Moon principal element inversion method based on deep learning and self-supervision

Abstract

The invention discloses a lunar principal element inversion method based on deep learning and self-supervision. The lunar principal element inversion method comprises the following steps: preprocessing a lunar orbiter remote sensing spectral image; constructing a first data set and a second data set based on the preprocessed remote sensing spectral image; performing one-dimensional convolution auto-encoder self-supervised learning through the first data set to obtain a feature extractor; performing feature extraction on the second data set by using a feature extractor, and performing regression training in combination with a support vector machine to obtain a lunar principal element inversion model; and inputting the full-moon hyperspectral data into the lunar major element inversion model, and inverting and drawing a lunar major element abundance distribution diagram. According to the method, deep learning and self-supervised learning are fused, features can be directly learned from the lunar orbit remote sensing spectral image, dependence on data labels and the difference between laboratory data and remote sensing data are effectively overcome, and the precision and efficiency of lunar surface principal element inversion are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of lunar major element inversion, and more specifically to a lunar major element inversion method based on deep learning and self-supervision. Background Art

[0002] As human beings continue to explore space, the geological structure and material composition of the moon, as the closest neighbor of the earth, are increasingly attracting attention. Understanding the element content on the lunar surface is of great significance for understanding the formation, evolution, geological activities and potential resource distribution of the moon. This not only helps to deepen our scientific understanding of the moon, but also provides an important basis for future exploration and development of lunar resources.

[0003] In the field of lunar exploration, traditional methods rely on limited lunar sample analysis results to infer the elemental content of the lunar surface. However, these samples often only cover a very small area of ​​the lunar surface and are expensive to obtain, making it difficult to fully reveal the overall material composition characteristics of the moon. In addition, traditional methods are also subject to various limitations in the sample collection, preservation and analysis process, which leads to doubts about the accuracy and representativeness of the results.

[0004] In recent years, with the development of remote sensing technology, especially the application of hyperspectral imagers carried by orbiters, a new way has been provided for the inversion of the element content on the lunar surface. Hyperspectral imagers can obtain continuous spectral information on the lunar surface. By analyzing these spectral data, the material composition and content of the lunar surface can be inferred. However, hyperspectral data has the characteristics of high data dimension, data redundancy, and high correlation between adjacent bands. Traditional inversion methods such as parameter ratio method are unable to cope with these data. These methods are often based on simplified models of physical and chemical properties, which make it difficult to accurately capture the deep-level features in hyperspectral data, resulting in inaccurate inversion results.

[0005] With the rise of deep learning technology, it has achieved remarkable success in image processing, natural language processing and other fields. Researchers have begun to try to apply deep learning technology to the inversion of the element content on the lunar surface. However, due to the scarcity of lunar sample data, deep learning methods based on supervised learning often face the problem of insufficient training samples during the training process. In addition, there are differences in light source conditions, instrument response, etc. between the spectral data obtained in the laboratory and the orbital remote sensing spectral data, which further increases the difficulty and uncertainty of model training.

[0006] Therefore, how to design a lunar major element inversion method based on deep learning and self-supervision that can make full use of the information in hyperspectral data, overcome the differences between data sources, and improve the accuracy and efficiency of lunar major element inversion is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] In view of this, the present invention provides a lunar major element inversion method based on deep learning and self-supervision, which uses a self-supervised learning method to directly extract features from orbital remote sensing spectral images, reduces dependence on data labels, and avoids the differences between laboratory data and remote sensing data. At the same time, through the combination of deep learning and support vector machines, the inversion effect is further improved, providing more reliable and accurate data support for lunar geological research and lunar exploration projects.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A lunar major element inversion method based on deep learning and self-supervision includes the following steps:

[0010] S1. Preprocessing the lunar orbiter remote sensing spectral image to obtain a preprocessed remote sensing spectral image;

[0011] S2. constructing a first data set and a second data set based on the preprocessed remote sensing spectral image;

[0012] S3, performing self-supervised learning of a one-dimensional convolutional autoencoder on the first data set to obtain a feature extractor;

[0013] S4, extracting features from the second data set using a feature extractor, and performing regression training in combination with a support vector machine to obtain an inversion model of the lunar major elements;

[0014] S5. Input the full-month hyperspectral data into the lunar major element inversion model, invert and draw the lunar major element abundance distribution map.

