A method for screening characteristic bands of minerals on the surface of Mars

Abstract

This invention provides a method for screening characteristic bands of Martian surface minerals, comprising: acquiring a standard hyperspectral dataset of Martian minerals, dividing it into a training hyperspectral dataset and a test hyperspectral dataset according to a preset division ratio, and performing data preprocessing on each dataset; constructing a band screening supervised learning network model, iteratively training it on the preprocessed training hyperspectral dataset to obtain an optimized band screening supervised learning network model, and determining the band screening results; performing band screening on the preprocessed test hyperspectral dataset based on the band screening results, and determining the test classification accuracy; and, if the test accuracy meets a preset accuracy condition, performing band screening on the Martian mineral hyperspectral dataset to be screened according to the optimized band screening supervised learning network model to obtain the target screening dataset. This method can accurately screen characteristic bands of Martian surface minerals, thereby improving the accuracy of Martian surface mineral identification and classification.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral remote sensing technology, specifically to a method for screening characteristic bands of minerals on the surface of Mars. Background Technology

[0002] Hyperspectral remote sensing is a primary method for identifying and classifying minerals on the Martian surface. While Martian hyperspectral remote sensing images possess high spectral dimensionality, they contain significant noise and redundant bands, with only a small portion containing mineral information. In practical applications of mineral identification and classification on the Martian surface, using the entire spectrum for mineral classification is insufficient for extracting mineral information and may lead to overfitting and other problems, resulting in unsatisfactory classification results. Band filtering of hyperspectral remote sensing images can effectively extract important information from the spectral data, thereby improving classification performance.

[0003] Currently, unsupervised learning-based band selection methods are commonly used, including clustering-based, ranking-based, and search-based methods. While these methods can effectively extract mineral spectral and structural features, they are relatively independent of spectral classification, meaning the extracted features do not contain mineral category information. This makes it difficult to accurately select characteristic bands for Martian surface minerals from hyperspectral data, thus reducing the accuracy of Martian surface mineral identification and classification. Summary of the Invention

[0004] The technical problem to be solved by this invention is that it is impossible to accurately screen the characteristic bands of Martian surface minerals in hyperspectral data, which leads to a decrease in the accuracy of Martian surface mineral identification and classification.

[0005] To address the aforementioned technical problems, this invention provides a method for screening characteristic bands of minerals on the Martian surface, specifically employing the following technical solution:

[0006] This invention provides a method for screening characteristic bands of minerals on the Martian surface, comprising: First, acquiring a standard hyperspectral dataset of Martian minerals and dividing it into a training hyperspectral dataset and a test hyperspectral dataset according to a preset division ratio. Then, preprocessing the training and test hyperspectral datasets respectively to obtain preprocessed training and test hyperspectral datasets. Second, constructing a band screening supervised learning network model. Iteratively training the band screening supervised learning network model using the preprocessed training hyperspectral dataset to obtain an optimized band screening supervised learning network model. Further, determining the band screening results based on the optimized band screening supervised learning network model. Performing band screening on the preprocessed test hyperspectral dataset based on the band screening results to obtain a test screening dataset. Classifying minerals based on the test screening dataset and determining the test classification accuracy. Then, if the test accuracy meets a preset accuracy condition, acquiring a Martian mineral hyperspectral dataset to be screened. Finally, performing band screening on the Martian mineral hyperspectral dataset to be screened using the optimized band screening supervised learning network model to obtain a target screening dataset.

[0007] This method employs a supervised learning network model to filter bands in Martian surface hyperspectral data, achieving dimensionality reduction of the original data. It fully utilizes the contextual information and spectral category information of different minerals within the hyperspectral data. While extracting the spectral and structural features of the Martian hyperspectral data, it uses spectral category information as the basis for band selection, ensuring that the selected bands contain mineral category information, thus facilitating mineral classification. Furthermore, the method can perform more targeted band selection for hyperspectral data in different application scenarios, demonstrating strong adaptability. This improves the accuracy of mineral identification and classification on the Martian surface.

