Deep learning approaches for delineating wetlands on historical topographic maps

Abstract Historical topographic maps are an important source of a visual record of the landscape, showing geographical elements such as terrain, elevation, rivers and water bodies, roads, buildings, and land use and land cover (LULC). Historical maps are scanned to their digital representation, a raster image. To quantify different classes of LULC, it is necessary to transform scanned maps to their vector equivalent. Traditionally, this has been done either manually, or by using (semi)automatic methods of clustering/segmentation. With the advent of deep neural networks, new horizons opened for more effective and accurate processing. This article attempts to use different deep‐learning approaches to detect and segment wetlands on historical Topographic Maps 1: 10000 (TM10), created during the 50s and 60s. Due to the specific symbology of wetlands, their processing can be challenging. It deals with two distinct approaches in the deep learning world, semantic segmentation and object detection, represented by the U‐Net and Single‐Shot Detector (SSD) neural networks, respectively. The suitability, speed, and accuracy of the two approaches in neural networks are analyzed. The results are satisfactory, with the U‐Net F1 score reaching 75.7% and the SSD object detection approach presenting an unconventional alternative.