[0015] Furthermore, in S1, the preprocessing includes: reflectivity conversion, bad pixel repair, bad line repair, fringe removal, flat field correction, reflectivity inspection and accuracy estimation.

[0016] Furthermore, in S2, constructing a first data set includes:

[0017] Based on the preprocessed remote sensing spectral images, areas covering different latitudes and terrain types were selected, and known lunar sampling sites were identified;

[0018] Based on the lunar sampling points in the selected area, remote sensing spectral images are selected to construct the first data set; the size of the remote sensing spectral image is 128*128 pixels or 512*512 pixels; the total number of pixels of the remote sensing spectral image is at least 5 times the network parameters of the one-dimensional convolutional autoencoder.

[0019] Furthermore, in S2, constructing a second data set includes:

[0020] Determine sampling points in the preprocessed remote sensing spectral image based on sampling point coordinates;

[0021] The average spectrum of the sampling point and its surrounding eight pixels is selected as the true spectrum of the sampling point;

[0022] The element content analysis result of the returned sample laboratory corresponding to the real spectrum is used as the label of the real spectrum;

[0023] The real spectra and their labels are augmented by generative adversarial networks to obtain the second dataset.

[0024] Furthermore, the S3 includes:

[0025] S31, dividing the first data set into a training set and a validation set;

[0026] S32, performing maximum and minimum value normalization processing on the spectral data in the training set and the validation set;

[0027] S33, based on the spectral data after maximum and minimum value normalization processing, performing one-dimensional convolutional autoencoder training; wherein the one-dimensional convolutional autoencoder includes: an encoder, a bottleneck layer and a decoder;

[0028] S34. After training is completed, the encoder and bottleneck layer are selected as feature extractors.

[0029] Furthermore, in S33, one-dimensional convolutional autoencoder training is performed, including:

[0030] Extract high-dimensional spectral features through the one-dimensional convolutional layer and downsampling layer of the encoder;

[0031] Compressing the high-dimensional spectral features into low-dimensional features at the bottleneck layer;

[0032] The low-dimensional features are restored through a one-dimensional deconvolution layer and an upsampling layer of a decoder to obtain a reconstructed spectrum;

[0033] Compare the reconstructed spectrum with the original spectrum and calculate the mean square error (MSE) or binary cross entropy (BCE) as the loss function to measure the reconstruction error.

[0034] Combined with the Adam optimizer, the network parameters and weights are iteratively updated until the preset conditions are met to obtain the optimized one-dimensional convolutional autoencoder.

[0035] Furthermore, in S3, an early stopping mechanism is also combined; the early stopping mechanism includes:

[0036] During the training process, a minimum improvement threshold and a patience value N are set; if the reconstruction error does not exceed the minimum improvement threshold in N consecutive iterations, the training process is stopped and the model parameters with the minimum reconstruction error are saved.

[0037] Furthermore, the S4 includes:

[0038] S41, extracting features from the second data set using a feature extractor;

[0039] S42, performing standardization processing on the extracted features;

[0040] S43, inputting the standardized features into a support vector machine for regression training, and searching for the best parameter configuration of the support vector machine in combination with a genetic algorithm to obtain an optimized support vector machine;

[0041] S44. In combination with the feature extractor, a lunar major element inversion model is obtained.

[0042] Furthermore, in S43, searching for the optimal parameter configuration of the support vector machine in combination with a genetic algorithm includes:

[0043] Set the genetic algorithm parameters to generate an initial population containing a variety of parameter configurations;

[0044] Train the SVM with each parameter configuration and evaluate the model performance as individual fitness;

[0045] According to the fitness value, individuals with high fitness are selected as parents;

[0046] Perform a crossover operation on the parent generation to generate offspring containing new gene combinations;

[0047] Perform mutation operations on the offspring, randomly changing some genes to explore new parameter spaces;

[0048] Merge the parent and offspring to form a new population, and repeat the evaluation and evolution until the conditions are met;

[0049] Select the parameter configuration with the highest final fitness and output the optimized support vector machine model.

