An open set hyperspectral alteration mineral identification method and system based on hierarchical filtering network

Abstract

This invention discloses an open-set hyperspectral alteration mineral identification method and system based on a hierarchical filtering network, comprising: acquiring known category features of the source domain, background sample features, and target domain sample features; fusing the known category centroids of the known category features of the source domain and the background sample centroids of the background sample features; determining the decision boundary and classifying the result based on the pseudo-unknown class sample features and the known category features of the source domain; aligning the known category features of the source domain with the known category features of the target domain using a first domain discriminator; aligning the pseudo-unknown class sample features of the source domain with the unknown category features of the target domain using a second domain discriminator; and jointly training the first and second domain discriminators using the aligned known category features and the aligned unknown category features. This invention can expand the feature space of unknown categories, improve the perception and recognition capabilities of unknown categories, and achieve cross-domain feature transfer with generalization capabilities.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an open-set hyperspectral alteration mineral identification method and system based on a hierarchical filtering network. Background Technology

[0002] Cross-domain open-set hyperspectral alteration mineral identification aims to not only identify known categories in the target domain but also accurately discover and distinguish unknown categories. This is crucial for improving the security, reliability, and generalization ability of image recognition models in practical applications. However, this task is challenging to achieve, primarily because existing image recognition methods rely heavily on known category information in the source domain. During training, models mainly focus on feature learning and classification optimization for known categories, lacking the ability to effectively model the features of unknown categories. Therefore, in cross-domain scenarios, models often struggle to make accurate judgments when faced with unknown samples not encountered during training. It is particularly important to note that in cross-domain open-set identification tasks, unknown categories are unpredictable during training, making it impossible to directly obtain their prior distribution or category features. This poses a significant challenge to learning unknown category identification strategies.

[0003] Currently, a clear decision boundary is typically constructed between known and unknown categories, and uncertainty is predicted using a quantified model, combined with a threshold determination mechanism to identify the unknown category. For example, related methods usually determine whether the current sample should be classified as an unknown category based on the entropy value of the predicted probability distribution, the maximum class confidence, or a pre-set confidence threshold. However, the above methods usually rely on the premise that the model prediction results have high reliability, and are highly dependent on the accuracy of the model output. In the scenario of cross-domain open set hyperspectral alteration mineral identification, when the unknown category in the target domain shows high similarity with two or more known categories in the feature space, the model often struggles to effectively distinguish the unknown category and is prone to misclassifying it as a known category, resulting in high-confidence erroneous predictions and causing both the unknown category to be missed and the known category to be misclassified simultaneously.

[0004] Furthermore, existing methods typically rely on simple noise perturbations, reconstruction error simulations, or adversarial example generation to produce pseudo-unknown class samples when constructing training samples for unknown categories. However, these pseudo-unknown class samples often fail to adequately represent the complexity and diversity of real unknown categories in terms of spectral distribution, spatial structure, and category morphology. This leads to discrepancies between the unknown class discrimination features learned by the model and the actual scene, thereby reducing the model's adaptability and generalization performance in complex real-world environments.

[0005] Therefore, to improve the accuracy and robustness of cross-domain open set hyperspectral alteration mineral identification, it is urgent to construct a more complex, diverse, and realistic pseudo-unknown class generation mechanism to more fully simulate unknown targets that may appear in the target domain. Simultaneously, it is also necessary to further design more refined category decision boundaries and utilize the generated sample set with "unknown" characteristics to participate in model training, enabling the model to more accurately distinguish between known and unknown categories. In contrast, traditional open set classification methods focus more on utilizing known category information from the source domain, lacking targeted design for modeling unknown category features. Consequently, they struggle to effectively identify unknown categories in cross-domain scenarios, easily leading to increased misclassification, decreased recognition performance, and insufficient generalization ability. Summary of the Invention

[0006] This invention provides an open set hyperspectral alteration mineral identification method and system based on a hierarchical filtering network. It addresses the problem of misjudgment and degraded generalization performance in open set image recognition tasks due to over-reliance on known source domain category information during training and lack of ability to model unknown category features. The invention provides a solution using a hybrid centroid interpolation mechanism and a multi-decision boundary category screening strategy.

