Drainage ditch network extraction from lidar data using deep convolutional neural networks in a low relief landscape

Du, L ORCID: https://orcid.org/0000-0002-2628-9557, McCarty, GW ORCID: https://orcid.org/0000-0001-7064-7166, Li, X ORCID: https://orcid.org/0000-0003-4595-9170, Zhang, X ORCID: https://orcid.org/0000-0001-7844-593X, Rabenhorst, MC, Lang, MW, Zou, Z, Zhang, X and Hinson, AL ORCID: https://orcid.org/0000-0002-4231-4820 (2024) Drainage ditch network extraction from lidar data using deep convolutional neural networks in a low relief landscape. Journal of Hydrology, 628. 130591. ISSN 0022-1694

Preview

Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (14MB) | Preview

Abstract

Drainage networks composed of small, channelized ditches are very common in the eastern United States. These are human-made features commonly constructed for wetland drainage and constitute the headwater portion of permanent hydrographic networks. Accurate information on the drainage ditch location can help define where wetlands have been drained and evaluate impacts of artificial drainage patterns on hydrologic changes. Traditional water channel extraction approaches often cannot accurately identify small ditches especially in low-relief agricultural landscapes. In this study, we employed a state-of-the-art deep learning (DL) approach to extract drainage ditches using light detection and ranging (lidar) data in a low-relief agricultural landscape within the Delmarva area. First, we adopted a deep convolutional neural network based on U-Net architecture to classify ditches from different combinations of aerial optical and lidar derived (i.e., topographic and return intensity) features. The classification results were compared with a typical pixel-oriented machine learning classifier, random forest (RF). Next, we improved the connectivity of ditch networks through a minimum-cost approach and a further incorporation of FA to connect with natural drainage networks. Finally, we evaluated the connected drainage networks against flowlines derived from typical flow routing method (D8), an open-source channel network extraction tool (GeoNet), and the U.S. Geological Survey National Hydrography Dataset High Resolution data at 1:24,000 scale. Our results show that the DL model significantly outperformed the RF model, and the lidar derived topographic features were the most important input for ditch classification. The connected drainage networks extracted with DL exhibited pronouncedly higher precision (0.88) and recall (0.89) and a higher positional accuracy (within one pixel) than other flowline products. Overall, this study demonstrates the utility of DL approaches for automated extraction of ditch networks and the important contribution of lidar-derived topographic data for operational drainage network mapping at local and regional scales.