Modeling alpine grassland forage phosphorus based on hyperspectral remote sensing and a multi-factor machine learning algorithm in the east of Tibetan Plateau, China

The accurate and effective retrieval of forage phosphorus (P) content can provide significant information for the management of pastoral agriculture and grazing livestock. In this study, we constructed 39 models to estimate the forage P of alpine grassland in the east of Tibetan Plateau based on hyperspectral remote sensing and multiple factors (topography, soil, vegetation and meteorology) using a machine learning algorithm. The results show that (1) first derivative (FD) and continuum removal (CR) spectra can retrieve more feature bands that are mainly located in the near infrared (NIR) and shortwave infrared (SWIR) regions than log transformed (Log (1/R)) and original (OR) spectra for the forage P estimation; (2) in terms of the model precision, the combination of important bands (IBs) and important factors (longitude and monthly mean temperature) increase the accuracy of forage P estimation as compared with the models that used IBs alone; and (3) considering the precision, stability and simplicity of the model comprehensively, the FD-IBs + support vector machine (SVM) model is the optimum forage P inversion model, which presents coefficient of determination (R-2) and root mean squared error (RMSE) values of 0.67 and 0.0472%, respectively, and standard deviations (SDs) of 0.2386 and 0.0050%, respectively. This model can account for 88% of the variation of forage P in alpine grassland. This study demonstrates the importance of using a multi-factor modeling approach and spectral transformation techniques for estimating the forage P of grasslands and provides a scientific basis for the reasonable use and management of alpine grassland resources.