[0007] In a first aspect, the present invention provides an open-set hyperspectral alteration mineral identification method based on a hierarchical filtering network, the method comprising: Feature extraction was performed on the source domain hyperspectral image, background sample, and target domain hyperspectral image respectively, and the known category features of the source domain, background sample features, and target domain sample features were obtained accordingly. The known class centroids of the known class features of the source domain and the background sample centroids of the background sample features are fused to obtain pseudo-unknown class sample features of the source domain; Based on the pseudo-unknown class sample features in the source domain and the known class features in the source domain, the decision boundary of the multi-head binary classifier is determined, and based on the decision boundary, the target domain sample features are determined to be either known class features or unknown class features in the target domain; the classification result is determined according to the known class features or unknown class features in the target domain. Using a first domain discriminator, the distributions of known category features in the source domain and known category features in the target domain are aligned to obtain aligned known category features; using a second domain discriminator, the distributions of pseudo-unknown class sample features in the source domain and unknown category features in the target domain are aligned to obtain aligned unknown category features. The first domain discriminator and the second domain discriminator are jointly trained using the aligned known category features and the aligned unknown category features.

[0008] In some embodiments of the present invention, the known class centroids of the known class features of the source domain and the background sample centroids of the background sample features are fused to obtain pseudo-unknown class sample features of the source domain, including: Linear interpolation is performed between the known class centroid and the background sample centroid to obtain the source domain pseudo-unknown class sample features.

[0009] In some embodiments of the present invention, linear interpolation is performed between the known class centroid and the background sample centroid to obtain source domain pseudo-unknown class sample features, including: Using the following formula, linear interpolation is performed between the known class centroid and the background sample centroid to obtain the features of the pseudo-unknown class samples in the source domain: ; In the formula, This represents the features of pseudo-unknown class samples in the source domain. It is a centroid of a known category randomly selected from the set of known category centroids. It is a background sample centroid selected from the set of background sample centroids. These are coefficients sampled from the Beta distribution.

[0010] In some embodiments of the present invention, the multi-head binary classifier includes a class corresponding to a known class. A first category header and a second category header corresponding to the unknown category; The step of determining whether the target domain sample features are known category features or unknown category features based on the decision boundary includes: The target domain sample features are input into the first classification head and the second classification head respectively, and the first classification result and the second classification result are obtained accordingly. If the second classification result indicates that the category is unknown, the current target domain sample feature is determined to be the target domain unknown category feature; If the second classification result indicates that the category is not unknown, then The second classification result with the highest probability value among the second classification results is taken as the target result. If the target result is characterized as having an unknown category, the current target domain sample feature is determined as the target domain unknown category feature; if the target result is characterized as having a known category, the current target domain sample feature is determined as the target domain known category feature.

[0011] In some embodiments of the present invention, the multi-head binary classifier is trained based on the following loss function: ; In the formula, This is the total loss value. Given the number of known categories, It is a sample Sample features obtained by the feature extractor Sample features The first classification result is obtained through the first classification head prediction. Sample features The second classification result is obtained through prediction by the second classification head. It is a binary label for a known category. It is a binary label for unknown categories.

[0012] In some embodiments of the present invention, the loss value of the first domain discriminator Characterized as: ; In the formula, Expressing expectations; and They represent from Samples obtained from sampling and from Samples obtained from sampling; and express , After feature extractor The source domain and target domain have known category features. This represents a set of data with known categories in the source domain. This represents a set of data with known categories in the target domain. The loss value of the second domain discriminator Characterized as: ; In the formula, and They represent from Samples obtained from sampling and from Samples obtained from sampling; and These are the features of pseudo-unknown class samples in the source domain and the features of unknown class samples in the target domain, respectively. This represents a dataset of pseudo-unknown categories in the source domain. This represents the dataset of samples in the target domain that are identified as unknown categories.

[0013] In some embodiments of the present invention, the first domain discriminator corresponds to a first generator, and the second domain discriminator corresponds to a second generator; The loss value of the first generator is characterized as follows: ; The loss value of the second generator is characterized as follows: .

[0014] Secondly, the present invention also provides an open-set hyperspectral alteration mineral identification system based on a hierarchical filtering network, the system comprising: The feature extraction module is used to extract features from the source domain hyperspectral image, background sample, and target domain hyperspectral image respectively, and obtain the known category features of the source domain, background sample features, and target domain sample features accordingly. The fusion module is used to fuse the known class centroids of the known class features in the source domain and the background sample centroids of the background sample features to obtain pseudo-unknown class sample features in the source domain. The classification module is used to determine the decision boundary of a multi-head binary classifier based on the features of pseudo-unknown class samples in the source domain and the features of known class samples in the source domain, and to determine whether the features of the target domain samples are known class features or unknown class features in the target domain based on the decision boundary; and to determine the classification result based on the known class features or unknown class features in the target domain. The alignment module is used to align the distributions of known category features in the source domain with those in the target domain using a first domain discriminator to obtain aligned known category features; and to align the distributions of pseudo-unknown class sample features in the source domain with those in the target domain using a second domain discriminator to obtain aligned unknown category features. The joint training module is used to jointly train the first domain discriminator and the second domain discriminator using the aligned known category features and the aligned unknown category features.

