Deep learning-enhanced detection of road culverts in high-resolution digital elevation models: Improving stream network accuracy in Sweden

Sweden, a mostly forested country with many small forest roads obstructing topographical modelling of shallow groundwater and streams. Maps have traditionally been constructed from aerial photos, but dense forest canopies often obscure these streams from view. Topographical modelling is a widely adopted method for mapping small streams and has been proposed as a potential solution. However, road embankments can disrupt flow paths, acting like dams in the digital landscape. This study presents a novel method where a unique dataset of 28,512 culverts was used to develop a deep learning method to map road culverts, enabling the correction of digital elevation models and enhancing the accuracy of topographically derived stream networks. The deep learning model successfully mapped 87 % of all culverts in the test data, and integrating these predicted culverts into topographical models slightly improved the precision of stream networks extracted from high-resolution digital elevation models in the region. These findings demonstrate the possibility for UNet models to enhance hydrological modelling and stream network mapping but further research should focus on reducing the number of false positive culvert predictions. • Deep learning maps 87 % of road culverts using topographical and orthophoto data. • Residual attention UNets are better at detecting road culverts than standard UNets. • Further research is required to reduce the number of false positive culvert detections. • Topographically derived streams outperform current maps in Sweden.