[0050] Furthermore, the S5 includes:

[0051] S51, preprocessing the hyperspectral data of the whole month;

[0052] S52, normalizing the maximum and minimum values ​​of the preprocessed hyperspectral data for the whole month;

[0053] S53, extracting spectral features from the normalized data using a feature extractor;

[0054] S54, performing standardization processing on the extracted features;

[0055] S55, inputting the standardized features into a support vector machine regression model to predict element content;

[0056] S56. According to the prediction results, fill in the pixel positions of the corresponding spectrum one by one and draw a distribution map of the abundance of major elements on the moon.

[0057] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:

[0058] Beneficial effects:

[0059] 1. This method introduces a one-dimensional convolutional autoencoder for self-supervised learning, which can automatically learn deep features from spectral images without a large amount of labeled data. It can not only capture more complex and subtle data structures, but also greatly reduce the dependence on manually labeled data, and improve the generalization ability and prediction accuracy of the model.

[0060] 2. By using a generative adversarial network to expand the second dataset, the problem of scarce samples at the lunar sampling points was effectively solved. The generated additional spectral data and its labels increased the diversity and representativeness of the training data, making the model more stable and accurate in different environments. It helps to simulate more diverse spectral characteristics and further enhances the model's adaptability to unknown data.

[0061] 3. Combine genetic algorithms to search for the best parameter configuration of support vector machines, overcoming the local optimal problem that traditional parameter adjustment methods may encounter. Genetic algorithms simulate the natural selection process, explore and find approximate global optimal solutions in a wide range of parameter spaces, thereby improving the prediction performance of support vector machines. At the same time, combined with the early stopping mechanism to prevent overfitting, ensure the effectiveness and efficiency of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0063] Figure 1 A flow chart of a lunar major element inversion method based on deep learning and self-supervision provided in an embodiment of the present invention;

[0064] Figure 2 A schematic diagram of TiO2 content distribution provided in an embodiment of the present invention;

[0065] Figure 3 A schematic diagram of a one-dimensional convolutional autoencoder self-supervised learning process provided by an embodiment of the present invention;

[0066] Figure 4A schematic diagram of a regression training process in combination with a support vector machine provided in an embodiment of the present invention;

[0067] Figure 5 A schematic diagram of the process of obtaining a lunar major element abundance distribution map based on a lunar major element inversion model provided in an embodiment of the present invention;

[0068] Figure 6 A schematic diagram of TiO2 abundance distribution provided in an embodiment of the present invention;

[0069] Figure 7 A schematic diagram of FeO abundance distribution provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] like Figure 1 As shown, this embodiment provides a lunar major element inversion method based on deep learning and self-supervision, comprising the following steps:

[0072] S1. Preprocessing the lunar orbiter remote sensing spectral image to obtain a preprocessed remote sensing spectral image;

[0073] S2. constructing a first data set and a second data set based on the preprocessed remote sensing spectral image;

[0074] S3, performing self-supervised learning of a one-dimensional convolutional autoencoder on the first data set to obtain a feature extractor;

[0075] S4, extracting features from the second data set using a feature extractor, and performing regression training in combination with a support vector machine to obtain an inversion model of the lunar major elements;

[0076] S5. Input the full-month hyperspectral data into the lunar major element inversion model, invert and draw the lunar major element abundance distribution map.

[0077] This method combines deep learning with self-supervised learning, uses a one-dimensional convolutional autoencoder to automatically extract spectral features, and uses a generative adversarial network (GAN) to amplify training data, significantly improving the accuracy and generalization ability of the lunar major element inversion. The genetic algorithm optimizes the support vector machine parameter configuration to ensure optimal model performance, while the early stopping mechanism prevents overfitting. The refined processing of the entire process ensures data quality and prediction reliability, providing strong technical support for the efficient and accurate analysis of the distribution of elements on the lunar surface.

[0078] The lunar major element inversion model in this implementation can be used for the Chang'e-1 Interference Imaging Spectrometer (IIM) data, SELENE (KAGUYA) Multiband Imager (MI) data, Chandrayaan-1 Moon Mineralogy Mapper (M3) data, and Clementine's Ultraviolet Visible (UVVIS) data.

