 Our paper presents a novel approach that uses Bayesian deep learning to model noise inherent in historical maps for semantic segmentation of hydrological features incorporating multi-scale contextual information through atro-spatial pyramid pooling, ASPP. The algorithm yields predictions with an average dice coefficient of 0.827 and outputs intuitively interpretable pixel-wise uncertainty maps that capture uncertainty in object boundaries, noise from drawing, aging, and scanning, as well as out-of-distribution designs. These uncertainty maps can potentially be used to refine segmentation results, locate rare designs, and select reliable features for future GIS analyses. This article was authored by CDWU, Magnus Heitzler, and Lorenz Herney.