Integrating GIS, field surveys, and machine learning models for enhanced culvert inventory management in Oklahoma.

Culverts are critical in transportation infrastructure and provide hydraulic conveyance beneath roadways. However, factors such as culvert age, urbanization, and climate change are believed to contribute to growing vulnerability. As culverts deteriorate over time, it is expected that their structural integrity and hydraulic capacity diminish, leading to potential road-stream crossing failures, increased maintenance cost, and adverse environmental impacts. To address these challenges, this study integrates Geographic Information Systems (GIS) analysis, field assessments, and machine learning (ML) models to improve culvert inventory management, assess culvert conditions, and develop data-driven methods for predicting culvert deterioration. The National Bridge Inventory (NBI) database is the primary data source for existing culverts in the United States. GIS analysis was conducted to identify vulnerable culverts for field visit. Field surveys were conducted at 100 vulnerable road-stream crossings with culvert conditions ranging from fair to failed. These site inspections involved verifying culvert characteristics, assessing culvert condition, and evaluating the feasibility of retrofitting or replacement. A standardized data collection protocol was implemented using a Microsoft form that incorporated elements from the Southeast Aquatic Resource Partnership (SARP), the Specifications for National Bridge Inventory (SNBI), and the Pavement Surface Evaluation and Rating (PASER) system. In addition to GIS analysis and field assessments, ML models were developed to analyze and predict culvert deterioration. The ML analysis was conducted on 3,191 culvert entries to identify culvert features that could affect culvert conditions. The dataset was split into 80% training and 20% testing. Sixteen ML models were evaluated across four data categories, including refined data, age-adjusted data, filtered data, and complete data. Gradient Boosting classifier achieve the highest accuracy of 82.66% on refined data and 82.48% on age-adjusted data, Extreme Gradient Boost 82.46% on filtered data, and CatBoost classifier 81.78% on complete data. This study proposes a comprehensive methodology for culvert inventory and inspection prioritization using the state of Oklahoma as a case study. The findings from this study informed the design of an Excel-based culvert prioritization tool. The tool can help agencies identify culverts in need of inspection, retrofitting, or replacement. It is anticipated that the findings from this work will enhance culvert asset management, improve culvert evaluation methods, and better inform decision-making for infrastructure resilience. The integration of GIS analysis, field surveys, and ML models provides a scalable framework that can be adapted in other regions.