基于核最小噪声分离变换的高光谱遥感影像特征提取研究

高光谱遥感影像具有高维非线性的特点,线性特征提取方法容易造成信息丢失和失真。在最小噪声分离变换(MNF)线性特征提取算法的基础上,引入核方法,提出核最小噪声分离变换(KMNF)高光谱遥感影像非线性特征提取方法。KMNF通过核函数,将样本映射到高维特征空间,在特征空间中运算线性MNF,实现原始空间中的非线性KMNF算法。进行基于KMNF的高光谱影像特征提取实验,分析样本个数对KMNF特征提取的效果,发现样本数量对KMNF特征提取的结果影响很小,较少的样本数即可达到较多样本时特征提取的效果。对比KMNF与MNF特征提取的效果,分析它们降维的效率与保留的信息量,发现KMNF总体降维效率与MNF相当,且体现出高光谱图像的非线性特征；在KMNF和MNF特征提取的基础上,利用SVM进行高光谱图像分类,KMNF+SVM的分类精度优于MNF+SVM。

Hyperspectral Remote Sensing Image Feature Extraction based on Kernel Minimum Noise Fraction Transformation

Hyperspectral remote sensing images are high-dimensional and non-linear,so information loss and distortion are easily caused by linear feature extraction.Based on the minimum noise fraction transformation (MNF) which is a hyperspectral remote sensing image linear feature extraction algorithm,kernel minimum noise fraction transformation(KMNF) is proposed by introducing the kernel method,so KMNF is a non-linear feature extraction method.Samples are mapped into high dimensional feature space through a kernel function,MNF is conducted in feature space,thus non-linear KMNF algorithm in the original space is realized.And hyperspectral remote sensing image feature extraction based on KMNF was carried out,the effects of sample amount to KMNF were analyzed,it was found that the sample number influences KMNF slightly,a small number of samples can get almost the same result as a large number of samples.KMNF and MNF feature extraction results were compared,and their dimension reduction efficiency and remaining information were analized,and it was found that KMNF can get almost the same dimensional reduction efficiency as MNF,and KMNF can extract non-linear information from hyperspectral remote sensing images.Using SVM for hyperspectral image classification based on KMNF and MNF,it is showed that classification accuracy of KMNF and SVM is higher than MNF and SVM.