论文部分内容阅读
以2005年深圳市福田区Quick Bird影像为主要数据源,根据面向对象多尺度分割结果,构建和分析了不同地类对象的光谱、形状和纹理信息特征,在此基础上利用逐步判别分析法(stepwise discriminant analysis,SDA)结合分类回归树(classification and regression trees,CART)构建多尺度、多变量分类模型。结果表明:1采用SDA在一定程度上能够客观、准确地进行特征子集预筛选,笔者从32个特征中筛选出27个特征用于构建CART模型,并按其区分地类能力进行了重要性排序,其中光谱特征与纹理特征排序比较靠前,形状特征中仅LengthWidth排在第5位,剩余特征比较靠后。2利用分类回归树模型可进一步优化特征选取,并智能化计算出分离阈值,基本实现面向对象的自动化分类。其中Mean_b3、Length Width、Ratio_b3、GLDVE和NDVI是重要分类节点。3逐步判别分析结合分类回归树构建的分类模型,可以在提高或者不显著降低影像分类精度的条件下实现特征降维。当取相关指数R2>0.2时,构建的分类模型效果最优,CART16模型训练和验证精度分别为94.44%和83.37%,其特征子集规模最小,与原始特征数量相比减少了一半。
Taking the Quick Bird image in Futian District of Shenzhen City as the main data source in 2005, the spectral, shape and texture features of different ground objects were constructed and analyzed according to the object-oriented multi-scale segmentation results. Based on this, stepwise discriminant analysis (SDA) to construct a multi-scale and multivariate classification model with classification and regression trees (CART). The results show that: 1 Using SDA can objectively and accurately pre-screening feature subset, the author selected 27 features from 32 features for building CART model, and according to the importance of their ability to distinguish between classes Sorting, spectral features and texture features ranked more front, the shape features only LengthWidth ranked No. 5, the remaining features are relatively backward. 2 Classification regression tree model can be further optimized feature selection, and intelligent calculation of the separation threshold, the basic realization of object-oriented automated classification. Mean_b3, Length Width, Ratio_b3, GLDVE and NDVI are important classification nodes. Step-by-step Discriminant Analysis Classification model combined with classification and regression tree can achieve feature dimensionality reduction with or without significantly reducing the accuracy of image classification. When the correlation index R2> 0.2, the constructed classification model has the best effect. The precision of training and verification of CART16 model is 94.44% and 83.37% respectively, and its feature subset is the smallest, which is reduced by half compared with the original feature.