基于Sentinel-2影像和机器学习算法的冬小麦秸秆覆盖度遥感估算  

Remote Sensing Estimation of Winter Wheat Residue Coverage Based on Sentinel-2 Images and Machine Learning Algorithm

作　　者：朱奇磊 梁栋[1,4] 徐新刚 安晓飞[3] 陈立平[3] 杨贵军[2] 黄林生[1,4] 许思喆 ZHU Qilei;LIANG Dong;XU Xingang;AN Xiaofei;CHEN Liping;YANG Guijun;HUANG Linsheng;XU Sizhe(School of Electronic and Information Engineering,Anhui University,Hefei,Anhui 230601,China;Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;Intelligent Equipment Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;National Engineering Research Center for Agro-Ecological Big Data Analysis and Application,Anhui University,Hefei,Anhui 230601,China)

机构地区：[1]安徽大学电子信息工程学院,安徽合肥230601 [2]北京市农林科学院信息技术研究中心,北京100097 [3]北京市农林科学院智能装备技术研究中心,北京100097 [4]安徽大学农业生态大数据分析与应用技术国家地方联合工程研究中心,安徽合肥230601

出　　处：《麦类作物学报》2023年第4期524-535,共12页Journal of Triticeae Crops

基　　金：国家重点研发计划项目(2019YFE0125300);北京市农林科学院科技创新能力建设专项(KJCX20210433);国家现代农业产业技术体系资助项目(CARS-03)。

摘　　要：为探究大范围小麦秸秆覆盖度(CRC)估测方法,以冬小麦秸秆为研究对象,基于Sentinel-2遥感卫星影像光谱指数、波段和纹理特征及其不同特征组合,利用灰色关联-随机森林(GRA-RF)敏感特征提取方法,结合高斯过程(GPR)、套索(LASSO)、岭回归(RR)和偏最小二乘(PLSR)等多种机器学习算法,开展小麦CRC估算的最优模型研究。结果表明,基于GRA-RF特征优选后的机器学习模型显著改善了小麦CRC的估算精度,LASSO算法总体对小麦CRC的估测效果最佳,并且针对不同的光谱特征组合表现出差异化的结果。其中,以光谱指数、波段和纹理信息构成的组合特征集构建的CRC遥感估算模型精度最优(r^(2)=0.65,RMSE=9.25%),以波段与纹理两者组合特征估算的CRC精度次之(r^(2)=0.63,RMSE=9.31%),仅利用单一的光谱指数、波段或者纹理特征估算冬小麦CRC的精度均劣于组合特征的结果。这说明应用GRA-RF组合筛选方法能够有效优选秸秆覆盖度的光谱特征;相比于单一特征,光谱指数、波段、纹理信息等构成的组合特征更能有效地监测小麦秸秆覆盖度。In order to explore the estimation method for crop residue coverage(CRC)in a large scale,winter wheat residue was taken as the research object.Based on the spectral index,band and texture features of Sentinel-2 remote sensing satellite image and their different feature combinations,the optimal model of wheat CRC estimation was studied by using the sensitive feature extraction method of grey correlation-random forest(GRA-RF),combined with a variety of machine learning algorithms,such as Gaussian process regression(GPR),LASSO,ridge regression(RR),and partial least squares regression(PLSR).The results showed that the machine learning model based on GRA-RF feature selection significantly improved the estimation accuracy of wheat CRC,and the LASSO algorithm had the best overall estimation effect on wheat CRC,and showed different results for different spectral feature combinations.The CRC remote sensing estimation model based on the combination feature set composed of spectral index,band and texture information had the best accuracy(r^(2)=0.65,RMSE=9.25%).The CRC accuracy estimated by the combination characteristics of band and texture was the second(r^(2)=0.63,RMSE=9.31%).Only using a single spectral index,band or texture feature to estimate the CRC of winter wheat,the model accuracy was inferior to the results of the combined features.It indicates that the GRA-RF integrated method can effectively select spectral features related to CRC,and the combined features had the better capability of estimating CRC.

关 键 词：秸秆覆盖度 灰色关联分析-随机森林 机器学习算法 特征变量筛选 

分 类 号：S512.1[农业科学—作物学] S311