Inversion method and system for surface oxide in lunar remote sensing image

Abstract

The invention relates to the technical field of satellite remote sensing image processing and application, in particular to an inversion method and system for surface oxides in lunar remote sensing images, and the method comprises the steps: integrating the ground real data of a Chang'e 6 lunar back surface return sample, the existing lunar front surface sample data and the remote sensing spectrum data of a SELENE multiband imager; constructing a complete data basis; and then updating the lunar surface oxide inversion model through a residual convolutional neural network to realize accurate quantification of the lunar surface oxide abundance. The new map generated based on the inversion result can refine the range of the earth body on the back of the moon, provides a key basis for deep cognition of the highland area on the back of the moon and deep substances exposed in the Artemikine Basin of Antarctic, is of great significance for revealing the geologic features and evolution process of the moon, and can provide a new guide map for future moon exploration tasks at the same time. In a deep space lunar surface oxide inversion task, the regression precision of the method is obviously superior to that of an existing model.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing image processing and application technology, specifically relating to a method and system for inverting surface oxides in lunar remote sensing images. Background Technology

[0002] As the closest celestial body to Earth, the Moon plays a crucial role in revealing the formation and evolution of the Earth-Moon system and the solar system. The lunar regolith, formed by space weathering, contains abundant ejecta sediments, disordered lunar crustal material, and may also contain lunar mantle material. Therefore, observing the composition of the lunar surface is a core approach to revealing the Moon's internal composition. Analysis of returned lunar samples reveals that the main elements on the lunar surface include aluminum (Al), calcium (Ca), iron (Fe), magnesium (Mg), silicon (Si), and titanium (Ti), and the abundance of their corresponding oxides has been clearly defined. Among these, ferrous oxide (FeO) and titanium dioxide (TiO2) are not only directly related to the optical properties of the lunar surface but also have significant value for the development and utilization of lunar metal oxide resources (such as ilmenite). The content of aluminum oxide (Al2O3) and FeO in the lunar surface shows a negative correlation, a phenomenon presumably related to the binary mixing of plagioclase and magnesian minerals within the lunar surface. Simultaneously, the magnesium molar ratio Mg / (Mg+Fe) (Mg#) is one of the key parameters for revealing subsequent lunar magmatic processes. In summary, the content of the aforementioned major oxides is crucial for understanding the chemical characteristics and evolutionary history of the Moon.

[0003] Quantifying elemental abundance from spectral information is the core objective of lunar spectral remote sensing technology. By combining returned lunar samples with remote sensing spectra from sampling points, a global elemental oxide abundance map of the Moon can be created. Representative existing research results include: obtaining FeO and TiO2 abundance maps based on image data from the Clementine ultraviolet-visible camera and the Kaguya Multispectral (MI) imager (see: "Lunariron and titanium abundance algorithms based on final processing of Clementine ultraviolet-visible images," J. Geophys. Res. Planets, vol. 105, no. E8, pp. 20297-20305, 2000); and combining data from the Lunar Prospector Gamma-ray Spectrometer (LPGRS) and the Kaguya Multispectral (MI) imager. Based on the aguya GRS data and hyperspectral images from the Chandrayaan-1 Lunar Mineralogy Mapper (M3), a map of the distribution of refractory elements on the Moon can be drawn (see: "Global mapping of flunar refractory elements: Multivariate regression vs. machine learning," Astronomy & Astrophysics, vol. 627, no. A15, 2019); using Christiansen features (CF) data from the Lunar Reconnaissance Orbiter (LRO) Diviner, a map of the global abundance of major elements on the Moon can be drawn (see: "Global estimates of flunar surface chemistry derived from LRO Diviner data," Icarus, vol. 371, pp. 114697, 2022).

