应用相关近邻局部线性嵌入算法的高光谱遥感影像分类

传统的局部线性嵌入（LLE）算法需用欧氏距离度量近邻，但欧氏距离只表示两点间的直线距离，在高维空间中不一定能反映数据间的真实空间分布，导致近邻选取不稳定。针对此问题，本文提出了相关近邻（CN）LLE（CN-LLE）和相关最近邻分类（CNN）算法。提出的算法首先利用相关系数度量数据间的近邻，实现更准确的局部重构，提取鉴别特征；然后用CNN对低维嵌入特征进行分类。在KSC和Indian Pine高光谱遥感数据集上的地物分类实验结果表明：本文提出的CN-LLE+CNN算法比LLE、LLE+CNN和CN-LLE等算法的总分类精度提升了2.11%~11.55%，Kappa系数提升了0.026~0.143。由于该算法增加了近邻为同类的概率，便于更有效地提取同类数据的鉴别特征，且有更好的稳定性，故能更有效地实现高光谱遥感数据的地物分类。

Classification of Hyperspectral remote sensing images using correlation neighbor LLE

Traditional Locally Linear Embedding (LLE) manifold learning algorithm uses Euclidean distance to measure neighbor points. However, Euclidean distance represents the straight line distance between two points and does not necessarily reflect the actual data distribution in a high dimension space, which leads to the instability of neighbor point selecttion. In order to solve this problem, an algorithm based on Correlation Neighbor LLE (CN-LLE) and Correlation Nearest Neighbor (CNN) classification is proposed. This algorithm uses the correlation coefficient of data to measure the neighbor points and to achieve more effective local reconstruction to extract the distinguishing character. Then, it uses the CNN to classify the reduced dimension data. The experiment results from KSC and Indian Pine hyperspectral remote sensing data sets show that the total accuracy of the proposed CN-LLE+CNN algorithm is improved by 2.11%-11.55% and the Kappa coefficient is improved 0.026-0.143 as compared with those of LLE, LLE+CNN and CN-LLE. The CN-LLE+CNN algorithm increases the probability of the same class neighbor, can extract the distinguishing characters of the same data effectively and has a better stability. This algorithm can effectively classify hyperspectral remote sensing data of ground objects.