Predicting Section Loss and Corrosion Severity in Steel Using Ensemble Learning on Visual Imagery Data

ABSTRACT The integration of emerging technologies such as unmanned aerial vehicles (UAVs), computer vision, and machine learning is transforming infrastructure inspections, offering a means to conduct remote, safe, and efficient condition assessments. Computer vision techniques have been developed to segment corrosion information in images; however, they are not yet able to accurately and robustly quantify section loss, a critical aspect of structural steel inspections. This study introduces a novel method to predict section loss in corroded steel based on visual image features and ensemble learning methods, providing a foundation for quantitative corrosion assessment during structural inspections. Steel coupons were corroded using an accelerated laboratory cyclic corrosion test, with section loss recorded and images captured at multiple intervals in controlled, semi-controlled, and uncontrolled light environments. Various image features, including color and texture descriptors, were quantified, and the data were balanced through Synthetic Minority Over-sampling Technique-Regression (SMOTE-R) augmentation. Tree-based ensemble algorithms were initially trained to predict section loss from image features. Feature importance analysis identified the most relevant predictors, and models were reevaluated with reduced feature sets. Results show that Random Forest consistently outperforms other approaches, achieving an average normalized root mean square error of 0.119 through fivefold cross validation. These findings demonstrate that quantitative values of section loss can be reliably estimated from image data alone, enabling a practical, image-based methodology for assessing corrosion severity. With further validation, this approach could be integrated into UAV-aided infrastructure inspections, improving the accuracy, efficiency, and objectivity of corrosion assessments and supporting data-driven structural retrofit and maintenance decisions.