Asteroid spectrum intelligent classification method based on deep learning

Abstract

The invention discloses an asteroid spectrum intelligent classification method based on deep learning, and belongs to the technical field of asteroid classification. The method comprises the following steps: S1, collecting original spectral data; s2, labeling the original spectral data in an One-Hot format; s3, performing data enhancement and expansion on the original spectrum data to obtain a spectrum training set, and taking the original spectrum data as a test set; s4, constructing a deep learning spectrum classification model A-Net, and substituting the spectrum training set into the constructed model for training; and S5, calculating the precision of each category and the average precision of the trained model on the test set, and evaluating the classification performance of the model. The deep learning spectrum classification model A-Ne based on the combination of the convolutional neural network and the Transform network is high in precision and good in robustness and generalization, and can effectively solve the problems of low classification precision and the like due to the fact that a traditional machine learning model is limited in nonlinear fitting capability and high-latitude feature extraction capability, poor in noise robustness and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of asteroid classification, and specifically relates to an asteroid spectral intelligent classification method based on deep learning. Background Art

[0002] The vast majority of asteroids are concentrated in the asteroid belt between the orbits of Mars and Jupiter. Asteroids are generally believed to have evolved from planetesimals during the formation of the solar system. They are the most discovered solar system objects. According to data from the Asteroid Center as of December 31, 2020, 1,026,572 asteroids have been confirmed in the solar system, and the number is still increasing by hundreds each year. Since changes in asteroid material are only related to the solar system environment and mutual collisions, asteroids contain information about the early formation and evolution of the solar system and are regarded as living fossils for studying the solar system. Research on asteroids is of great significance to understanding the formation and evolution of the solar system, planet formation, and possible development of space resources. Therefore, the detection of asteroids has always been the focus and hotspot of international deep space exploration activities.

[0003] Asteroid spectra are an important tool for studying asteroids. Before the probe arrives at the asteroid, the reflection spectrum of the asteroid can be observed using ground-based infrared and optical telescopes. By analyzing the spectral data obtained from the observation, we divide it into different categories and preliminarily determine its composition and mineral composition based on the reflection and absorption characteristics of the asteroid. At the same time, different types of asteroids reflect different evolutionary processes and environmental conditions. For example, carbon-type (C) asteroids may be the remnants of early materials in the formation of the solar system, while metallic (S) asteroids may be the remains of planetary fragments. By analyzing the spectral characteristics of different types of asteroids, their possible origins, evolutionary histories, and evolutionary environments can be inferred. In addition, asteroid spectral analysis can determine the types and distribution of resources on its surface, such as water, metal minerals, rare elements, etc. These resources are of great significance for the future development and utilization of space resources. Spectral data can help identify potential resource-rich areas and provide targets and guidance for future space mining missions.

[0004] Bus et al. analyzed asteroid spectra and classified existing asteroids into 24 categories, which is called the Bus-DeMeo (B-DM) classification method. B-DM is one of the latest and most mainstream asteroid classification methods.

[0005] In the past decade, in the study of asteroid classification, machine learning has been widely used in asteroid classification tasks based on reflectance spectra. Various machine learning methods such as principal component analysis (PCA), support vector machine (SVM), random forest, etc. can replace manual methods to perform intelligent feature extraction and analysis of spectra. However, due to the limited nonlinear fitting ability and high-dimensional feature extraction ability of traditional machine learning models, as well as poor noise robustness, the classification accuracy is low. Summary of the invention

[0006] In view of the above problems existing in the prior art, the purpose of the present invention is to provide an intelligent classification method for asteroid spectra based on deep learning, which is based on a deep learning spectral classification model A-Net that combines convolutional neural networks with Transformer networks. Compared with traditional machine learning methods and other deep learning models, A-Net has obvious advantages in the accuracy of asteroid spectral classification.

