多种分类器在华北地区土地覆盖遥感分类中的性能评价

Evaluation of Various Classifiers on Regional Land Cover Classification in Huabei Area

应用MODIS 250m分辨率遥感影像对中国华北地区分别采用最大似然法、Parzen窗、CART决策树、BP神经网络、Fuzzy ARTMAP神经网络等5种分类方法进行区域尺度上土地覆盖制图的比较试验.结果表明:(1)Parzen窗法分类性能最优,CART和BP其次,Fuzzy ARTMAP表现较差.(2)CART决策树具有较好鲁棒性,但缺点是样本代价较大;BP神经网络分类器能达到较高精度,但缺点是需较高质量的样本、网络结构参数难以确定,造成其稳健性较差;FuzzyARTMAP则未能表现出理想结果.(3)训练样本数量差异造成:最大似然法的分类精度差异值低于5%;Parzen窗法和Fuzzy ARTMAP差异为5%～10%;CART和BP差异在10%以上.

Five classification methods which are MLC （Maximum Likelihood Classifier）, Parzen window, CART decision tree, BP neural network and Fuzzy ARTMAP neural network are selected to map the land cover of Huabei Area in China using MODIS 250m data. The results show that Parzen window performs best in the five classifiers. And CART and BP have satisfactory accuracy whereas Fuzzy ARTMAP has unexpected bad accuracy in comparison with MLC. CART decision tree has better flexibility and robustness. However, it pursues high accuracy at the cost of the sample size. BP neural network has high accuracy but requires high-quality samples and it is hard to define its net structure parameters. The results also show the classification accuracy difference caused by the size of training samples on MLC, Parzen window and Fuzzy ARTMAP, CART and BP are below 5 %, 5 %-10 % and above 10%, respectively.