Image feature-based machine learning for non-destructive corrosion assessment of weathering steel along a railway line in southwestern China

Rapid corrosion assessment of weathering bridge steels is important for infrastructure durability management. Here, Q345qE and Q345qENH steels were exposed to cyclic wet–dry accelerated corrosion tests simulating four atmospheric environments along a railway line in southwestern China for up to 48 months of equivalent exposure. An image-based corrosion data model was developed using standardized rust-layer photographs. A total of 105 image descriptors were extracted, and the descriptor subset was re-selected within every training fold to build an SVR-RBF model for corrosion-rate prediction, achieving R 2 = 0.728 under specimen-wise leave-one-out cross-validation. SHAP analysis revealed that color and texture features reflected rust-layer heterogeneity, with a color saturation entropy boundary at approximately 5.0 separating rust surfaces that differ significantly in measured corrosion rate and corrosion potential. Q345qENH formed a denser rust layer than Q345qE and reduced cumulative weight loss by up to 31.1%. This study provides a data-model approach for rapid, non-destructive durability evaluation of weathering steel bridges.