[0004] Existing research indicates that establishing the correlation between the abundance of major elemental oxides and their spectral characteristics relies on real-world data from lunar samples returned from the Moon. Lunar samples collected from the near-side regolith by the six Apollo missions, three lunar missions, and the Chang'e-5 mission, as well as elemental abundance data analyzed by the Alpha Particle X-ray Spectrometer (APXS) aboard Chang'e-3 and Chandrayaan-3, have provided valuable real-world information about the near-side of the Moon. However, the far side of the Moon (especially the South Pole-Aitken Basin, SPA) is one of the most chemically complex and diverse regions on the Moon. Due to the lack of real-world data from previous missions, the identification and analysis of the composition of this region remains insufficient. To date, scholars both domestically and internationally have proposed various methods for retrieving lunar surface oxides, but these methods still have significant limitations, including: (1) Due to the lack of high-resolution images, the significant advantages of optical remote sensing technology have not been fully realized; and a large number of samples are collected from the same pixel range sensed by the orbital spectrometer, which makes it impossible to effectively distinguish the samples and affects the inversion accuracy. (2) Although the amount of real data returned by various exploration missions is increasing, the number of sampling points on the near side of the moon is still limited, making it difficult to fully reflect the compositional diversity of the global surface of the moon, which limits the applicability of inversion models based on these sampling points. (3) Existing studies mostly use empirical formulas, support vector machines, random forests and other methods. The network structure of the models used is relatively simple, which makes it difficult to fully explore the complex nonlinear relationship between oxide abundance and spectral features. This can easily lead to potential overfitting risks in the inversion model and insufficient robustness of the model.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] To address the aforementioned technical problems in the existing technology, this invention provides a method and system for inverting surface oxides in lunar remote sensing images, solving the problem of insufficient samples in the inversion task of lunar surface oxides from multispectral remote sensing images, as well as the problems in the extraction and utilization of spectral feature information in oxide inversion models.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: Firstly, a method for inverting surface oxides in lunar remote sensing images includes: S1. Collect real ground data from sampling points on the far side of the moon by Chang'e 6, integrate historical sample data from the near side of the moon, and the orbital spectral dataset from the SELENE multi-band imager to construct a training sample set; S2. Construct a residual convolutional neural network model, wherein the residual convolutional neural network model embeds residual blocks in the skip connections of a one-dimensional convolutional neural network; S3. Train the residual convolutional neural network model based on the training sample set, and use the trained model to invert the global abundance of six oxides and magnesium index on the lunar surface; S4. Based on the inversion results of step S3, generate a global chemical composition map of the moon and optimize the boundaries of geological units on the far side of the moon by combining the sampling data of Chang'e 6.

[0008] Furthermore, in step S1, the training sample set contains 116 samples from 56 lunar sampling points, among which the Chang'e 6 sampling point data is the first time that real ground data from the far side of the moon has been included. The orbital spectral data of the SELENE multi-band imager selects eight spectral bands, with center wavelengths of 415nm, 750nm, 900nm, 950nm, 1001nm, 1050nm, 1250nm, and 1550nm, respectively.

[0009] Furthermore, terrain shadow correction is performed on the raw data of the SELENE multi-band imager to obtain a 5-level reflectance map; when extracting the reflectance data corresponding to the 8 bands based on the actual latitude and longitude of the sample points, the bilinear interpolation method is used to calculate the weighted average of the four adjacent pixels around the sampling point.

[0010] Furthermore, in step S2, when constructing the residual convolutional neural network model, two spectral coefficients are introduced for the inversion of titanium dioxide and ferrous oxide, respectively. The structure of the residual convolutional neural network model includes an input layer, a multi-stage residual block group, a global pooling layer, and a linear transformation layer. The input layer is used to extract preliminary features from the spectral data. Each stage in the multi-stage residual block group contains two 3×3 convolutional layers and one skip connection. The global pooling layer is used to compress the feature dimension, and the linear transformation layer is used to output the abundance values ​​of oxides and magnesium index.

[0011] Furthermore, in step S2, leave-one-out cross-validation is used to validate the residual convolutional neural network model. The specific process is as follows: The training sample set is divided into n folds, where n is the total number of samples in the training sample set. During each validation, samples from n-1 folds are selected as the training subset to train the model, and samples from the remaining 1 fold are used as the test subset to validate the model performance. This process is repeated n times, so that each sample is used as the test subset once. Finally, the average of the n validation results is taken as the performance evaluation result of the Res-CNN model to suppress model overfitting.

