基于随机森林算法的城区土地覆盖分类研究

Research on urban land cover classification based on random forest algorithm

随机森林算法是一种高度灵活且易于使用的机器学习算法,目前在遥感影像分类中应用广泛。为了验证其在城市土地覆盖分类中的效果,本文对河南省洛阳市局部城区进行了土地覆盖分类实验,将Landsat 8(OLI)遥感影像的光谱波段、光谱指数和纹理特征相结合,构成多种特征组合进行随机森林算法分类比较,选择分类效果最佳方案,并与支持向量机方法进行比较。后利用随机森林算法对该组合特征变量高维数据进行降维处理,得到优化特征方案。实验结果表明:采用多源特征组合的随机森林算法的土地利用分类效果最佳,总体精度为90.54%,Kappa系数为0.890,比支持向量机方法的分类精度提高了3.1%;降维处理后的特征方案与随机森林结合在保证分类结果拥有高准确度的同时,减少了运算时间,实现了土地覆被类型信息的高效获取。表明随机森林算法在城区土地覆盖分类上有很好的适用性与稳定性。

Random forest algorithm is a highly flexible and easy-to-use machine learning algorithm.It is currently widely used in remote sensing image classification.In order to verify its effect in the classification of urban land cover,this study carried out a land cover classification experiment on local urban areas of Luoyang City,Henan Province,combining the spectral band,spectral index and texture features of Landsat 8(OLI)remote sensing images to form multiple a variety of feature combinations are used to perform random forest algorithm classification and comparison,and the best classification scheme is selected,and compared with the support vector machine method.Then the random forest algorithm is used to reduce the dimension of the combined feature variable high-dimensional data to obtain the optimal feature solution.The experimental results show that the multi-source feature combination random forest algorithm has the best land use classification effect,with an overall accuracy of 90.54%and a Kappa coefficient of 0.890,which is 3.1%higher than the classification accuracy of the support vector machine method;The feature scheme after dimension reduction and random forest are combined to ensure high accuracy of classification results,reduce operation time and achieve efficient acquisition of land cover type information.It shows that the random forest algorithm has good applicability and stability in urban land cover classification.