Evaluation of Random Forest and Adaboost tree-based ensemble classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery
Evaluation of Random Forest and Adaboost tree-based ensemble classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery
复制标题
DOI：
10.1016/j.rse.2008.02.011
复制
发表时间：
2008-06-16
影响因子：
13.5
通讯作者：
Paelinckx, Desire
中科院分区：
文献类型：
--
作者：
Chan, Jonathan Cheung-Wai;Paelinckx, Desire
详细的土地利用/土地覆被分类对环境评价具有重要意义。在这项研究中，我们探讨了利用航空高光谱图像进行生态群落分类的可能性。特别地，我们基于标准分类精度、训练时间和分类稳定性评估了两种基于树的集成分类算法：Adaboost和Random Forest。我们的结果表明，Adaboost和Random Forest的总体准确率几乎相同（接近70%），差异小于1%，两者都优于神经网络分类器（63.7%）。然而，随机森林的训练速度更快，也更稳定。两种集成分类器都被认为是处理高光谱数据的有效方法。在此基础上，提出了两种特征选择方法，即袋外策略和包装器方法，采用最佳优先搜索方法选择特征子集。两种方法选择的波段大部分集中在1.4 ~ 1.8 μ m之间的早期短波红外区。我们的波段子集分析还包括Thenkabail等人[Thenkabail， RS., Enclona, e.a., Ashton, m.s., and Van Der Meer， B.(2004)]建议的0.4 - 2.5 μ m之间的22个最佳波段。植被分析应用的高光谱波段性能精度评估。遥感与环境，1999,35 -376。由于目标类的相似性。本研究中考虑的所有三个波段子集都可以很好地与两种分类器一起工作，因为在大多数情况下，总体准确率仅下降了不到1%。通过将所有特征子集组合在一起创建了53个波段的子集，并与使用整个集合进行比较，Adaboost的总体精度相同，而Random Forest则提高了0.2%。使用一篮子波段选择方法的策略效果更好。一般来说，属于乔木类的生态生境比属于草类的生态生境分类好。建议对分类方案进行小幅度调整，以提高遥感方法在生态群落详细制图中的适用性。(C) 2008爱思唯尔公司版权所有。
Detailed land use/land cover classification at ecotope level is important for environmental evaluation. In this study, we investigate the possibility of using airborne hyperspectral imagery for the classification of ecotopes. In particular, we assess two tree-based ensemble classification algorithms: Adaboost and Random Forest, based on standard classification accuracy, training time and classification stability. Our results show that Adaboost and Random Forest attain almost the same overall accuracy (close to 70%) with less than 1% difference, and both outperform a neural network classifier (63.7%). Random Forest, however, is faster in training and more stable. Both ensemble classifiers are considered effective in dealing with hyperspectral data. Furthermore, two feature selection methods, the out-of-bag strategy and a wrapper approach feature subset selection using the best-first search method are applied. A majority of bands chosen by both methods concentrate between 1.4 and 1.8 mu m at the early shortwave infrared region. Our band subset analyses also include the 22 optimal bands between 0.4 and 2.5 mu m suggested in Thenkabail et al. [Thenkabail, RS., Enclona, E.A., Ashton, M.S., and Van Der Meer, B. (2004). Accuracy assessments of hyperspectral waveband performance for vegetation analysis applications. Remote Sensing of Environment, 91, 354-376.] due to similarity of the target classes. All of the three band subsets considered in this study work well with both classifiers as in most cases the overall accuracy dropped only by less than 1%. A subset of 53 bands is created by combining all feature subsets and comparing to using the entire set the overall accuracy is the same with Adaboost, and with Random Forest, a 0.2% improvement. The strategy to use a basket of band selection methods works better. Ecotopes belonging to the tree classes are in general classified better than the grass classes. Small adaptations of the classification scheme are recommended to improve the applicability of remote sensing method for detailed ecotope mapping. (C) 2008 Elsevier Inc. All rights reserved.