The performance of many artificial intelligence systems critically depends on classification methods. Unfortunately, most classifiers are highly sensitive to distortions in the input distribution, for example, due to sensor failures or a switch of the sensor system. In such a case, new labeled data has to be recorded to retrain the classifier under the distorted input distribution, which can be expensive. Transfer learning addresses this issue by reusing the original model and adapting it to the new domain using as few new data points as possible. In this contribution, we present a novel supervised transfer learning scheme for the Large Margin Nearest Neighbor classifier. We evaluate our approach on a real-world example of a transfer between two hyper-spectral sensors applied to chemical molecule classification. Our results show that transfer learning outperforms the original model as well as a newly trained model on the new data and requires only very little data to be trained.