基于MODIS数据的决策树分类方法研究与应用

Research and Application of the Decision Tree Classification Using MODIS Data

介绍了目前国际上流行的两种决策树算法———CART算法与C4·5算法,并引入了两种机器学习领域里的分类新技术———boosting和bagging技术,为探究这些决策树分类算法与新技术在遥感影像分类方面的潜力,以中国华北地区MODIS250m分辨率影像进行了土地覆盖决策树分类试验与分析。研究结果表明决策树在满足充分训练样本的条件下,相对于传统方法如最大似然法(MLC)能明显提高分类精度,而在样本量不足下决策树分类表现差于MLC;并发现在单一决策树生成中,分类回归树CART算法表现较C4·5算法具有分类精度和树结构优势,分类精度的提高取决于树结构的合理构建与剪枝处理;另外在决策树CART中引入boosting技术,能明显提高那些较难识别类别的分类准确率18·5%到25·6%。

Decision tree classification algorithms have significant potential in remote sensing data classification. In this research, two popular decision tree algorithms——CART and C4.5 are presented, and two techniques known as boosting and bagging in machine learning area are introduced. We examined these methods to maximize classification accuracies using these decision trees and techniques to map land cover of Huabei area in China from MODIS 250m data. The result indicates that decision tree with abundance training samples has higher classification accuracy than maximum likelihood classifier(MLC)in the land cover classification test, whereas insufficient samples resulted in a lower accuracy for decision tree than MLC. The result also shows CART algorithm has more advantageous than C4.5 algorithm in classification accuracy and tree structure. And the decision tree classification accuracy depends on the optimal structure and pruning process. We also tested the behaviour of boosting and bagging techniques combined with CART and the result shows that adding boosting technique to decision tree can increase classification accuracies by 18.5%—25.6% for the poorly separable classes in MLC.