The analysis about factors influencing the supervised classification accuracy for vegetation hyperspectral remote sensing imagery

摘要

Classification of hyperspectral imagery has drawn much attention in recent years since the development of hyperspectral sensor. The hyperspectral sensor may offer hundreds of contiguous and very narrow spectral channels which may detect more detailed classes and improve the classification accuracy. However, with the increasing dimensionalities of classification data space, the parameters in the supervised classification models are also increasing quickly, which will need much more numbers of training samples to ensure the parameters estimation accuracy. Most recent researches focus on developing novel classification algorithms to improve the hyperspectral imagery processing performance. There are still lack of systematic researches about the factors (such as numbers and distributions of training samples, classification algorithms, dimensionalities of feature space) influencing classification accuracy for vegetation hyperspectral imagery so far. To analyze the factors, firstly, two methods for dimensionality reduction are proposed based on correlation coefficient and spectrum curve inflection points; On this basis, we use region of interest (ROI) constructed by different training numbers of pixels and different distributions to classify OMIS aerial hyperspectral image of ZaoYuan Town of Yanan City by several often used supervised classification algorithms; then analyze the relationships among classifiers and training samples and feature space dimensionality. This research shows that the classifiers based on Mahal distance and Maximum Likelihood (ML) that use secondary moment are superior to Euclid distance (ED), parallelepiped (PP) and spectral angle mapper (SAM); when the numbers of training samples are large enough, dimensionality reduction makes little influence on classification accuracy; when the numbers of training samples are limited, the influence is distinct on Mahal distance and ML but not distinct on Euclidean distance, PP and SAM.