Mars surface mineral classification method based on hyperspectral data

Abstract

The invention provides a Mars surface mineral classification method based on hyperspectral data. The Mars surface mineral classification method comprises the following steps: dividing a preprocessed Mars surface mineral standard hyperspectral data set into a training data set, a verification data set and a test data set; then, constructing a mineral classification deep learning model; and performing training optimization on the mineral classification deep learning model according to the training data set and the verification data set. And the optimized mineral classification deep learning model is tested according to the test data set, and the test classification accuracy is determined. And finally, under the condition that the test accuracy meets a preset accuracy condition, inputting to-be-identified Mars surface mineral hyperspectral data into the optimized mineral classification deep learning model, and outputting to obtain a Mars surface mineral classification result. Thus, richer feature representation in the hyperspectral data of the Mars surface minerals can be learned through the mineral classification deep learning model, more accurate mineral classification is achieved, and therefore the accuracy of classification of the Mars surface minerals can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a method for classifying Martian surface minerals based on hyperspectral data. Background Art

[0002] Hyperspectral remote sensing technology is the main means of classifying minerals on the Martian surface. Hyperspectral remote sensing data contains rich spectral and spatial information, and has the characteristics of a unified map and spectrum. Different types of minerals have their own unique spectral characteristics, and mineral types and mineral components can be identified based on the diagnostic spectral characteristics of minerals. Therefore, Mars hyperspectral data has been widely used in the study of mineral classification on the Martian surface.

[0003] At present, the classification methods of Martian surface minerals based on hyperspectral data are usually based on spectral similarity comparison, browsing products and machine learning. Although these methods have shown good results in the classification of Martian surface minerals, they still have certain limitations. For example, they have insufficient feature extraction capabilities, low automation, and low recognition accuracy. The existence of these limitations has led to the existing Martian surface mineral classification methods having problems such as low classification accuracy and low automation, making it difficult to achieve large-scale application. Summary of the invention

[0004] The technical problem to be solved by the present invention is the inability to accurately classify the minerals on the surface of Mars.

[0005] In order to solve the above technical problems, the present invention provides a method for classifying Martian surface minerals based on hyperspectral data, which specifically adopts the following technical solutions:

[0006] The present invention provides a method for classifying minerals on the surface of Mars based on hyperspectral data, comprising: first, obtaining a standard hyperspectral data set of minerals on the surface of Mars. Preprocessing the standard hyperspectral data set of minerals on the surface of Mars to obtain a preprocessed standard hyperspectral data set of minerals on the surface of Mars. Then, the preprocessed standard hyperspectral data set of minerals on the surface of Mars is divided into: a training data set, a verification data set, and a test data set according to a preset division ratio. Next, a mineral classification deep learning model is constructed, and the mineral classification deep learning model is used to identify and classify minerals on the surface of Mars. The mineral classification deep learning model is trained according to the training data set and the verification data set to obtain an optimized mineral classification deep learning model. The optimized mineral classification deep learning model is tested according to the test data set, and the test classification accuracy is determined. Finally, when the test accuracy meets the preset accuracy condition, the hyperspectral data of the minerals on the surface of Mars to be identified is input into the optimized mineral classification deep learning model, and the classification result of the minerals on the surface of Mars is obtained by output.

[0007] In this method, the importance of the hyperspectral data channel of Martian surface minerals can be obtained through the mineral classification deep learning model, the internal correlation of the data can be captured, and the valuable information in the hyperspectral data of Martian surface minerals can be selectively extracted, while ignoring unimportant information. In this way, the mineral classification deep learning model can learn richer feature representations in the hyperspectral data of Martian surface minerals, and achieve more accurate mineral classification, thereby effectively improving the accuracy of Martian surface mineral classification, which has important guiding significance for the development of Martian resource exploration projects and the development and utilization of deep space resources.

[0008] In an optional implementation, the mineral classification deep learning model includes: a channel selector and a classifier, and the channel selector and the classifier are connected. Among them, the channel selector is composed of a plurality of channel attention modules connected in series, and the channel attention module is composed of four first fully connected layers, a channel-by-channel weighted layer and a residual connection layer connected in series; the first fully connected layer is connected with a corresponding first activation function and a normalization layer. The structural type of the classifier is: a multi-layer perceptron structure, or a convolutional neural network structure.

