基于随机森林算法的模糊坡位分类  

Fuzzy Slope Position Segmentation Based on Random Forest

作　　者：王磊[1,2] 马风华 武伟 刘洪斌[1,2] WANG Lei;MA Feng-hua;WU Wei;LIU Hong-bin(School of Resources and Environment,Southwest University,Chongqing 400716,China;School of Computer and Information Science,Southwest University,Chongqing 400715,China;Chongqing Key Laboratory of Digital Agriculture,Chongqing 400716,China)

机构地区：[1]西南大学资源环境学院,重庆400716 [2]重庆市数字农业重点实验室,重庆400716 [3]西南大学计算机与信息科学学院,重庆400715

出　　处：《西南师范大学学报（自然科学版）》2018年第1期10-17,共8页Journal of Southwest China Normal University(Natural Science Edition)

基　　金：国家科技支撑计划课题(2008BADA4B10);中央高校基本科研业务费专项(XDJK2016D041)

摘　　要：基于相似度的模糊坡位分类算法使用最小值算子来综合不同地形因子的相似度,忽略了不同的地形因子对不同坡位的影响程度的差异.该文使用随机森林算法分析地形因子与坡位类型之间的关系,计算出不同地形因子相对于各类坡位的重要性评分,并以此作为加权因子,通过加权因子原则,综合计算待分类位置与典型位置的相似度.结果显示,基于随机森林算法的分类准确度和Kappa系数分别达到了0.97和0.96,而基于最小值算子算法的分类准确度和Kappa系数分别为0.88和0.83.相比原有方法,新方法的分类准确度提高了0.09,Kappa系数提高了0.13,由此证明,使用本文提出的新方法进行模糊坡位分类,在一定程度上提高了坡位分类的效果.The similarity-based algorithm of fuzzy slope position segmentation has been adopted to identify slope position.However,this algorithm which uses minimal operator(MIN)to calculate the comprehensive similarity ignores varied influence of terrain indicators on different slope positions.Thus,the purpose of this study is to improve this algorithm by introducing random forest method.The work has been conducted in a hilly area(about 0.6 km 2)of Chongqing,southwest China.Relative position index,profile curvature,plan curvature,slope,and elevation were derived from a digital elevation model with a spatial resolution of 2 m.Typical slope positions including ridge,slope shoulder,backslope,footslope,and channel were identified by prior knowledge of local experts.These locations were divided into training and validation sets.The relationship between the terrain indicators and slope positions was investigated and fuzzy membership functions were determined.The relative importance of terrain indicators to identify different slope positions were calculated using random forest method.These values were the weighting factors of the terrain indicators for distinguishing the slope positions.The results show that the new proposed method(NRF)outperforms the MIN algorithm.The overall accuracy and kappa coefficient of NRF and MIN are 0.97,0.96,0.88 and 0.83,respectively.The improvement of overall accuracy and kappa coefficient of NRF over MIN are 10%and 16%,respectively.These suggest that the relative importance of terrain indicators is suitable for identifying slope positions.

关 键 词：随机森林 地形因子 加权因子原则 模糊坡位分类 相似度 

分 类 号：P91[天文地球—自然地理学]