The objective of the present paper is to identify healthy wheat and disease wheat by using hyeprspectral remote sensing as soon as possible. The canopy spectral reflectance of winter wheat infected by different severity yellow rust was measured and the disease indices (DI) were investigated in the field respectively. Smoothing the canopy spectra and calculating the first derivative values, the two methods were used to calculate the red edge position (REP) and yellow edge position (YEP) of the first derivative values: (a) maximum of the first derivative value; (b) Cho and Skidmore method. The result showed that REP gradually shifted to short-wave band, and the YEP gradually shifted to long-wave band with disease severity increasing, however, REP-YEP quickly became smaller. Analyzing and comparing the prediction precision of REP, YEP and REP-YEP for DI, the result indicated that the model REP-YEP as variable has the best estimation precision for DI than REP and YEP, the model estimation error is 6.22, and relative error is 14.3%, and it could identify healthy and disease wheat 12 days before the disease symptom apparently appeared. Therefore, this study not only can provide theory and technology for large areas monitoring of wheat disease by using hyperspectral remote sensing in the future, but also has the important meaning and practical application value for implementing precision agriculture.