Stainless steel sheet pile walls are increasingly adopted in agricultural drainage channels as a corrosion-resistant alternative to conventional steel sheet piles. However, their high passive-film stability means that measurable thickness reduction does not occur within approximately ten years of installation, rendering conventional ultrasonic thickness measurement ineffective for early-stage condition assessment. This study proposes a deep learning-based automated detection system for pitting corrosion on stainless steel sheet pile surfaces using visible images acquired with a standard smartphone camera.

Two martensitic and ferritic stainless steel grades, SUS410 (Pitting Index: 11) and SUS430 (Pitting Index: 16), were sampled from an agricultural drainage channel in Niigata Prefecture, Japan, after five years of exposure in a brackish water environment with chloride ion concentrations of approximately 120 mg/L. Pixel-level annotation was performed on cropped regions of approximately 60 mm × 270 mm, and the U-Net semantic segmentation model was adopted as the detection model. Bayesian hyperparameter optimization using Optuna with the Tree-structured Parzen Estimator was applied across 100 independent trials to ensure robust performance evaluation. Data augmentation using vertical flip and blur operations was incorporated to improve generalization on the limited dataset.

The deep learning approach achieved F1-scores of 0.831 (SUS410) and 0.808 (SUS430), substantially outperforming the conventional binary thresholding baseline (F1-scores: 0.407 and 0.329, respectively). Data augmentation contributed improvements of approximately 1.2–2.8 percentage points. The results also confirmed the superior pitting resistance of SUS430, which exhibited markedly lower pit density and area ratio relative to SUS410. The proposed method enables non-destructive, quantitative assessment of early-stage pitting corrosion using readily available imaging equipment, offering a practical and cost-effective tool for infrastructure maintenance and long-term durability evaluation of agricultural water management facilities.