[0007] The present invention provides the following technical solution: a method for intelligent classification of asteroid spectra based on deep learning, comprising the following steps:

[0008] S1. Collect raw data: collect various types of spectral data of asteroids in visible and near-infrared bands;

[0009] S2. Annotate the collected raw spectral data in One-Hot format;

[0010] S3, adding random Gaussian noise to the original spectral data in the space where the spectral information originally resides and the PCA space for data enhancement and expansion, to obtain a spectral training set, and the original spectral data is used as a test set;

[0011] S4. Construct a deep learning spectral classification model A-Net based on the combination of convolutional neural network CNN and Transformer network, and bring the spectral training set into the constructed model for training;

[0012] S5. Calculate the accuracy and average accuracy of each category of the trained model on the test set, and compare them with the learning algorithm to evaluate the model classification performance.

[0013] Furthermore, the deep learning spectral classification model A-Net is an encoding-decoding structure, the encoder includes three one-dimensional convolutional layers, the front end of the decoder includes three fully connected layers, and the back end is a classifier, including a transformer module, three fully connected layers and a softmax layer; the encoder encodes the input information, and inputs the encoded feature information into the decoder, which is further classified by the decoder.

[0014] Furthermore, in the deep learning spectral classification model A-Net, a differential module Diff is introduced to expand the input information, and the differential information and the original spectral data are input into the subsequent encoder as different feature channels to enhance the accuracy and robustness of the network; the differential module is calculated as follows:

[0015] f'(x)=f(x)-f(x-1).

[0016] Where x is the horizontal coordinate of the spectral data, that is, the albedo of the asteroid at the corresponding wavelength sampling point.

[0017] Furthermore, the activation function of the encoder is LeakyReLu, which is expressed as follows:

[0018]

[0019] In the formula, α is a manually set hyperparameter, and α is set to 0.01 in this model.

[0020] Furthermore, the softmax layer outputs the probability of each category, and the calculation formula is as follows:

[0021]

[0022] Where: x i , x j are the i-th and j-th elements of the input vector, and n represents the dimension of the vector.

[0023] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] The present invention combines the convolutional neural network CNN with the Transformer network to construct a neural network for asteroid spectral classification with deeper layers and more sophisticated design. The network can extract higher-dimensional semantic features, which helps to improve the accuracy of the network in spectral classification. A Diff difference module is added to the input end to extract additional input information, which helps the network to extract spectral gradient information and enhance the network's sensitivity to spectral peaks and troughs. At the same time, the data set is enhanced by adding random Gaussian noise to the original spectral space and PCA space. Using the enhanced data set to train the network can improve the generalization of the network and the classification robustness under noise interference. The neural network constructed by the present invention has obvious advantages in classification accuracy, generalization, and robustness over existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the method of the present invention;

[0026] Figure 2 A-Net network structure diagram of the present invention;

[0027] Figure 3 A schematic diagram of data generation for the simulation of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments of the specification. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0029] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0030] See also Figure 1-3 , an intelligent asteroid spectral classification method based on deep learning, which combines convolutional neural network (CNN) with Transformer network to construct a deep learning spectral classification model A-Net.

[0031] The network structure of A-Net is as follows Figure 2 shown.

[0032] Specifically, A-Net is designed as an encoding-decoding structure.

[0033] In the data input part, this embodiment introduces a difference module Diff to expand the input information. The difference module is calculated as follows:

[0034] f'(x)=f(x)-f(x-1)

[0035] The differential information and the original spectrum are input into the subsequent encoder as different feature channels to enhance the accuracy and robustness of the network. The encoder consists of three one-dimensional convolutional layers, and the activation function is LeakyReLu, which is expressed as follows:

[0036]

[0037] The encoder encodes the input information, and the encoded feature information is input into the decoder, which further classifies it. The decoder is designed to be similar to the U-Net structure. The front end consists of three fully connected layers with 128, 256, and 512 neurons, which further extract and refine the input feature information. The second half is the classifier, which includes a transformer module, three fully connected layers, and a softmax layer. The number of neurons in the three fully connected layers is 512, 256, and 11, respectively. The softmax layer outputs the probability of the final categories. The calculation formula of the softmax layer is as follows:

[0038]

[0039] The specific process of the intelligent classification method is as follows:

[0040] Step 1: Data collection, model training and testing. The original data used is from SMASS public asteroid visible light and near-infrared spectral data. The band range is 0.45 microns to 2.45 microns, the number of bands is 200, and each 0.01 micron is a band. Based on the Bus classification method, it is divided into four categories: S, X, C, and Other. The number of spectra in each category of the data set is shown in Table 2:

[0041] Table 2 Dataset

[0042] category Number of spectra C 66 X 47 S 178 Other 149

[0043] Step 2: Label the obtained data in One-Hot format.

