为了有效利用已标记与未标记样本提高高光谱遥感影像分类精度, 提出一种新的半监督流形学习方法——半监督稀疏鉴别嵌入算法(SSDE)。该算法结合了近邻流形结构及稀疏性的优点, 不仅保留样本间的稀疏重构关系, 而且通过引入少量有标记的训练样本以及大量无标记训练样本来获得高维数据的内在属性以及低维流形结构, 实现鉴别特征提取, 提高分类精度。在Washington DC Mall和Indian Pine数据集上的分类识别实验表明, 该算法能够较为有效地发现高维空间中数据的内蕴结构, 分类性能比其他算法有明显的提升。在随机选取8个有类别标记和60个无类别标记的数据作为训练样本的情况下, 本文提出的SSDE算法在上述两个数据集上的分类精度分别达到了77.36%和97.85%。

To improve the classification accuracy of hyperspectral remote sensing images by utilizing labeled and unlabeled samples, a new semi-supervised manifold learning method called Semi-supervised Sparse Discriminant Embedding (SSDE) is proposed. By combining the advantages of manifold structure among classes and sparsity, the algorithm not only preserves the sparse reconstruction relationship between the samples, but also gets the intrinsic attribute of high dimensional data and the manifold structure of low dimensional data by introducing a few labeled training samples and a large number of unlabeled training samples. So,it extracts the discriminant feature of data and improves the classification accuracy. The classification experiments in Washington DC Mall and Indian Pine data set show that the method is a more effective way to find the internal structure of data in a high dimensional space. Compared to other methods, the SSDE obviously improves the classification performance. By taking randomly selected 8 training samples with classification labels and 60 ones without classification labels as examples, the highest classification precision of SSDE respectively reach 77.36% in Indian Pine and 97.85% in Washington DC Mall data set.