多种深度学习方法组合应用于小样本空间目标分类研究

随着近年来光谱探测仪器灵敏度、精确度和易用度的不断提升，光谱技术已经深入到各行各业的物质成分的鉴定与分析中。对于空间目标的光谱观测是传统光学观测的重要拓展之一，因其具有的非接触、无损伤等优点而备受关注，然而由于观测条件所限，空间目标的光谱数据量极小，通过传统方法对其进行分类分析达不到较好效果，必须探求提高分类精度的方法。首先，通过1.2 m空间目标光学望远镜上搭载的光谱相机终端获取空间目标高光谱图像；再通过天文学测光IRAF方法，提取空间目标的一维光谱数据；为对空间目标光谱进行分类，提出一种结合多种深度学习方法解决小样本数据量的空间目标分类问题。该方法应用密度聚类方法将空间目标粗糙分类，一维生成对抗网络方法增加空间目标数据，一维卷积神经网络方法将空间目标精细分类，三者组合进而达到较好的实验效果，整体精度约为79.1%（基于密度聚类、过采样、一维卷积神经网络方法组合、基于K-means、一维生成对抗网络、一维卷积神经网络方法组合和基于K-means、过采样、一维卷积神经网络方法组合的整体精度分别约为78.4%，77.9%和77.2%）。粗糙分类模型中，密度聚类方法比K-means方法整体精度平均高出约为0.67%；数据增广模型中，一维生成对抗网络方法比过采样方法整体精度平均高出约为1.52%；精细分类模型中，一维卷积神经网络方法二层网络比三层网络整体精度平均仅高出约为0.003%，但是运算时间更长。四种组合方法精度均高于单一方法。实验结果表明本文提出的组合方法在小样本空间目标类别未知情况下，可实现细分类且精度较高，为实现空间目标极小数据量下的图谱一体化分析，提供一定参考价值。

A Combination of Multiple Deep Learning Methods Applied to Small-Sample Space Objects Classification

With the continuous improvement of the sensitivity, accuracy and easy use of spectral detection instruments in recent years, spectral technology has penetrated the identification and analysis of material components in all walks of life. Spectral observation of space targets is one of the important extensions of traditional optical observations. It has attracted much attention due to its non-contact and damage-free advantages. However, due to the limited observation conditions, the amount of spectral data of space targets is minimal. Traditional methods cannot achieve better results in classification analysis. In this paper, Firstly, the hyperspectral image of the space target is obtained through the spectroscopic camera terminal mounted on the 1.2 m space target optical telescope; Secondly, the one-dimensional spectral data of the space target is extracted through the astronomical photometric IRAF method; Finally, the combination of multiple deep learning methods, classify the spectral data of space targets. Accordingly, this paper proposes a combination of multiple deep learning methods to solve small sample data’s spatial object classification problem. This method uses Density Clustering method to roughly classify spatial targets, one-dimensional Generative Adversarial Network method to generate spatial target data, one-dimensional Convolutional Neural Network method to finely classify spatial targets, the combination of three methods can achieve relatively good experimental results and overall accuracy is about 79.1% (Based on the combination of Density Clustering, Oversampling, one-dimensional Convolutional Neural Network methods; Based on the combination of K-means, one-dimensional Generative Adversarial Network, one-dimensional Convolutional Neural Network methods; Based on the combination of K-means, Oversampling, One-dimensional Convolutional Neural Network methods, the overall accuracy is about 78.4%, 77.9%, 77.2%). In the rough classification model, the overall accuracy of the Density Clustering method is about 0.67% higher than the K-means method; In the data augmentation model, the overall accuracy of the one-dimensional Generative Adversarial Network method is about 1.52% higher than the Oversampling method; In the fine classification model, the two-layer network of the one-dimensional Convolutional Neural Network method has an average accuracy of only about 0.003% higher than the three-layer network, but the calculation time is longer. The accuracy of the four combined methods are higher than the single method. The experimental results show that the combination method proposed in this paper can achieve fine classification and high accuracy when the small sample space target category is unknown. It provides a certain reference value for realizing the integrated analysis of the map under the minimal data volume of the space target.