It's pointed out by this paper that the priori knowledge of random rough soil surface is needed in soil moisture inversion by using passive remote sensing, so the precision of priori knowledge has effect on the inversion result. Several kinds of inversion error are defined in this paper to depict the influence of three kinds of priori knowledge on the inversion, these three kinds of priori knowledge are the type of soil surface height distribution, soil temperature, and soil texture. Simulation result shows that it's feasible to invert soil moisture by neural network (NN) based on bi-spectrum model (BSM) . Using two kinds of emissivity data of two polarizations as the input of NN, the inversion error of soil moisture is allowable even there is some uncertainty on these three kinds of priori knowledge.