Combining UAV multispectral imagery and ecological factors to estimate leaf nitrogen and grain protein content of wheat

Abstract Nitrogen is an essential element of wheat growth and grain quality. Leaf nitrogen content (LNC), a critical monitoring indicator of crop nitrogen status, plays a reference role for later estimations of grain protein content (GPC). Developments in unmanned aerial vehicle (UAV) platforms and multispectral sensors have provided new approaches for LNC monitoring and GPC estimation, with great convenience for assessing the nutritional status of plants and grains without traditional destructive sampling. The objective of this study was to evaluate the feasibility of wheat LNC monitoring and GPC estimation based on UAV multispectral imagery. Wheat experiments were carried out in Xinghua, Kunshan and Suining of Jiangsu Province during 2018−2019 and in Rugao of Jiangsu Province during 2020−2021 with different varieties and nitrogen application rates. Remote sensing images were obtained by a multi-rotor UAV carrying a multispectral camera. The destructive sampling method was used to collect LNC, GPC and other field data. Wheat LNC monitoring and GPC estimation models were established after selection of the optimal indicators. Different modelling methods were used for the comparative analysis, including unitary linear regression, multiple linear regression and artificial neural network (ANN) methods. Three techniques were adopted to improve the GPC prediction accuracy: (1) multiple factors were substituted for single factor for the prediction; (2) texture information was added through further imagery mining; and (3) ecological factors were considered to improve the prediction mechanism. The results showed that the use of UAV-based Airphen multispectral imagery had a good effect on wheat LNC monitoring and GPC estimation. The vegetation indices constructed by red-edge and near-infrared bands had good performances in LNC monitoring and GPC estimation. The addition of texture information and ecological factors further improved the modelling accuracy. In this study, the optimal wheat GPC estimation model was established by NDVI (675, 730) at the jointing stage, NDVIT (730mea., 850) at the booting stage, NDVIT (730mea., 850) at the flowering stage and NDVI (730, 850) at the early filling stage. The modelling R2, validation R2 and relative root mean square error (RRMSE) reached 0.662, 0.7445 and 0.0635, respectively. The results provide a reference for crop LNC monitoring and GPC estimation based on UAV multispectral imagery.