Estimation of winter wheat chlorophyll content by combing canopy spectrum red edge parameters with random forest machine learning

The sudden increase in vegetation canopy reflectance from low reflectance in red band to near infrared band forms the red-edge spectral characteristics, which is unique to healthy vegetation. Many parameters that can describe this characteristic have been designed and developed as important indicators of crop growth and nutrition status. However, few studies have systematically compared and evaluated the applicability of these red-edge parameters to estimate winter wheat leaf CHL values at different growth stages and nitrogen application levels. In this study, canopy spectral reflectance and leaf CHL of winter wheat at 4 key growth stages (jointing stage, heading stage, flowering stage and filling stage) and 3 nitrogen application levels were obtained through a 4-year field experiment. The sensitivity of 47 spectral red-edge parameters to CHL was evaluated, and the relative importance of spectral red-edge parameters was used to optimize the random forest machine learning model to estimate winter wheat CHL. The results showed that the sensitivity of spectral red edge parameter to CHL was affected by the growth period and nitrogen application level of winter wheat, and the correlation R² between the best red edge parameter and CHL in a single growth period was between 0.39 and 0.89. The best red-edge parameter in the whole growth period was NDDRmid, and the coefficient of determination between it and CHL was 0.76. The sensitivity was the highest in filling stage, and the coefficient of determination between the best red edge parameter RVI5 and CHL was 0.89 and the R² between red edge parameters REPRpi, NDDRmid, RVI2, RVI4, RVI5, RVI6, NDRE, RVI12 and RVI13 with CHL were all higher than 0.80. Nitrogen application level increased the sensitivity of red-edge parameter to CHL. At single nitrogen application level, the coefficient of determination between best red edge parameter and CHL was between 0.75 and 0.81. At N1 and N2 conditions the best red-edge parameter is NDDRmid, and at N3 condition the best red-edge parameter is RIDRfd (R²=0.81). The sensitivity of the red edge parameters NDDRmid, RVI5, RVI12 and DIDA to CHL at different growth stages and nitrogen application levels were all in the best 10 red-edge parameters. In the four single growth stages, the best accuracy of the random forest model was achieved, when 10 to 30 relative importance red-edge parameters were used as inputs, with R² between 0.40 and 0.89 and RMSE between 3.07 and 5.29. At four individual growth stage, the best accuracy of the model was achieved at jointing stage when 10 red edge parameters were used. At the three nitrogen application levels, the model performed best when 10 to 30 relatively important red-edge parameters were used, with R² between 0.78 and 0.87 and RMSE between 2.79 and 4.47. With the increase of nitrogen application level, the model accuracy was enhanced. At N3 condition, the model performed best when 30 relatively important red edge parameters were used with R² = 0.87 and RMSE=2.79. Step by step selection of relative importance red edge parameter features as input to optimize the random forest machine learning model improved the estimation accuracy of CHL. The best estimation accuracy in the whole growth period was R² =0.80 and RMSE=4.25. With the increase of nitrogen application level, the estimation accuracy of the model was improved, R² =0.87 and RMSE=2.79 at the condition of N3. At different growth stages and nitrogen application levels, the red-edge parameters DIDA and RVI13 were used as important features to construct the best model. The results revealed the potential of spectral red edge parameter in estimating CHL of winter wheat under different growth stages and nitrogen application conditions, and also provided a reference for the detection of chlorophyll content of other crops based on the characteristics of red edge parameter.