基于旋转森林的Landsat-8影像森林植被分类

摘要

Based on Landsat 8OLI imageries of Walagan forest farm,Daxing'an Mountains,Heilongjiang Province in 2013,we added NDVI (Normalized Difference Vegetation Index),textural features and topological features to spectral bands,combined them as three feature groups (F1:OLI bands and NDVI;F2:OLI bands,NDVI and texture features;F3:OLI bands,NDVI,texture and topographic features),and implemented forest vegetation classification using Rotation Forest (RoF).With the optimal feature group,RoF was compared with MLC (Maximum Likelihood Classifier) and SVM (Support Vector Machine).F3 feature group yielded the highest classification accuracy using RoF,that is,87.54％,higher than F1 and F2 group by 11.08％ and 3.39％,respectively.Comparing different classification methods,RoF yielded higher classification accuracy than MLC and SVM by 13.24％ and 5.39％,respectively.Due to the good stability of RoF,hill shadow had little influence on forest vegetation classification,thus,RoF could provide the best classification map with highest accuracy and least "pepper and salt" effect among the three classification methods.%以黑龙江大兴安岭塔河林业局瓦拉干林场2013年的Landsat8OLI影像为数据源,在光谱特征基础上,增加归一化植被指数、纹理特征和地形特征,得到3种特征组合(光谱特征和NDVI(F1);光谱特征、NDVI和纹理特征结合(F2);光谱特征、NDVI、纹理特征和地形特征结合(F3)),将旋转森林算法分别应用于3种特征组合下的森林植被分类,获得分类精度最高的特征组合;之后利用最佳特征组合将旋转森林与最大似然分类法和支持矢量机2种分类方法进行对比和精度验证分析.结果表明:利用旋转森林算法并结合光谱特征、NDVI、纹理特征和地形特征的特征组合分类精度最高,为87.54％,比F1和F2特征组合的精度分别提高了11.08％和3.39％.比较不同分类方法,旋转森林算法进行森林植被的分类精度比最大似然法和支持矢量机方法的分类精度分别提高了13.24％和5.39％.由于旋转森林算法稳定性好,在植被分类中受山地阴影的影响较少,因此在分类图中“椒盐”现象最少,图像更加清晰,分类效果最好.