[0009] In an optional implementation, the above-mentioned training of the mineral classification deep learning model according to the training data set and the verification data set to obtain an optimized mineral classification deep learning model includes: first, the training data set and the verification data set are respectively input into the channel selector, and the output is the first channel selector output result corresponding to the training data set, and the second channel selector output result corresponding to the verification data set. Then, the first channel selector output result and the second channel selector output result are respectively input into the classifier, and the output is the first predicted classification result corresponding to the first channel selector output result, and the second predicted classification result corresponding to the second channel selector output result. Next, a loss function is constructed, and the hyperparameters of the mineral classification deep learning model are iteratively optimized according to the first predicted classification result, so that the loss function converges to obtain the hyperparameter training result. Finally, the hyperparameter training result is optimized and adjusted according to the second predicted classification result to obtain an optimized mineral classification deep learning model.

[0010] In an optional implementation, the expression of the above channel selector is:

[0011] x′=S(x;θ s );

[0012] Among them, S represents the channel selector, x′ represents the output result of the channel selector, x represents the input hyperspectral data, θ s Represents the set of parameters in the channel selector. When the channel selector is composed of N channel attention modules in series, the expansion of the channel selector is:

[0013]

[0014] Among them, f (i) represents the i-th channel attention module, t (i) represents the output of the i-th channel attention module, θ s (i) represents the parameters of the i-th channel attention module, i=1,2…N.

[0015] In an alternative implementation, the expression of the above channel attention module is:

[0016]

[0017] Among them, I represents the input of the channel attention module, O represents the output of the channel attention module, and z (1) Represents the output result of the first fully connected layer, z (2) Represents the output result of the second first fully connected layer, z (3) Represents the output result of the third first fully connected layer, z (4) represents the output result of the 4th first fully connected layer, θ sc (1) represents the parameters of the first fully connected layer, θ sc (2) represents the parameters of the second first fully connected layer, θ sc (3) represents the parameters of the third first fully connected layer, θ sc (4) represents the parameters of the 4th first fully connected layer, γ1 represents the first activation function, σ represents the Sigmoid activation function, L1 represents the first fully connected layer, and LN1 represents the first normalized layer. Represents element-wise multiplication.

[0018] In an optional implementation, the expression of the above classifier is:

[0019]

[0020] Among them, C represents the classifier, x′ represents the output result of the channel selector, represents the classification result of Martian surface minerals output by the classifier, θ c Represents a set of parameters in a classifier.

[0021] In an optional implementation, when the structure type of the classifier is a multi-layer perceptron structure, the classifier includes: 4 second fully connected layers, and the second fully connected layers are connected with corresponding second activation functions. Then the expansion of the classifier is:

[0022]

[0023] Among them, x′ represents the output result of the channel selector, Represents the Martian surface mineral classification results output by the classifier, Represents the output result of the first second fully connected layer, Represents the output result of the second second fully connected layer, Represents the output result of the third second fully connected layer, θ c1 (1) represents the parameters of the first second fully connected layer, θ c1 (2) represents the parameters of the second fully connected layer, θ c1 (3) represents the parameters of the third second fully connected layer, θ c1 (4) represents the parameters of the 4th second fully connected layer, γ2 represents the second activation function, σ represents the Sigmoid activation function, L2 represents the second fully connected layer, and LN2 represents the second normalization layer.

