遥感决策树模型毕业论文

Multi-spectral remote sensing image classification decision tree automatically generated method Abstract: remote sensing image classification study of remote sensing technology is the key issue is the use of remote sensing images on other aspects of the premise. Current remote sensing images have been classified by the previous visual interpretation, the computer has become automatic classification. How to design a convenient and easy classification accuracy of the computer automatically higher classification of remote sensing image classification is a hot area of research. Pattern Recognition in the decision tree is classified as a way to the peak or distribution of a wide range of issues, this approach is particularly convenient. Use of decision tree can be a complex multi-category classification problem into a number of simple classification issues to resolve. It is not attempting to use an algorithm, a decision-making rules to make a number of separate categories, but in the form of adopting a classification, the classification issues have been settled gradually. These characteristics of the decision tree is very suited to the needs of remote sensing image classification, a common method of remote sensing classification. In this paper, automatic classification of remote sensing imaging technology and the theory of decision tree to do about the system, mainly by a multi-spectral remote sensing image classification decision tree automatically generated methods, the methods for the final evaluation, decision tree and the use of remote sensing Image classification techniques of the future.

Remote sensing image classification study of remote sensing technology is the key issue is the use of remote sensing images on other aspects of the premise. At present the classification of remote sensing images from the past has been the visual interpretation, the computer has become automatic classification. How to design a convenient and easy classification accuracy of the computer automatically higher classification of remote sensing image classification is a hot area of Recognition in the decision tree is classified as a way to the peak or distribution of a wide range of issues, this approach is particularly convenient. Use of decision tree can be a complex multi-category classification problem into a number of simple classification issues to resolve. It is not attempting to use an algorithm，A decision-making rules to make a number of separate categories, but in the form of adopting a classification, the classification issues have been settled gradually. These characteristics of the decision tree is very suitable for the needs of remote sensing image classification, a common method of remote sensing classification. In this paper, automatic classification of remote sensing imaging technology and the theory of decision tree to do about the system, mainly by a multi-spectral remote sensing image classification decision tree automatically generated methods, the methods for the final evaluation, decision tree and the use of remote sensing The image classification techniques the future.

[5] 贾永红，李芳芳. 一种新的湿地信息遥感提取方法研究[J]. 华中师范大学学报（自然科学版），2007，41（04）. [6] 李芳芳，贾永红. 一种基于 TM 影像的湿地信息提取方法及变化检测[J]. 测绘科学，2008，33（2）. [10] 赵英时. 遥感应用分析原理[M]. 北京：科学出版社，2003. 这几篇文献里有提到~~~ 在遥感图像分类中，往往需要深入研究地物的总体规律及内在联系，理顺其主次因果关系，建立一种树状结构的框架。即建立所谓的分类树来说明地物关系，并根据分类树所描述的判断准则，对图像中的各像元进行逐级分层的识别和归类，通过若干次中间判别，最终把研究目标一一区分，这就是决策树分类法，也叫分层分类法[10]。 决策树分类的基本思想是逐步从原始影像中分离并掩膜每一种目标作为一个图层或树枝，目标间关系被大大简化。因在分类树的各个中间节点上，只存在较少的类别。面对较少的对象就能选择更有效的判别函数或有针对性的分类方法，避免此目标对其他目标提取时造成干扰和影响，同时利用辅助数据进行复合处理，最终将所有图层符合以实现图像的自动分类。 决策树分类法根据不同目的要求进行层层深化，相互关系明确，局部细节描述得更为清楚，针对性更强，每个节点上只需要考虑与区分目标有关的最佳变量，因而分类精度提高的同时，也避免了数据的冗余。

第一篇我们主要关注了根结点及内部结点的选择 第二篇主要关注如何处理“过拟合”现象 参考

个性化 与 泛化 是一个相互矛盾概念，就像个体化诊疗与指南的矛盾一样。 决策树对训练数据可以得到很低的错误率，但是运用到测试数据上却得到非常高的错误率，这就是“过拟合现象”。 具体解释如下：对于决策树，我们希望每个叶子节点分的都是正确的答案，所以在不加限制的情况下，决策树倾向于把每个叶子节点单纯化，那如何最单纯呢？极端情况下，就是每个叶子节点只有一个样本，那这样，这个模型在建模集的准确率就非常高了。但是，这又带来了一个问题——过拟合，这会导致该模型在建模集效果显著，但是验证集表现不佳。 这可能有以下几个原因： 1、训练集里面有噪音数据，干扰了正常数据的分支 2、训练集不具有特征性 3、特征太多

使用信息增益来种树时，为了得到最优的决策树，算法会不惜带价倾向于将熵值降为最小（可能的话甚至为0），这颗树会显得非常的冗杂。

通过限制复杂度参数（complexity parameter），抓主要矛盾，来防止模型的过拟合。具体的计算过程可以参考<医学僧的科研日记>，这里我直接引用

剪枝（pruning）则是决策树算法对付过拟合的主要手段，剪枝的策略有两种如下：

定义：预剪枝就是在构造决策树的过程中，先对每个结点在划分前进行估计，如果当前结点的划分不能带来决策树模型泛化性能的提升，则不对当前结点进行划分并且将当前结点标记为叶结点。

相比于预剪枝，后剪枝往往应用更加广泛，