基于K近邻非线性分类器的高光谱遥感数据分类研究

A Nonlinear K-Nearest Neighbor Classifier for Hyperspectral Remote Sensing Imagery Classification

K近邻等传统分类算法在高光谱遥感影像数据上进行分类时,由于其高维度、非线性特点,分类效果会受到严重影响。本文利用核函数方法,融合K近邻分类算法与Isomap非线性降维算法,提出了一种新的K近邻非线性分类器。该分类器无需通过降维预处理,并具备处理非线性数据的能力。在实验中,通过交叉验证与参数验证证明该方法在高光谱遥感影像上的分类效果明显优于原始K近邻分类算法以及结合主成分分析法的K近邻分类法。

High dimensionality and nonlinearity are two main factors of Hyperspectral remote sensing data that will decrease the classification accuracy for most existing classification algorithms ,such as K-nearest neighbor （KNN）. This paper proposed a new nonlinear KNN classifier, which fuses the original KNN algorithm and Isomap algorithm by Kernel trick. This classifier does not need explicitly dimensionality reduction but still has the ability to analyze the nonlinearity by taking advantage of the Isomap algorithm. By cross-validation and parameter analysis in the experiments with hyperspectral test data,this new method has been proven to out-perform the original KNN and KNN with PCA algorithm in Classification Accuracy.