欢迎访问林业科学,今天是

林业科学 ›› 2017, Vol. 53 ›› Issue (2): 26-34.doi: 10.11707/j.1001-7488.20170204

• 论文与研究报告 • 上一篇    下一篇

基于元胞自动机和BP神经网络算法的Landsat-TM遥感影像森林类型分类比较

田静, 邢艳秋, 姚松涛, 曾旭婧, 焦义涛   

  1. 东北林业大学 森林作业环境研究中心 哈尔滨 150040
  • 收稿日期:2016-04-15 修回日期:2016-11-06 出版日期:2017-02-25 发布日期:2017-03-23
  • 通讯作者: 邢艳秋
  • 基金资助:
    林业公益性行业科研专项(201504319)。

Comparison of Landsat-TM Image Forest Type Classification Based on Cellular Automata and BP Neural Network Algorithm

Tian Jing, Xing Yanqiu, Yao Songtao, Zeng Xujing, Jiao Yitao   

  1. Research Center of Forest Operations and Environment, Northeast Forestry University Harbin 150040
  • Received:2016-04-15 Revised:2016-11-06 Online:2017-02-25 Published:2017-03-23

摘要: [目的] 针对森林资源遥感监测效果往往受森林类型识别分类方法的影响,提出一种基于元胞自动机的遥感影像森林类型分类方法,以提高Landsat-TM遥感影像的分类精度,为森林资源遥感监测提供技术支持。[方法] 以小兴安岭带岭林业经营管理局为研究区,基于2010年Landsat5-TM影像数据和2012年森林资源二类调查数据,采用窗口法获取TM第5波段各待分类别的像元均值作为聚类中心,以元胞自动机的Moore模型为框架,以元胞为基本单位,以像元均值为对象,利用最小距离法求取进化规则(判断准则是中心元胞周围的8个元胞距每类聚类中心的距离最近且像元数量最多,则中心元胞属于该类别),充分考虑影像及地物之间的空间特征,采用元胞自动机分类方法进行森林类型的识别分类。同时,以相同的样本数,采用3层BP神经网络模型对TM遥感影像进行分类试验,并比较2种方法的分类效果。[结果] 基于元胞自动机的分类方法总体分类精度为88.712 1%,Kappa系数为0.829 1,针叶林、阔叶林和针阔混交林的用户精度分别为73.60%,92.94%和94.13%,达到了区分针叶林、阔叶林和针阔混交林的分类目的。BP神经网络算法的总体分类精度为86.671 3%,Kappa系数为0.798 4,针叶林、阔叶林和针阔混交林的用户精度分别为69.22%,93.37%和90.76%。2种分类方法均可有效识别森林类型信息。[结论] 元胞自动机模型应用于遥感影像森林类型识别分类可弥补因TM影像空间分辨率较低造成的遥感影像分类精度过低的问题,提高分类精度。在森林分布破碎、种类类型多样且结构复杂的带岭林区,该研究结果有助于森林资源监测与管理,可为大区域尺度的森林动态信息监测提供更好的数据及技术支持。

关键词: 元胞自动机, BP神经网络, 森林类型分类, 像元值, Landsat 5-TM影像

Abstract: [Objective] The results of forest resources monitoring with remote sensing are often affected by the method of forest type classification, so this study proposed a forest type classification method based on cellular automata (CA), and the proposed method was tested with a Landsat-TM imagery, aiming at improve the classification accuracy, obtained results can be provide a technique support for remote sensing monitoring of forest resource.[Method] The study area was located in Dailing Forestry Management Bureau of Lesser Khingan Range, based on a good quality Landsat5-TM imagery covering the study area acquired in 2010 and forest inventory data of the study area in 2012, by using window method to obtain the 5th band mean pixel value of all classes as center cluster, and the Moore model as the framework, the cell as the basic unit, the pixel mean value as the object, and the minimum distance method to obtain evolutionary rules, the judging criteria is that the distance between 8 cell which around the center cell and the each center cluster is the nearest as well as the number of pixels is largest which the center cell belongs to the this category. By compared to BP neural network classificationto illustrate the classification accuracy of the cellular automata.[Result] The forest type classification based on cellular automata resulted in an overall accuracy and Kappa coefficient of respectively 88.712 1% and 0.829 1, especially user accuracy of needle forest, broad-leaved forest and mixed forest respectively 73.60%, 92.94% and 94.13%, this results show that forest type identification is obtained the ideal result. BP neural network classification method achieved the overall accuracy of 86.671 3%, Kappa coefficient of 0.798 4.So the two classification methods can effectively identify the forest type information.[Conclusion] The cellular automata model used in remote sensing image forest type classification can make up the problem of the low classification accuracy caused by the low spatial resolution of TM imagery, it alsocan effectively improve the classification accuracy. And the cellular automata classification method can contribute to monitor the forest resource change and also provide an effective way to improve the efficiency of forest resources monitoring in large-scale areas.

Key words: cellular automata, BP neural network, forest type classification, pixel value, landsat5-TM imagery

中图分类号: