Planetary surface typical target high-precision positioning identification system and method

Abstract

The invention discloses a planetary surface typical target high-precision positioning identification system and method, and the method comprises the steps: a data enhancement module receives hyperspectral data, carries out the data enhancement of the hyperspectral data, and builds a contrast learning frame based on an enhanced sample; extracting a background sample and a prior sample in the hyperspectral data, enhancing the sample and extracting a space-spectrum combined feature, combining the space-spectrum combined feature with a contrast learning framework to calculate a contrast loss, and performing sample reconstruction on the background sample; the weight reconstruction module calculates a reconstruction error of an original hyperspectrum and a prior sample, calculates a deviation vector according to the reconstruction error, converts the deviation vector into a reconstruction weight, and dynamically adjusts the comparison loss by using the reconstruction weight; the system trained by the modules performs feature extraction and calculates feature similarity to generate a target detection image; according to the method, high-precision and robust typical target positioning and recognition can be realized in a planetary surface environment which is lack of labeled samples and has spectral change.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target positioning and identification, and particularly relates to a high-precision positioning and identification system for typical targets on a planetary surface.

[0002] The present application also relates to a high-precision positioning and identification method for typical targets on a planetary surface. BACKGROUND

[0003] Hyperspectral imaging technology can simultaneously obtain spatial and spectral information of the ground surface in multiple continuous spectral bands, reflecting the spectral characteristics and spatial distribution of the ground surface material. This technology has been widely used in fields such as earth observation, mineral exploration and planetary surface exploration. In planetary exploration missions, Hyperspectral Target Detection (HTD) aims to locate specific targets on the planetary surface at the pixel or sub-pixel level according to prior spectral characteristics. Unlike traditional target detection based on rectangular boxes, HTD can separate targets and backgrounds at a fine granularity level.

[0004] Due to the complex planetary environment, significant differences in illumination and terrain, and the phenomena of same-object-different-spectrum and same-spectrum-different-object in hyperspectral data, that is, the spectral response of the same target changes with the conditions, and the spectral curves of different substances may be similar. This spectral variability brings significant difficulties to target detection. Early hyperspectral target detection methods mainly include algorithms based on spectral matching and statistical assumptions. Spectral matching methods such as Euclidean distance, spectral angle mapping, and spectral information divergence distinguish targets and backgrounds by calculating spectral similarity, which is simple in principle but susceptible to noise and illumination. Statistical models such as matched filtering, constrained energy minimization, and orthogonal subspace projection perform detection under the assumption that the spectral distribution conforms to the Gaussian model, but their performance is limited in complex nonlinear scenes. In addition, some representation learning methods model pixels as linear combinations of dictionary atoms and use reconstruction error as the basis for detection, which does not require distribution assumptions but is difficult to capture the nonlinear characteristics of hyperspectral data.

[0005] In recent years, deep learning methods have become an important research direction in hyperspectral target detection due to their strong non-linear feature extraction capabilities. However, the high dimensionality and spectral variability of hyperspectral images pose significant challenges to deep learning-based methods, such as the scarcity of labeled samples and the inconsistency of intra-class spectra. Currently, various algorithms based on convolutional neural networks (CNN), Transformers, and state-space models (SSM) have made important progress in addressing these issues.

[0006] The prior art is as follows: Existing deep learning-based hyperspectral target detection algorithms can be mainly divided into two categories: 1) Supervised learning-based detection methods.

[0007] Such methods distinguish the target from the background by training a binary classification network with existing labeled samples. To alleviate the problem of insufficient labeled samples, detection algorithms usually introduce data augmentation strategies based on prior target spectra, such as automatic encoder generation, diffusion model generation, random mask strategy expansion, etc. However, due to the problem of over-augmentation, the augmented target samples may deviate from the real data distribution, and the model may introduce noise in the supervised feature learning stage, reducing the detection accuracy.

[0008] 2) Detection methods based on self-supervised contrastive learning.

[0009] To reduce the dependence on labeled samples, self-supervised learning achieves feature extraction under the condition of no label by constructing auxiliary tasks or sample pairs. Contrastive Learning (CL) is a typical representative. This method generates different forms of augmented samples for the same data sample, constructs positive and negative sample pairs, and uses a contrastive loss function to pull positive samples and push negative samples in the feature space, thereby achieving effective feature learning. CL can construct diverse positive sample pairs without human annotation, enabling the model to learn semantic consistent feature representations under spectral variation conditions.

