Research On Monitoring Method By Satellite Remote Sensing For Biomass And LAI Of Wheat From Jointing Stage To Flowering Stage

Satellite remote sensing has a unique effect on the investigation,monitoring and management of large-scale agricultural production due to its advantages such as wide coverage and high speed.Previous models of remote sensing monitoring and prediction mainly rely on single remote sensing variable or the selection of input variable needs to be studied,and did not effectively combine the comprehensive characteristics of each input variable.In order to improve the accuracy of the remote sensing monitoring model for biomass and LAI of wheat from jointing to flowering,this study took 10 counties including Gaoyou,Yizheng,Xinghua and Zhangjiagang in central Jiangsu province and along the Yangtze River as the survey area.Firstly,the wheat planting area in the central plain of Jiangsu Province and along the Yangtze River was extracted.Secondly,the main structural parameters,such as biomass and LAI,were obtained in the field at the three key growth stages of wheat jointing,booting and flowering in each experimental area,and the remote sensing data of ESA Sentinel-2 satellite were obtained simultaneously to extract the remote sensing vegetation index.Combining Filter,Embedded and Wrapper three kinds of machine learning feature engineering technology for optimal variable selection.Three classic machine learning algorithms,namely ridge regression,Lasso and elastic net regression,were used to construct the model,and the best fitting parameters were determined by grid search and cross-validation methods.Integrate the best feature screening technology and model construction algorithm combination,and combine the measured data for model evaluation.Finally,the goal of remote sensing monitoring of biomass and LAI of wheat from jointing to flowering period was realized.The main research results are as follows:(1)Extraction of wheat planting areas in central Jiangsu and along the Yangtze River.The images at the jointing stage were used for wheat field extraction.The field survey samples in each county were used as training samples and verification samples,and the Maximum Likelihood Estimate,Minimum Distance Estimate,Support Vector Machine Estimate and Neural Net Estimate were used to extract the wheat area of each test county.Meanwhile,the Kappa coefficients of the four methods were compared.The results showed that the Kappa coefficient of each study area extraction effect based on the Support Vector Machine Estimate had the best overall performance.Therefore,the Support Vector Machine Estimate was selected as the wheat area extraction method in central Jiangsu and along the Yangtze River.(2)The number of variables used in the modeling of the wheat biomass at jointing stage using Filter,Embedded and Wrapper were 24,19,and 10/5/5,respectively;The number of variables used in the modeling of the wheat LAI at jointing stage using Filter,Embedded and Wrapper were 24,10,and 5/4/9,respectively.The best combined models for monitoring biomass and LAI at the jointing stage were the combination of Wrapper-Ridge regression and the Embedded-Ridge regression.The R2 on the modeling set and verification set reached 0.70,0.67 and 0.66,0.65,respectively.Compared with the ordinary univariate linear model,the R2 on the modeling set and verification set of the biomass monitoring model increased by 0.2470 and 0.2306,respectively,and the RMSE decreased by 279.28 kg/ha and 267.03 kg/ha,respectively.The R2 on the modeling set and verification set of the LAI monitoring model increased by 0.2172 and 0.2225,respectively,and the RMSE decreased by 0.1175 and 0.1209,respectively.(3)Three feature selection methods were used to screen out the feature numbers of the biomass and LAI modeling at the booting stage of wheat as 16,12,12/6/14 and 15,7,7/5/7.The best combined models for monitoring biomass and LAI at the booting stage were the combination of Wrapper-Lasso and the Filter-Ridge regression.The R2 on the modeling set and verification set reached 0.67,0.65 and 0.69,0.65,respectively.Compared with the ordinary univariate linear model,the R2 on the modeling set and verification set of the biomass monitoring model increased by 0.1881 and 0.1797,respectively,and the RMSE decreased by 317.46 kg/ha and 306.86 kg/ha,respectively.The R2 on the modeling set and verification set of the LAI monitoring model increased by 0.2556 and 0.2350,respectively,and the RMSE decreased by 0.1233 and 0.1171,respectively.(4)Three feature selection methods were used to screen out the feature numbers of the biomass and LAI modeling at the flowering stage of wheat as 16,21,12/9/18 and 24,8,6/4/10.The best combined models for monitoring biomass and LAI at the flowering stage were the combination of Wrapper-Ridge regression and the Embedded-Lasso.The R2 on the modeling set and verification set reached 0.68,0.64 and 0.66,0.64,respectively.Compared with the ordinary univariate linear model,the R2 on the modeling set and verification set of the biomass monitoring model increased by 0.2058 and 0.1998,respectively,and the RMSE decreased by 525.07 kg/ha and 515.97 kg/ha,respectively.The R2 on the modeling set and verification set of the LAI monitoring model increased by 0.2400 and 0.2295,respectively,and the RMSE decreased by 0.1203 and 0,1143,respectively.In view of the current selection methods of remote sensing variables mainly based on empirical selection and modeling methods that still need to be reformed,this study used machine learning feature selection technology combined with regression algorithms to effectively improve the accuracy of the monitoring model of biomass and LAI of wheat during the key growth period,thereby providing technical reference for effective services in precision agriculture.Combining the best monitoring model for biomass and LAI of wheat during the key growth period and remote sensing images of the same period,remote sensing monitoring thematic maps at different scales could be drawn,so as to achieve multi-scale monitoring of the wheat field conditions at the provincial,municipal,county levels,and field levels.Provide fast,accurate and continuous monitoring information for agricultural production activities.