Inversion of nitrogen and phosphorus contents in cotton leaves based on the Gaussian mixture model and differences in hyperspectral features of UAV

The nitrogen (N) and phosphorus (P) contents in cotton leaves can directly reflect growth conditions. Rapid and nondestructive acquisition of the N and P content in cotton leaves at the field scale is essential for rational fertilization strategies and precision agriculture. However, traditional direct destructive sampling in the field is performed at the sample point scale, which cannot rapidly obtain cotton leaf N and P content from the entire field. In this study, we propose that post-classification modeling based on differences in spectral features is beneficial for improving the prediction of N and P contents in cotton leaves. To test this hypothesis, we first used principal component analysis to downscale the hyperspectral data and then used Gaussian mixture modeling (GMM) to segment the hyperspectral data for spectral differences. The in-situ measured data was then combined with the random forest model to establish N and P prediction models for cotton leaves with spectral differences and full samples. Finally, the predictive model was utilized for leaf N and P spatial mapping of cotton in the field using UAV hyperspectral images as the input data. The results demonstrate that the spectral reflectance features of the different clusters classified by the GMM differ significantly in intensity and shape. The accuracy of the cotton leaf N and P prediction model based on the spectral differences was attributed to the full sample. The results validate the existence of spectral differences between crop leaf content by UAV hyperspectroscopy, and modeling based on spectral differences can improve the accuracy of predicting the spatial distribution of nitrogen and phosphorus in cotton leaves in the field.