Automatic extraction of road intersection points from USGS historical map series using deep convolutional neural networks

Road intersection data have been used across a range of geospatial analyses. However, many datasets dating from before the advent of GIS are only available as historical printed maps. To be analyzed by GIS software, they need to be scanned and transformed into a usable (vector-based) format. Because the number of scanned historical maps is voluminous, automated methods of digitization and transformation are needed. Frequently, these processes are based on computer vision algorithms. However, the key challenges to this are (1) the low conversion accuracy for low quality and visually complex maps, and (2) the selection of optimal parameters. In this paper, we used a region-based deep convolutional neural network-based framework (RCNN) for object detection, in order to automatically identify road intersections in historical maps of several cities in the United States of America. We found that the RCNN approach is more accurate than traditional computer vision algorithms for double-line cartographic representation of the roads, though its accuracy does not surpass all traditional methods used for single-line symbols. The results suggest that the number of errors in the outputs is sensitive to complexity and blurriness of the maps, and to the number of distinct red-green-blue (RGB) combinations within them.