Angular effect of algorithms for monitoring leaf nitrogen concentration of wheat using multi-angle remote sensing data

Abstract

Leaf nitrogen concentration (LNC) is a key indicator of crops’ growth status, and the timely and accurate large-scale area monitoring of LNC is crucial for guiding field management. Here, we collected LNC data and hyperspectral data at the canopy scale, from 13 view zenith angles (VZAs) and two locations (Zhengzhou and Shangshui, in China) across 2 years (2012–2014). Four data processing methods—vegetation indices (VIs), back-propagation neural network (BPNN), eXtreme gradient boost (XGBoost), and partial least squares regression (PLSR)—were applied and compared for their ability to estimate LNC under 13 VZAs. These results revealed consistent trends in the performance of the four algorithms at 13 VZAs: the nadir direction is better than the extreme angle, and the best performance is obtained for the range of –30° to 0° (R2 ≥ 0.83; RMSE ≤ 0.41%). The number of bands is a critical factor affecting the accuracy of LNC monitoring: including red edge bands can alleviate angular effect to some extent. The accuracy of PLSR for monitoring LNC is not only superior near the nadir direction (R2 = 0.91), but also better than that provided VIs (16%–17%), BPNN (15%–16%), or XGBoost (29%–58%) at ± 60°. Therefore, it is strongly recommended the PLSR algorithm be used to process multi-angle remote sensing data to estimate the LNC of field crops. This work provides a timely reference basis for selecting the appropriate flight angle and the number and positioning of bands for unmanned aerial vehicle (UAV) and satellite applications in the future.