Failure prediction and presence cues for budgeted review of Saudi road images

Abstract

Limited review capacity makes report selection consequential for automated pavement inspection. This study tested whether regional failure prediction improves the recovery of complete pothole-detector misses beyond image-presence controls. A compact Faster R-CNN detector was evaluated on 4,179 Saudi reports, including 2,802 source-positive reports. Three optimization seeds crossed with three protected folds produced nine correlated configurations. A regional predictor combined spatial features, detector summaries and no-box-first ordering. Residual risk was the proportion of source-positive reports with no matched target after review, assuming perfect correction against supplied annotations. At 10% review (417 reports), candidate risk averaged 24.63%, compared with 25.04% and 25.01% for center- and letterbox-presence controls. Candidate-minus-control differences were -0.412 and -0.385 percentage points, with adjusted 97.5% paired protected-group bootstrap intervals of [-0.849, +0.107] and [-0.864, +0.087], respectively. Superiority over both controls was not established. Fine-tuned presence had the lowest descriptive deployable-policy risk, 24.33%. At this budget, all candidate reviews contained no emitted box, leaving a mean 346.44 emitted-but-incorrect complete misses unreviewed. Unchanged transfer across seven external acquisition settings was heterogeneous: Japan favored the candidate, while external detector F1 ranged from 0.0000 to 0.0919. The contribution is a reproducible, capacity-matched benchmark that separates detection quality from review allocation and identifies how fixed queue priority can exclude a failure type from correction. The findings motivate joint validation of queue allocation and emission thresholds with independent pavement assessment; source-reference oracle recovery does not establish human inspection or safety benefits.