摘要
针对干旱区果园大面积遥感提取困难、识别精度低等问题,本研究基于GEE(Google earth engine)平台,综合应用Sentinel-1/Sentinel-2影像构建特征集.通过对比原始特征组合、Jeffries-Matusita(J-M)距离、属性重要度3 种优化方式,结合随机森林(Random forest,RF)分类方法,对比得到最佳优化方式,探索果园最优分类特征集.结果表明:识别效果最好的方案为G17JM,总体精度为 91.25%,kappa系数为 0.89,面积精度为 82.55%.最优特征集为B8 asm、B8 ent、B8 idm、NDVIre3、B6、B7、a、e、b、EVI、B11、B8A、B8、VV.使用J-M距离进行特征集优化,有效降低数据量、提高计算效率,更有利于精确遥感识别果园种植面积.表明GEE快速、准确获取果园种植面积的可行性,为获取果园动态变化提供强有力的基础.
Abstract
Aiming at the problems of difficult extraction and low recognition accuracy of orchards in arid areas,based on the Google earth engine(GEE)platform,this study comprehensively applied Sentinel-1/Sentinel-2 images to con-struct feature sets.By comparing the three optimization methods of original feature combination,Jeffries-Matusita(J-M)distance and attribute importance,combined with random forest(RF)classification method,the best optimization method was obtained,and the optimal classification feature set of orchard was explored.The results showed that the best recognition scheme was G17JM,the overall accuracy was 91.25%,kappa coefficient was 0.89,and area accuracy was 82.55%.The op-timal feature set was B8 asm,B8 ent,B8 idm,NDVIre3,B6,B7,a,e,b,EVI,B11,B8A,B8,VV.Using J-M dis-tance to optimize the feature set can effectively reduce the amount of data and improve the computational efficiency,which was more conducive to the accurate identification of orchard planting area.It shows that GEE is feasible to obtain orchard planting area quickly and accurately,and provides a strong basis for obtaining orchard dynamic changes.
基金项目
新疆自然科学计划(自然科学基金)联合基金项目(2021D01C055)
维普
万方数据