[0012] Furthermore, in step S3, when using the trained residual convolutional neural network model for inversion, the reflectance data of eight spectral bands of each pixel point globally are first extracted and input into the model to obtain the oxide abundance of each pixel point; then the whole-month remote sensing image is divided into several sub-images, and inversion prediction is performed on each sub-image. After all sub-images are predicted, the inversion results of each sub-image are mosaicked to form a global map of oxide and magnesium index abundance at the whole-month scale.

[0013] Furthermore, in step S3, the global distribution of the six major oxides obtained from the inversion exhibits a ternary model: The average content of ferrous oxide in the lunar maria region is 15.36 wt%, the average content of titanium dioxide is 2.44 wt%, and the standard deviation of the spatial distribution variation of ferrous oxide is 4.63. The average content of alumina in the highland area is 24.99 wt%, and the average content of calcium oxide is 15.10 wt%, and the alumina and calcium oxide are evenly distributed in the highland area. The Antarctic Aitken Basin has the highest average silica content at 46.38 wt%, while the average ferrous oxide content is 12.28 wt% and the average alumina content is 18.19 wt%, with the content of both falling between that of the lunar maria and the highlands.

[0014] Furthermore, in step S4, the optimization of the geological unit boundaries on the far side of the moon includes the optimization of the South Pole-Aitken Basin region: Using the magnesium index threshold, the extent of the magnesia-pyroxene ring in the Antarctic-Aitken Basin was updated, while the extent of the compositional anomaly region in the Antarctic-Aitken Basin was narrowed. The Chang'e 6 landing site is located at the boundary between the magnesia-pyroxene ring and the compositional anomaly region in the Antarctic-Aitken Basin.

[0015] Furthermore, in step S4, in the geological analysis at the landing area scale, a false-color composite map is created by combining the inverted full-moon chemical composition map, and the geological features of the Chang'e 6 landing area are analyzed through the false-color composite map.

[0016] Secondly, an inversion system for surface oxides in lunar remote sensing images includes: Data acquisition and integration module: used to collect real ground data from sampling points on the far side of the moon by Chang'e 6, integrate historical sample data from the near side of the moon, and the orbital spectral dataset from the SELENE multi-band imager to build a training sample set; Model building module: used to build residual convolutional neural network models, wherein residual blocks are embedded in the skip connections of a one-dimensional convolutional neural network; Oxide Inversion Module: Used to train the residual convolutional neural network model constructed by the model building module based on the training sample set output by the data acquisition and integration module, and to invert the global abundance of six oxides and magnesium index on the lunar surface through the trained model. Geological Unit Optimization Module: This module receives the inversion results output by the oxide inversion module, generates a global chemical composition map of the moon based on the inversion results, and optimizes the boundaries of geological units on the far side of the moon by combining the data from the Chang'e 6 sampling points.

[0017] Compared with existing technologies, the present invention provides a method and system for inverting surface oxides in lunar remote sensing images. Utilizing the SELENE (Kaguya) multi-band imager (MI) dataset and combining it with the first real data of the far side of the moon acquired by Chang'e 6, this invention conducts an inversion study of six major oxides on the lunar surface. To update and improve the global distribution of these six oxides and the magnesium index, this study proposes a residual convolutional neural network model. This model effectively enhances robustness and improves inversion accuracy by integrating residual blocks into each skip connection of the one-dimensional convolutional neural network. Furthermore, based on the newly obtained global chemical composition distribution map of the moon, this study further compares the inversion results from different spectral sensors and discusses the compositional patterns and spatial distribution characteristics of the lunar surface in depth, providing support for lunar geological evolution analysis. Attached Figure Description