[0010] The disadvantages of the prior art are as follows: (1) The data augmentation strategy is single and it is difficult to balance model robustness and feature consistency; existing detection methods based on self-supervised CL usually rely on fixed sample pair construction strategies. When the augmentation operation is too strong, the sample distribution deviates, introducing noise; when the augmentation is too weak, it is difficult to improve the model robustness; the existing methods lack a mechanism to adaptively balance different augmentation intensities, making it difficult for the model to balance semantic consistency and robustness to spectral perturbations.

[0011] (2) The feature extraction network fails to effectively cope with spectral variation and lacks multi-scale spatial-spectral joint modeling capability; the planet surface is significantly affected by differences in terrain, illumination, and reflection characteristics, leading to frequent occurrences of same-object different-spectrum and same-spectrum different-object phenomena. Most existing network structures use one-way convolution or Transformer structure, which is difficult to capture multi-directional and long-distance spatial-spectral features at low complexity, limiting the detection robustness and generalization performance of the model in complex scenes.

[0012] (3) The existing contrastive learning framework fails to fully utilize hidden positive sample information, leading to scattered feature clustering; in existing contrastive learning methods, semantically similar target samples (hidden positive sample pairs) are often incorrectly classified as negative samples, causing the model to excessively distinguish similar classes in the embedding space, disrupting intra-class consistency, and resulting in unstable feature distribution and reduced detection accuracy.

[0013] In summary, the prior art still has obvious deficiencies in robustness under spectral variation conditions, self-supervised feature consistency, and hidden positive sample identification, and there is an urgent need for a planetary surface typical target positioning identification method that can balance efficiency and accuracy. SUMMARY

[0014] The purpose of the present application is to provide a planetary surface typical target high-precision positioning identification system and method, specifically a planetary surface typical target positioning identification system and method that still has robust feature learning ability under the condition of lacking labeled samples, can adapt to spectral variation, avoid hidden positive sample interference, and balance spatial-spectral modeling and lightweight.

[0015] The technical solution adopted by the present application is as follows: a planetary surface typical target high-precision positioning identification system, the system comprising: A data enhancement module for receiving hyperspectral data and performing data enhancement on the hyperspectral data, obtaining enhanced samples, and constructing a contrast learning framework based on the enhanced samples; A data preprocessing module for receiving hyperspectral data and extracting background samples and prior samples; A feature extraction module for receiving the enhanced samples and extracting spatial-spectral joint features, combining the spatial-spectral joint features with the contrast learning framework to calculate a contrast loss, and reconstructing the background samples; A weight reconstruction module for calculating the reconstruction error of the original hyperspectral data and the prior samples, calculating a bias vector based on the reconstruction error, converting the bias vector into a reconstruction weight, and dynamically adjusting the contrast loss using the reconstruction weight; A result output module for feature extraction by the system trained by the above modules and generating a target detection map output after calculating the feature similarity.

[0016] Further, the data enhancement module is specifically a strong and weak frequency domain interference enhancement module for applying different intensities of interference to the input hyperspectral data in the frequency domain to generate strong and weak enhanced samples.

[0017] Further, the feature extraction module is specifically a dual-flow four-way scanning feature extraction network, specifically including a spatial flow module and a spectral flow module, for receiving the enhanced samples to extract spatial structure features and spectral distribution features through the spatial flow module and the spectral flow module, respectively, to obtain spatial-spectral joint features.

[0018] Further, the background samples are received by a DSQM encoder, and the background samples are reconstructed by a decoder, and then the reconstruction error of the original hyperspectral data and the prior samples is calculated.

[0019] Further, the weight re-modeling module is specifically a reconstruction weight-driven adaptive temperature module, configured to dynamically adjust the temperature parameter between sample pairs in the contrast loss according to the reconstruction weight.

[0020] Further, the contrast learning framework is a balanced ternary contrast learning framework.

