摘要
在高光谱影像地物分类应用中时常因光谱波段数多而导致'维数灾难'问题,提出了一种鉴别稀疏保持嵌入的维数约简算法。该方法利用稀疏表示的自然鉴别力,分别构建了类内e_1图和类间e_1图;在低维嵌入空间中,保持同类数据的内在稀疏流形结构,同时分离开非同类数据,提取出鉴别特征。DSPE不仅继承了稀疏表示的优点,而且增加了非同类数据间的可分性。在PaviaU和Urban高光谱数据集上的地物分类实验结果表明,该方法的总体分类精度分别提高到87.53%和80.49%。提出的方法能自适应地揭示出数据间的内在关系,更有效地提取出鉴别特征,改善地物分类精度。
Aiming at the problem that in hyperspectral image land cover classification apphcation, a large number of spectral bands lead to "curse of dimensionality", a new dimensionality reduction algorithm called discriminant sparse preserving embedding (DSPE) is proposed in this paper. This method constructs an intra-class e1 graph and an inter-class l1 graph based on the natural discriminating power of sparse representation. In a low-dimensional embedding space, the algorithm preserves the sparse manifold structure of the data from the same class, separates the data from different classes and extracts the discriminating features. DSPE algorithm not only inherits the merits of sparse representation but also improves the separability among the data from different classes. The land cover classification experiments on PaviaU and Urban hyperspectral data sets were conducted, and the results show that using the DSPE algorithm the overall land cover classification accuracies are improved to 87.53% and 80.49% , respectively. The proposed method can adaptively reveal the intrinsic relationship of the data, effectively extract the discriminating features and improve the land cover classification accuracy.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2016年第1期177-183,共7页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(41371338
61101168)
重庆市基础与前沿研究计划(cstc2013jcyjA40005)
中央高校基本科研业务费(106112013CDJZR125501
106112013CDJZR120004)
重庆市研究生科研创新项目(CYB15052)项目资助
关键词
高光谱遥感
维数约简
地物分类
鉴别学习
稀疏图
hyperspectral remote sensing
dimensionality reduction
land cover classification
discriminating learning
sparse graph