[0018] Figure 1 A flowchart of the method for inverting surface oxides in lunar remote sensing images based on residual networks provided in an embodiment of the present invention; Figure 2 A model structure diagram provided for an embodiment of the present invention; Figure 3 This is a distribution diagram of FeO provided in an embodiment of the present invention; Figure 4 This is a distribution diagram of TiO2 provided in an embodiment of the present invention; Figure 5 Al2O3 distribution diagram provided for embodiments of the present invention; Figure 6 The MgO distribution diagram provided in the embodiments of the present invention; Figure 7 CaO distribution diagram provided for embodiments of the present invention; Figure 8 This is a distribution diagram of SiO2 provided in an embodiment of the present invention; Figure 9 This is a Mg# distribution diagram provided for an embodiment of the present invention; Figure 10 This is an updated diagram of different components in the SPA region provided in an embodiment of the present invention; Figure 11 This is an updated map of different components of the landing area provided in an embodiment of the present invention. Detailed Implementation

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

[0020] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0021] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0022] Example 1 See Figure 1 , Figure 1 This is a flowchart of a method for inverting surface oxides in lunar remote sensing images proposed in this invention. Specific steps may include: S1. Collect real ground data from sampling points on the far side of the moon by Chang'e-6 (CE-6), integrate historical sample data from the near side of the moon, and the orbital spectral dataset from the SELENE multi-band imager to construct a training sample set. The training sample set contains 116 samples from 56 lunar sampling points, among which the CE-6 sampling point data is the first real ground data from the far side of the moon to be included. Specifically, this includes the following steps: S11. Using the topographically shading-corrected SELENE (Kaguya) Multispectral (MI) five-level reflectance map, the MI data consists of nine spectral bands, including five ultraviolet and visible bands centered at 415 nm, 750 nm, 900 nm, 950 nm, and 1001 nm, and four near-infrared bands centered at 1000 nm, 1050 nm, 1250 nm, and 1550 nm. Due to the similarity between the 1000 nm and 1001 nm bands, this study applied eight spectral bands (415 nm, 750 nm, 900 nm, 950 nm, 1001 nm, 1050 nm, 1250 nm, and 1550 nm) to calculate the oxide content. S12. Considering the limited number of returned samples and the spatial resolution of the MI data, this study integrated sample data from the CE-6 mission to further enrich the test dataset, collecting a total of 116 samples from 56 lunar sampling points. S13. Based on the actual latitude and longitude of the sample points, extract the reflectance data for the corresponding 8 bands. The center wavelengths of the 8 spectral bands are 415nm, 750nm, 900nm, 950nm, 1001nm, 1050nm, 1250nm, and 1550nm, respectively. The following examples will further illustrate the implementation of SELENE (Kaguya) multispectral remote sensing imagery and sample point matching, preparing for subsequent model input: The acquired multispectral imagery data originated from images of SELENE (Kaguya) and was acquired by the Japan Aerospace Exploration Agency (JAXA). This is a spectral image derived from the Kaguya multiband imager, developed in collaboration with the University of Hawaii. It was created from topographically corrected MI reflectance data obtained from the Level 2 processing version of the Kaguya Archive. The multiband imager has nine spectral bands: 415nm, 750nm, 900nm, 950nm, 1000nm, 1001nm, 1050nm, 1250nm, and 1550nm, with a resolution of 59m, covering a range of 65 degrees north and south latitude.

[0023] In the specific implementation process, eight spectral features are combined, and then the reflectance is extracted from the latitude and longitude values ​​of the sampling points to form the input feature data. The latitude and longitude data of the sampling points are shown in Table 1.

[0024] During the extraction process, bilinear interpolation is applied to extract spectral data. The primary aim is to address the potential spatial mismatch between the sampling point coordinates and the pixels in the remote sensing image, thereby improving the accuracy of reflectance extraction and suppressing noise. Specifically, for each sampling point, we first identify four neighboring pixels in the image. Then, the reflectance value of the sampling point is calculated as a weighted average of these four neighboring pixels, where the weights are determined from the sampling point to the center of each pixel (closer pixels have higher weights). This method not only reduces the impact of random noise from individual pixels but also provides more representative and stable spectral reflectance values ​​for subsequent analysis, improving the reliability of the analysis. Combinations of different bands can provide sufficient spectral feature information, which is helpful for subsequent deep learning models to fully extract information and perform deep learning.