[0079] The following is a further detailed description of each step and related features in the above method using IIM data:

[0080] In this embodiment S1, the lunar orbiter remote sensing spectral image is preprocessed to obtain a preprocessed remote sensing spectral image;

[0081] Specifically, in the preprocessing stage of the lunar major element inversion method, the original remote sensing spectral image first maps the digital value to the physical quantity reflectance through reflectance conversion to ensure that the data has physical meaning and comparability, and directly improve the quality of the spectral characteristics. Then, the spatial domain or frequency domain filtering technology is used to repair the bad points, eliminate the isolated outliers caused by sensor failure or transmission errors, and ensure data consistency and integrity; at the same time, the average value of adjacent rows / columns or other interpolation methods are used to repair the bad lines formed by the damage of the entire row or column, and restore the continuity of the image.

[0082] Furthermore, in order to remove the periodic noise caused by the non-uniform response of the detector or external environmental factors, a specific algorithm is applied to identify and subtract these systematic errors to improve the image quality. Taking into account the gain differences between different bands and the non-uniformity of the detector response during the imaging process, a flat field correction is implemented to standardize the response characteristics of each band and reduce the deviation caused by the instrument. After completing various corrections, the reflectivity test verifies the rationality of the correlation between the bands and whether there are obvious outliers, and confirms the validity and reliability of the data. Finally, the preprocessing effect is quantified through precision estimation, and the signal-to-noise ratio, root mean square error and other indicators of the images before and after processing are calculated. The known ground truth samples are compared to further evaluate the effects of reflectivity conversion and other correction measures to ensure that the final data accurately reflects the true spectral characteristics of the lunar surface. This series of preprocessing steps significantly improves the quality of the input data and lays a solid foundation for the subsequent construction of a high-precision lunar major element inversion model.

[0083] In this embodiment S2, based on the preprocessed remote sensing spectral image, a first data set and a second data set are constructed;

[0084] Wherein, constructing the first data set includes:

[0085] Based on the preprocessed remote sensing spectral images, 512*512 pixel images were selected with the real landing sampling points of Apollo, Luna and Chang'e missions as the center. In addition to the sampling point area, the training data includes high, medium and low latitudes, and covers the lunar seas and high and low areas.

[0086] The IIM data does not cover the Apollo_15 sampling points. The final sampling points selected are Apollo_11, Apollo_12, Apollo_14_LM, Apollo_14_cone, Apollo_16_LM, Apollo_16_S11, Apollo_16_S13, Apollo_16_S1, Apollo_16_S2, Apollo_16_S4, Apollo_16_S5, Apollo_16_S6, Apollo_16_S8, Apollo_16_S9, Apollo_17_LM, Apollo_17_LRV1, Apollo_17_LRV2, Apollo_17_LRV3, Apollo _17_LRV4、Apollo_17_LRV5、Apollo_17_LRV6、Apollo_17_LRV7、Apollo_17_LRV8、Apollo_17_LRV9、Apollo_17_LRV10、Apollo_17_LRV11、Apollo_17_LRV12、Apollo _17_S1、Apollo_17_S2、Apollo_17_S3、Apollo_17_S5、Apollo_17_S6、Apollo_17_ S7, Apollo_17_S8, Apollo_17_S9, Luna_16, Luna_20, Luna_24, Chang'e 3, Chang'e 5, a total of 40 sampling points.

[0087] In addition, by comparing the relative reflectivity of the IIM spectral data with that of the ground-based telescope spectrum, it is found that the first 9 bands of the IIM data fluctuate significantly, the last 3 bands deviate significantly from the ground-based telescope spectrum, and there is obvious noise in bands 1 to 4, 31, and 32 of the IIM data. Therefore, only bands 11 to 29 of the IIM data (a total of 19 bands) are selected as valid data to participate in the construction of the first data set.

[0088] The size of the above remote sensing spectral images is 128*128 pixels or 512*512 pixels. The selection of this size balances the requirements of computational efficiency and spatial resolution, and can capture sufficient local details without increasing unnecessary computational burden due to excessive image size.

