基于形态学属性剖面的高光谱图像分类特征提取

63 浏览量更新于2024-08-27 收藏 356KB PDF 举报

"基于形态属性剖面的高光谱图像分类特征提取" 本文是一篇研究论文，主要探讨了高光谱图像分类中的特征提取方法。传统的方法通常仅依赖原始光谱特征，而忽略了图像的空间特性。针对这一问题，作者提出了一个新颖的基于形态属性剖面的高光谱图像分类方法。在该研究中，首先采用了独立成分分析（Independent Component Analysis, ICA）进行维数约简。ICA是一种统计方法，用于从混合信号中分离出独立的、非高斯分布的源信号，这有助于减少数据的复杂性，同时保留关键信息。通过ICA，原始高光谱图像被转换到一个新的子空间中，这个子空间更好地揭示了数据的内在结构。接着，研究提出了一种扩展的多属性剖面（Extended Multiple Attribute Profiles, EMAP）算法来在这个子空间中提取空间特征。EMAPs是一种利用图像形态学特性的方法，它能够捕获图像中的形状、纹理和结构信息，这对于空间信息的利用至关重要。这种方法的创新之处在于它能有效地利用ICA诱导的子空间中的信息，以提高特征的表示能力。为了实现精确的分类，文章进一步提出了一个贝叶斯最大后验（Bayesian Maximum A Posteriori）框架，该框架结合了基于EMAP的特征提取和类条件概率，同时引入了马尔可夫随机场（Markov Random Field, MRF）作为先验知识。MRF是一种统计模型，常用于处理图像分割和分类问题，因为它能够考虑像素间的空间连通性和相似性，从而提高分类的稳定性和准确性。实验结果表明，所提出的基于EMAP和MRF的方法显著优于传统的和最先进的方法，它倾向于产生更平滑的分类地图，误分类的异常点显著减少。这表明，结合了空间信息的特征提取策略对于高光谱图像分类具有极大的提升作用，对于改善分类性能和减少错误分类有显著贡献。关键词：高光谱图像，分类，形态属性剖面，独立成分分析，扩展多属性剖面，贝叶斯最大后验，马尔可夫随机场

Feature Extraction Based on Morphological Attribute Profiles for

Classification of Hyperspectral Image

Zhen Ye, Yuchan Yan, Lin Bai, Meng Hui

School of Electronics and Control Engineering, Chang'an University, Xi’an, China,

yezhen525@126.com

ABSTRACT

Traditional hyperspectral image classification typically uses raw spectral signatures without considering the spatial

characteristics. In this paper, we proposed a novel method for hyperspectral image classification based on morpho-

logical attribute profiles. We employed independent component analysis for dimensionality reduction and designed an

extended multiple attribute profiles (EMAP) to extract spatial features in ICA-induced subspaces. For accurate clas-

sification, we proposed a Bayesian maximum a posteriori formulation that couples EMAPs-based feature extraction for

the class-conditional probability with an MRF-based prior. Experimental results show that the proposed method

substantially outperforms traditional and state-of-the-art methods tending to result in smoother classification maps

with fewer erroneous outliers.

Keywords: Hyperspectral image, classification, morphological attribute profiles (AP), Markov random field (MRF).

1. INTRODUCTION

Due to the high spectral resolution, hyperspectral imagery (HSI) has been widely used to provide finer classification in

various fields, such as agriculture, military and mineralogy. In traditional systems, classification of HSI has focused on

the pixelwise application of classifiers operating on spectral information exclusively. However, there has been in-

creasing interest in recent years in HSI classification that exploits spatial information. There are usually two types of

approaches for spatial information extraction. The first one considers spatial and contextual dependence within a

predefined neighborhood system [1], such as a Bayesian maximum a posteriori formulation coupled with a Markov

random field (MRF) (e.g. [2, 3]). Specifically, for an n-sample dataset

 

Xx=

, we wish to find the image of

class labels,

 

y =

, where

 

1,2, ,

yk K

. By Bayes’ rule, the posterior density of the class labels is

( ) ( | )

( | )

()

y X y

(1)

Since

is given,

()p X

is a constant and can be disregarded, while

()p y

is the prior probability of the

classes, and

()p Xy

is the class-condition probability of

. Applying a Bayesian maximum a posteriori formu-

lation to (1), we arrive at the estimated class image being

 

1, ,

arg max log ( ) log ( )



=+y X y y

(2)

In the typical application of (2) to spatial-spectral HSI classification,

()p y

is considered as a spatial class prior

and is usually provided by an MRF which promotes the assigning of neighboring pixels to the same class. Naturally,

then, the class-conditional probability,

()p Xy

, is designed to exploit spectral information, typically in the form of

a pixelwise probability estimate driven by the spectral signature of the pixel being classified. To this end, support

vector machine (SVM) [4] is commonly used to provide

()p Xy

for (2). In each of these formulations, however,

this class-conditional probability is devoid of spatial information and is, instead, derived exclusively from spectral

signatures. While the partitioning of (2) into a spectral-only term and a spatial-only term makes some sense concep-

tually, it is not necessarily optimal. Indeed, the MRF may override otherwise correct results from the spectrally-driven

class-conditional probability if the latter provides only relatively weak support (i.e.,

()p Xy

is maximum for the

correct class labeling but not by a wide margin over other labelings).

To address the shortcomings of the first type of approaches, an adaptive and given neighborhood system can be

considered for classification, such as morphological profiles (MP)-based approaches [1, 5-8]. M. Pesaresi and J. A.

Benediktsson used morphological transformations to build MP [5]. Specifically, bright and dark spatial structures of

image were isolated by morphological opening and closing, respectively. Recent years, the strategy has been extended

to HSI analysis by principal component analysis (PCA) to reduce the dimensionality of HSI, which is named extended

morphological profiles (EMP) [6, 7]. Since the structuring element (SE) of MP cannot characterize in formation

下载后可阅读完整内容，剩余5页未读，立即下载

weixin_38685694

粉丝: 4
资源: 899

基于形态学属性剖面的高光谱图像分类特征提取

Feature Extraction and Image Processing for Computer Vision(Third Edition 2012)

Hyperspectral image classification based on spectral-spatial feature extraction

An algorithm for movement related potentials feature extraction based on transfer learning

Endmember extraction for hyperspectral image based on integration of spatial-spectral information(结合空间信息的高光谱影像端元光谱提取方法)

Comparison of sEMG-Based Feature Extraction and Hand Motion Classification Methods

Sparse representation for tumor classification based on feature extraction using latent low-rank representation.

Dilemma and Solution of Traditional Feature Extraction Methods Based on Inertial Sensors

Magnetic dispersive solid-phase extraction based on modified magnetic nanoparticles for the detection of cocaine and cocaine metabolites in human urine by HPLC-MS

武汉大学-转子数据集Data for Feature Extraction of Rotor Fault Based on EE

Best basis-based wavelet packet entropy feature extraction and hierarchical EEG classification for epileptic detection

最新资源