Deep Learning-Based Spatial Detection of Drainage Structures using Advanced Object Detection Methods

Hydrologic connectivity plays a critical role in understanding and managing environmental processes. The spatial characterization of hydrologic connectivity often relies on hydro-topographic delineation using Geographic Information Systems (GIS) and digital terrain models (DEMs). Recent advancements in LiDAR technology have provided high-resolution DEMs that accurately represent topographic conditions. However, accurately delineating hydrologic connectivity using LiDAR DEMs faces challenges, particularly in the presence of virtual flow barriers such as roads and bridges which impede water flow and act as “digital dams.” This paper addresses the need for an efficient and effective approach to detect the locations of drainage structures, such as roads and bridges, which significantly impact hydrologic connectivity. While previous studies have shown that incorporating drainage structures improves the delineation of drainage flows, the availability of consistent and high-quality drainage structure datasets remains limited. Therefore, this study aims to develop a methodology that utilizes deep learning (DL) frame-works to detect drainage structures by leveraging their unique topographic patterns on LiDAR DEMs and supplemental GIS datasets. The paper explores multiple advanced deep learning-based object detection models, including Faster RCNN, DINO, DETR:DINO and YOLOv5, to analyze the distinctive patterns exhibited by drainage structures. These models are trained to spatially detect the locations of drainage structures by recognizing the specific “signatures” present in their topographic patterns. The investigation of these state-of-the-art DL frameworks for drainage structure detection represents a novel approach that extends the current understanding of utilizing DL techniques in the field of hydrologic connectivity analysis. We performed both quantitative and qualitative analyses, and propose a novel evaluation framework to demonstrate that DINO:DETR and Faster-RCNN methods are both capable of correctly identifying culvert locations and outperform DETR and YOLOv5 methods.