[0024] In an optional implementation, when the structure type of the classifier is a convolutional neural network structure, the classifier includes: 3 convolutional layers, 2 third fully connected layers and 1 Flatten layer, and the third fully connected layer is connected with a corresponding third activation function. Then the expansion of the classifier is:

[0025]

[0026] Among them, x′ represents the output result of the channel selector, Represents the Martian surface mineral classification results output by the classifier, Represents the output result of the first convolutional layer, Represents the output of the second convolutional layer. Represents the output result of the third convolutional layer. Represents the output result of the Flatten layer. represents the output of the third fully connected layer, θ c2 (1) represents the parameters of the first convolutional layer, θ c2 (2) represents the parameters of the second convolutional layer, θ c2 (3) represents the parameters of the third convolutional layer, θ c2 (4) represents the parameters of the first third fully connected layer, θ c2 (5)represents the parameters of the second third fully connected layer, γ3 represents the third activation function, σ represents the Sigmoid activation function, L3 represents the third fully connected layer, LN3 represents the third normalization layer, CV represents the convolutional layer, MP represents the maximum pooling layer, and FL represents the Flatten layer.

[0027] In an optional implementation, the above loss function is a cross entropy loss function, and the hyperparameters of the mineral classification deep learning model include: the number of channel attention modules and the structural type of the classifier.

[0028] In an alternative implementation, the expression for testing classification accuracy is:

[0029]

[0030] in,

[0031]

[0032] Among them, ACC represents the test classification accuracy, d′ represents the size of the test data set, Represents the results of the Mars surface mineral classification test, y′ i Represents the standard results of the Martian surface mineral classification.

[0033] In an optional implementation, the above-mentioned Martian mineral standard hyperspectral data set includes: iron / magnesium montmorillonite spectral data, magnesium carbonate spectral data and fayalite spectral data; the preset division ratio is: training data set: verification data set: test data set is 8:1:1. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of the principle of a method for classifying Martian surface minerals based on hyperspectral data provided by an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of a process for classifying Martian surface minerals based on hyperspectral data provided by an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of the structure of a deep learning model for mineral classification provided by an embodiment of the present invention;

[0037] Figure 4 A schematic diagram of the mineral mapping results provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following embodiments do not represent all implementations consistent with the present application. They are only examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

[0039] Hyperspectral remote sensing technology is the main means of classifying minerals on the Martian surface. Hyperspectral remote sensing data contains rich spectral and spatial information, and has the characteristics of a unified map and spectrum. Different types of minerals have their own unique spectral characteristics, and mineral types and mineral components can be identified based on the diagnostic spectral characteristics of minerals. Therefore, Mars hyperspectral data has been widely used in the study of mineral classification on the Martian surface.

[0040] At present, the classification methods of Martian surface minerals based on hyperspectral data are usually based on spectral similarity comparison, browsing products and machine learning. Although these methods have shown good results in the classification of Martian surface minerals, they still have certain limitations. For example, they have insufficient feature extraction capabilities, low automation, and low recognition accuracy. The existence of these limitations has led to the existing Martian surface mineral classification methods having problems such as low classification accuracy and low automation, making it difficult to achieve large-scale application.

[0041] In order to solve the above problems, the present invention provides a method for classifying Martian surface minerals based on hyperspectral data. Figure 1 A schematic diagram of the principle of the method for classifying Martian surface minerals based on hyperspectral data provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method can divide the preprocessed Mars surface mineral standard hyperspectral data set into a training data set, a validation data set, and a test data set. Then, the mineral classification deep learning model is optimized by training the training data set and the validation data set to obtain an optimized mineral classification deep learning model. Further, the optimized mineral classification deep learning model is tested by the test data set. Finally, when the test classification accuracy meets the preset accuracy condition, the hyperspectral data of the Mars surface minerals to be identified is subjected to mineral identification and classification by the optimized mineral classification deep learning model, and the Mars surface mineral classification result is output. In this way, the importance of the Mars surface mineral hyperspectral data channel can be obtained through the mineral classification deep learning model, the internal correlation of the data can be captured, and the valuable information in the Mars surface mineral hyperspectral data can be selectively extracted, while ignoring unimportant information, thereby effectively improving the accuracy of the Mars surface mineral classification.

[0042] The solution provided by the embodiment of the present application is introduced below in conjunction with the accompanying drawings.

[0043] For details, see Figure 2 , which is a flow chart of a method for classifying Martian surface minerals based on hyperspectral data provided by an embodiment of the present invention, such as Figure 2 As shown, the method for classifying Martian surface minerals based on hyperspectral data provided by the present invention includes the following steps S101-S107:

[0044] S101. Obtain a standard hyperspectral dataset of Martian surface minerals.

