Research On Remote Sensing Monitoring Of Wheat Leaf Rust Based On Ground Hyperspectral And UAV Images

Wheat is one of the main refined grain crops in China,which plays an important role in national grain provision and economic development.The frequent occurrence of wheat leaf rust caused huge grain yield losses and production cost increases,simultaneously increased the risk of poisonous and harmful materials residue in agricultural products.Traditional monitoring methods were time-consuming,laborious,subjective and lagging,while remote sensing monitoring technology could achieve rapid,real-time and non-destructive monitoring,so as to prevent the occurrence of leaf rust.In this study,wheat affected by leaf rust was taken as the research object.The near-ground hyperspectral and UAV image technology wes used to monitor wheat leaf rust respecti vely,and a prediction model of wheat leaf rust was established based on multi-source remote sensing data.The main contents of this research were as follows:(1)The spectral reflectances of different filed plots were measured by analytical spectral devices(ASD)Field SpecPro FR spectroradiometer,and the canopy spectral response characteristics of wheat leaf rust with different disease indexes(DI)were compared by first-order differential data transformation.The results showed that in the blue,green,and red bands(530～730 nm),the reflectivity increases significantly with the increase of DI;in the near-infrared band(750～900 nm),the reflectivity decreases gradually with the increase of DI.After the damage of leaf rust,the first order differential spectrum curve showed the characteristic of’double peak’.Moreover,with the increase of DI,the first-order differential value gradually increases in the range of the green edge(about 520 nm).The red margin decreased gradually in the range of about 730 nm,and the red margin showed an obvious ’blue shift’ phenomenon at both the early and middle grain filling stages.(2)Correlation analysis showed that the red band(615～705 nm)and near-infrared band(743～920 nm)could be used as the sensitive band at the early stage of grout filling.The red band(595～706 nm)and near infrared band(740～900 nm)can be used as the sensitive band in the middle grouting stage.The bands with high correlation were selected to construct 28 hyperspectral characteristic parameters,from which 23 hyperspectral characteristic parameters with extremely significant correlation were selected to construct a one-dimensional linear,stepwise regression and BPNN prediction model,and the accuracy was verified.The research showed that the traditional linear prediction model based on Rg/Rr and NPCI was the best,and the model based on the stepwise regression method was better than the linear estimation model.However,the prediction model based on BPNN had higher accuracy than the prediction model based on unary linear and stepwise regression.And the inversion effect of the early filling stage was better than that of the middle filling stage.The R2 and RMSE of the modeling set were 0.9 and 4.7%,respectively,and the R2 and RMSE of the verification set were 0.82 and 4.6%,respectively.Therefore,the best time to monitor wheat leaf rust by using the BPNN algorithm in the near field was in the early stage of grain filling.(3)The RGB and near-infrared images of the wheat canopy were obtained by unmanned aerial vehicle.Based on the correlation between the original bands of red,green,blue and near-infrared and DI indexes,a one-dimensional linear,step-by-step regression and BPNN prediction model was established and the accuracy was verified.Meanwhile,the difference in accuracy was compared and analyzed.The results showed that there was a very significant correlation between the red band and DI,and the unitary linear prediction model had the worst fitting degree.Both R2 and RMSE of BPNN model were superior to stepwise regression model,and could obtain more reliable inversion values.(4)Correlation analysis was conducted between the 25 color feature indexes and DI,and the optimal estimation model of wheat leaf rust at different growth stages was screened out through unitary linear,stepwise regression and BPNN modeling method.The results of the study showed that there were 23 kinds of color characteristic indexes that reached a significant correlation with the DI of wheat leaf rust.In the early stage of wheat grain filling,the BPNN model had a better prediction effect than the linear and the stepwise regression model.In the middle stage of wheat grain filling,the overall prediction accuracy of stepwise regression model was higher.In addition,the inversion effect of early grouting was better than that of middle grouting.The R2 and RMSE of the modeling set were 0.91 and 4.58%,and the R2 and RMSE of the validation set were 0.74 and 5.58%,respectively.