特征选择是面向对象信息提取的关键步骤之一.本文针对分离阈值(separability and threshold,SEaTH)这一特征选择方法在实际应用中存在的不足,例如未考虑特征值的离散程度,仅利用J-M距离评判单一特征,特征间可能存在较强相关性,以及无法有效确定出分类顺序,提出了一种改进的SEaTH算法(optimized SEaTH,OPSEaTH).OPSEaTH 算法首先在J-M 距 离基础上构建了一类特征评价指标(E值),有效解决了特征值的离散度问题;然后,基于E值构建出特征组合评价指标(C,值),可有效评估得到每种地物的最佳特征组合并自动确定出地物的分类顺序;最后基于eCognition等分类器可完成对地物对象的最终有效分类.利用高分二号遥感影像数据对本文方法进行了测试,并将结果分别与SEaTH算法、DPC、OIF和最近邻分类器的分类结果进行了对比,结果表明:OPSEaTH算法不仅能有效降低特征维数、优化特征空间,还能够对分类顺序进行自动化合理确定,总体精度和Kappa系数及其他精度指标,均显著优于基于SEaTH算法的特征选择结果.本文方法无论从特征降维效果、分类结果精度还是计算效率方面均优于DPC、OIF和最近邻分类器结果.OPSEaTH是一种更优的特征选择方法.
An feature optimization selection method of SEaTH considering discretization degree
Feature selection is one of the key steps in object-oriented information extraction.In view of the fact that the separa-bility and threshold(SEaTH),a feature selection method,does not consider the discrete degree of eigenvalues,only uses the J-M distance to judge a single feature,there may be strong differences between features,and the inability to effectively deter-mine the limitations of the classification order in practical application.Therefore,the extraction of ground objects cannot achieve the optimal effect,the extraction rules of ground objects are also complex,and the portability of the classification mod-el is still poor.To solve these problems,an improved SEaTH algorithm(optimized SEaTH,OPSEaTH)was developed in this study.First,a feature evaluation index(E)is constructed by OPSEaTH based on J-M distance,which can effectively solve the dispersion of eigenvalues.Further,a feature combination evaluation index(Ce)is constructed based on E value,which can ef-fectively evaluate the best feature combination of each feature and automatically determine the classification order of features.Then,the effective classification of feature objects can be completed based on eCognition and other classifiers.In this study,the new algorithm is tested by using GF-2 remote sensing image data,and compared with the classification results of SEaTH algorithm,DPC(density peaks cluster),OIF(optimal index factor),and the nearest neighbor classifier,respectively.The re-sults show that:OPSEaTH algorithm can not only effectively reduce the feature dimension and optimize the feature space,but also automatically and reasonably determine the classification order.The overall accuracy,Kappa coefficient and other accuracy indexes of the OPSEaTH algorithm are significantly better than the feature selection results based on SEaTH algorithm.In ad-dition,the OPSEaTH algorithm is superior to DPC,OIF and the nearest neighbor classifier in terms of feature dimension re-duction effect,classification accuracy and computational efficiency.OPSEaTH algorithm is a better feature selection method.
万方数据
维普