[0045] In the embodiment of the present application, the standard hyperspectral dataset of Martian surface minerals may be a compact reconnaissance imaging spectromer for Mars (CRISM) dataset.

[0046] In one implementation, the Martian surface mineral standard hyperspectral dataset can be obtained from a standard spectral library of minerals identified through CRISM analysis (MICA).

[0047] In some embodiments, the Mars surface mineral standard hyperspectral data set includes: ferromagnesian montmorillonite spectral data, magnesium carbonate spectral data, and fayalite spectral data.

[0048] S102. Preprocess the standard hyperspectral dataset of Martian surface minerals to obtain a preprocessed standard hyperspectral dataset of Martian surface minerals.

[0049] In some embodiments, data preprocessing is performed on the standard hyperspectral dataset of Martian surface minerals, including at least one of the following: clipping, smoothing, continuum removal, and standardization.

[0050] Specifically, the clipping process is to intercept the hyperspectral data of the target spectral wavelength. For example, the spectral wavelength range in the MICA spectral library is 0-4μm. Through the clipping process, the 1.0-2.6μm part can be intercepted as the preprocessed Mars surface mineral standard hyperspectral dataset. The smoothing process is to perform noise reduction on the Mars surface mineral standard hyperspectral dataset to remove the peaks in the hyperspectral data. The continuum removal process is to remove the continuum of the hyperspectral data to highlight the absorption characteristics. The standardization process is to normalize the value of the hyperspectral data to near 0, so that the hyperspectral data in the Mars surface mineral standard hyperspectral dataset is more regular.

[0051] S103, dividing the preprocessed Mars surface mineral standard hyperspectral dataset into: a training dataset, a validation dataset, and a test dataset according to a preset division ratio.

[0052] Specifically, to facilitate the training and optimization of the deep learning model for mineral classification, the embodiment of the present application may divide the preprocessed Martian surface mineral standard hyperspectral dataset into: a training dataset, a validation dataset, and a test dataset.

[0053] In some embodiments, the preset division ratio may be: training data set: validation data set: test data set is 8:1:1. In this way, the mineral classification deep learning model can be effectively trained through the training data set with a relatively large proportion, so as to improve the accuracy of the mineral classification deep learning model in classifying Martian surface minerals.

[0054] In one implementation, the training data set can be represented by X1, Among them, d1 is the number of hyperspectral data contained in the training data set, and k1 is the dimension, that is, the number of spectral channels. A single training data in the training data set is represented by x1. Therefore, the training dataset can be expressed as The representation of the validation set X2 and the test set X3 is similar to that of the training set X1 and will not be repeated here.

[0055] In one implementation, the annotation of the training data set (i.e., the classification result of the Martian surface minerals corresponding to the training data set) can be represented by Y1, Among them, 4 are the four types of mineral labels after one-hot encoding. For example, mineral labels 1, 2, 3, and 4 can represent iron / magnesium montmorillonite, magnesium carbonate, iron olivine, and background pixels, respectively. Among them, the single data label in the annotation of the training data set is represented by y1, y∈R 1X4 Therefore, the annotation of the training dataset can be expressed as

[0056] It is understandable that the way of expressing the annotations of the validation dataset and the test dataset is similar to the way of expressing the annotations of the training dataset, and will not be described in detail here.

[0057] S104. Construct a deep learning model for mineral classification.

[0058] Among them, the mineral classification deep learning model can be used to identify and classify minerals on the surface of Mars. In some embodiments, Figure 3 A schematic diagram of the structure of the mineral classification deep learning model provided by the embodiment of the present invention is shown in FIG. Figure 3 As shown, the mineral classification deep learning model includes: a channel selector (channel selector, S) and a classifier (classifier, C), and the channel selector and the classifier are connected.