[0008] In one alternative implementation, the aforementioned band selection supervised learning network model includes: a weight learning unit, a band selection unit, and a category learning unit. The weight learning unit is connected to the band selection unit, and the band selection unit is connected to the category learning unit. Specifically, the weight learning unit consists of a normalized layer (LN), a fully connected layer (FCN), and a ReLU activation function arranged sequentially; the category learning unit consists of a convolutional layer (CONV), a pooling layer (Pooling), a fully connected layer (FCN), and a ReLU activation function arranged sequentially.

[0009] In this implementation, the band-selection supervised learning network model with the above structure can be used to filter and classify Martian mineral hyperspectral datasets.

[0010] In one alternative implementation, the above-mentioned iterative training of the band selection supervised learning network model based on the preprocessed training hyperspectral dataset to obtain an optimized band selection supervised learning network model includes: First, inputting the preprocessed training hyperspectral dataset into the weight learning unit to obtain band weight combinations. Then, inputting the band weight combinations into the band selection unit to sort the band weight combinations in descending order of weight and determine the training band subset, wherein the data bands in the training band subset are the bands corresponding to a predetermined number of weights before weight sorting. Further, inputting the training band subset into the category learning unit for mineral classification to determine the training mineral classification results. Finally, constructing a loss function based on the training mineral classification results and iteratively optimizing the parameters to make the loss function converge, thereby obtaining the optimized band selection supervised learning network model.

[0011] In one alternative implementation, the expression for the band weights in the above band weight combination is:

[0012] bw = f(x; θ) a );

[0013] Where bw represents the band weight, f represents the weight learning unit, x represents the data in the preprocessed training hyperspectral dataset, and θ a Represents the parameters of the weighted learning unit;

[0014] Specifically, the expansion of f is:

[0015]

[0016] θ a The expression is:

[0017]

[0018] Among them, z i ,i=1,2…na represents intermediate variables, σ i ,i=1,2…na represents the ReLU activation function of the weight learning unit, H i ,i=1,2…na represents na fully connected layers, norm() represents layer normalization, w i ,i=1,2…na represents the weights of the fully connected layer, b i ,i=1,2…na represents the bias of the fully connected layer, and na is the number of layers of weight learning units.

[0019] In one alternative implementation, the output expression of the above category learning unit is:

[0020]

[0021] in, The output of the category learning unit is represented by f, where f represents the category learning unit, x′ represents the data in the training band subset, and θ represents the output of the category learning unit. c Parameters representing category learning units;

[0022] Specifically, the expansion of g is:

[0023]

[0024] θ c The expression is:

[0025]

[0026] Among them, s i ,i=1,2…nc represents intermediate variables, α i ,i=1,2…nc represents the ReLU activation function of the category learning unit, h′ i ,i=1,2…nc represents nc fully connected layers, w′ i ,i=1,2…nc represent the weights of each fully connected layer, b′ i ,i=1,2…nc represents the bias of each fully connected layer, and nc is the number of layers for class learning units.

[0027] In one alternative implementation, the loss function described above is the cross-entropy loss function.

[0028] In one alternative implementation, the above-mentioned mineral classification based on the test screening dataset and determination of the test classification accuracy includes: first, classifying minerals using category learning units based on the test screening dataset to obtain mineral classification test results; then, determining the test classification accuracy based on the mineral classification test results and the mineral classification standard results.

[0029] In one alternative implementation, the expression for the above test classification accuracy is:

[0030]

[0031] in,

[0032]

[0033] Where ACC represents the test classification accuracy, and n′ represents the size of the preprocessed test hyperspectral dataset. This represents the mineral classification test result, y′ i This indicates the results of the mineral classification standards.

[0034] In one alternative implementation, the above-mentioned data preprocessing of the training hyperspectral dataset and the test hyperspectral dataset includes at least one of the following: cropping, smoothing, continuum removal, and normalization.

