Scaling effects of chlorophyll content in walnut leaves estimations with coupling Boruta algorithm and machine learning model

Accurate and efficient acquisition of vegetation chlorophyll content (CLC) is of great significance for precision agriculture. In this study, multi-spectral unmanned aerial vehicle (UAV) images at three flight altitudes were used to explore the impact of scale effects on the accuracy of CLC estimation. Using multi-spectral images of UAV at different flight heights, 17 spectral indices were calculated. Boruta algorithm was used to select the features of the 17 spectral indices. The screened spectral indices were taken as input variables, and BP neural network regression algorithm was used to construct the CLC estimation model. Coefficient of determination (R2), root mean square error (RMSE), normalized root mean square error (NRMSE) and interquartile distance (RPIQ) were used as evaluation indexes. The results showed that: (1) the correlation between the spectral index obtained at different flight altitudes and CLC was significantly different. The correlation between the spectral index obtained at 120m flight altitude and CLC was the highest, followed by 80m flight altitude and 100m flight altitude. (2) Different flight altitudes have a great impact on the estimation accuracy of CLC. The estimation accuracy of 100m is the worst, followed by 80m flight altitude, and 120m flight altitude has the best estimation accuracy. R
²
, RMSE and NRMSE in the modeling set are 0.69, 2.67($\mu$g/cm
²
) and 5.47%, respectively. In the verification set, the values were 0.54, 3.52($\mu$g/cm
²
), 7.21%, and RPIQ was 2.21. The above results show that the regression model coupled with Boruta and BP neural network can be used as the best model to estimate CLC at 120m flight altitude. This study can provide important reference for flight height selection when using UAV images to estimate growth parameters of economic tree species.