[0089] The total number of pixels in the remote sensing spectral image is at least 5 times the number of parameters of the one-dimensional convolutional autoencoder network. This ratio ensures that there are enough samples in the training process to fully optimize the network weights, avoid the problem of underfitting the model due to insufficient data, and also helps to improve the model's generalization ability to new data.

[0090] Furthermore, a second data set is constructed, including:

[0091] From the preprocessed lunar orbit remote sensing hyperspectral image, the pixel with the sampling point coordinates as the center is extracted, and the average spectrum of nine pixels including the neighboring pixels of the center pixel is calculated. The average spectrum represents the remote sensing spectrum of the sampling point, and the analysis results of the laboratory composition of the sample returned by the sampling point (six major oxides of TiO2, FeO, Al2O3, CaO, MgO, and SiO2 and the mineral parameter magnesium index Mg#) are used as the label corresponding to the spectrum.

[0092] The real spectra and their labels are amplified by generative adversarial networks. On the one hand, there are only 40 spectral data of the original IIM sampling points, which is far from enough to support the training of the regressor. On the other hand, the content (label) of each element corresponding to the original data is extremely unbalanced, such as Figure 2 As shown in the figure, taking TiO2 as an example, the data distribution is mainly concentrated in the part with extremely low titanium content (<1%), and the data points gradually decrease as the TiO2 content increases. This unbalanced data distribution may cause the model to over-focus on data-intensive areas when training a regression model, thereby affecting the prediction accuracy of data in sparse areas.

[0093] For the training of the regressor, data balance is very important. The regressor takes the overall regression accuracy as the learning goal. When the data distribution is unbalanced, too much attention will be paid to the majority of the data in the interval, which will reduce the regression performance of the minority of data in the interval. The spectral data of the original IIM sampling points were amplified using a generative adversarial network. The spectral data at each content was amplified to 30, and finally the data in the second data set was 1200.

[0094] In this way, the original data can be expanded to a large enough scale to meet the needs of machine learning model training. At the same time, because the data generated by GAN is highly realistic and diverse, the expanded data set can better simulate the complex situations in the real world, which helps to improve the generalization ability and robustness of the model.

[0095] like Figure 3 As shown, in this embodiment S3, a one-dimensional convolutional autoencoder is self-supervised learned through the first data set to obtain a feature extractor; specifically, it includes:

[0096] S31, dividing the first data set into a training set and a validation set;

[0097] S32, performing maximum and minimum value normalization processing on the spectral data in the training set and the validation set;

[0098] S33, based on the spectral data after maximum and minimum value normalization processing, performing one-dimensional convolutional autoencoder training; wherein the one-dimensional convolutional autoencoder includes: an encoder, a bottleneck layer and a decoder;

[0099] Perform one-dimensional convolutional autoencoder training, including:

[0100] Extract high-dimensional spectral features through the one-dimensional convolutional layer and downsampling layer of the encoder;

[0101] Compressing the high-dimensional spectral features into low-dimensional features at the bottleneck layer;

[0102] The low-dimensional features are restored through a one-dimensional deconvolution layer and an upsampling layer of a decoder to obtain a reconstructed spectrum;

[0103] Compare the reconstructed spectrum with the original spectrum and calculate the mean square error (MSE) or binary cross entropy (BCE) as the loss function to measure the reconstruction error.

[0104] Combined with the Adam optimizer, the network parameters and weights are iteratively updated until the preset conditions are met to obtain the optimized one-dimensional convolutional autoencoder. In the Adam optimizer, the learning rate is 0.001, the number of epochs is 100, and the batch_size is 512.

[0105] It also combines an early stopping mechanism; the early stopping mechanism includes: during the training process, setting a minimum improvement threshold and a patience value N; if the reconstruction error does not exceed the minimum improvement threshold in N consecutive iterations, the training process is stopped and the model parameters with the minimum reconstruction error are saved.

[0106] S34. After training is completed, the encoder and bottleneck layer are selected as feature extractors.

