结合无人机多光谱数据和机器学习算法的春小麦叶面积指数反演  

Using multispectral spectrometry and machine learning to estimate leaf area index of spring wheat

作　　者：刘琦 屈忠义[1,2] 白燕英 杨威[1] 方海燕 白巧燕 杨旖璇 张如鑫 LIU Qi;QU Zhongyi;BAI Yanying;YANG Wei;FANG Haiyan;BAI Qiaoyan;YANG Yixuan;ZHANG Ruxin(Inner Mongolia Agricultural University,Water Conservancy and Civil Engineering College,Hohhot 010018,China;Inner Mongolia University of Science&Technology,School of Energy and Environment,Baotou 014010,China;Hetao Irrigation District Water Resources Development Centre-Jiefangzha Subcentre,Bayannur 015400,China;Water Conservancy Service Center of Bayannur City,Bayannur 015000,China)

机构地区：[1]内蒙古农业大学水利与土木建筑工程学院,呼和浩特010018 [2]内蒙古科技大学能源与环境学院,内蒙古包头014010 [3]河套灌区水利发展中心解放闸分中心,内蒙古巴彦淖尔015400 [4]内蒙古巴彦淖尔市水利事业服务中心,内蒙古巴彦淖尔015000

出　　处：《灌溉排水学报》2024年第11期63-73,共11页Journal of Irrigation and Drainage

基　　金：国家重点研发计划项目(2021YFC3201205);内蒙古自治区高校青年科技英才项目(NJYT22045)。

摘　　要：【目的】探究基于无人机多光谱数据反演大田春小麦叶面积指数(LAI)的最优机器学习建模方法。【方法】以内蒙古沿黄流域的土默川平原春小麦为对象,利用大疆P4M无人机采集了3个关键生育时期(拔节期、孕穗期、灌浆期)多光谱影像数据,提取植被指数同实测LAI进行相关性分析,用以植被指数的筛选,筛选后的植被指数进行主成分分析(PCA),将新主成分因子作为模型输入变量结合多元线性回归(MLR)、决策树回归(DTR)、BP神经网络回归(BPNN)、梯度提升树回归(GBDT)、支持向量机回归(SVR)和随机森林回归(RFR)6种建模方法,分别构建不同生育时期LAI估算模型,经精度验证,确定LAI的最优估算模型。【结果】归一化植被指数(NDVI)、改进的简单比值植被指数(MSR)、比值植被指数(RVI)、差值植被指数(DVI)、土壤调节植被指数(SAVI)和归一化差异红边指数(NDRE)与LAI均存在显著相关性,但重归一化植被指数(RDVI)在孕穗期和灌浆期与LAI的相关系数仅为0.23和0.21。在反演模型方面,拔节期的BPNN模型表现最优,验证集R^(2)为0.822,RMSE为0.305,MAE为0.257。而在孕穗期、灌浆期以及整个生育期内,RFR模型表现最佳,验证集R^(2)分别为0.613、0.811和0.834,对应的RMSE分别为0.189、0.150和0.174,MAE分别为0.126、0.121和0.133,且以多生育时期数据建立的RFR模型精度大于单生育时期模型。【结论】利用无人机多光谱数据计算的植被指数结合机器学习算法可以较好反映春小麦各生育时期LAI分布情况,其中,使用多生育时期数据建立的模型估算精度大于单生育时期模型,单生育时期中,拔节期的预测精度最高,其次是灌浆期和孕穗期。在算法方面,BPNN算法构建的LAI估算模型在拔节期的反演精度最高RFR算法构建的LAI估算模型是准确估算春小麦中后期LAI表型参数的优选方法,相比之下GBDT模型在各生育时期的反演精度均较低,不推�【Objective】The leaf area index(LAI)is an important trait of plant canopies but challenging to measure accurately at large scales.We studied the feasibility of using multispectral imaging and machine learning to estimate the LAI of spring wheat.【Method】The experiment was conducted in a spring wheat field on the Tumochuan Plain in the Yellow River Basin,Inner Mongolia.Images of the spring wheat at the jointing,booting,and grain-filling stages were acquired using a multispectral camera mounted on a DJI P4M UAV.Selected vegetation indices were subjected to principal component analysis(PCA),and the resulting components were used to estimate LAI.We compared six models:multiple linear regression(MLR),decision tree regression(DTR),backpropagation neural network regression(BPNN),gradient boosting decision tree regression(GBDT),support vector machine regression(SVR),and random forest regression(RFR).LAI was calculated separately for each growth stage using different vegetation indices.【Result】LAI was significantly correlated with the normalized difference vegetation index(NDVI),modified simple ratio(MSR),ratio vegetation index(RVI),difference vegetation index(DVI),soil-adjusted vegetation index(SAVI),and normalized difference red edge index(NDRE).It showed a weak correlation with the renormalized difference vegetation index(RDVI)during the heading and grain-filling stages,with their correlation coefficients being 0.23 and 0.21,respectively.The BPNN model was most accurate during the jointing stage,with R^(2),RMSE,and MAE being 0.822,0.305,and 0.257,respectively.In contrast,the RFR model performed best during the heading,grain-filling,and entire growth periods,with R^(2)being 0.613,0.811 and 0.834,RMSE being 0.189,0.150 and 0.174,and MAE being 0.126,0.121 and 0.133,respectively.Additionally,the RFR model constructed using data from all three stages was more accurate than models derived from data at individual growth stages.【Conclusion】Multispectral data acquired via UAV,combined with machine learning algorit

关 键 词：无人机 多光谱 春小麦 叶面积指数 机器学习算法 植被指数 

分 类 号：S127[农业科学—农业基础科学]