[0015] In the open-set hyperspectral alteration mineral identification method and system based on hierarchical filtering networks provided in this invention, the model's ability to perceive unknown classes is improved by fusing the known class centroids of known category features in the source domain with the background sample centroids of background sample features. A multi-head binary classifier is used to construct a decision boundary for the source domain pseudo-unknown class sample features and the source domain known category features. Based on the decision boundary, the target domain sample features are classified to obtain the target domain known category features and the target domain unknown category features, achieving multi-level filtering of unknown samples and improving the model's accuracy in identifying unknown classes. A first domain discriminator aligns the distributions of the source domain known category features and the target domain known category features; a second domain discriminator aligns the distributions of the source domain pseudo-unknown class sample features and the target domain unknown category features, achieving more refined cross-domain feature transfer and improving the model's generalization ability and stability in the target domain. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is one of the flowcharts of the open set hyperspectral alteration mineral identification method based on hierarchical filtering network provided in the embodiments of the present invention;

[0018] Figure 2 This is the second flowchart of the open set hyperspectral alteration mineral identification method based on hierarchical filtering network provided in this embodiment of the invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0022] The use of "applies to" or "configured to" in this invention implies an open and inclusive language, which does not exclude the applicability to or configuration to devices performing additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more conditions or values ​​may in practice be based on additional conditions or values ​​beyond those conditions.

[0023] In this invention, the term "exemplary" is used to mean "serving as an example, illustration, or description." Any embodiment described as "exemplary" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0024] The following describes, with reference to the accompanying drawings, an open-set hyperspectral alteration mineral identification method and system based on a hierarchical filtering network provided by an embodiment of the present invention.

[0025] like Figure 1 As shown in the figure, this invention provides an open-set hyperspectral alteration mineral identification method based on a hierarchical filtering network, which includes the following steps:

[0026] S101 extracts features from the source domain hyperspectral image, background sample, and target domain hyperspectral image respectively, and obtains the known category features of the source domain, background sample features, and target domain sample features accordingly.

[0027] S102, the known class centroids of the known class features in the source domain and the background sample centroids of the background sample features are fused to obtain the pseudo-unknown class sample features in the source domain.

[0028] S103, based on the source domain pseudo-unknown class sample features and the source domain known class features, determine the decision boundary of the multi-head binary classifier, and based on the decision boundary, determine the target domain sample features as target domain known class features or target domain unknown class features; determine the classification result according to the target domain known class features or target domain unknown class features, that is, determine the classification accuracy of the target domain in the target domain hyperspectral image belonging to each known class and unknown class.

[0029] S104. Using the first domain discriminator, align the distribution of known category features in the source domain with the known category features in the target domain to obtain aligned known category features; using the second domain discriminator, align the distribution of pseudo-unknown category features in the source domain with the unknown category features in the target domain to obtain aligned unknown category features.

[0030] S105, by aligning known category features and aligning unknown category features, the first domain discriminator and the second domain discriminator are jointly trained.

[0031] It is understandable that joint training maximizes the discriminative power of both the first-domain discriminator and the second-domain discriminator.

[0032] The open-set hyperspectral alteration mineral identification method based on hierarchical filtering networks provided in this invention improves the model's ability to perceive unknown classes by fusing the known class centroids of known category features in the source domain with the background sample centroids of background sample features. A multi-head binary classifier is used to construct a decision boundary between the source domain pseudo-unknown class sample features and the source domain known category features. Based on this decision boundary, target domain sample features are classified to obtain target domain known category features and target domain unknown category features, achieving multi-level filtering of unknown samples and improving the model's accuracy in identifying unknown classes. A first domain discriminator aligns the distributions of source domain known category features and target domain known category features; a second domain discriminator aligns the distributions of source domain pseudo-unknown class sample features and target domain unknown category features, achieving more refined cross-domain feature transfer and improving the model's generalization ability and stability in the target domain.

[0033] In some embodiments of the present invention, S102 includes the following steps:

[0034] Linear interpolation is performed between the known class centroids and the background sample centroids to obtain the features of pseudo-unknown class samples in the source domain. .

