In this paper, we present experiments on POS tagging historical texts that contain spelling variation. The experiments are conducted in a low-resource scenario with a small amount of training data (here: 12,000 tokens). We investigate different ways of dealing with spelling variation in such a situation on different variants of historical German. Firstly, we add character n-grams as features to the tagger to enable it to learn spelling variation. Our tagging experiments show that this improves accuracy when there is enough variation in the data, but leads to a decrease in accuracy if the amount of variation is low. Secondly, we preprocess the data before training and applying the tagger, reducing spelling variation by normalization, rule-based simplification and substitution of spelling variants. All three methods improve tagging accuracy in comparable levels. Since normalization has the drawback of requiring additional resources, we recommend rule-based simplification and substitution of spelling variants for low-resourced settings. Finally, we evaluate the utility of additional unlabeled data to create word embeddings and employing external resources, which we use to further improve tagging accuracy.