[0021] The second technical solution adopted by the present application is a high-precision positioning and identification method for typical targets on the surface of a planet, which adopts the high-precision positioning and identification system for typical targets on the surface of a planet, and the method comprises a training phase and a testing phase. The training phase comprises: The input hyperspectral data is divided into original image blocks. The original image blocks are extracted through a data preprocessing module to obtain background samples and prior samples. The original image blocks are input into a strong-weak frequency domain interference enhancement module to obtain strong and weak enhanced samples, and a balanced ternary contrast learning framework is constructed based on the strong and weak enhanced samples. The strong and weak enhanced samples are input into a double-flow four-way scanning feature extraction network to extract spatial structure features and spectral distribution features, then the spatial structure features and the spectral distribution features are fused to obtain spatial-spectral joint features, and a contrast loss is calculated based on the spatial-spectral joint features and the balanced ternary contrast learning framework. The background samples are input into a DSQM encoder and a decoder to reconstruct the background, and the reconstruction error of the original image blocks and the prior samples is calculated. The reconstruction error is input into a reconstruction weight-driven adaptive temperature module, a deviation vector is calculated according to the reconstruction error vector, and the deviation vector is converted into reconstruction weight to dynamically adjust the temperature parameter between sample pairs in the contrast loss. The testing phase comprises that the test samples and the prior samples are simultaneously input into the system trained in the training phase, and a target detection map is generated through nonlinear background suppression processing.

[0022] Further, in the data preprocessing module, the spectral angle between the prior spectral vector and the center pixel of each original image block is first calculated, and the image block with the smallest angle value is selected as the prior sample.

[0023] Further, in the data preprocessing module, the original image blocks are processed through a constrained energy minimization algorithm and a density-based clustering method to obtain the background samples.

[0024] Further, in the testing phase, the cosine similarity of the feature vectors between the test samples and the prior samples is calculated, then the sample background is suppressed by using a nonlinear function, and finally a target detection map is generated.

[0025] In summary, due to the adoption of the technical solutions described above, the present application has the following beneficial effects: 1. No need for a large amount of labeled data to improve training efficiency: The hyperspectral target detection framework based on self-supervised triple contrast learning eliminates the dependence on a large amount of labeled data, realizes efficient training under the condition of no labeling, significantly reduces the data labeling cost and improves the training efficiency.

[0026] 2. Balancing robustness and feature consistency: The system designs S-WFI technology, generates enhanced samples by applying strong and weak interference in the frequency domain, and constructs a triple contrast framework. Strong and weak enhanced samples are dynamically constrained to avoid over-enhancement in data augmentation. S-WFI effectively balances model robustness and feature consistency, improving robust detection performance under spectral variation conditions.

[0027] 3. Spectral variation processing and real-time performance: The system designs a DSQM feature extraction network to effectively model spatial-spectral discriminative features to cope with spectral variation phenomena. In addition, DSQM adopts a low-complexity design of four-way scanning and space-spectrum state space model to ensure real-time inference ability, significantly improving the real-time performance and efficiency of target detection, and is suitable for scenarios such as planetary exploration that require efficient and fast response.

[0028] 4. Fine intra-class features, highlight targets and suppress background: The present application reconstructs the weight-driven adaptive temperature adjustment module RWAT to dynamically adjust the temperature parameters of hidden positive samples HP, effectively enhancing intra-class cohesion and inter-class separability, thereby improving the background and target separation capability and further improving the target detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 Fig. 1 is a structural schematic diagram of a planetary surface typical target high-precision positioning and recognition system of the present application; Figure 2 Fig. 2 is a flowchart of the training stage in the planetary surface typical target high-precision positioning and recognition method of the present application; Figure 3 Fig. 3 is a flowchart of the test stage in the planetary surface typical target high-precision positioning and recognition method of the present application; Figure 4 Fig. 4 is an architecture diagram of a double-flow four-way scanning feature extraction network DSQM in the planetary surface typical target high-precision positioning and recognition system of the present application. DETAILED DESCRIPTION

[0030] The present application will be described in detail below with reference to the accompanying drawings.