[0035] In some examples, pre-trained feature extractors are utilized. Extract known categories (total) from the source domain hyperspectral image. Features of each category (i.e., features of known categories in the source domain). For each known category The known set of category features in the source domain is .

[0036] Each known category was calculated using the K-means clustering algorithm. Characteristic centroid That is, the class centroid is known. It is a representative point in the feature space of this category, which is determined by minimizing the in-cluster sum of squares (WCSS):

[0037] ;

[0038] in, Denotes the centroid vector. Represents the set of known category features in the source domain. One of the feature vectors, This represents the square of the Euclidean norm.

[0039] Background samples that do not overlap with any known categories are selected from the source domain hyperspectral image, and a pre-trained feature extractor is used. The background sample feature set is extracted from the background samples. And calculate its eigencentroid. .

[0040] Using the following formula, linear interpolation is performed between the known class centroid and the background sample centroid to obtain the features of the pseudo-unknown class samples in the source domain:

[0041] ;

[0042] In the formula, This represents the features of pseudo-unknown class samples in the source domain. It is a centroid of a known category randomly selected from the set of known category centroids. It is a background sample centroid selected from the set of background sample centroids. From Beta distribution The coefficients of the sampling in the middle, Let be the two shape parameters of the Beta distribution.

[0043] In some embodiments of the present invention, relying solely on a Softmax-based multi-classifier can easily misclassify truly unknown samples into a known category. To alleviate this overfitting phenomenon, a multi-decision boundary classification is employed, the core of which is to construct an independent binary classification decision boundary for each known category.

[0044] Multi-head binary classifier Including those corresponding to known categories A first category header and a second category header corresponding to the unknown category. Both the first and second classification heads are binary classification heads. The source domain hyperspectral image corresponds to... There are 1 known categories, each corresponding to a first category head, and the first category heads are independent of each other.

[0045] Accordingly, S103 includes the following sub-steps:

[0046] S1031, Input the target domain sample features into the first classification head and the second classification head respectively, and obtain the first classification result and the second classification result accordingly.

[0047] Understandably, the first classification header outputs two logical values, logits, which represent the categories to which the features of the target domain sample belong. Possibility and not belonging to category The possibility of this classification process can be expressed as:

[0048] ;

[0049] in, Indicates the first The first category head is responsible for classifying the first category. Categories Indicates the first The logical values ​​of each category, i.e., the first classification result.

[0050] S1032, when the second classification result indicates that the category is unknown, determine the current target domain sample features as target domain unknown category features.

[0051] S1033, when the second classification result indicates that the category is not unknown, The second classification result with the highest probability value among the second classification results is taken as the target result. When the target result is characterized as an unknown category, the current target domain sample feature is determined to be an unknown category feature of the target domain; when the target result is characterized as a known category, the current target domain sample feature is determined to be a known category feature of the target domain.

[0052] In some examples, a target domain sample feature is input. In the first category head, the corresponding obtained The first classification result. The final class determination of the target domain sample features is a two-stage composite decision, without setting an additional uniform threshold for the known class probabilities.

[0053] First, from Extract each known category from the first classification result. The logical value for "belongs to this category" in all known categories. Find the category with the highest "belongs to" logical value. .

[0054] Simultaneously, the second classification header outputs two logical values: one representing "not unknown" and the other representing "unknown". (Judgment) Which logical value is higher in the first classification result: If the logical value representing "unknown" is higher, then the sample is classified as "unknown" by this independent head.

[0055] If the independent unknown category header classifies a sample as "unknown," then regardless of how the known category headers classify it, the sample will ultimately be classified as "unknown." Otherwise, it will be classified into the category with the highest confidence among the known categories.

[0056] In some examples, the multi-head binary classifier is trained based on the following loss function:

[0057] ;

[0058] In the formula, This is the total loss value. Given the number of known categories, It is a sample Sample features obtained by the feature extractor Sample features The first classification result is obtained through the first classification head prediction. Sample features The second classification result is obtained through prediction by the second classification head. It is a binary label for a known category. It is a binary label for unknown categories.

[0059] It is understandable that binary labels are used for known categories. Specifically, if the sample Authentic Labels For category Then binary tags :if (Including other known categories and pseudo-unknown categories), then (Indicates that it does not belong to the category) ).

[0060] In some examples, binary cross-entropy Defined as:

[0061] ;

[0062] in, It is a real binary label vector. It is the value representing the positive category (i.e., belonging to the category). ), These are the predicted labels output by the model. This is the predicted value corresponding to the positive category. It is the Sigmoid activation function, used to convert logical values ​​into probabilities.

