GeoAI-based drainage crossing detection for elevation-derived hydrographic mapping

Abstract

The increasing availability of High-Resolution Digital Elevation Models (HRDEMs) allows accurate delineation of stream and drainage flowlines at the field scale. However, the presence of digital flow barriers like roads effectively impedes hydrological connectivity represented on the HRDEMs. Conventional methods for locating these artificial barriers such as on-screen digitization and field surveying are cost prohibitive over large geographic areas. Thus, a database of drainage crossings under roads is a crucial input for refining flowlines derived from HRDEMs. In this study, we developed advanced deep learning models for detecting the locations of drainage crossing structures in agricultural areas. Our method assesses the performance of a two-stage object detector, Faster R-CNN and a single-stage object detector, YOLOv5. The models were trained using random HRDEM tiles and ground truth labels developed for the West Fork Big Blue Watershed, Nebraska. The Faster R-CNN and YOLOv5 achieved an average F1-score of 0.78. The best-fit models in Nebraska were then transferred to three other watersheds in Illinois, North Dakota, and California. These findings show effective spatial detection of these drainage crossing features, attributed to their distinct topographic patterns. Such spatial object detection approaches offer a promising avenue for automated integration of drainage crossings into HRDEMs with minimal manual interventions, thereby enhancing the delineation of elevation-derived hydrographic features for regional applications.