[0044] Step 3: To solve the imbalance problem of data volume of each type, the spectral training set is augmented and expanded by adding random Gaussian noise to the original spectral data in the original space and PCA space, and 200 simulated data of each type are obtained as the training set. The original SMASS data is used as the test set to train and test the model. Figure 2 This is an example diagram of data simulation enhancement, in which 200 simulated data are generated based on the original measured spectrum.

[0045] Step 4: Calculate the accuracy and average accuracy of each category of the trained model on the test set, and compare it with other deep learning algorithms and machine learning algorithms to evaluate the model classification performance.

[0046] In step 4, the average accuracy of each category of the final model and the comparison results with other algorithms are shown in Tables 3 and 4. Since the amount of data in each category in the test set SMASS data is unbalanced, the accuracy (Acc) is used to evaluate the model performance. The accuracy is the total number of correctly classified spectra divided by the total number of spectra, expressed as follows:

[0047]

[0048] Where: T is the correctly classified spectrum, and F is the incorrectly classified spectrum.

[0049] Table 3 A-Net classification accuracy test results

[0050]

[0051] Table 4 Comparison results of A-Net and other algorithms

[0052] algorithm PCA+SVM Random Forest CNN-Net ResNet-50 Transformer A-Net Acc 89.64 86.24 90.93 92.74 90.13 97.28

[0053] From the results, we can see that in the task of intelligent classification of asteroid spectra, deep learning methods have mobility advantages over traditional machine learning methods, and A-Net has a higher accuracy rate than several other deep learning methods, with an accuracy rate of up to 97.28%, which can achieve high-precision intelligent classification of asteroid spectra, proving the practicality and effectiveness of the present invention.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for intelligent classification of asteroid spectra based on deep learning, characterized in that: The following steps are involved: S1. Collect raw data: collect various types of spectral data of asteroids in visible and near-infrared bands; S2. Annotate the collected raw spectral data in One-Hot format; S3, adding random Gaussian noise to the original spectral data in the original space and PCA space for data enhancement and expansion, obtaining a spectral training set, and using the original spectral data as a test set; S4. Construct a deep learning spectral classification model A-Net based on the combination of convolutional neural network CNN and Transformer network, and bring the spectral training set into the constructed model for training; S5. Calculate the accuracy and average accuracy of each category of the trained model on the test set, and compare them with the learning algorithm to evaluate the model classification performance.

2. According to claim 1, a method for intelligent classification of asteroid spectra based on deep learning is characterized in that: The deep learning spectral classification model A-Net is an encoding-decoding structure. The encoder includes three one-dimensional convolutional layers, the front end of the decoder includes three fully connected layers, and the back end is a classifier, including a transformer module, three fully connected layers and a softmax layer; the encoder encodes the input information and inputs the encoded feature information into the decoder, which is further classified by the decoder.

3. The method for intelligent classification of asteroid spectra based on deep learning according to claim 2, characterized in that: In the deep learning spectral classification model A-Net, the differential module Diff is introduced to expand the input information, and the differential information and the original spectral data are input into the subsequent encoder as different feature channels to enhance the accuracy and robustness of the network; the differential module is calculated as follows: f'(x)=f(x)-f(x-1). Where x is the horizontal coordinate of the spectral data, that is, the albedo of the asteroid at the corresponding wavelength sampling point.

4. The method for intelligent classification of asteroid spectra based on deep learning according to claim 3, characterized in that: The activation function of the encoder is LeakyReLu, which is expressed as follows: In the formula, α is a manually set hyperparameter, and α is set to 0.01 in this model.

5. The method for intelligent classification of asteroid spectra based on deep learning according to claim 4, characterized in that: The softmax layer outputs the final probability of each category, which is calculated as follows: Where: x i , x j are the i-th and j-th elements of the input vector, and n represents the dimension of the vector.

Images