A CLOSER LOOK AT SEGMENTATION UNCERTAINTY OF SCANNED HISTORICAL MAPS

Before modern earth observation techniques came into being, historical maps are almost the exclusive source to retrieve geo-spatial information on Earth. In recent years, the use of deep learning for historical map processing has gained popularity to replace tedious manual labor. However, neural networks, often referred to as "black boxes", usually generate predictions not well calibrated for indicating if the predictions are trustworthy. Considering the diversity in designs and the graphic defects of scanned historical maps, uncertainty estimates can benefit us in deciding when and how to trust the extracted information. In this paper, we compare the effectiveness of different uncertainty indicators for segmenting hydrological features from scanned historical maps. Those uncertainty indicators can be categorized into two major types, namely aleatoric uncertainty (uncertainty in the observations) and epistemic uncertainty (uncertainty in the model). Specifically, we compare their effectiveness in indicating erroneous predictions, detecting noisy and out-of-distribution designs, and refining segmentation in a two-stage architecture.