融合空间-光谱信息的无监督高光谱图像分类算法

151 浏览量更新于2024-08-27 收藏 449KB PDF 举报

"无监督超光谱图像分类算法通过融合空间-光谱信息" 本文提出了一种集成的空间-光谱信息算法，用于超光谱图像分类。该算法利用了空间像素关联（SPA）和光谱信息散度（SID），旨在通过减少区域数量来提升分类精度。此外，还开发了一种类边界校正方法，以最小化每个类别边缘的误分类像素，解决类别合并的问题。通过超光谱数据的实验展示了该方法在分类准确性上的有效性和优势，与传统方法相比有显著提升。 1. 引言超光谱成像技术的发展使得获取高光谱分辨率的图像成为可能，这些图像包含丰富的光谱信息，可用于地物识别、环境监测等多个领域。然而，由于大量的光谱维数和复杂的地物混合，超光谱图像的分类仍然是一个挑战。传统的分类方法如主成分分析（PCA）和线性判别分析（LDA）往往无法充分利用这些信息。 2. 方法论本文的方法首先通过SPA利用像素间的光谱相似性，通过SID检测和量化光谱差异，揭示像素之间的空间关系。接着，采用光谱聚类方法将相似的像素分组，降低数据的复杂性。这种方法旨在保持光谱信息的同时，减少因过度细分导致的类别混淆。 3. 类边界校正为了解决分类边缘的误分类问题，论文提出了一种类边界校正策略。该策略通过分析相邻像素的分类结果，调整分类边界，以减少误分类像素的数量，并防止类别之间的不必要合并。 4. 实验与结果实验部分使用了多种超光谱数据集进行验证，结果表明，与传统的基于纯光谱或纯空间特征的分类方法相比，该方法在提高分类准确性和减少错误率方面表现出优越性。同时，类边界校正有效地改善了分类边缘的精确性。 5. 结论结合空间-光谱信息的无监督分类算法成功地提升了超光谱图像处理的性能，特别是在减少误分类和优化类边界方面。这一成果对超光谱图像分析和应用具有重要的理论和实践意义，为未来的研究提供了新的思路和工具。关键词：类边界；分类精度；框架；超光谱图像；超光谱数据；超光谱图像分类；信息算法；光谱聚类；光谱信息散度

Unsupervised Hyperspectral Image Classification Algorithm

By Integrating Spatial-Spectral Information

Belkacem Baassou, Mingyi He,Shaohui Mei, Yifan Zhang

Shaanxi Provincial Key Laboratory of Information Acquisition and Processing (IAP)

School of Electronics and Information, Northwestern Polytechnical University, Xi'an, 710129, China

myhe@nwpu.edu.cn,gaga05@hotmail.fr

Abstract

An integrated spatial-spectral information algorithm

for hyper spectral image classification is proposed,

which uses spatial pixel association (SPA)by exploiting

spectral information divergence (SID),and spectral

clustering to reduce regions number and improve

classification accuracy .Moreover, a class boundary

correction method is also developed to minimize the

misclassified pixels at the edge of each class and to

solve the problem of merged classes. Experiments with

hyper spectral data demonstrate the effectiveness and

advantages of the proposed frame work over some

traditional methods in term of classification accuracy.

1. Introduction

The development of image sensor technology has

made it possible to capture image data in hundreds of

contiguous and very narrow spectral channels covering

a broad spectrum of wavelength, providing a large

amount of detailed information about the spatial and

spectral characteristics of the materials that are present

in the scene. Recently, interest in hyper spectral

imagery has increased; it is viewed as an emerging

valuable remote sensing technology in both military and

civilian applications

[1]

such as wide-area

reconnaissance, vegetation classification, scientific

sensing, environ-mentalmonitoring, lunar soil

identification, targets search, rescue operations, and

also in biomedical applications such as skin cancer

diagnosis.

Hyperspectral image classification is the process of

assigning all pixels in a digital image to particular

classes according to their characteristics

[2]

. As a result

we obtain a thematic map in which each pixel belongs

to a particular class. There are two main classification

schemes, the Supervised and

theUnsupervisedClassification.Supervised classification

can be defined as the process of using samples of

known identity to classify pixels of unknown identity,

while unsupervised classification can be defined as the

identification of natural groups or structures within the

data. It clusters pixels in a data set based only on their

statistics without using previous knowledge about the

spectral classes presented in the image

[3]

. The K-means

algorithm is one of the most commonly used

unsupervised classification schemes

[4]

Recent studies show that besides the spectral

information, the spatial information may also provide a

useful knowledge and can greatly improve the

performance of the classification of hyperspectral

images

[5]

.Unfortunately, the spatial information has not

got much attention in the previous research.

In this paper, both spatial and spectral information

are exploited to achieve a better clustering performance.

A spatial pixel association(SPA) process

[6]

isused to

fully exploit the spatial similarity between pixels and

their neighbors by using the Spectral Information

Divergence criterion in order to characterize the

spectral properties provided in a single pixel vectorand

to measure spectral similarity between adjacent

pixels.Moreover, a neighborhood correction algorithm

is carried out as a post processing step to further

improve the performance of classification and to orient

misclassified pixels to take their suitable classes. After

correction a spectral clustering process has been applied

to exploit the information provided by the spectral

dimension in the aim of increasing the classification

performance. Finally, experiments on AVIRIS

hyerspectral data are conducted to verify the

performance of the proposed approachand for

comparison with some traditional methods.

2. Methodology

In a hyperspectral image, pixels from a

homogeneous ground object often present in adjacent

areas, which enable us to use the information provided

by the spatial adjacent pixels location for better

下载后可阅读完整内容，剩余5页未读，立即下载

weixin_38647517

粉丝: 2
资源: 964

融合空间-光谱信息的无监督高光谱图像分类算法

Unsupervised-Attention-guided-Image-to-Image-Translation-master.zip

Unsupervised Learning of Human Action Categories Using Spatial-TemporalWords

C:\Users\86157\Desktop\Unsupervised-Classification-master\Unsupervised-Classification-master\requirements.txt，路径在这儿

EnvironmentFileNotFound: 'C:\Users\86157\Desktop\Unsupervised-Classification-master\Unsupervised-Classification-master\environment.yml' file not found

C:\Users\86157\Desktop\Unsupervised-Classification-master\Unsupervised-Classification-master\data\imagenet.py，这是什么意思啊

unsupervised-image-segmentation-by-WNet-with-NormalizedCut:WNet的Tensorflow实现，用于PASCAL VOC 2012数据集上的无监督图像分割

均值漂移聚类matlab代码-unsupervised-image-segmentation-using-CNN:使用从卷积自动编码器中学到的

DualGAN: Unsupervised Dual Learning for Image-to-Image Translation

最新资源