[0059] Among them, the channel selector is composed of multiple (N) channel attention modules (CAM) connected in series. Specifically, the channel attention module can be composed of 4 first fully connected layers (FCN), channel-wise weighting layers and skip connect layers connected in series; the first fully connected layer is connected with the corresponding first activation function and normalization layer (LN). The structural type of the classifier is: multilayer perception (MLP) structure, or convolutional neural network (CNN) structure.

[0060] like Figure 3 As shown in the figure, in the mineral classification deep learning model, first, the Mars surface mineral hyperspectral data X is input into the channel selector, and the channel selector outputs the channel selector output result data X'. Then, the channel selector output result data X' is input into the classifier, and the classifier outputs the Mars surface mineral classification result.

[0061] In some embodiments, the expression of the channel selector may be:

[0062] x′=S(x;θ s );

[0063] Among them, S represents the channel selector, x′ represents the output result of the channel selector, x represents the input hyperspectral data, θ s Represents a set of parameters in a channel selector;

[0064] In the case where the channel selector consists of N channel attention modules in series, the expansion of the channel selector is:

[0065]

[0066] Among them, f (i) represents the i-th channel attention module, t (i) represents the output of the i-th channel attention module, θ s (i) represents the parameters of the i-th channel attention module, i=1,2…N.

[0067] In some embodiments, the expression of the channel attention module can be:

[0068]

[0069] Among them, I represents the input of the channel attention module, O represents the output of the channel attention module, and z (1) Represents the output result of the first fully connected layer, z (2) Represents the output result of the second first fully connected layer, z (3) Represents the output result of the third first fully connected layer, z (4) Represents the output result of the 4th first fully connected layer. θ sc (1) represents the parameters of the first fully connected layer, θ sc (2) represents the parameters of the second first fully connected layer, θ sc (3) represents the parameters of the third first fully connected layer, θ sc (4) represents the parameters of the 4th first fully connected layer. γ1 represents the first activation function, which can be, for example, a Relu activation function. σ represents the Sigmoid activation function, L1 represents the first fully connected layer, LN1 represents the first normalized layer, Represents element-wise multiplication.

[0070] In some embodiments, the expression of the above classifier may be:

[0071]

[0072] Among them, C represents the classifier, x′ represents the output result of the channel selector, represents the classification result of Martian surface minerals output by the classifier, θ c Represents a set of parameters in a classifier.

[0073] In some embodiments, when the structural type of the classifier is a multi-layer perceptron structure, the classifier includes: 4 second fully connected layers, and the second fully connected layers are connected with corresponding second activation functions.

[0074] Then the expansion of the classifier can be:

[0075]

[0076] Among them, x′ represents the output result of the channel selector, Represents the Martian surface mineral classification results output by the classifier, Represents the output result of the first second fully connected layer, Represents the output result of the second second fully connected layer, Represents the output result of the third second fully connected layer. θ c1 (1) represents the parameters of the first second fully connected layer, θ c1(2) represents the parameters of the second fully connected layer, θ c1 (3) represents the parameters of the third second fully connected layer, θ c1 (4) represents the parameter of the 4th second fully connected layer. γ2 represents the second activation function, which may be, for example, a Relu activation function. σ represents the Sigmoid activation function, L2 represents the second fully connected layer, and LN2 represents the second normalization layer.

[0077] In some embodiments, when the structural type of the classifier is a convolutional neural network structure, the classifier includes: 3 convolutional layers, 2 third fully connected layers and 1 Flatten layer, and the third fully connected layer is connected to the corresponding third activation function.

[0078] Then the expansion of the classifier can be:

[0079]

[0080] Among them, x′ represents the output result of the channel selector, Represents the Martian surface mineral classification results output by the classifier, Represents the output result of the first convolutional layer, Represents the output result of the second convolutional layer, Represents the output result of the third convolutional layer. Represents the output result of the Flatten layer. Represents the output of the third fully connected layer. θ c2 (1) represents the parameters of the first convolutional layer, θ c2 (2) represents the parameters of the second convolutional layer, θ c2 (3) represents the parameters of the third convolutional layer, θ c2 (4) represents the parameters of the first third fully connected layer, θ c2 (5) represents the parameters of the second third fully connected layer. γ3 represents the third activation function, which can be, for example, a Relu activation function. σ represents the Sigmoid activation function, L3 represents the third fully connected layer, LN3 represents the third normalization layer, CV represents the convolutional layer, MP represents the maximum pooling layer, and FL represents the Flatten layer.

