Accurate water flow modeling from DEM requires accounting for man-made features like culverts and associated influential structures, which are often unrepresented within digital elevation models (DEMs). Additional hydro-enforcement treatments applied to DEMs are crucial for accurate modeling and informed decision-making in the water infrastructure management domain. The challenge, however, of detecting and accurately incorporating into the model, features like culverts, can be particularly daunting when dealing with large geographic areas and complex topography.

This presentation provides an approach with an updated methodology for culvert detection that addresses this challenge as it is applied to LiDAR-based elevation-derived hydrography (EDH). Building on early methodological developments and drawing on existing data from almost 10 years ago, the approach is refined with some novel open-source data processing elements. The success of these image detection methods for culverts is reliant on multidirectional Sky View Factor (SVF) imagery created from the DEM as analyzed through the Relief Visualization Toolbox (RVT) Python library. The SVF images are subdivided and tessellated into small image subsets to maintain maximum feature fidelity or resolution during a subsequent custom trained deep-learning approach. This tiling allows the trained model to rapidly predict small features such as culverts over vast geographic areas while maintaining the input high-resolution detail from the initial DEM input SVF data. The versatility and potential of SVF imagery as a predictor of such infrastructure is also demonstrated in detection of related man-made features, including berms and cattleguards.

The transferability of the model trained with culvert data from a 3000 square mile Santa Fe County, New Mexico 3DEP LiDAR dataset is also demonstrated through application over a geographically distinct large 40 billion raster cell elevation dataset. Details are presented to show the performance and computational demands of these predictions on a variety of computing platforms and operating systems. While this culvert feature prediction technique certainly highlights some achievements in the detection, it also uncovers some potential drawbacks including overprediction which require additional geospatial data management techniques. Employed techniques include spatial proximity filtering and data-driven culvert endpoint refinements for determining a “best-use” culvert cutline suitable for ultimate DEM enforcement in EDH applications.

Notwithstanding potential geospatial data management challenges, the approach's ability to rapidly detect culverts, independent of any potential drawbacks, serves as a powerful accelerant to the subsequent flow modeling stage, enabling timely access to the necessary ingredients of hydro-enforcement. Accurate 3D culvert feature geometry cutlines then inform DEM hydro-enforcement. Next, iterative flow modeling workflow procedures in GRASS hydrological toolsets can be deployed. Here again, the computational demands of hydrological processes can be challenging for large area data processing, especially as related to computational memory requirements. The scalability limits of these core GRASS processes were tested with results presented comparing computational benchmarks between 10 billion raster cell elevation versus 40 billion raster cell elevation datasets on various hardware platforms.

The presented culvert detection workflow in LiDAR-based elevation-derived hydrography has significant implications for water infrastructure management. As high-resolution datasets and advanced sensor technologies continue to expand, it will be imperative to deliver faster, more efficient, and effective means to process and extract information from data. The highlighted approach facilitates more rapid detection of elements that impact flow modeling through elevation-derived hydrography processing. Combining cutting-edge technology with growing societal needs for informed data models, these infrastructure detection and modeling approaches do enhance the ability of geospatial professionals to deliver timely and accurate information for resiliency planning in support of the water infrastructure management community.