Computational study demonstrates reliable defect segmentation and calibrated uncertainty quantification in underwater bridges, highlighting specific environmental interference mechanisms.
Key Points
To develop an uncertainty-aware Bayesian segmentation framework integrating uncertainty quantification and defect attribution analysis for reliable underwater bridge inspection under adverse environmental conditions.
Designed a Bayesian segmentation architecture to quantify both epistemic (model-related) and aleatoric (data-noise) uncertainties under underwater interference.
Developed an attribution-based interpretation strategy to trace uncertainty sources back to specific physical environmental factors.
Evaluated segmentation accuracy and uncertainty calibration using mixed underwater inspection datasets.
The framework achieved high segmentation mean Intersection over Union (MIoU) alongside well-calibrated uncertainty estimations under environmental interference.
Attribution analysis identified that water turbidity and biological attachments mainly degrade exposed rebar detection, whereas rough concrete textures predominantly disrupt concrete spalling identification.