[0025] Table 1

[0026] S2. Construct a Residual Convolutional Neural Network (Res-CNN) model, in which residual blocks are embedded in the skip connections of a one-dimensional convolutional neural network; specifically, this may include: S21, in addition to the eight spectral bands (415nm, 750nm, 900nm, 950nm, 1001nm, 1050nm, 1250nm and 1550nm), two dedicated spectral coefficients are also introduced simultaneously, respectively for titanium dioxide ( ) and ferrous oxide ( The inversion of spectral coefficients is performed; this design can suppress the interference of spatial weathering on the inversion results to a certain extent. The calculation expressions for the two spectral coefficients are as follows: For Inverted spectral coefficients: ;in, The reflectivity is in the 415nm band. The reflectivity is for the 750nm wavelength band; For Inverted spectral coefficients: ;in, The reflectivity is in the 950nm band. The reflectivity is for the 750nm band.

[0027] The feature processing and output process of S22 and Res-CNN models may specifically include: S221. Preliminary Feature Extraction: The model input layer first performs preliminary feature extraction on the input spectral data (including reflectance in 8 bands and 2 spectral coefficients) to provide basic features for subsequent deep processing. S222, Multi-stage Residual Block Group Processing: Initial features are fed into multi-stage residual block groups. Each stage in the residual block group contains two 3×3 convolutional layers and one skip connection. The residual block is a neural network structure that directly passes the input to the deep layers through the skip connection. This structure enables the network to learn residuals more easily. Its skip connection allows the gradient to be directly backpropagated to the shallow layers, effectively alleviating the gradient vanishing problem that is prone to occur in deep networks. At the same time, the residual block can also extract multi-scale features and enhance the ability to preserve detailed information. S223. Feature Compression and Output: The features processed by the multi-stage residual block group are sequentially passed through the global pooling layer (compressing the feature dimension and reducing computational complexity) and the linear transformation layer, and finally output the abundance result of lunar surface oxides.

[0028] S23. To suppress overfitting, leave-one-fold cross-validation is used to validate the performance of the Res-CNN model. Specifically, this may include: S231, leave-one-fold cross-validation is a commonly used model performance evaluation method and a special form of cross-validation. Its core idea is to make full use of limited data to estimate the generalization ability of the model as objectively as possible, which is especially suitable for scenarios with extremely small sample sizes. S232. Assuming the dataset to be validated contains n samples, firstly, the dataset is split into n folds, each containing only one sample; each time, samples corresponding to n-1 folds are selected from the n folds as the training set to train the Res-CNN model; the samples corresponding to the remaining 1 fold are used as the test set to evaluate the model performance (such as computational error, accuracy, etc.); the above training and evaluation process is repeated n times to ensure that each sample is used as a test set exactly once; finally, the average of the n evaluation results is taken as the final performance estimate of the model; This method avoids performance evaluation fluctuations caused by random sample partitioning, and since almost all samples participate in training and testing, it can more accurately reflect the model's generalization ability in practical applications.

[0029] S3. Train a residual convolutional neural network model based on the training sample set, and use the trained model to invert the six oxides on the lunar surface. ) and magnesium index ( The global abundance of ) can be specifically included in: S31. When performing inversion using the trained Res-CNN model, first extract the reflectance data of 8 spectral bands of each pixel in the global range, input the data into the model to obtain the oxide abundance of each pixel; since there are many pixels in the global image, the global image is first divided into segments, and the network processes one row of data at a time through a loop, and finally the output results are stitched together to form a complete inversion result image. S32. Divide the full-month remote sensing image into several sub-images, perform inversion prediction on each sub-image, and after all sub-images are predicted, mosaic the inversion results of each sub-image to form a global map of oxide and magnesium index abundance at the full-month scale.

