【作者】 吴凯；

【导师】 曹治国；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2007， 硕士

【摘要】 自从第一颗地球资源技术卫星发射升空以来,遥感技术有了突飞猛进的发展,景物分类技术则是其中的一个热点,在国防建设和社会发展各个方面发挥越来越重要的作用。目前制约遥感图像景物分类精度和实现图像自动化分类的主要因素之一就是没有系统的理论指导选用合适的特征。分析评价各种特征在不同谱段、不同成像时间下的重要性和类别区分能力,指导选用何种有效和性质优秀的特征成为了当务之急。本文以国内某市多谱段(可见光、中波红外、长波红外)遥感图像和作者所拍摄的三谱段图像为信息源,详细介绍了目前使用较多的特征及其提取方法,包括随气候和时间改变而变化显著的光谱特征,反映图像亮度的空间变化情况的纹理特征等。分析评价了原始光谱特征中灰度均值的聚类分布,并采用了监督分类中较为成熟的最近邻规则对各谱段组合灰度特征的分类性质进行了实验和分析。研究结果表明可见光和长波红外灰度信息组合效果最好,中波红外和长波红外的组合则相对最差。在用特征选择方法进行特征评价时,针对高维特征集往往存在大量冗余特征和不相关特征的情况,提出了一种3级串联式组合特征选择方法,先过滤消除冗余和不相关特征,最后使用次优搜索法选择所需数量的特征,实验证明该方法能有效地选择出对分类重要性高的特征并降低特征维数。约减得到特征子集后,这些特征对分类的贡献度仍有高低之分,所以本文提出了一种简单直观的特征排序算法对其进行排序和评价。为了对不同时刻下特征性能的变化给出评价,本文使用上述的特征选择和排序算法,分析连续24个时刻点的特征选择结果,然后按特征种类统计出其数目并给出时间性分布曲线。本文最后对研究工作给出了总结,提出了今后的进一步研究方向。更多还原

【Abstract】 Remote sensing technology has been developing rapidly after the first landsat sent into outer space. Scene classification,one hotspots of remote sensing , becomes more and more important in the area of national defence and social development. However, having no systemic theoretics guiding how to choose appropriate features has become one of the main limitations in accurate and automatic classification currently. It’s exigent to research choosing the effectual and eximious features with analysing the significance and classi- fication ability of feature by multi-spectral or asynchronism remote sensing image.Using the multi-spectral (visible, medium Infrared Ray,long Infrared Ray)remote sensed image of a city in China and some of image which was shot by author,as the main data source,itemize the features used frequently which contain spectral feature alters rapidly following the change of climate or time and texture spectrum expresses the dime- nsional character of image lum. Clustering distribution of average gray has been analysed and assessed. And then the precision of classification with one-spectral and multi- spectral average gray also was analysed by nearest neight rule,a classical supervised classification method.The resu- lts show that the gray information combined with visible and long Infrared Ray(IR) spectral gets the best classification,however long IR with medium IR spectral does the worst. For supervised high-dimensional feature selection, the author presents a three-stage select model, firstly reduces remove the irrelevant and redundant features in the original set, while chooses the required number features at last. The experi- mental results proved the method can drasti- cally reduce the dimension of selected feature set. There also has distinction of classific- ation in selected feature set, therefore a simple feature compositor is presented.For asse- ssing the capability of classification as the time changes, the author analysed the hourly results all-day using the above feature selected model. In addition, distributing curve has been drawing according to the feature species.In the end, this paper concludes by summarizing the research and indicating its fiuture work.更多还原