[0107] In this step, by implementing the self-supervised learning of the one-dimensional convolutional autoencoder, highly representative spectral features can be efficiently extracted from the first data set. First, by normalizing the spectral data in the training set and the validation set to the maximum and minimum values, the consistency and stability of the input data are ensured, laying a good foundation for subsequent model training. Then, during the training process of the one-dimensional convolutional autoencoder, the encoder effectively captures the high-dimensional spectral features through the convolution layer and the downsampling layer, and compresses it to the low-dimensional representation of the bottleneck layer, achieving a significant reduction in the feature dimension without losing key information. The decoder is responsible for restoring these low-dimensional features to reconstructed spectra, and by calculating the reconstruction error and combining the Adam optimizer to continuously iterate and update the network parameters, and finally obtain the optimized autoencoder. In particular, the early stopping mechanism is introduced to prevent overfitting, ensuring that the performance of the model on the validation set is optimal, thereby improving the generalization ability and robustness of the model.

[0108] After the training is completed, the encoder and bottleneck layer are selected as feature extractors. This design not only retains the important features in the original spectral data, but also greatly reduces the feature dimension and improves the computational efficiency. This feature extractor can provide a more compact and informative feature representation in subsequent tasks, such as element content prediction, which greatly promotes the learning effect and prediction accuracy of the model. In addition, since the autoencoder is pre-trained on a large amount of unlabeled data, it can better adapt to different spectral modes and enhance the ability to process unknown data, providing strong technical support for the inversion of the main elements of the moon.

[0109] like Figure 4 As shown, in this embodiment S4, a feature extractor is used to extract features from the second data set, and regression training is performed in combination with a support vector machine to obtain a lunar major element inversion model; specifically, the following steps are included:

[0110] S41, extracting features from the second data set using a feature extractor;

[0111] S42, performing standardization processing on the extracted features;

[0112] S43, inputting the standardized features into a support vector machine for regression training, and searching for the optimal parameter configuration of the support vector machine in combination with a genetic algorithm to obtain an optimized support vector machine; wherein searching for the optimal parameter configuration of the support vector machine in combination with a genetic algorithm includes:

[0113] Set the genetic algorithm parameters to generate an initial population containing a variety of parameter configurations;

[0114] Train the SVM with each parameter configuration and evaluate the model performance as individual fitness;

[0115] According to the fitness value, individuals with high fitness are selected as parents;

[0116] Perform a crossover operation on the parent generation to generate offspring containing new gene combinations;

[0117] Perform mutation operations on the offspring, randomly changing some genes to explore new parameter spaces;

[0118] Merge the parent and offspring to form a new population, and repeat the evaluation and evolution until the conditions are met;

[0119] Select the parameter configuration with the highest final fitness and output the optimized support vector machine model.

[0120] S44. In combination with the feature extractor, a lunar major element inversion model is obtained.

[0121] Specifically, in this embodiment, the autoencoder is composed of three one-dimensional convolutional layers (Conv1D), three one-dimensional maximum pooling layers (MaxPooling1D), a flattening layer (Flatten), a bottleneck layer (bottleneck), two one-dimensional transposed convolutional layers (Conv1DTranspose) and two downsampling layers (UpSampling1D). It should be noted that when the autoencoder is applied to other lunar spectral remote sensing data outside IIM, due to the different number of bands and noise levels of different data, the number of layers and parameters of the autoencoder network need to be modified in a targeted manner.

[0122] In this step, by combining feature extractors, support vector machines, and genetic algorithms, the construction and optimization of the lunar major element inversion model was realized, significantly improving the prediction accuracy and generalization ability of the model. The feature extractor trained in the self-supervised learning stage was used to extract features from the second data set, effectively capturing the key information in the spectral data and providing high-quality feature input for subsequent regression training. By standardizing the extracted features, the dimensional differences between different features were eliminated, further improving the training efficiency and prediction performance of the support vector machine.

[0123] On this basis, a genetic algorithm is introduced to search for the optimal parameter configuration of the support vector machine. By simulating natural selection and genetic mechanisms, the population is continuously iterated and evolved, gradually approaching the global optimal solution. In this process, the genetic algorithm not only takes into account the training error of the model, but also explores new parameter spaces through crossover and mutation operations, effectively avoiding local optimal solutions, so that the final support vector machine model shows higher fitness and prediction accuracy on the validation set. Combining the feature extractor and the optimized support vector machine, the constructed lunar major element inversion model can accurately and efficiently predict the element content of each pixel in the full-moon hyperspectral image, providing strong data support for lunar geological research.

