云的光谱和纹理特征统计分析

利用静止卫星图像资料建立了夏季白天中低纬地区的11 种云/ 表面类型的样本集, 从中随机 挑选656 个样本, 提取116 个光谱和纹理特征参数并进行统计分析, 通过特征选择组成特征向量, 带入逐个修改聚类和模糊聚类的分类器进行敏感性试验。结果发现, 在反映云特征方面, 光谱特征 是云分类最基本的特征, 比纹理特征明显, 是云分类识别的主要依据; 除去水汽通道的标准差以外 其它光谱特征都比较明显, 红外和水汽通道的特征明显好于可见光通道, 尤其是对中低云和卷云的 描述。纹理特征在反映云特征方面也有一定的代表性, 特别是一阶概率特征中四通道的惯量及水汽 通道的逆差距; 纹理特征引入后分类准确率显著提高, 但在引入一阶概率特征基础上引入灰度级差 矢量特征效果改善并不明显。

Staticical Analysis on Spectral and Textural Features of Clouds

Data sets of 11 cloud/surface classes are collected from daytime geostationary satellite imagery data over middle latitude and low latitude regions in summer.656 samples are randomly selected,and then 116 spectral and textural features derived from these samples are statistically investigated,curves of all features are drawn and analyzed.After feature selection,sensitivity tests of two classifiers-stepwise cluster and fuzzy cluster are explored by using of more significant feature arrays.For spectral features,results of feature curves analysis indicate that spectral characteristics are more significant than those of textural characteristics,the spectral features of IR or WV except Std show more distinguishingly than those of VIS,especially for low- or mid-level and cirrus;both results of sensitivity tests from two classifiers suggest that spectral features provide the major information of cloud classification,which are main indexes of cloud pattern discrimination.Results of textural feature curves indicate first order probability vectors(FOPV) demonstrate more obviously than those of gray level difference vectors(GLDV),especially CON of four channels and HOM of WV;results of tests show that textural features play an important role in cloud type identification,both classifiers give higher accuracies after adding textural features,combing FODV and spectral features significantly improve classification results achieved using only spectral features,but the accuracy of adding GLDV based on FODV and spectral features has little distinction with those of adding FOPV based on spectral features,which indicates that textural characteristics of FODV have more encouraging ability to distinguish between cloud/surfaces than those of GLDV,consisting with results of feature curves analysis.