在观测空间目标时, 往往会受到地基观测仪器等因素的制约, 导致无法利用目标图像信息从外形上进行识别。 根据不同空间目标表面组成材料不同, 其产生的反射光谱会存在差异这一特性, 可利用空间目标特有的光谱信息进行识别分类。 基于此, 从光谱学角度对空间目标识别算法进行研究, 在K最近邻算法(KNN)的基础上, 采用了一种自适应权重局部超平面方法(AWKH), 算法主要在计算预测样本与超平面距离时加入对特征权重的考虑, 构建了以样本特征组间差与组内差的比值作为特征权重值的超平面模型, 从而提高了分类效果和分类效率。 为验证算法的分类效果, 本文进行了四组验证实验, 第一组实验将美国地质勘探局数据库中提取出的九种常用材料光谱随机选出三种混合成多类进行识别； 第二、 三组实验将四种常用空间目标材料的光谱作为纯物质光谱, 分别从可见光和近红外波段对其混合物质进行分类； 第四组实验通过实测四个方形模型样本六个面的光谱对其进行识别分类。 实验过程中将实验结果与目前常用的支持向量机（SVM）进行对比, 对比结果表明改进后的AWKH算法在识别精度和样本适用范围上具有更高的优越性。

Because of ground observation instruments and other factors, we can not recognize the space target only from the external shape in the image. Since the reflection spectrum of the space target is determined by the surface material of space object, spectral analysis technique can be used for classifying the space objects. Based on the K-nearest neighbor algorithm (KNN), a method called adaptive weight k-local hyperplane (AWKH) is proposed in this paper. The main improvement of the algorithm is that weight discrimination is added in the processes of calculating the hyperplane distance between predicted samples. The algorithm constructs a hyperplane model by using the difference between the groups and within group ratio for the weights of features. In order to verify the classification effectiveness and efficiency of the algorithm, this paper carried out four sets of verification experiments. In the first set of experiments, 9 kinds of common materials were extracted from the database of United State Geological Survey. Then 3 kinds of these materials were mixed into multi-class objections. In the second and third sets of experiments, the spectra of four normal space target materials were mixed in different classes. Then these classes were identified from the visible and near-infrared wave bands. In the fourth set of experiments, four square models of hexahedron were classified by the spectra of their surface material. The experimental results indicate that the AWKH algorithm has more advantages in identification accuracy and effectiveness of the complex samples by comparing with the support vector machine (SVM) method.