Nitrogen Nutrition Estimation of Maize Based on UAV Spectrum and Texture Information

Crop nitrogen nutrition status is a key indicator to characterize the green degree and health status of maize canopy. In order to compare the accuracy of single spectral index model and texture information fusion model in maize nitrogen nutrition estimation model, this investigated the accuracy and reliability of maize nitrogen nutrition estimation model based on UAV multispectral information and texture information fusion.

Matrice-300 RTK multi-rotor aircraft equipped with MS600 Pro multi-spectral sensor was used to obtain multi-spectral images of maize tasseling-silking stages under six nitrogen levels in two years. By extracting vegetation index and texture features, the correlation between vegetation index, single texture feature, combined texture index and fusion information of vegetation index and texture index, was comprehensively analyzed. The vegetation index, normalized difference texture index (NDTI) and their combined parameters with the largest amount of information were selected. Four nitrogen nutrition parameters of maize leaf nitrogen content (LNC), plant nitrogen content (PNC), leaf nitrogen accumulation (LNA), and plant nitrogen accumulation (PNA) were compared and estimated by multiple stepwise regression (MSR), random forest (RF), support vector machine (SVM), and grey wolf optimized convolutional neural network (GWO-CNN).

(1) There were differences in the original spectral reflectance of maize under different nitrogen treatments, and the differences in the red band R (660 nm), blue band B (450 nm) and near-infrared band NIR (840 nm) were significant. (2) The vegetation indices (EVI, GARI, REOSAVI, SIPI, and MCARI), single texture features (var₄₅₀, var₆₆₀, mean₈₄₀, dis₇₂₀, and hom₈₄₀) and combined texture index NDTI extracted from UAV multispectral images could be used for LNC, PNC, LNA and PNA estimation of maize in VT-R1 stage. The GWO-CNN model based on vegetation index had better estimation effect on LNC, PNC, LNA and PNA than single texture feature and texture index model, and its R² were 0.831, 0.761, 0.826 and 0.770, respectively. (3) The accuracy of GWO-CNN model with vegetation index and texture index for LNC, PNC, LNA and PNA estimation was significantly higher than that of vegetation index and texture index, and its R² was 0.921, 0.901, 0.917 and 0.892, respectively, which was 9.77%, 15.54%, 9.92% and 13.68% higher than that of single spectral information optimal estimation model.

Fusion of multi-spectral vegetation index and texture index could effectively improve the estimation accuracy of maize nitrogen nutrition, and better evaluate the distribution of maize nitrogen distribution, which provided new ideas for precise maize nitrogen fertilizer management based on UAV platform at field scale.