[0063] In some embodiments of the present invention, the loss value of the first domain discriminator Characterized as:

[0064] ;

[0065] In the formula, Expressing expectations; and They represent from Samples obtained from sampling and from Samples obtained from sampling; and express , After feature extractor The obtained features, that is Given known category features in the source domain, The target domain has known category features; This represents a set of data with known categories in the source domain. This represents the set of data with known categories in the target domain, that is, the set of samples in the target domain that are identified as known categories with high confidence.

[0066] Loss value of the second-domain discriminator Characterized as:

[0067] ;

[0068] In the formula, and They represent from Samples obtained from sampling and from Samples obtained from sampling; and These are the features of pseudo-unknown class samples in the source domain and the features of unknown class samples in the target domain, respectively. This represents a dataset of pseudo-unknown categories in the source domain. This represents the dataset of samples in the target domain that are identified as unknown categories.

[0069] In some embodiments of the present invention, the first domain discriminator corresponds to a first generator, and the second domain discriminator corresponds to a second generator;

[0070] The loss value of the first generator is characterized as follows:

[0071] .

[0072] The goal of the first generator in adversarial training is to minimize the discriminative power, that is, to minimize the "source domain" classification confidence of the target domain features, which is equivalent to minimizing its negative log-likelihood.

[0073] The loss value of the second generator is characterized as follows:

[0074] .

[0075] The training objective of the second generator is to minimize the discriminative power.

[0076] It should be noted that the first generator and the second generator provided in the embodiments of the present invention are the aforementioned feature extractors.

[0077] On the other hand, embodiments of the present invention also provide an open set hyperspectral alteration mineral identification system based on a hierarchical filtering network, the system including a feature extraction module, a fusion module, a classification module, an alignment module, and a joint training module.

[0078] The feature extraction module is used to extract features from the source domain hyperspectral image, background sample, and target domain hyperspectral image respectively, and obtain the known category features of the source domain, the background sample features, and the target domain sample features accordingly.

[0079] The fusion module is used to fuse the known class centroids of the known class features in the source domain with the background sample centroids of the background sample features to obtain pseudo-unknown class sample features in the source domain.

[0080] The classification module is used to determine the decision boundary of the multi-head binary classifier based on the features of pseudo-unknown class samples in the source domain and the features of known class samples in the source domain, and to determine whether the features of the target domain samples are known class features or unknown class features in the target domain based on the decision boundary; and to determine the classification result based on the known class features or unknown class features in the target domain.

[0081] The alignment module is used to align the distribution of known class features in the source domain with the known class features in the target domain using the first domain discriminator, thereby obtaining aligned known class features; and to align the distribution of pseudo-unknown class sample features in the source domain with the unknown class features in the target domain using the second domain discriminator, thereby obtaining aligned unknown class features.

[0082] The joint training module is used to jointly train the first domain discriminator and the second domain discriminator by aligning known category features with unknown category features.

[0083] The open set hyperspectral alteration mineral identification system based on hierarchical filtering network provided in this embodiment corresponds to the open set hyperspectral alteration mineral identification method based on hierarchical filtering network provided in the above embodiment, and will not be described again here.

[0084] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0086] The above provides a detailed description of an open-set hyperspectral alteration mineral identification method and system based on a hierarchical filtering network provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying open-set hyperspectral alteration minerals based on hierarchical filtering networks, characterized in that, The method includes: Feature extraction was performed on the source domain hyperspectral image, background sample, and target domain hyperspectral image respectively, and the known category features of the source domain, background sample features, and target domain sample features were obtained accordingly. The known class centroids of the known class features of the source domain and the background sample centroids of the background sample features are fused to obtain pseudo-unknown class sample features of the source domain; Based on the pseudo-unknown class sample features in the source domain and the known class features in the source domain, the decision boundary of the multi-head binary classifier is determined, and based on the decision boundary, the target domain sample features are determined to be either known class features or unknown class features in the target domain; the classification result is determined according to the known class features or unknown class features in the target domain. Using a first domain discriminator, the distributions of known category features in the source domain and known category features in the target domain are aligned to obtain aligned known category features; using a second domain discriminator, the distributions of pseudo-unknown class sample features in the source domain and unknown category features in the target domain are aligned to obtain aligned unknown category features. The first domain discriminator and the second domain discriminator are jointly trained using the aligned known category features and the aligned unknown category features.