[0035] In one alternative implementation, the aforementioned Mars mineral standard hyperspectral dataset is the Compact Mars Reconnaissance Imaging Spectroradiometer (CRISM) dataset. The Mars mineral standard hyperspectral dataset includes a ratio of I radiance to F incident solar radiation. The preset split ratio is 4:1 for the training hyperspectral dataset and the test hyperspectral dataset. Attached Figure Description

[0036] Figure 1 A schematic flowchart of the Martian surface mineral feature band screening method provided in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the structure of the band-screening supervised learning network model provided in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the band selection results provided in an embodiment of the present invention. Detailed Implementation

[0039] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0040] Hyperspectral remote sensing is a primary method for identifying and classifying minerals on the Martian surface. While Martian hyperspectral remote sensing images possess high spectral dimensionality, they contain significant noise and redundant bands, with only a small portion containing mineral information. In practical applications of mineral identification and classification on the Martian surface, using the entire spectrum for mineral classification is insufficient for extracting mineral information and may lead to overfitting and other problems, resulting in unsatisfactory classification results. Band filtering of hyperspectral remote sensing images can effectively extract important information from the spectral data, thereby improving classification performance.

[0041] Currently, unsupervised learning-based band selection methods are commonly used, including clustering-based, ranking-based, and search-based methods. While these methods can effectively extract mineral spectral and structural features, they are relatively independent of spectral classification, meaning the extracted features do not contain mineral category information. This leads to reduced accuracy in identifying and classifying minerals on the Martian surface.

[0042] To address the aforementioned issues, this application provides a method for screening characteristic bands of minerals on the Martian surface. This method employs supervised learning, extracting spectral and structural features from Martian hyperspectral data while using spectral category information as the basis for band screening. This ensures that the selected bands contain mineral category information, which is more conducive to mineral classification and thus improves the accuracy of identifying and classifying minerals on the Martian surface.

[0043] The solutions provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0044] For details, see Figure 1 The above is a schematic flowchart of the Martian surface mineral feature band screening method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the Martian surface mineral feature band screening method based on hyperspectral data provided by the present invention includes the following steps S101-S109:

[0045] S101. Obtain the standard hyperspectral dataset of Martian minerals, and divide the standard hyperspectral dataset of Martian minerals into a training hyperspectral dataset and a test hyperspectral dataset according to a preset division ratio.

[0046] In this embodiment of the application, the Mars mineral standard hyperspectral dataset can be a Compact Reconnaissance Imaging Spectromer for Mars (CRISM) dataset.

[0047] In one implementation, the Martian mineral standard hyperspectral dataset can be obtained from a standard spectral library obtained through CRISM analysis (MICA). The Martian mineral standard hyperspectral dataset can include 24 mineral classes.

[0048] In some embodiments, the Mars mineral standard hyperspectral dataset includes a ratio I / F, which is the ratio of radiance to incident solar radiation. The ratio I / F is least affected by background information in remotely sensed images and exhibits the most pronounced absorption characteristics related to mineral categories. Therefore, the Mars mineral standard hyperspectral dataset can employ the ratio I / F.

[0049] In some embodiments, the preset partitioning ratio can be set according to the needs of the actual application. For example, the preset partitioning ratio could be: a 4:1 ratio of training hyperspectral dataset to test hyperspectral dataset. This can improve the training effect of the band selection supervised learning network model.

[0050] S102. Perform data preprocessing on the training hyperspectral dataset and the test hyperspectral dataset respectively to obtain the preprocessed training hyperspectral dataset and the preprocessed test hyperspectral dataset.

[0051] In some embodiments, the training hyperspectral dataset and the test hyperspectral dataset are respectively preprocessed, including at least one of the following: cropping, smoothing, continuum removal, and normalization.