[0124] like Figure 5 As shown, in this embodiment S5, the full-month hyperspectral data is input into the lunar major element inversion model, and the lunar major element abundance distribution map is inverted and drawn, including:

[0125] S51, preprocessing the full-month hyperspectral data; the preprocessing process here is the same as the preprocessing process in the above step 1, ensuring that the data quality of the input model is consistent and reliable.

[0126] S52, normalize the maximum and minimum values ​​of the preprocessed hyperspectral data for the whole month, so that the values ​​of all spectral data are distributed between 0 and 1, ensuring the consistency of spectral data in different regions.

[0127] S53. Use a feature extractor to extract spectral features from the normalized data. Use the encoder's one-dimensional convolutional layer and downsampling layer to extract useful features from the spectrum, and combine the bottleneck layer to compress the high-dimensional data into a low-dimensional representation.

[0128] S54. Standardize the extracted features to ensure that all features are at the same scale.

[0129] S55. Input the standardized features into the support vector machine regression model to predict the element content; the support vector machine separates data points of different categories by finding the optimal hyperplane, thereby achieving accurate prediction of the element content.

[0130] S56. According to the prediction results, fill in the pixel positions of the corresponding spectra one by one to obtain the abundance distribution maps of the six main oxides TiO2, FeO, Al2O3, CaO, MgO, and SiO2 and the mineral parameter magnesium index Mg#.

[0131] Specifically, taking the inversion results of TiO2 and FeO as an example, Figure 6 As shown in the figure, different colors represent different concentrations of TiO2 content, the unit is weight percentage (wt.%); Figure 7 As shown, different colors in the figure represent the FeO content in weight percentage (wt.%);

[0132] exist Figure 6 and Figure 7 In the figure, the right side is the abundance color bar. The color changes from blue to red, indicating that the abundance gradually increases. The darker the red, the higher the abundance of TiO2 and FeO, and the darker the blue, the lower the abundance of TiO2 and FeO.

[0133] This embodiment provides a systematic method for inverting the main elements of the moon, which breaks through the bottleneck that the traditional parameter ratio method can only invert the abundance of a limited number of elements, uses the lunar remote sensing spectrum for self-supervised learning, and breaks the problem of inaccurate transmission of factual differences between laboratory and remote sensing data in the past, effectively improving the ability to extract useful information from lunar orbit remote sensing spectrum images, thereby improving the accuracy of the main element inversion of the lunar surface. The accurate distribution of the main elements of the whole moon inverted by this method can provide a basis for the study of the geological history and evolution of the moon, improve the accuracy of the division of lunar geological units, and provide effective support for the site selection of landing sites and in-situ utilization of resources for subsequent lunar exploration projects.

[0134] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0135] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lunar major element inversion method based on deep learning and self-supervision, characterized in that: The following steps are involved: S1. Preprocessing the lunar orbiter remote sensing spectral image to obtain a preprocessed remote sensing spectral image; S2. constructing a first data set and a second data set based on the preprocessed remote sensing spectral image; S3, performing self-supervised learning of a one-dimensional convolutional autoencoder on the first data set to obtain a feature extractor; S4, extracting features from the second data set using a feature extractor, and performing regression training in combination with a support vector machine to obtain an inversion model of the lunar major elements; S5. Input the full-month hyperspectral data into the lunar major element inversion model, invert and draw the lunar major element abundance distribution map.

2. The method for lunar major element inversion based on deep learning and self-supervision according to claim 1, characterized in that: In S1, the preprocessing includes: reflectivity conversion, bad pixel repair, bad line repair, fringe removal, flat field correction, reflectivity inspection and accuracy estimation.

3. The method for lunar major element inversion based on deep learning and self-supervision according to claim 1, characterized in that: In S2, constructing a first data set includes: Based on the preprocessed remote sensing spectral images, areas covering different latitudes and terrain types were selected, and known lunar sampling sites were identified; Based on the lunar sampling points in the selected area, remote sensing spectral images are selected to construct the first data set; the size of the remote sensing spectral image is 128*128 pixels or 512*512 pixels; the total number of pixels of the remote sensing spectral image is at least 5 times the network parameters of the one-dimensional convolutional autoencoder.