[0030] S33, see reference Figure 2-9 Based on the inverted global distribution map, combined with and The content, calculate Specifically, Magnesium content is an important indicator for measuring the relative enrichment of magnesium in rocks or minerals, and is widely used to characterize magma evolution, mineralogy, or rock type. and The ratio, usually expressed as a percentage (range 0-100), indicates a higher relative magnesium content. The global distribution of major oxides follows a ternary pattern (lunar maria, highlands, and the Antarctic-Aitken (SPA) basin): Yuehaifu (15.36wt%) and (2.44wt%), reflecting the evolution of basaltic magma, The spatial variation is the greatest (std=4.63), revealing the compositional diversity of different lava flows.

[0031] Highland region rich (24.99wt%) and CaO (15.10wt%) are relatively evenly distributed, representing the original anorthosite lunar crust.

[0032] SPA Basin The highest (46.38wt%) (12.28wt%) and (18.19wt%) Located between the lunar maria and highlands, suggesting exposure of brilliance-rich material in the upper mantle of the lower lunar crust. Highlands on the far side. (24.49wt%) and (14.89wt%) Approaching global highlands, but (7.21 wt%) slightly higher than the highland average, reflecting a wider incorporation of magnesian rocks. Global The average distribution is approximately 65.0, with lower values ​​in lunar maria, which is associated with basalt evolution, and higher values ​​in highlands. High-altitude, mainly composed of plagioclase, SPA Medium-sized, mixed with deep materials, with relatively clear boundaries between the three major regions.

[0033] S4. Based on the inversion results of step S3, generate a global lunar chemical composition map and optimize the boundaries of geological units on the far side of the moon using data from the CE-6 sampling point. Optimization of the boundaries of geological units on the far side of the moon includes optimization of the South Pole-Aitken Basin region: using a magnesium index threshold, update the extent of the magnesia-pyroxene zonation zone within the South Pole-Aitken Basin, while simultaneously reducing the extent of the compositional anomaly zone in the South Pole-Aitken Basin, with the CE-6 landing site located at the boundary between the magnesia-pyroxene zonation zone and the compositional anomaly zone in the South Pole-Aitken Basin. Specifically, this includes: S41, see reference Figure 10-11 For the SPA region, which has a more complex composition, the extent of the magnesite zonation zone and the SPA compositional anomaly zone within the SPA basin has been updated.

[0034] Magnesia-pyroxene zoning: Based on the new Mg# map, using two thresholds of 60 and 64, the boundary of MPA was found to expand in the northern part of the SPA basin. The most magnesium-rich exposed area ( )lie in Meteorite crater, Crater walls, Crater walls and some smaller crater walls. This indicates that the SPA impact excavated and exposed a wider range of magnesium-rich subsurface material than previously thought, and that the near-end ejecta was less mixed with the lunar crustal basement than initially inferred, preserving a wider range of pristine excavated lower lunar crust and upper mantle material.

[0035] The shrinking of the SPA compositional anomaly zone, with an average Mg# value of approximately 56.56, is attributed to the differentiation of the SPA impact melt layer, the formation of a smooth plain in the nearby basin, or the deposition of ancient high-alumina basalt. This shrinking extent suggests that the calc-iron pyroxene is confined to the vicinity of the central depression and isolated volcanic deposits, implying that the lateral extent of the post-impact alteration may have been limited by pre-existing topographic lowlands or tectonic boundaries. Furthermore, the CE-6 landing site lies on the boundary between the magnesian pyroxene zonation zone and the SPA compositional anomaly zone.

[0036] S42. In the geological analysis at the landing area scale, a false-color composite map is created by combining the inverted full-moon chemical composition map, and the geological characteristics of the CE-6 landing area are analyzed through the false-color composite map.

[0037] Analysis of the CE-6 landing site's geology and distribution reveals that it is located within the SPA Basin, south of the Apollo Basin, and exhibits complex regional composition and lithology. Some crater walls contain localized rich deposits. The lithology indicates the presence of numerous plagioclase lithologies beneath the weathered layer on the surface of the mafic material; the landing area The values ​​are heterogeneous. The largest landing area in the image The value was 69.0, significantly higher than that of typical lunar mare basalts in this region, which is consistent with the fine-grained nephrite debris found in the CE-6 sample. Consistent with the data, the landing area is predominantly composed of eutectic material. This high-magnesium material is interpreted as deep plutonic rocks exposed by the SPA or Apollo impact event. The SPA impact may have penetrated the crust-mantle boundary, excavating and melting lunar mantle-derived material to form a magma ocean, which then differentiated and stratified.