[0052] Specifically, the cropping process involves extracting hyperspectral data at the target spectral wavelength. For example, the spectral wavelength range in the MICA spectral library is 0-4 μm; cropping can extract the 1.0-2.6 μm portion as the preprocessed training and test hyperspectral datasets. Smoothing involves denoising both the training and test hyperspectral datasets to remove sharp peaks. Continuum removal removes continuum from the hyperspectral data to highlight absorption features. Standardization normalizes the hyperspectral data values ​​to near 0, making the hyperspectral data in both the training and test datasets more regular.

[0053] S103. Construct a supervised learning network model for band selection.

[0054] In some embodiments, Figure 2 This is a schematic diagram of the structure of the band-screening supervised learning network model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the band selection supervised learning network model includes: a weight learning unit, a band selection unit, and a category learning unit. The weight learning unit is connected to the band selection unit, and the band selection unit is connected to the category learning unit.

[0055] Specifically, the weight learning unit consists of a normalized layer (LN), a fully connected layer (FCN), and a ReLU activation function arranged in sequence; the category learning unit consists of a convolutional layer (CONV), a pooling layer (Pooling), a fully connected layer (FCN), and a ReLU activation function arranged in sequence.

[0056] like Figure 2 As shown, the weight learning unit can be used to determine the band weight combination based on the preprocessed training hyperspectral dataset as input. The band selection unit can be used to sort the band weight combinations in descending order of weight and determine the band selection result, i.e., the band subset. The category learning unit can be used to determine the mineral classification result based on the band subset.

[0057] S104. The band selection supervised learning network model is iteratively trained based on the preprocessed training hyperspectral dataset to obtain an optimized band selection supervised learning network model.

[0058] In some embodiments, with Figure 2 Taking the band selection supervised learning network model shown as an example, S104 can specifically include the following steps S1041-S1044:

[0059] S1041. Input the preprocessed training hyperspectral dataset into the weight learning unit to obtain the band weight combination.

[0060] For example, if the preprocessed training hyperspectral dataset is denoted by D, then D∈R nXb Where n is the number of data in the preprocessed training hyperspectral dataset, and b is the number of spectral channels, i.e., the number of bands, for example, 2^32.

[0061] After the preprocessed training hyperspectral dataset D enters the weight learning unit, it will sequentially pass through multiple combined structures. Each combined structure consists of a normalized layer (LN), a fully connected layer (FCN), and a ReLU activation function.

[0062] In some embodiments, the expression for the band weights in the band weight combination is:

[0063] bw = f(x; θ) a );

[0064] Where bw represents the band weight, f represents the weight learning unit, x represents the data in the preprocessed training hyperspectral dataset, and θ a This represents the parameters of the weighted learning unit.

[0065] Specifically, the expansion of f is:

[0066]

[0067] θ a The expression is:

[0068] Among them, z i ,i=1,2…na represents intermediate variables, σ i ,i=1,2…na represents the ReLU activation function of the weight learning unit, h i ,i=1,2…na represents na fully connected layers, norm() represents layer normalization, w i ,i=1,2…na represents the weights of the fully connected layer, b i ,i=1,2…na represents the bias of the fully connected layer. The weights and biases together constitute the parameters θ of the weight learning unit. a na represents the number of layers containing weighted learning units. Specifically, na is an adjustable hyperparameter.

[0069] S1042. Input the band weight combination into the band filtering unit to sort the band weight combination in descending order of weight, and determine the training band subset. The data bands in the training band subset are: the bands corresponding to a preset number of weights before weight sorting.

[0070] Specifically, the band selection unit can sort the band weight combinations in descending order of weight, and use the data of the bands corresponding to the top preset number of weights with the largest weights as the training band subset.

[0071] For example, the band weights obtained by the weight learning unit can be expressed as:

[0072] bw = [bw1, bw2, ..., bw] b ];

[0073] Sort the elements in bw from largest to smallest. After sorting, their indices in bw are:

[0074] i j j = 1, 2, ..., k, i j ∈[1,b];

[0075] Furthermore, a transformation matrix T∈R can be constructed. bXk The elements of T satisfy:

[0076]

[0077] The result of band selection, i.e., the training band subset output by the band selection unit, is:

[0078] D′=D·T∈R nXk ;

[0079] After band selection, each training data point is represented by x′, where x′∈R 1Xk Let B′ represent the set formed by the selected bands.