[0081] S105. Train the mineral classification deep learning model according to the training data set and the verification data set to obtain an optimized mineral classification deep learning model.

[0082] In some embodiments, Figure 3 Taking the mineral classification deep learning model shown in FIG. 1 as an example, S105 may specifically include the following steps S1051-S1054:

[0083] S1051. Input the training data set and the verification data set into the channel selector respectively, and output a first channel selector output result corresponding to the training data set and a second channel selector output result corresponding to the verification data set.

[0084] The first channel selector output result and the second channel selector output result can be determined based on the expression of the channel selector.

[0085] S1052, inputting the output result of the first channel selector and the output result of the second channel selector into a classifier respectively, and outputting a first predicted classification result corresponding to the output result of the first channel selector and a second predicted classification result corresponding to the output result of the second channel selector.

[0086] The first predicted classification result and the second predicted classification result can be determined based on the expression of the above classifier.

[0087] S1053. Construct a loss function, and iteratively optimize the hyperparameters of the mineral classification deep learning model according to the first predicted classification result so that the loss function converges and obtains the hyperparameter training result.

[0088] In some embodiments, the loss function may be a cross entropy loss function. Exemplarily, the cross entropy loss function may be expressed as:

[0089]

[0090] Where r represents the model dimension of the mineral classification deep learning model, represents the test results of Mars surface mineral classification (such as the first predicted classification results and the second predicted classification results), y′ i Represents the standard results of the Martian surface mineral classification.

[0091] In some embodiments, the hyperparameters of the mineral classification deep learning model include: the number of channel attention modules and the structural type of the classifier.

[0092] S1054. Optimize and adjust the hyperparameter training results according to the second predicted classification results to obtain an optimized mineral classification deep learning model.

[0093] The second prediction classification result can be used to further optimize and adjust the hyperparameter training results, determine the optimized hyperparameters, and then obtain an optimized mineral classification deep learning model.

[0094] S106. Test the optimized mineral classification deep learning model according to the test data set and determine the test classification accuracy.

[0095] Furthermore, the test data set can be input into the optimized mineral classification deep learning model to obtain a third predicted classification result, which is then compared with the standard classification result in the test data set to determine the test classification accuracy.

[0096] In some embodiments, the expression for testing classification accuracy may be:

[0097]

[0098] in,

[0099]

[0100] Among them, ACC represents the test classification accuracy, d′ represents the size of the test data set, represents the test result of Mars surface mineral classification (such as the third predicted classification result mentioned above), y′ i Represents the standard results of Martian surface mineral classification (such as the standard classification results mentioned above).

[0101] S107. When the test accuracy meets the preset accuracy condition, the hyperspectral data of the Martian surface minerals to be identified is input into the optimized mineral classification deep learning model, and the Martian surface mineral classification results are output.

[0102] Specifically, taking the above-mentioned test classification accuracy ACC as an example, the higher the ACC is, the more accurate the classification results of the optimized mineral classification deep learning model are.

[0103] 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 Mars regions. The preset accuracy conditions can be, for example, greater than 85%, or greater than 90%, etc. When the test accuracy meets the preset accuracy conditions, it can be determined that the optimized mineral classification deep learning model can meet the application requirements.

[0104] Furthermore, when the test accuracy meets the preset accuracy conditions, the hyperspectral data of the Martian surface minerals to be identified can be input into the optimized mineral classification deep learning model to output the Martian surface mineral classification results.

[0105] In some embodiments, the Martian surface mineral classification results may be mineral mapping results marked with different colors, where each color corresponds to a Martian mineral type.