4. The method for lunar major element inversion based on deep learning and self-supervision according to claim 1, characterized in that: In S2, constructing a second data set includes: Determine sampling points in the preprocessed remote sensing spectral image based on sampling point coordinates; The average spectrum of the sampling point and its surrounding eight pixels is selected as the true spectrum of the sampling point; The element content analysis result of the returned sample laboratory corresponding to the real spectrum is used as the label of the real spectrum; The real spectra and their labels are augmented by generative adversarial networks to obtain the second dataset.

5. The method for lunar major element inversion based on deep learning and self-supervision according to claim 1, characterized in that: The S3 includes: S31, dividing the first data set into a training set and a validation set; S32, performing maximum and minimum value normalization processing on the spectral data in the training set and the validation set; S33, based on the spectral data after maximum and minimum value normalization processing, performing one-dimensional convolutional autoencoder training; wherein the one-dimensional convolutional autoencoder includes: an encoder, a bottleneck layer and a decoder; S34. After training is completed, the encoder and bottleneck layer are selected as feature extractors.

6. The method for inversion of lunar major elements based on deep learning and self-supervision according to claim 5, characterized in that: In S33, one-dimensional convolutional autoencoder training is performed, including: Extract high-dimensional spectral features through the one-dimensional convolutional layer and downsampling layer of the encoder; Compressing the high-dimensional spectral features into low-dimensional features at the bottleneck layer; The low-dimensional features are restored through a one-dimensional deconvolution layer and an upsampling layer of a decoder to obtain a reconstructed spectrum; Compare the reconstructed spectrum with the original spectrum and calculate the mean square error (MSE) or binary cross entropy (BCE) as the loss function to measure the reconstruction error. Combined with the Adam optimizer, the network parameters and weights are iteratively updated until the preset conditions are met to obtain the optimized one-dimensional convolutional autoencoder.

7. The method for lunar major element inversion based on deep learning and self-supervision according to claim 6, characterized in that: In S3, an early stopping mechanism is also combined; the early stopping mechanism includes: During the training process, a minimum improvement threshold and a patience value N are set; if the reconstruction error does not exceed the minimum improvement threshold in N consecutive iterations, the training process is stopped and the model parameters with the minimum reconstruction error are saved.

8. The method for lunar major element inversion based on deep learning and self-supervision according to claim 1, characterized in that: The S4 comprises: S41, extracting features from the second data set using a feature extractor; S42, performing standardization processing on the extracted features; S43, inputting the standardized features into a support vector machine for regression training, and searching for the best parameter configuration of the support vector machine in combination with a genetic algorithm to obtain an optimized support vector machine; S44. In combination with the feature extractor, a lunar major element inversion model is obtained.

9. The method for lunar major element inversion based on deep learning and self-supervision according to claim 8, characterized in that: In S43, searching for the optimal parameter configuration of the support vector machine in combination with a genetic algorithm includes: Set the genetic algorithm parameters to generate an initial population containing a variety of parameter configurations; Train the SVM with each parameter configuration and evaluate the model performance as individual fitness; According to the fitness value, individuals with high fitness are selected as parents; Perform a crossover operation on the parent generation to generate offspring containing new gene combinations; Perform mutation operations on the offspring, randomly changing some genes to explore new parameter spaces; Merge the parent and offspring to form a new population, and repeat the evaluation and evolution until the conditions are met; Select the parameter configuration with the highest final fitness and output the optimized support vector machine model.

10. The method for lunar major element inversion based on deep learning and self-supervision according to claim 1, characterized in that: The S5 comprises: S51, preprocessing the hyperspectral data of the whole month; S52, normalizing the maximum and minimum values ​​of the preprocessed hyperspectral data for the whole month; S53, extracting spectral features from the normalized data using a feature extractor; S54, performing standardization processing on the extracted features; S55, inputting the standardized features into a support vector machine regression model to predict element content; S56. According to the prediction results, fill in the pixel positions of the corresponding spectrum one by one and draw a distribution map of the abundance of major elements on the moon.

Images