[0080] S1043. Input the training band sub-dataset into the category learning unit to classify minerals and determine the training mineral classification results.

[0081] In some embodiments, the output expression of the category learning unit is:

[0082]

[0083] in, The output of the category learning unit is represented by f, where f represents the category learning unit, x′ represents the data in the training band subset, and θ represents the output of the category learning unit. c The parameters represent the category learning units.

[0084] Specifically, the expansion of g is:

[0085]

[0086] θ c The expression is:

[0087]

[0088] Among them, s i ,i=1,2…nc represents intermediate variables, α i ,i=1,2…nc represents the ReLU activation function of the category learning unit, h′ i ,i=1,2…nc represents nc fully connected layers, w′ i , i = 1, 2…nc represents the weights of each fully connected layer, b i Let ', i = 1, 2…nc represent the bias of each fully connected layer. The weights and biases of the fully connected layers constitute the parameters of the category learning unit, and nc is the number of layers for the category learning unit. Specifically, nc is an adjustable hyperparameter.

[0089] S1044. Construct a loss function based on the training mineral classification results, iteratively optimize the parameters to make the loss function converge, and obtain the optimized band screening supervised learning network model.

[0090] In some embodiments, the loss function is the cross-entropy loss function.

[0091] Specifically, the cross-entropy loss function can be expressed as:

[0092]

[0093] Where c represents the number of categories, for example, in this embodiment, c can be 24. i This represents the class label in the classification results of the training data. The model parameters used after training are... express.

[0094] S105. Determine the band selection results based on the optimized band selection supervised learning network model.

[0095] For example, the band selection result can be represented as:

[0096]

[0097] In this dataset, the elements of bw′ are sorted from largest to smallest, and the band data corresponding to the largest preset number of weights are used as the training band subset B′. For example, the preset number can be 50. The preset number can be set according to the needs of actual applications, and this application does not impose a specific limitation on it.

[0098] S106. Based on the band selection results, perform band selection on the preprocessed test hyperspectral dataset to obtain the test selected dataset.

[0099] For example, the preprocessed test hyperspectral dataset can be represented by S, S∈R n′Xb Where n′ represents the number of data points in the preprocessed hyperspectral test dataset, and b represents the number of spectral channels, i.e., the number of bands, for example, 2^32. Each test data point is represented by x, where x∈R. 1Xb Let B represent the set consisting of the original bands.

[0100] Based on the training band subset B′, construct the transformation matrix T∈R. bXk Then the test filtering dataset S′ after band filtering is:

[0101] S′=S·T∈R nXk ;

[0102] In S′, each test data point is represented by x′, where x′∈R 1Xk .

[0103] S107. Classify minerals based on the test dataset and determine the test classification accuracy.

[0104] Specifically, firstly, mineral classification can be performed using category learning units on the selected test dataset to obtain mineral classification test results. Then, based on the mineral classification test results and the mineral classification standard results, the test classification accuracy is determined.

[0105] For example, by passing the above x′ through the trained category learning unit, the mineral classification test result can be obtained.

[0106]

[0107] Will By comparing the result y′ with the mineral classification standard result, the accuracy of the test classification can be obtained. In one implementation, the mineral classification standard result y′ can be obtained from the MICA standard spectral library.

[0108] In some embodiments, the expression for testing classification accuracy is:

[0109]

[0110] in,

[0111]

[0112] Where ACC represents the test classification accuracy, and n′ represents the size of the preprocessed test hyperspectral dataset. This represents the mineral classification test result, y′ i This indicates the results of mineral classification standards. For example, the results of mineral classification standards can be mineral category labels, which can be obtained from the MICA standard spectral library.