[0106] For example, the hyperspectral data of the Martian surface minerals to be identified is an MTRDR image derived from CRISM. The MTRDR image data is represented by X_im, and the Martian surface mineral classification result is represented by Y_im. X_im is input into the optimized mineral classification deep learning model, and Y_im can be output. The mineral category result corresponding to each pixel of Y_im is marked with different colors in the MTRDR image, and the mineral mapping result can be obtained. Figure 4 Schematic diagram of mineral mapping results provided by an embodiment of the present invention. Figure 4 As shown, in the mineral mapping result, the mineral category represented by the blue area 401 is: iron / magnesium montmorillonite, the mineral category represented by the green area 402 is: magnesium carbonate, and the mineral category represented by the red area 403 is: fayalite.

[0107] It should be noted that the embodiments of the present application provide Figure 4 The colors in the figure are for illustrative purposes only. The colors corresponding to different types of Martian minerals can be set according to actual application requirements, and this application does not make any specific limitations on this.

[0108] By using the Mars surface mineral classification method based on hyperspectral data provided by the embodiment of the present invention, the importance of the Mars surface mineral hyperspectral data channel can be obtained through the mineral classification deep learning model, the internal correlation of the data can be captured, and the valuable information in the Mars surface mineral hyperspectral data can be selectively extracted, while ignoring unimportant information. In this way, the mineral classification deep learning model can learn richer feature representations in the Mars surface mineral hyperspectral data, and achieve more accurate mineral classification, thereby effectively improving the accuracy of Mars surface mineral classification, which has important guiding significance for the development of Mars resource exploration projects and the development and utilization of deep space resources.

[0109] Through the description of the above implementation methods, technical personnel in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned 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.

[0110] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0111] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0112] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0113] Similar parts between the embodiments provided in this application can be referenced to each other. The specific implementation methods provided above are only a few examples under the general concept of this application and do not constitute a limitation on the protection scope of this application. For those skilled in the art, any other implementation methods expanded based on the scheme of this application without creative work belong to the protection scope of this application.

Claims

1. A method for classifying Martian surface minerals based on hyperspectral data, characterized in that: include: Acquire a standard hyperspectral dataset of Martian surface minerals; Performing data preprocessing on the Mars surface mineral standard hyperspectral dataset to obtain a preprocessed Mars surface mineral standard hyperspectral dataset; Dividing the preprocessed Mars surface mineral standard hyperspectral dataset into: a training dataset, a validation dataset, and a test dataset according to a preset division ratio; Constructing a mineral classification deep learning model, wherein the mineral classification deep learning model is used to identify and classify minerals on the surface of Mars; Training the mineral classification deep learning model according to the training data set and the verification data set to obtain an optimized mineral classification deep learning model; Testing the optimized mineral classification deep learning model according to the test data set and determining the test classification accuracy; When the test accuracy meets the preset accuracy condition, the hyperspectral data of the Martian surface minerals to be identified is input into the optimized mineral classification deep learning model, and the Martian surface mineral classification results are output.

2. The method according to claim 1, characterized in that The mineral classification deep learning model includes: a channel selector and a classifier, wherein the channel selector is connected to the classifier; The channel selector is composed of a plurality of channel attention modules connected in series, and the channel attention module is composed of four first fully connected layers, a channel-by-channel weighted layer and a residual connection layer connected in series; the first fully connected layer is connected with a corresponding first activation function and a normalization layer; The structure type of the classifier is: a multi-layer perceptron structure, or a convolutional neural network structure.

3. The method according to claim 2, characterized in that The step of training the mineral classification deep learning model according to the training data set and the verification data set to obtain an optimized mineral classification deep learning model includes: Inputting the training data set and the verification data set into the channel selector respectively, and outputting a first channel selector output result corresponding to the training data set and a second channel selector output result corresponding to the verification data set; Inputting the first channel selector output result and the second channel selector output result into the classifier respectively, and outputting a first predicted classification result corresponding to the first channel selector output result and a second predicted classification result corresponding to the second channel selector output result; Constructing a loss function, and iteratively optimizing the hyperparameters of the mineral classification deep learning model according to the first prediction classification result, so that the loss function converges and obtains a hyperparameter training result; The hyperparameter training result is optimized and adjusted according to the second prediction classification result to obtain the optimized mineral classification deep learning model.

