基于无人机多光谱影像和机器学习方法的玉米叶面积指数反演研究  

Comparing different machine learning methods for maize leaf area index(LAI)prediction using multispectral image from unmanned aerial vehicle(UAV)

作　　者：马俊伟 陈鹏飞[2,4] 孙毅 谷健[3] 王李娟[1] MA Jun-Wei;CHEN Peng-Fei;SUN Yi;GU Jian;WANG Li-Juan(School of Geography,Geomatics and Planning,Jiangsu Normal University,Xuzhou 221116,Jiangsu,China;State Key Laboratory of Resource and Environmental Information System,Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences,Beijing 100101,China;Institute of Applied Ecology,Chinese Academy of Sciences,Shenyang 110016,Liaoning,China;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application,Nanjing 210023,Jiangsu,China)

机构地区：[1]江苏师范大学地理测绘与城乡规划学院,江苏徐州221116 [2]中国科学院地理科学与资源研究所/资源与环境信息系统国家重点实验室,北京100101 [3]中国科学院沈阳应用生态研究所,辽宁沈阳110016 [4]江苏省地理信息资源开发与利用协同创新中心,江苏南京210023

出　　处：《作物学报》2023年第12期3364-3376,共13页Acta Agronomica Sinica

基　　金：中国科学院先导A专项(XDA28040502);国家自然科学基金项目(41871344);江苏师范大学研究生科研创新计划项目(2022XKT0070)资助。

摘　　要：为实现基于机器学习方法和无人机影像的叶面积指数(leaf area index,LAI)准确估测。本研究对比了人工神经网络法(Artificial Neural Network algorithm,ANN)、高斯过程回归法(Gaussian Process Regression algorithm,GPR)、支持向量回归法(Support Vector Regression algorithm,SVR)和梯度提升决策树法(Gradient Boosting Decision Tree,GBDT)等几种主流的机器学习方法在基于无人机影像的玉米LAI反演中的优劣。为此,开展了不同有机肥、无机肥、秸秆还田以及种植密度处理的玉米田间试验,在不同生育期获取了无人机多光谱影像和LAI数据。基于这些数据,首先通过相关性分析,选择对LAI敏感的光谱指数作为估测变量,然后分别耦合偏最小二乘法(Partial Least Squares Regression,PLSR)和ANN、GPR、SVR、GBDT建立LAI反演模型,并对它们进行对比分析。结果表明,PLSR+GBDT法构建的LAI反演模型精度最高,稳定性最好,建模R_(cal)^(2)和RMSE_(cal)为0.90和0.25,验证R_(val)^(2)和RMSE_(val)为0.90和0.29;与PLSR+GBDT模型结果最接近的是基于PLSR+GPR法建立的模型,其建模R_(cal)^(2)和RMSE_(cal)为0.86和0.30,验证R_(val)^(2)和RMSE_(val)为0.89和0.29,且具有训练速度快,并能给出反演结果不确定度的优势;PLSR+ANN法的建模R_(cal)^(2)和RMSE_(cal)为0.85和0.31,验证R_(val)^(2)和RMSE_(val)为0.89和0.30;PLSR+SVR法的建模R_(cal)^(2)和RMSE_(cal)为0.86和0.32,验证R_(val)^(2)和RMSE_(val)为0.90和0.33。因此,PLSR+GBDT法和PLSR+GPR法被推荐作为玉米LAI反演模型构建的最优方法。To make an accurate estimation of leaf are index(LAI)based on machine learning methods and images from UAV,we compared the several mainstream machine learning methods for maize LAI prediction,such as Artificial Neural Network method(ANN),Gaussian Process Regression method(GPR),Support Vector Regression method(SVR),and Gradient Boosting Decision Tree(GBDT).For this purpose,field experiments that considering apply of different amount of organic fertilizer,different amount of inorganic fertilizer,different amount of crop residue,and different planting density were carried out.Based on these experiments,field campaign were conducted to obtain UAV multispectral images and LAI data at different growth stages in maize.Based on above data,firstly,correlation analysis was used to select LAI-sensitive spectral indices,and then the Partial Least Squares Regression method(PLSR)and ANN,GPR,SVR,GBDT were coupled to design the LAI prediction models,respectively,and their performance for LAI prediction were compared.The results showed that the LAI prediction model constructed by PLSR+GBDT method had the highest accuracy and the best stability.The models of R^(2) and RMSE values were 0.90 and 0.25,and the verified R^(2) and RMSE values were 0.90 and 0.29 during validation,respectively.The model based on PLSR+GPR model was followed,with R^(2) and RMSE values of 0.86 and 0.30 during calibration,and R^(2) and RMSE values of 0.89 and 0.29 during validation,respectively.Besides,it had faster training speed and could give the uncertainty of the prediction.The model designed by PLSR+ANN method had R^(2) and RMSE values of 0.85 and 0.31 during calibration,and R^(2) and RMSE values of 0.89 and 0.30 during validation,respectively.The model designed by PLSR+SVR method had R^(2) and RMSE values of 0.86 and 0.32,and R^(2) and RMSE values of 0.90 and 0.33,respectively.Therefore,PLSR+GBDT method and PLSR+GPR method are recommended as the optimal methods for designing maize LAI prediction models.

关 键 词：叶面积指数 机器学习 无人机 多光谱影像 玉米 

分 类 号：S513[农业科学—作物学]