[0113] S108. If the test accuracy meets the preset accuracy conditions, obtain the Martian mineral hyperspectral dataset to be screened.

[0114] Specifically, taking the above-mentioned test classification accuracy ACC as an example, the higher the ACC, the more beneficial the band screening results are for mineral classification processing.

[0115] In some embodiments, corresponding preset accuracy conditions can be set according to different application scenarios and application requirements. For example, corresponding preset accuracy conditions can be set according to different Martian regions. Preset accuracy conditions can be, for example, greater than 85%, greater than 90%, etc. If the test accuracy meets the preset accuracy conditions, it can be determined that the optimized band screening supervised learning network model can meet the application requirements. Furthermore, a hyperspectral dataset of Martian minerals to be screened can be obtained for band screening.

[0116] S109. Based on the optimized band-screening supervised learning network model, band screening is performed on the hyperspectral dataset of Martian minerals to be screened to obtain the target screening dataset.

[0117] Specifically, the hyperspectral dataset of Martian minerals to be screened can be input into an optimized band screening supervised learning network model, which will then pass through a weight learning unit and a band screening unit in sequence, and finally output the target screening dataset.

[0118] The Martian surface mineral feature band screening method based on hyperspectral data provided in this invention employs a supervised learning network model to perform band screening on Martian surface hyperspectral data, achieving dimensionality reduction of the original data. This method fully utilizes the contextual information and spectral category information of different minerals in the hyperspectral data. While extracting the spectral and structural features of the Martian hyperspectral data, it uses spectral category information as the basis for band screening, ensuring that the selected bands contain mineral category information, which is more conducive to mineral classification. Furthermore, the method can perform more targeted band screening on hyperspectral data in different application scenarios, exhibiting strong adaptability. This improves the accuracy of mineral identification and classification on the Martian surface.

[0119] In some embodiments, to verify the effectiveness of the Martian surface mineral feature band screening method based on hyperspectral data provided in this invention, the CRISM Martian mineral standard hyperspectral dataset was obtained through the MICA standard spectral library. This dataset was divided into a training hyperspectral dataset and a test hyperspectral dataset at a 4:1 ratio, and then pruned, smoothed, removed continuum, and normalized, respectively. Then, a dataset was constructed as follows... Figure 2 The band selection supervised learning network model shown includes a weight learning unit, a band selection unit, and a category learning unit. The preprocessed training hyperspectral dataset is then input into the band selection supervised learning network model for iterative training. This includes constructing a loss function, iteratively optimizing parameters to gradually reduce the loss function until convergence, thus completing model training. In this embodiment, the loss function is the cross-entropy loss function, and the number of training epochs is set to 20. Next, the preprocessed test hyperspectral dataset is input into the trained and optimized band selection supervised learning network model. A subset of Martian mineral spectral bands of size 50 is output from the band selection unit to obtain the selection dataset. Figure 3 This is a schematic diagram of the band selection results provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the original spectral curve is curve B, and the curve segment on curve B that coincides with region A is the spectral curve corresponding to the band screening result. The band screening method provided in this embodiment of the invention can effectively screen the characteristic bands of minerals on the Martian surface, thereby improving the accuracy of mineral identification and classification on the Martian surface.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0121] In the description of this invention, it should be understood that 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 indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0122] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0123] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0124] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. A method for screening characteristic bands of minerals on the surface of Mars, characterized in that, include: Obtain a standard hyperspectral dataset of Martian minerals, and divide the standard hyperspectral dataset of Martian minerals into a training hyperspectral dataset and a test hyperspectral dataset according to a preset division ratio; The training hyperspectral dataset and the test hyperspectral dataset are preprocessed to obtain the preprocessed training hyperspectral dataset and the preprocessed test hyperspectral dataset, respectively. Construct a supervised learning network model for band selection; The band selection supervised learning network model is iteratively trained based on the preprocessed training hyperspectral dataset to obtain an optimized band selection supervised learning network model. Based on the optimized band selection supervised learning network model, the band selection result is determined; Based on the band selection results, the preprocessed test hyperspectral dataset is subjected to band selection to obtain the test selected dataset. Based on the test screening dataset, mineral classification is performed, and the test classification accuracy is determined; Under the condition that the test accuracy meets the preset accuracy condition, obtain the Martian mineral hyperspectral dataset to be screened; The optimized band-selection supervised learning network model is used to perform band selection on the Martian mineral hyperspectral dataset to be selected, thereby obtaining the target selection dataset; The band selection supervised learning network model includes: a weight learning unit, a band selection unit, and a category learning unit; the weight learning unit is connected to the band selection unit, and the band selection unit is connected to the category learning unit. The weight learning unit consists of a normalized layer (LN), a fully connected layer (FCN), and a ReLU activation function arranged in sequence; the category learning unit consists of a convolutional layer (CONV), a pooling layer (Pooling), a fully connected layer (FCN), and a ReLU activation function arranged in sequence. The step of iteratively training the band selection supervised learning network model based on the preprocessed training hyperspectral dataset to obtain an optimized band selection supervised learning network model includes: The preprocessed training hyperspectral dataset is input into the weight learning unit to obtain the band weight combination. The band weight combination is input into the band filtering unit to sort the band weight combination in descending order of weight and determine the training band subset. The data bands in the training band subset are: bands corresponding to a preset number of weights before weight sorting. The training band sub-dataset is input into the category learning unit for mineral classification to determine the training mineral classification result; Based on the training mineral classification results, a loss function is constructed, and the parameters are iteratively optimized to make the loss function converge, thereby obtaining the optimized band screening supervised learning network model.