[0031] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0032] Embodiment 1 This embodiment provides a high-precision positioning and identification system for typical targets on the surface of a planet, as shown in the figure, which specifically comprises a strong-weak frequency domain interference enhancement module S-WFI, a data preprocessing module, a double-stream four-way scanning feature extraction network DSQM, a reconstruction weight driven adaptive temperature module RWAT and a nonlinear background suppression module, forming a complete identification system; wherein the enhanced samples output by the S-WFI module are input into the DSQM network for feature extraction and calculation of ternary contrast loss; the reconstruction error information is transmitted to the RWAT module, and the temperature parameters are adaptively adjusted after weighting, so as to finally realize robust feature representation and high-precision target detection, as follows: Figure 1 The strong-weak frequency domain interference enhancement module S-WFI: based on the random quantization and neighborhood correlation calculation method, different intensities of interference are applied to the input original image patches in the frequency domain to generate strong and weak enhanced samples; a ternary contrast learning framework is constructed using the strong and weak enhanced samples; the strong-weak frequency domain interference enhancement strategy proposed in this embodiment realizes adaptive balance of model robustness and feature consistency by applying hierarchical interference in the frequency domain and constructing a ternary contrast learning framework.

[0033] The data preprocessing module: the spectral angle between the prior spectral vector and the center pixel of each original patch is calculated, and the patch with the smallest angle value is selected as the prior sample to represent the typical spectral features of the target; in addition, the hyperspectral data is processed by combining the constrained energy minimization algorithm and the density-based clustering method, and reliable background samples patches are selected to train the background reconstruction autoencoder to model the background distribution.

[0034] The double-stream four-way scanning feature extraction network DSQM: as shown in the figure, the network includes spatial stream and spectral stream two parts, which extract spatial structure features and spectral distribution features respectively; through the four-way scanning mechanism (head-tail, tail-head and double-center outward direction), multi-scale information is fused to obtain discriminative spatial-spectral joint features; then combined with the DSQM encoder, the reconstruction error of each background sample patch is extracted to provide a reference basis for subsequent adaptive temperature adjustment. The double-stream four-way scanning SSM network structure designed in this embodiment captures multi-scale spatial-spectral features with linear complexity, effectively alleviating the influence of spectral variation on detection performance Figure 4 ​​The Reconstruction Weight-Driven Adaptive Temperature Module (RWAT) calculates the deviation vector based on the reconstruction error vector and converts it into reconstruction weights. It then dynamically adjusts the temperature parameters between sample pairs in the contrast loss based on these reconstruction weights, enabling adaptive mining and optimization of hidden positive samples, thereby enhancing intra-class feature alignment. This embodiment introduces a reconstruction weight-driven adaptive temperature adjustment mechanism, dynamically adjusting the contrast temperature parameters based on the reconstruction error between the target and background data. This enhances the feature alignment of hidden positive sample pairs and improves intra-class aggregation and inter-class separability.

[0035] Nonlinear background suppression module: Calculates the cosine similarity of feature vectors between test samples and prior samples, uses a nonlinear function to further suppress the background, and finally generates a target detection map to complete typical target detection.

[0036] Example 2 This embodiment provides a high-precision positioning and identification method for typical targets on the planetary surface, specifically including a training phase and a testing phase; Among them, such as Figure 2 As shown, the training phase specifically includes: The system first divides the input hyperspectral data into raw patches, and then extracts reliable background patches and prior sample patches through the data preprocessing module; The original patches are processed by the S-WFI module to generate enhanced samples of different intensities. The enhanced samples are then used to construct a balanced ternary contrast learning framework. The enhanced samples are then input into the DSQM network to extract multi-scale spatial-spectral features, and the contrast loss is calculated using the balanced ternary contrast learning framework. Then, the background samples are reconstructed using the DSQM encoder and decoder, and the reconstruction error between the original patches and the prior samples is calculated. The reconstruction error is input into the reconstruction weight-driven adaptive temperature module (RWAT). The deviation vector is calculated based on the reconstruction error vector, and the deviation vector is converted into reconstruction weights. The reconstruction weight-driven adaptive temperature module (RWAT) dynamically adjusts the temperature parameters of the contrast loss based on the error weights to improve the feature alignment capability of hidden positive samples.

[0037] Among them, such as Figure 3 As shown, the testing phase specifically includes: In this system, the DSQM encoder remains frozen, and test samples and prior sample patches are input into the system to extract features. The final target detection result is generated through similarity calculation and nonlinear background suppression.

