Coefficient of variation method combined with XGboost ensemble model for wheat growth monitoring

Obtaining wheat growth information accurately and efficiently is the key to estimating yields and guiding agricultural development.

This paper takes the precision agriculture demonstration area of Jiaozuo Academy of Agriculture and Forestry in Henan Province as the research area to obtain data on wheat biomass, nitrogen content, chlorophyll content, and leaf area index. By using the coefficient of variation method, a Comprehensive Growth Monitoring Indicator (CGMI) was constructed to perform fractional derivative processing on drone spectral data, and correlation analysis was performed on the fractional derivative spectra with a single indicator and CGMI, respectively. Then, grey correlation analysis was carried out on differential spectral bands with high correlation, the grey correlation coefficients between differential spectral bands were calculated, and spectral bands with high correlation were screened and taken as input variables for the model. Next, ridge regression, random forest, and XGboost models were used to establish a wheat CGMI inversion model, and the coefficient of determination (R²) and root mean squared error (RMSE) were adopted for accuracy evaluation to optimize the wheat optimal growth inversion model.

The results of the study show that: using the data of wheat biomass, nitrogen content, chlorophyll content and leaf area index to construct the comprehensive growth monitoring indicators, the correlation between the wheat growth monitoring indicators and the spectra was calculated, and the results showed that the correlation between the comprehensive growth monitoring indicators and the single indicator correlation had different degrees of increase, and the growth rate could reach 82.22%. The correlation coefficient between the comprehensive growth monitoring indexes and the differential spectra reached 0.92 at the flowering stage, and compared with the correlation coefficient with the original spectra at the same period, the correlation coefficients increased to different degrees, which indicated that the differential processing of spectral data could effectively enhance the spectral correlation. The three models of Random Forest, Ridge Regression and XGBoost were used to construct the wheat growth inversion model with the best effect at the flowering stage, and the XGBoost model had the highest inversion accuracy when comparing in the same period, with the training and test sets reaching 0.904 and 0.870, and the RMSEs were 0.050 and 0.079, so that the XGBoost model can be used as an effective method of monitoring the growth of wheat. To sum up, this study demonstrates that the combination of constructing comprehensive growth monitoring indicators and differential processing spectra can effectively improve the accuracy of wheat growth monitoring, bringing new methods for precision agriculture management.