The aim of this research is to realize the rapid measurement of soil available P content. The suitable proportion of available P could promote the crops grow. Taking ‘Lou’ soil as sample, the soil diffusion reflectance spectrum in 900~1700nm under different observation heights were collected by using the portable spectrographs. Firstly, five observation heights (5, 7, 10, 12, 15cm) were compared, and 10cm was considered to be the best. The abnormal samples were identified and eliminated by using 3 times standard deviation and principal component analysis method. That effectively improved the model precision. Then, the effect of four different wavelengths selecting methods (SPA, CARS, sCARS, RF) on modeling was analyzed. The result showed that sCARS was the best. Finally,the different nonlinear modeling methods (RBF neural network, WNN, LSSVM) were experimented. The results proved that LSSVM had the best result. When the observation height was 10cm, the modeling prediction correlation coefficient was 0.8581, and the prediction root mean square error was 10.8801. The results showed a high accuracy and feasibility of soil available P content prediction.