods of the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A spectral identification system for targets of interest on a planetary surface, characterized in that, The system includes: The data preprocessing module is used to receive hyperspectral images and divide the hyperspectral images to obtain multi-category sample sets, which include support sets and query sets. The spectral enhancement sensing module is used to receive the sample set, extract kernel attention class features and kernel dispersion features through dual-branch features and apply consistency constraints to obtain sample features; The graph construction module uses the sample features to obtain the prototype vector of each category. With the prototype vector as the center, it constructs a bipartite graph structure between the sample and the prototype, and models the geometric distance information of the sample-level features and the non-geometric semantic information of the category level in a unified manner. The prototype vector is used as the category representation node, and the weight of each edge in the graph is calculated based on the spectral distance and probability correlation between the sample and the prototype. The gradient-aware measurement module calculates the gradient information between each sample feature and the corresponding class prototype vector, constructs a gradient force field, and then applies adaptive constraints by setting loss functions for positive and negative samples respectively. Specifically, it applies corresponding positive and negative gradients to difficult-to-distinguish positive and negative samples respectively, adaptively adjusts the distance relationship between the sample and the prototype, and outputs the recognition result.

2. The spectral identification system for targets of interest on planetary surfaces according to claim 1, characterized in that, The dual-branch feature extraction specifically includes a central branch and peripheral branches; Specifically, samples from the support set are used as input, and kernel attention class features are extracted through the central branch; After spectral reconstruction and spectral sampling, the samples supporting the set are input into the peripheral branches to extract kernel dispersion features.

3. A method for spectral identification of targets of interest on a planetary surface, employing the spectral identification system for targets of interest on a planetary surface as described in any one of claims 1 to 2, characterized in that, The method includes: Data preprocessing involves acquiring hyperspectral images and performing meta-task partitioning on the hyperspectral images to obtain multi-category sample sets, which include support sets and query sets. Spectral enhancement sensing receives the sample set and extracts kernel attention class features and kernel dispersion features through bi-branch feature extraction, wherein the bi-branch feature extraction includes a central branch and a peripheral branch; Specifically, samples from the support set are used as input, and kernel attention class features are extracted through the central branch; After spectral reconstruction and spectral sampling, the samples supporting the set are input into the peripheral branches to extract kernel dispersion features; Finally, the kernel attention features and kernel dispersion features are constrained by constructing a loss function to obtain the fused sample features; The prototype vector-guided graph construction utilizes the sample features to obtain the prototype vector for each category. Centered on the prototype vector, a bipartite graph structure is constructed between the sample and the prototype, unifying the geometric distance information of the sample-level features with the non-geometric semantic information of the category level. The prototype vector is used as the category representation node, and the weights of each edge in the graph are calculated based on the spectral distance and probability correlation between the sample and the prototype. The gradient-aware metric function calculates the gradient information between each sample feature and the corresponding class prototype vector, constructs a gradient force field, and then applies adaptive constraints to positive and negative samples by setting loss functions. Specifically, it applies corresponding positive and negative gradients to difficult-to-distinguish positive and negative samples respectively, adaptively adjusts the distance relationship between the sample and the prototype, and outputs the recognition result.

4. The method for spectral identification of targets of interest on planetary surfaces according to claim 3, characterized in that, In the data preprocessing, the hyperspectral image is cut into several image blocks, and the image blocks are then divided into meta-tasks. In each meta-task, N categories are selected, and a small number of K labeled samples are extracted from each category to form a support set for building category prototypes. Several samples are extracted from the remaining data of the same category to form a query set for model validation.

5. The method for spectral identification of targets of interest on planetary surfaces according to claim 4, characterized in that, The method also includes parameter updates, as detailed below: Pseudo-labels are generated based on the query set, and these pseudo-labels are used as monitoring signals to update the parameters of the entire method using the cross-entropy function.

6. The method for spectral identification of targets of interest on a planetary surface according to claim 3, characterized in that, In the prototype vector-guided graph construction, the prototype vector for each category is calculated through metric learning based on sample features.

Images