Insufficiency of phosphorus could greatly effect rice production, thus it is significant to adopt quick and nondestructive diagnosis of phosphorus content. The present paper focused on first expanded leaves with different phosphorus fertilization levels, comprehensively extracted 26 features’ spectral information such as color, texture and shape etc. Single feature index analysis was conducted. Then features were collected to integrate CfsSubsetEval+Scattersearch method for optimizing, evaluation and choosing. Based on the feature selection for different leave positions, leaves in different phosphorus fertilization levels were finally classified into three grades (extremly insufficient, significant insufficient and normal) according to rough set theory. Results showed that the accuracy of recognition was very high while few phosphorus contained in the leaves. Moreover, the third expanded leaf is the best part for phosphorus-nutrient diagnosis.