随机森林算法在干旱区土地利用遥感分类中的应用研究

Application of random forest algorithm in the remote sensing classification of land use in arid region

为了验证随机森林算法在干旱区土地利用遥感分类中的效果,本文采用随机森林算法,结合Landsat8遥感影像以及DEM、NDVI等辅助数据,解译了干旱区典型流域玛纳斯河流域的土地利用图。分析结果表明:(1)分析决策树数量(k)和分类变量数量(m)对分类精度具有很大影响。通过优化2个参数得到最优随机森林模型,当k取103、m取6时,模型分类精度可达95%;(2)通过土地利用分类精度的影响因子分析发现,海拔高程和归一化植被指数对土地利用分类的影响程度比坡向的影响更大。(3)通过分类结果对比分析发现,应用随机森林算法分类的精度比用最大似然法的分类精度高9%,利用变量重要性筛选出的遥感波段构建优化随机森林模型,能有效降低遥感数据源数据量,而Kappa系数保持在0.97不变。随机森林算法可以在干旱区土地利用分类中广泛应用。

The aim of this resesearch is to test the random forest algorithm in the remote sensing classification in arid area.The land use map of the Manasi River Basin,a typical watershed in arid area,had been interpreted using this method based on Landsat 8,DEM and NDVI data.The results show that：（1） both the number of decision trees（k） and the number of classification variables（m） have effect on the classification accuracy.When k,m are 103,6,respectively,the classification accuracy reaches 95%;（2） through the comparative analysis,DEM and NDVI have more influence on the classification accuracy than the slope;（3） Based on the results of classification,the random forest algorithm accuracy is higher than 9% with the maximum likelihood method..The random forest algorithm could reduce data redundancy and keep the accuracy with 0.97 of Kappa coefficient.The method could be applied in land use classification of remote sensing in the arid area.