2. The open-set hyperspectral alteration mineral identification method based on hierarchical filtering networks according to claim 1, characterized in that, The known class centroids of the known class features in the source domain and the background sample centroids of the background sample features are fused to obtain pseudo-unknown class sample features in the source domain, including: Linear interpolation is performed between the known class centroid and the background sample centroid to obtain the source domain pseudo-unknown class sample features.

3. The open-set hyperspectral alteration mineral identification method based on hierarchical filtering networks according to claim 2, characterized in that, Linear interpolation is performed between the known class centroids and the background sample centroids to obtain the source domain pseudo-unknown class sample features, including: Using the following formula, linear interpolation is performed between the known class centroid and the background sample centroid to obtain the features of the pseudo-unknown class samples in the source domain: ； In the formula, This represents the features of pseudo-unknown class samples in the source domain. It is a centroid of a known category randomly selected from the set of known category centroids. It is a background sample centroid selected from the set of background sample centroids. These are coefficients sampled from the Beta distribution.

4. The open-set hyperspectral alteration mineral identification method based on hierarchical filtering networks according to claim 1, characterized in that, The multi-head binary classifier includes classes corresponding to known categories. A first category header and a second category header corresponding to the unknown category; The step of determining whether the target domain sample features are known category features or unknown category features based on the decision boundary includes: The target domain sample features are input into the first classification head and the second classification head respectively, and the first classification result and the second classification result are obtained accordingly. If the second classification result indicates that the category is unknown, the current target domain sample feature is determined to be the target domain unknown category feature; If the second classification result indicates that the category is not unknown, then The second classification result with the highest probability value among the second classification results is taken as the target result. If the target result is characterized as having an unknown category, the current target domain sample feature is determined as the target domain unknown category feature; if the target result is characterized as having a known category, the current target domain sample feature is determined as the target domain known category feature.

5. The open-set hyperspectral alteration mineral identification method based on hierarchical filtering networks according to claim 4, characterized in that, The multi-head binary classifier is trained based on the following loss function: ； In the formula, This is the total loss value. Given the number of known categories, It is a sample Sample features obtained by the feature extractor Sample features The first classification result is obtained through the first classification head prediction. Sample features The second classification result is obtained through prediction by the second classification head. It is a binary label for a known category. It is a binary label for unknown categories.

6. The method for identifying open-set hyperspectral alteration minerals based on hierarchical filtering networks according to claim 1, characterized in that, Loss value of the first domain discriminator Characterized as: ； In the formula, Expressing expectations; and They represent from Samples obtained from sampling and from Samples obtained from sampling; and express , After feature extractor The source domain and target domain have known category features. This represents a set of data with known categories in the source domain. This represents a set of data with known categories in the target domain. The loss value of the second domain discriminator Characterized as: ； In the formula, and They represent from Samples obtained from sampling and from Samples obtained from sampling; and These are the features of pseudo-unknown class samples in the source domain and the features of unknown class samples in the target domain, respectively. This represents a dataset of pseudo-unknown categories in the source domain. This represents the dataset of samples in the target domain that are identified as unknown categories.

7. The open-set hyperspectral alteration mineral identification method based on hierarchical filtering networks according to claim 6, characterized in that, The first domain discriminator corresponds to a first generator, and the second domain discriminator corresponds to a second generator; The loss value of the first generator is characterized as follows: ； The loss value of the second generator is characterized as follows: 。 8. An open-set hyperspectral alteration mineral identification system based on a hierarchical filtering network, characterized in that, The system includes: The feature extraction module is used to extract features from the source domain hyperspectral image, background sample, and target domain hyperspectral image respectively, and obtain the known category features of the source domain, background sample features, and target domain sample features accordingly. The fusion module is used to fuse the known class centroids of the known class features in the source domain and the background sample centroids of the background sample features to obtain pseudo-unknown class sample features in the source domain. The classification module is used to determine the decision boundary of a multi-head binary classifier based on the features of pseudo-unknown class samples in the source domain and the features of known class samples in the source domain, and to determine whether the features of the target domain samples are known class features or unknown class features in the target domain based on the decision boundary; and to determine the classification result based on the known class features or unknown class features in the target domain. The alignment module is used to align the distributions of known category features in the source domain with those in the target domain using a first domain discriminator to obtain aligned known category features; and to align the distributions of pseudo-unknown class sample features in the source domain with those in the target domain using a second domain discriminator to obtain aligned unknown category features. The joint training module is used to jointly train the first domain discriminator and the second domain discriminator using the aligned known category features and the aligned unknown category features.

Images