基于随机森林与特征选择的藏东南土地覆被分类方法及精度评价

Land cover classification based on random forest and feature optimism in the Southeast Qinghai-Tibet Plateau

由于云污染、实地验证点的匮乏,以及地形地貌的复杂、破碎化,多云山区土地覆被的准确分类较难实现。以藏东南这一典型的多云山区及生态过渡区为研究区,基于Google Earth Engine(GEE)平台和野外实测数据,结合多光谱数据、雷达数据、高程数据、辅助数据,提取光谱特征、纹理特征、地形特征等信息,利用递归特征消除法对特征进行优化,并采用随机森林算法构建分类模型,以期有效利用多源遥感数据提高土地覆被分类精度。结果表明:(1)并非特征越多分类精度越高,特征选择后数量由58个减至38个,分类精度(总体精度93.96%,Kappa系数0.92)较未优化前(总体精度93.11%,Kappa系数0.92)略有提升。(2)地形特征及雷达特征对藏东南土地覆被分类具有重要作用,地形特征对多数土地覆被类型的分类精度具有影响,而雷达数据对裸地、建设用地、灌丛影响较大,分类过程中如不考虑地形及雷达特征,总体精度分别降至88.98%,92.48%。纹理特征以及时序特征仅对提高具有明显纹理以及时序变化的土地覆被类型的精度有帮助。结合随机森林和特征优化算法,能够在保证土地覆被分类精度的同时,高效整合多源数据信息,加快模型运算速度,为多云山区土地覆被分类提供切实可行的方法。

Obtaining accurate land cover information in cloudy mountain areas are severely impacted by cloud contaminations,the scarcity of field validation points,and the complexity and fragmentation of landforms.Taken the Southeast Qinghai-Tibet Plateau,a typical cloudy mountainous area and ecological transition zone,as study area,this research first extract the spectral features,radar features,textual features,topographic features through the spectral data,radar data,DEM data and auxiliary data based on the Google Earth Engine and filed observed data.Then we built the random forest model and made feature reduction using recursive feature elimination,in order to improve the accuracy of land cover classification by using multi-source remote sensing data effectively.Results showed that:1)The feature numbers reduced from 58 to 38 after feature optimization,classification accuracy(overall accuracy 93.96%,Kappa coefficient 0.92)slightly improved compared to unoptimized(overall accuracy 93.11%,Kappa coefficient 0.92);2)Topographic and radar features played an essential role in the land cover classification of mountainous cloudy areas.If the topographic features and radar feature were excluded,accuracies would decrease to 88.98%and 92.48%,respectively.Topographic features would influence the classification accuracy of most land cover types,while radar features had more impacts on bare lands,construction lands,and shrublands.Textual features and sequential features could only help to increase the accuracy of land cover type with significant textual features and temporal variations.More accurate land cover information can be detected by combing the random forest and feature optimization algorithm,while also provide a more efficient and faster way of integrating multisource data,thus making contribution to land cover classification of the cloudy and mountainous area.