[0038] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used for helping to understand the method of the present application and its core idea. It should be pointed out that, for ordinary skilled in the art, without departing from the principles of the present application, the present application can be improved and modified in several ways, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A high-precision positioning and identification system for typical targets on the planetary surface, characterized in that, The system includes: The data augmentation module is used to receive hyperspectral data and perform data augmentation on the hyperspectral data to obtain augmented samples, and to construct a contrastive learning framework based on the augmented samples. The data preprocessing module is used to receive hyperspectral data and extract background and prior samples; The feature extraction module is used to receive the enhanced samples and extract spatial-spectral joint features, combine the spatial-spectral joint features with a contrastive learning framework to calculate the contrastive loss, and reconstruct the background samples. The weighted reconstruction module is used to calculate the reconstruction error between the original hyperspectral image and the prior sample, calculate the deviation vector based on the reconstruction error, convert the deviation vector into reconstruction weights, and dynamically adjust the contrast loss using the reconstruction weights. The output module performs feature extraction on the system trained by the above modules, calculates feature similarity, and generates a target detection map for output.

2. The high-precision positioning and identification system for typical targets on the planetary surface according to claim 1, characterized in that, The data augmentation module is specifically a strong and weak frequency domain interference augmentation module, which is used to apply interference of different intensities to the input hyperspectral data in the frequency domain to generate strong and weak enhanced samples.

3. The high-precision positioning and identification system for typical targets on the planetary surface according to claim 1, characterized in that, The feature extraction module is specifically a dual-stream four-directional scanning feature extraction network, which includes a spatial flow module and a spectral flow module. It is used to receive the enhanced sample and extract spatial structure features and spectral distribution features through the spatial flow module and the spectral flow module respectively to obtain spatial-spectral joint features.

4. The high-precision positioning and identification system for typical targets on the planetary surface according to claim 1, characterized in that, The background sample is received using a DSQM encoder and reconstructed using a decoder. Then, the reconstruction error between the original hyperspectral data and the prior sample is calculated.

5. The high-precision positioning and identification system for typical targets on the planetary surface according to claim 4, characterized in that, The weight reconstruction module is specifically a weight-driven adaptive temperature module, which is used to dynamically adjust the temperature parameters between sample pairs in the contrast loss according to the reconstructed weights.

6. The high-precision positioning and identification system for typical targets on the planetary surface according to claim 1, characterized in that, The contrastive learning framework is a balanced ternary contrastive learning framework.

7. A method for high-precision positioning and identification of typical targets on a planetary surface, employing the high-precision positioning and identification system for typical targets on a planetary surface as described in any one of claims 1 to 6, characterized in that... The method includes a training phase and a testing phase; The training phase includes: The input hyperspectral data is divided into raw image patches; The original image patch is processed by a data preprocessing module to extract background and prior samples. The original image patch is input into the strong and weak frequency domain interference enhancement module for sample enhancement, strong and weak enhanced samples are obtained, and a balanced ternary contrast learning framework is constructed based on the strong and weak enhanced samples. The strong and weak enhanced samples are input into a dual-stream four-way scanning feature extraction network to extract spatial structure features and spectral distribution features. Then, the spatial structure features and spectral distribution features are fused to obtain spatial-spectral joint features. Based on the spatial-spectral joint features and the balanced ternary contrastive learning framework, the contrastive loss is calculated. Background samples are input into the DSQM encoder and decoder for background reconstruction, and the reconstruction error between the original image patch and the prior sample is calculated. The reconstruction error is input into the reconstruction weight driving adaptive temperature module, the deviation vector is calculated based on the reconstruction error vector, and the deviation vector is converted into reconstruction weights to dynamically adjust the temperature parameters between sample pairs in the contrast loss. The testing phase includes simultaneously inputting test samples and prior samples into the system trained in the training phase, and generating a target detection map through nonlinear background suppression processing.

8. The high-precision positioning and identification method for typical targets on the planetary surface according to claim 7, characterized in that, In the data preprocessing module, the spectral angle between the prior spectral vector and the center pixel of each original image block is first calculated, and the image block with the smallest angle value is selected as the prior sample.

9. The high-precision positioning and identification method for typical targets on the planetary surface according to claim 7, characterized in that, In the data preprocessing module, the original image patches are processed using a constrained energy minimization algorithm and a density-based clustering method to obtain background samples.

10. The high-precision positioning and identification method for typical targets on the planetary surface according to claim 7, characterized in that, In the testing phase, the cosine similarity of feature vectors between the test sample and the prior sample is calculated, and then a nonlinear function is used to suppress the sample background, finally generating a target detection map.

Images