Modelling and fitting the reflectance anisotropy of land surfaces is one of the most important issues in remote sensing studies. In the traditional linear kernel-driven model, the most widely used kernel functions are derived from radiative transfer model of vegetation canopy. Therefore, it is not validate to represent the forward scattering effect of snow/ice surfaces. We proposed a method by adding a forward kernel function to the traditional linear kernel-driven model, and validate it with in situ measured bidirectional reflectance factor (BRF) data. The validation results show that this method is efficient for fitting the BRF of snow/ice surfaces (R2=0.997 5, RMSE=0.022 6). We also compared it with empirical functions and the traditional linear kernel-driven model. The results show that: (1) The fitting results of linear kernel-driven model are better than those of empirical functions; (2) The fitting results can be significantly improved by adding the forward kernel function; (3) The fitting results of the improved linear kernel-driven model are stable at different wavelengths.