南京大学学报(自然科学版)
南京大學學報(自然科學版)
남경대학학보(자연과학판)
JOURNAL OF NANJING UNIVERSITY(NATURAL SCIENCES)
2009年
5期
665-670
,共6页
遥感图像%分类%高斯过程%空间相关性%核函数
遙感圖像%分類%高斯過程%空間相關性%覈函數
요감도상%분류%고사과정%공간상관성%핵함수
remote sensing imagery%classification%Gaussian processes%spatial relations%kernel function
高光谱遥感图像分类是遥感图像处理的一项重要内容.高光谱遥感图像具有非线性属性.图像中不同方位光谱特征的变化将使得仅从标记训练样本得到的分类器分类精度不会太高.为了提高分类的精度,一方面应对光谱信息的合理利用;另一方面,对空间信息的利用也非常重要.高斯过程(Gaussion process,GP)是一种贝叶斯统计学习方法,能够建立概率模型,并且使得分类结果更易于解释.传统GP分类方法中核函数的构造仅利用光谱信息.本文提出了一种加入空间关系的新分类方法.利用遥感图像空间相关性,在GP分类方法中通过构造新的核函数(spatial Gauss kernel,SGK)来实现空间约束,部分消除了同物异谱和同谱异物造成的分类错误.实验结果表明,该方法对于提高高光谱遥感图像的分类精度具有积极意义.
高光譜遙感圖像分類是遙感圖像處理的一項重要內容.高光譜遙感圖像具有非線性屬性.圖像中不同方位光譜特徵的變化將使得僅從標記訓練樣本得到的分類器分類精度不會太高.為瞭提高分類的精度,一方麵應對光譜信息的閤理利用;另一方麵,對空間信息的利用也非常重要.高斯過程(Gaussion process,GP)是一種貝葉斯統計學習方法,能夠建立概率模型,併且使得分類結果更易于解釋.傳統GP分類方法中覈函數的構造僅利用光譜信息.本文提齣瞭一種加入空間關繫的新分類方法.利用遙感圖像空間相關性,在GP分類方法中通過構造新的覈函數(spatial Gauss kernel,SGK)來實現空間約束,部分消除瞭同物異譜和同譜異物造成的分類錯誤.實驗結果錶明,該方法對于提高高光譜遙感圖像的分類精度具有積極意義.
고광보요감도상분류시요감도상처리적일항중요내용.고광보요감도상구유비선성속성.도상중불동방위광보특정적변화장사득부종표기훈련양본득도적분류기분류정도불회태고.위료제고분류적정도,일방면응대광보신식적합리이용;령일방면,대공간신식적이용야비상중요.고사과정(Gaussion process,GP)시일충패협사통계학습방법,능구건립개솔모형,병차사득분류결과경역우해석.전통GP분류방법중핵함수적구조부이용광보신식.본문제출료일충가입공간관계적신분류방법.이용요감도상공간상관성,재GP분류방법중통과구조신적핵함수(spatial Gauss kernel,SGK)래실현공간약속,부분소제료동물이보화동보이물조성적분류착오.실험결과표명,해방법대우제고고광보요감도상적분류정도구유적겁의의.
Classification of hyperspectral remote sensing imagery is an important issue of remote sensing images processing. Hyperspectral remote sensing images have nonlinear property. A classifier derived from labeled samples may not perform well for a specific sub-region if the spectral signatures of classes vary across the image. In order to improve accuracy of classification, not only spectral information of images should be utilized, but spatial information is necessary for classification as well. Gaussian process (GP) is a Bayesian statistics learning method. GP bears a full Bayesian formulation, thus enable explicitly probabilistic modeling and makes results easily interpretable. Usually, only spectral information is used for kernel construction in the traditional GP. In this paper, we explore the effectiveness of the Bayesian Gaussian process approach for classifying Hyperspectral remote sensing images. Furthermore, a new GP based classification method is proposed in which spatial information is considered. The method is a Bayesian kernel-based nonlinear method, so it is suitable for nonlinear data classification and it canreduce the uncertainty by computation of posterior label probabilities. By constructing a new spatial kernel function (SGK) in GP, spatial relations in remote sensing imagery is included, so that classification error partially caused by "same material different spectral" and "same spectral different material" can be eliminated. Experiment results show that this method is effective in improving accuracy of hyperspectral remote sensing imagery classification.