Summer Maize Growth Monitoring Based On UAV Multi-spectral Remote Sensing

In recent years,the research on the theory and practice of precision agriculture has started.It is an urgent need for precision agriculture to evaluate the growth situation quantitatively.Unmanned aerial vehicle（UAV）has the advantages of convenience,flexibility and high temporal and spatial resolution,so it has become an important data source of agricultural remote sensing at farm scale.In this paper,a two-year field experiment was carried out in Yangling to study the summer corn.There were seven treatments in this experiment,named CK,N1-1,N1-2,N2-1,N2-2,N3-1 and N3-2,respectively.Among them,CK（no fertilization）,N1(105 kg hm^-2),N2(210 kg hm^-2)and N3(315 kg hm^-2)were different nitrogen application rates.Under the same nitrogen application rate,it was divided into no topdressing（N1-1,N2-1,N3-1）and topdressing in jointing stage（N1-2,N2-2,N3-2）.In jointing stage,tasseling stage and filling stage,the multi-spectral images of unmanned aerial vehicles and the growth indicators such as leaf area index,aboveground biomass,SPAD and photosynthetic parameters were collected.Comprehensive growth index（CGI）was constructed by combining leaf area index,aboveground biomass,SPAD and photosynthetic parameters.After calculating the vegetation indexes from multispectral images,the correlations between growth indicators（leaf area index,aboveground biomass,SPAD,photosynthetic parameters,comprehensive growth index）and vegetation indexes were analyzed to select sensitive vegetation index as input variable of the model.The estimation models of growth indicators in different growth periods were constructed by partial least squares regression and random forest.Furthermore,the coefficient of determination（R²）,root mean square error（RMSE）and relative prediction deviation（RPD）were used to evaluate the accuracy to optimize the best estimation model,so as to achieve the purpose of monitoring the growth of summer corn and provide support for the development of precision agriculture.The main results of this study were as follows:（1）The vegetation indices（NDVI,RVI,OSAVI,MSR）constructed by red band and near infrared band had good consistency in characterizing crop canopy characteristics.The vegetation indices of CK were the smallest,followed by N1-1 and N1-2,and the vegetation indices of the other four treatments had no obvious difference.The vegetation indices（MCARI,NRI）,which were dominated by visible light band,had poor stability,and they often reached the maximum in CK treatment.After topdressing,CIre could better characterize the topdressing effect of nitrogen application levels of 210 kg hm^-2 and 315 kg hm^-2（i.e.N2-2 and N3-2）.（2）The correlations between growth indicators（leaf area index,aboveground biomass,SPAD,photosynthetic parameters,comprehensive growth index）and the selected vegetation indices had a good consistency.NDVI,GNDVI,RVI,OSAVI,MSR,CIre showed a strong and stable positive correlation with the growth indicators,but MCARI and NRI showed a negative correlation,and the correlation coefficients were poor.Except MCARI and NRI,the vegetation indices were highly correlated with the growth indicators basically.（3）For leaf area index,aboveground biomass and SPAD,with the increase of nitrogen application rate,leaf area index and aboveground biomass increased,while SPAD at N1was obviously lower,and there was little difference between N2 and N3.The estimation models of leaf area index,aboveground biomass and SPAD in different growth periods were established.The results showed that RF method was more suitable for estimating LAI.The accuracy of RF model in the whole growth period was the highest.The validation results showed that R² was 0.82,RMSE was 0.32,and RPD was 2.38.As for the aboveground biomass,the accuracy of aboveground biomass model was similar to that of leaf area index.RF model in the whole growth period had the best effect,but in filling stage,PLS model and RF model could not estimate the aboveground biomass effectively.For SPAD,the accuracy of RF model in jointing stage was the highest,followed by RF model in the whole growth stage.The validation results of RF model in the whole growth stage showed that R² was 0.83,RMSE was 2.23,and RPD was 2.40.RF models of leaf area index,aboveground biomass and SPAD in the whole growth period had great estimation accuracy.In practical application,RF model in the whole growth period could be preferred for convenience.（4）As for photosynthetic parameters,the changes of photosynthetic parameters in different treatment zones showed that topdressing at jointing stage could promote plant growth and delay plant senescence.The correlation analysis between photosynthetic parameters and vegetation index showed that GNDVI could better reflect the low photosynthetic intensity.PLS and RF could be used to establish relatively great estimation models of photosynthetic parameters in jointing stage,tasseling stage and filling stage,but for the whole growth period,both methods could not estimate photosynthetic parameters effectively.Generally speaking,the empirical statistical models of photosynthetic parameters（PLS model and RF model）based on the vegetation indexes had relatively poor effects.In contrast,PLS model was superior to RF model.For net photosynthetic rate,PLS model in tasseling stage was the best,followed by PLS model in jointing stage.For transpiration rate,the accuracy of RF model in filling stage was the highest;For stomatal conductance,the best model is the PLS model in tasseling stage,which has a good prediction ability.Both the modeling set and the validation set R² of the model were 0.7.For the intercellular carbon dioxide concentration,PLS model in the tasseling stage was the best.In jointing stage and filling stage,PLS and RF could not estimate the intercellular carbon dioxide concentration.（5）For the comprehensive growth indicators,the change trends of CGI1 and CGI2were consistent with the measured growth indicators,therefore,CGI1 and CGI2 could reflect the crop growth.Compared with CGI1,CGI2 had little change in tasseling stage,but decreased in jointing stage and filling stage.CGI2 can show the difference in different growth stages more obviously.By analyzing the correlation between comprehensive growth index and vegetation index,we could get that NDVI and OSAVI had high correlation coefficients with comprehensive growth index in each growth period.The correlation coefficient between CGI1 and vegetation index was slightly larger than that between CGI2 and vegetation index.Compared with the single growth indicator,the correlation between comprehensive growth index and vegetation index was stronger.Comparing the estimation models of comprehensive growth indicators in different growth stages,it was found that the accuracy of CGI2 model was lower than that of CGI1.For CGI1,RF model in jointing stage had the highest accuracy,followed by RF model in the whole growth stage.The accuracy of RF model in the whole growth stage was slightly lower than RF model in jointing stage.For CGI2,RF model in tasseling stage was the best,the validation results of which showed that R² was 0.68,RMSE was 0.11,and RPD was1.76.Compared with the single growth indicator,the model accuracy of CGI1 and CGI2was between phenotypic parameters（leaf area index,aboveground biomass,SPAD）and photosynthetic parameters.