4. The method according to claim 2, characterized in that: The expression of the channel selector is: x′=S(x;θ s ); Among them, S represents the channel selector, x′ represents the output result of the channel selector, x represents the input hyperspectral data, θ s Represents a set of parameters in a channel selector; In the case where the channel selector is composed of N channel attention modules connected in series, the expansion of the channel selector is: Among them, f (i) represents the i-th channel attention module, t (i) represents the output of the i-th channel attention module, θ s (i) represents the parameters of the i-th channel attention module, i=1,2…N.

5. The method according to claim 2, characterized in that: The expression of the channel attention module is: Among them, I represents the input of the channel attention module, O represents the output of the channel attention module, and z (1) Represents the output result of the first fully connected layer, z (2) Represents the output result of the second first fully connected layer, z (3) Represents the output result of the third first fully connected layer, z (4) represents the output result of the 4th first fully connected layer, θ sc (1) represents the parameters of the first fully connected layer, θ sc (2) represents the parameters of the second first fully connected layer, θ sc (3) represents the parameters of the third first fully connected layer, θ sc (4) represents the parameters of the 4th first fully connected layer, γ1 represents the first activation function, σ represents the Sigmoid activation function, L1 represents the first fully connected layer, LN1 represents the first normalized layer, Represents element-wise multiplication.

6. The method according to claim 2, characterized in that The expression of the classifier is: Among them, C represents the classifier, x′ represents the output result of the channel selector, represents the classification result of Martian surface minerals output by the classifier, θ c Represents a set of parameters in a classifier.

7. The method according to claim 2, characterized in that In the case where the structural type of the classifier is a multi-layer perceptron structure, the classifier includes: 4 second fully connected layers, and the second fully connected layers are connected with corresponding second activation functions; Then the expansion of the classifier is: Among them, x′ represents the output result of the channel selector, Represents the Martian surface mineral classification results output by the classifier, Represents the output result of the first second fully connected layer, Represents the output result of the second second fully connected layer, Represents the output result of the third second fully connected layer, θ c1 (1) represents the parameters of the first second fully connected layer, θ c1 (2) represents the parameters of the second fully connected layer, θ c1 (3) represents the parameters of the third second fully connected layer, θ c1 (4) represents the parameters of the 4th second fully connected layer, γ2 represents the second activation function, σ represents the Sigmoid activation function, L2 represents the second fully connected layer, and LN2 represents the second normalization layer.

8. The method according to claim 2, characterized in that: When the structural type of the classifier is a convolutional neural network structure, the classifier includes: 3 convolutional layers, 2 third fully connected layers and 1 Flatten layer, and the third fully connected layer is connected with a corresponding third activation function; Then the expansion of the classifier is: Among them, x′ represents the output result of the channel selector, Represents the Martian surface mineral classification results output by the classifier, Represents the output result of the first convolutional layer, Represents the output result of the second convolutional layer, Represents the output result of the third convolutional layer. Represents the output result of the Flatten layer. represents the output of the third fully connected layer, θ c2 (1) represents the parameters of the first convolutional layer, θ c2 (2) represents the parameters of the second convolutional layer, θ c2 (3) represents the parameters of the third convolutional layer, θ c2 (4) represents the parameters of the first third fully connected layer, θ c2 (5) represents the parameters of the second third fully connected layer, γ3 represents the third activation function, σ represents the Sigmoid activation function, L3 represents the third fully connected layer, LN3 represents the third normalization layer, CV represents the convolutional layer, MP represents the maximum pooling layer, and FL represents the Flatten layer.

9. The method according to claim 3, characterized in that: The loss function is a cross entropy loss function, and the hyperparameters of the mineral classification deep learning model include: the number of channel attention modules and the structural type of the classifier.

10. The method according to claim 1, characterized in that The expression of the test classification accuracy is: in, Among them, ACC represents the test classification accuracy, d ′ represents the size of the test dataset, Represents the results of the Mars surface mineral classification test, y i ′ Represents the standard results of the Martian surface mineral classification.

Images