Estimation Of Eco-physiological Parameters Of Alfalfa Based On PROSAIL Model And Machine Learning

Alfalfa（Medicago sativa L.）as an important forage in China.Its eco-physiological parameters can reflect the growth and health of alfalfa.The prediction of alfalfa eco-physiological parameters can provide data support for the dynamic adjustment of alfalfa production plan in Dingxi region.In addition,the remote sensing technology is used for nondestructive monitoring,and its simplicity,reliability and popularization are in line with the characteristics of Chinese agricultural modernization.This study takes alfalfa in Dingxi Loess Plateau as the research object.Based on the PROSAIL physical model,combined with local sensitivity analysis and out of bag data analysis,the sensitivity indices of eco-physiological parameters of alfalfa are selected.Finally,support vector regression（SVR）,random forest（RF）,adaptive boosting algorithm（Ada Boost）,Bayesian ridge regression（BR）,and linear regression（LR）algorithms are used to establish estimation models for eco-physiological parameters of alfalfa.R²,root mean square error（RMSE）,mean absolute error（MAE）,and standard deviation（SD）are used as evaluation indicators for model accuracy.The specific results are as follows:（1）The spectral reflectance of alfalfa gradually decreases with the increase of growth period,and the model spectral reflectance is slightly higher than the measured spectrum.Equivalent water thickness（EWT）,leaf mass per area（LMA）,leaf area index（LAI）,chlorophyll content（Cab）corresponding to the sensitive wavelength range of950～2400 nm,850～2400 nm,400～2400 nm,500～670 nm,respectively.（2）EWT sensitive indices include WI,NDWI,NDII and MSI,and the best index is WI.LMA sensitive indices include MCARI1,RDVI,SLMAI1,SLMAI3,and the best index is MCARI1.LAI sensitive indices include NDVI,RDVI,RENDVI and SLAII3,and the best index is NDVI.The sensitive indices of Cab include TGI,TCARI,MCARI and CI,and the best index is TGI.Based on RF off-bag data,the sensitive indices of yield include RDVI（0st）,m NDVI705（1st）,NDVI705（1st）,MCARI1（1st）,NDVI（1st）,DVI（1st）,PRI（0st）,NDVI（0st）,NDGI（0st）,and the best index is RDVI（0st）.Based on Ada Boost out-of-bag data,the sensitive indices of yield include RDVI（0st）,NDVI（1st）,NDVI705（1st）,m NDVI705（1st）,MCARI1（1st）,NDGI（1st）,DVI（1st）,PRI（0st）,NDVI705（0st）,and the best index is RDVI（0st）.（3）The R² of the five EWT estimation models are greater than 0.86,the RMSE is not greater than 0.0120 g·cm^-2,and the MAE is not greater than 0.0077 g·cm^-2.The best model is EWT-RF.The R² of the five LMA estimation models are greater than 0.82,the RMSE is not greater than 0.0020 g·cm^-2,and the MAE is not greater than 0.0014 g·cm^-2.The best model is LMA-SVR.The R² of the five LAI estimation models are greater than 0.86,the RMSE is not greater than 0.18,and the MAE is not greater than 0.12.The best model is LAI-RF.The R² of the five Cab estimation models are greater than 0.78,and the RMSE is not greater than 2.85μg·cm^-2,MAE are not greater than 1.86μg·cm^-2.The best model is Cab-SVR.The R² range of the ten yield estimation models are0.15～0.88,the RMSE is not greater than 0.0635 kg·m^-2,and the MAE is not greater than 0.0540 kg·m^-2.The best model is B-LR.In summary,the PROSAIL model is suitable for estimating the eco-physiological parameters of alfalfa.After coupling sensitivity analysis and out of bag data analysis to screen sensitivity indices,high-precision alfalfa growth and health monitoring models EWT-RF,LMA-SVR,LAI-RF,Cab-SVR,and yield monitoring model B-LR were obtained,which can provide a certain method reference for improving the production efficiency,growth and health status,and yield monitoring work of alfalfa industry in Dingxi and similar areas.