A Leaf-Colour- and Shadow-Structure-Guided Method for Correcting Wheat Canopy Reflectance

Variation in illumination and canopy geometry can affect the comparability of RGB measurements acquired at different times of day. This study proposes a wheat canopy reflectance correction method guided by leaf-colour and shadow-structure features. UAV RGB images were acquired under different nitrogen rates, sampling dates and acquisition times. SP, HSV, Lab and BDI features were extracted to build four input schemes, which were fed into multiple neural networks for correction. The Fullgrade_noflat input, which retains the BDI grade order, combined with 1D-CNN gave the most accurate predictions of the plot-level daily mean RGB reference among the tested configurations (R² = 0.8773, RMSE = 0.0213 and MAE = 0.0161). The within-day CV decreased for all three RGB channels. The model also yielded lower errors than three RGB shadow-compensation implementations evaluated on the same dataset. These results indicate improved within-day consistency under the conditions studied; performance in other environments and effects on agronomic trait analysis require further evaluation.