[0038] The experiments above demonstrate that this invention, based on SELENE (Kaguya) multispectral data and combined with the latest CE-6 data and other experimentally sampled data, extracts corresponding reflectance to form a dataset. Simultaneously, the proposed Res-CNN model extracts spectral feature information for the inversion of six major lunar surface oxides. Experimental results verify the superiority of this method; in the inversion task of deep-space lunar surface oxides, the regression accuracy is significantly better than existing models.

[0039] Example 2 This invention proposes an inversion system for surface oxides in lunar remote sensing images, comprising: Data acquisition and integration module: used to collect real ground data from sampling points on the far side of the moon by Chang'e 6, integrate historical sample data from the near side of the moon, and the orbital spectral dataset from the SELENE multi-band imager to build a training sample set; Model building module: used to build residual convolutional neural network models, which embed residual blocks in the skip connections of one-dimensional convolutional neural networks; Oxide Inversion Module: Based on the training sample set output by the data acquisition and integration module, the module trains the residual convolutional neural network model built by the model building module, and uses the trained model to invert the global abundance of six oxides and magnesium index on the lunar surface. Geological Unit Optimization Module: This module receives the inversion results output by the oxide inversion module, generates a global chemical composition map of the moon based on the inversion results, and optimizes the boundaries of geological units on the far side of the moon by combining the data from the Chang'e 6 sampling points.

[0040] In summary, the present invention has the following advantages: 1. By integrating real ground data from sampling points on the far side of the moon by Chang'e 6, combined with samples from the near side of the moon and high-resolution spectral datasets from the SELENE multi-band imager, a dataset of 116 samples covering 56 sampling points worldwide was constructed, providing a cross-regional, multi-scale training foundation for deep learning models and significantly improving the model's generalization ability in unexplored areas. 2. By using a fusion inversion framework based on residual convolutional neural networks, optimizations are achieved in terms of algorithm architecture and performance. Residual blocks are embedded in the skip connections of the one-dimensional convolutional neural network, and leave-one-out cross-validation is combined to suppress the gradient vanishing and overfitting problems caused by the network depth of traditional CNNs, thereby achieving high-precision inversion. 3. By using new oxide inversion mapping and combining it with the magnesium index to improve the boundaries of the three major geological units, the extent of the magnesia zonation zone and compositional anomaly zone in the Antarctic-Aitken Basin was corrected, revealing the geological background around the landing area and providing key guidance for the exploration of deep lunar materials.

[0041] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for inverting surface oxides in lunar remote sensing imagery, characterized in that, The application relates to a method for generating a global lunar chemical composition map, comprising the following steps: S1, collecting ground true data of Chang'e-6 lunar back sampling points, integrating historical lunar front sample data, and constructing a training sample set by using orbit spectral data of a SELENE multi-band imager; S2, constructing a residual convolutional neural network model, wherein a residual block is embedded in a skip connection of a one-dimensional convolutional neural network; S3, training the residual convolutional neural network model based on the training sample set, and inversing global abundances of six oxides and a magnesium index on the lunar surface by using the trained model; and S4, generating a global lunar chemical composition map based on the inversion result of step S3, and optimizing a lunar back geological unit boundary in combination with Chang'e-6 sampling point data. In step S1, the training sample set contains 116 samples of 56 lunar sampling points, wherein Chang'e-6 sampling point data are first lunar back ground true data; The orbit spectral data of the SELENE multi-band imager are selected from eight spectral bands, and the center wavelengths of the eight spectral bands are 415 nm, 750 nm, 900 nm, 950 nm, 1001 nm, 1050 nm, 1250 nm and 1550 nm. Terrain shadow correction is performed on the original data of the SELENE multi-band imager to obtain a five-level reflectivity map; when the reflectivity data of the eight bands corresponding to the actual latitude and longitude of the sample point are extracted, a bilinear interpolation method is adopted to calculate the weighted average value of the four adjacent pixels around the sampling point. In step S2, when the residual convolutional neural network model is constructed, two spectral coefficients are introduced for the inversion of titanium dioxide and ferrous oxide; 2. The method of claim 1, wherein, The structure of the residual convolutional neural network model comprises an input layer, a multi-stage residual block group, a global pooling layer and a linear transformation layer in sequence, wherein the input layer is used for extracting preliminary features of spectral data, each stage of the multi-stage residual block group contains two 3*3 convolutional layers and a skip connection, the global pooling layer is used for compressing the feature dimension, and the linear transformation layer is used for outputting the abundance values of oxides and a magnesium index. In step S2, the residual convolutional neural network model is verified by using a leave-one-out cross-validation method, and the specific process is as follows:

3. The method of claim 2, wherein, The training sample set is divided into n folds, n is the total number of samples in the training sample set, n-1 folds of samples are selected as a training subset to train the model each time, the remaining 1 fold of samples is used as a test subset to verify the model performance, the process is repeated n times, each sample is used as the test subset once, and finally the average value of the n verification results is taken as the performance evaluation result of the Res-CNN model to inhibit model overfitting.

4. The method of claim 1, wherein, In step S3, when the trained residual convolutional neural network model is used for inversion, the eight spectral band reflectivity data of each pixel point in the global range are extracted and input into the model to obtain the oxide abundance of each pixel point; the global remote sensing image is cut into a plurality of sub-images, each sub-image is inversely predicted, after the prediction of all the sub-images is completed, the inversion results of the sub-images are inlaid and spliced to form a global oxide and magnesium index abundance map of the full moon scale. ​ 5. The method of claim 4, wherein, ​ ​ 6. The method of claim 1, wherein, ​ 7. The method of claim 6, wherein, In step S3, the global distribution of the six major oxides obtained by inversion presents a ternary pattern: The average content of ferrous oxide in the mare region is 15.36wt%, and the average content of titanium dioxide is 2.44wt%, and the spatial distribution variation standard deviation of ferrous oxide is 4.63; In the high region, the average content of alumina is 24.99wt%, and the average content of calcium oxide is 15.10wt%, and the alumina and calcium oxide are uniformly distributed in the high region; In the Antarctic Eitken basin, the average content of silicon dioxide is the highest, which is 46.38wt%, the average content of ferrous oxide is 12.28wt%, and the average content of alumina is 18.19wt%, and the content of the two is between the mare region and the high region.

8. The method of claim 1, wherein, In step S4, the optimization of the boundary of the geological unit on the back of the moon includes the optimization of the Antarctic Eitken basin region: The threshold of the magnesium index is used to update the range of the enstatite ring belt in the Antarctic Eitken basin, and at the same time, the range of the composition anomaly area in the Antarctic Eitken basin is reduced, and the Chang'e-6 landing point is located at the boundary of the enstatite ring belt and the composition anomaly area in the Antarctic Eitken basin.

9. The method of claim 8, wherein, In step S4, in the geological analysis at the landing area scale, a false color composite image is made by combining the global chemical composition map generated by inversion, and the geological features of the Chang'e-6 landing area are analyzed through the false color composite image.

10. A system for inverting surface oxides in lunar remote sensing imagery, comprising: Including: A data acquisition and integration module is used to collect the ground true data of the Chang'e-6 lunar back sampling points, integrate the historical lunar front sample data, and construct a training sample set by integrating the orbital spectral data set of the SELENE multi-band imager; A model construction module is used to construct a residual convolutional neural network model, which embeds a residual block in the skip connection of a one-dimensional convolutional neural network; An oxide inversion module is used to train the residual convolutional neural network model constructed by the model construction module based on the training sample set output by the data acquisition and integration module, and to invert the global abundance of the six oxides and the magnesium index on the lunar surface through the trained model; A geological unit optimization module is used to receive the inversion results output by the oxide inversion module, generate a global chemical composition map of the moon based on the inversion results, and optimize the boundary of the geological unit on the back of the moon in combination with the Chang'e-6 sampling point data.

Images