2. The method according to claim 1, characterized in that, The expression for the band weights in the band weight combination is: ； in, Indicates the band weight. Represents a weighted learning unit. This represents the data in the preprocessed training hyperspectral dataset. Represents the parameters of the weighted learning unit; Specifically, The expansion is: ； The expression is: ； in, Indicates intermediate variables. This represents the ReLU activation function of the weighted learning unit. express A fully connected layer, Representation layer normalization, Indicates the weights of the fully connected layer. Indicates the bias of the fully connected layer. This represents the number of layers for weighted learning units.

3. The method according to claim 1, characterized in that, The output expression of the category learning unit is: ； in, This represents the output of the category learning unit. Represents category learning units, This represents the data in the training band subset. Parameters representing category learning units; Specifically, The expansion is: ； The expression is: ； in, Indicates intermediate variables. Represents the ReLU activation function of the category learning unit. express A fully connected layer, This represents the weights of each fully connected layer. This represents the bias of each fully connected layer. The number of layers for category learning units.

4. The method according to claim 1, characterized in that, The loss function is the cross-entropy loss function.

5. The method according to claim 1, characterized in that, The step of classifying minerals based on the test dataset and determining the test classification accuracy includes: Based on the test-selected dataset, mineral classification is performed using the category learning unit to obtain mineral classification test results; The accuracy rate of the mineral classification test is determined based on the results of the mineral classification test and the results of the mineral classification standard.

6. The method according to claim 5, characterized in that, The expression for the test classification accuracy is: ； in, ； in, This indicates the test classification accuracy. This indicates the size of the preprocessed test hyperspectral dataset. This indicates the results of the mineral classification test. This indicates the results of the mineral classification standards.

7. The method according to claim 1, characterized in that, The data preprocessing performed on the training hyperspectral dataset and the test hyperspectral dataset includes at least one of the following: cropping, smoothing, continuum removal, and standardization.

8. The method according to claim 1, characterized in that, The Mars mineral standard hyperspectral dataset is the Compact Mars Reconnaissance Imaging Spectroradiometer (CRISM) dataset; the Mars mineral standard hyperspectral dataset includes the ratio I radiance / F incident solar radiation; the preset division ratio is: training hyperspectral dataset : test hyperspectral dataset = 4:1.

Images