Aiming to differentiate 13 varieties of corn, present paper proposes an effective approach to solving the parameter drift problem of spectrum instruments. Remarkable drift has been found among the inter-day data when using the identical spectrum instrument to acquire sample data at different times, modeling with the intra-day data, and testing with the rest. The correct recognition rate is reduced to only 7.69% in the condition of severe drift. To tackle this problem, this paper proposes a supervised feature-based inter-day combination modeling approach, at first, the representative sample data acquired at multiple times will be selected to make up the modeling set, and then the PLS+LDA algorithm will be applied to extract the feature of varieties which is independent on instrument parameter drift, and finally BPR will be used to identify the varieties. The experiment results indicate that this approach is effective to rectify the data drift at different times, can bring higher recognition rate, and also shows its stability in practice.