Zero-Shot Water Monitoring from Fixed Drainage Surveillance Imagery Across Day/Night and Rain/No-Rain Conditions: A SAM 3 Benchmark at 25 Urban Sites

Abstract

Urban drainage systems rely on fixed surveillance cameras that operate continuously under daytime, nighttime-infrared, rainy, and non-rainy conditions, yet most vision-based water-monitoring methods are validated only on daytime imagery and require costly pixel-level annotation. This paper investigates whether a promptable segmentation foundation model can monitor water surfaces and waterlogging zero-shot—without any task-specific training or labeling—across the full illumination–weather spectrum of real surveillance operations. We construct a benchmark of 3651 frames from 25 fixed cameras of an urban drainage monitoring network in Zhenjiang, China, spanning four scene categories (river/canal, road ponding, underpass, pump-station apron), seven days in July–September 2026, and all four illumination–weather quadrants, explicitly including non-rainy negative days. Using SAM 3 with the text prompt “water”, we extract per-frame water-area ratio and normalized waterline level, and evaluate (i) quadrant-wise segmentation stability, (ii) rain-event detection, (iii) false-alarm behavior on negative days, and (iv) pixel-level agreement with expert reference masks. Zero-shot segmentation met the scene-specific validity criterion on 96.9% of the 162 diurnal-cycle frames at the canal reference site (100% in both no-rain quadrants, 88.9% in the night-rain quadrant); failures concentrate in heavy-rain-obscured and sediment-turbid dusk frames. On 10 August 2026—a day with two visually verified heavy-rain events—the waterline indicator detected both events (peaks 72.9% and 71.4% vs. a 43–62% control-day band), and the same afternoon event was independently corroborated by severe ponding at road and underpass sites city-wide. A level-only alarm rule produced an 11.8% false-alarm rate on 102 negative-day frames, traced to a camera re-aiming event that shifted the normalized indicator distribution; a compound rule combining area anomaly and rise rate produced no alarms on the 102 examined negative-day frames (vs. 12/102 for the level-only rule) while preserving both event detections—a retrospective, in-sample result that we do not treat as a validated false-alarm rate. We therefore froze the complete pipeline and rules and evaluated them prospectively on 14 additional days (7 days with event-triggered 5-min capture and 7 routine hourly capture days; 1072 frames) that had never been used for threshold or rule design: on the 13 visually verified dry days (784 frames) the frozen compound rule produced only six alarm frames (0.77%), all at illumination-mode switchover transients, and a two-frame persistence variant eliminated all of them (0/13 days; upper 95% confidence bound on the day-level false-alarm rate 20.6%) while detecting both the nighttime rise and the afternoon flood stage of a newly encountered heavy-rain day. Pixel-level verification on an expanded stratified subset of 79 frames from 24 sites with expert reference masks yields a mean IoU of 0.633 (0.652 excluding two failure frames) and near-zero false positives on 15 water-free frames (mean 0.15% of pixels); the reported metrics change by at most 0.003 mIoU under ±5-pixel annotation boundary tolerance. Against two same-protocol baselines—an open-vocabulary detector (YOLOE, mIoU 0.268) and a classic HSV color rule (mIoU 0.453, 68% false-positive area on negative frames)—zero-shot SAM 3 is the only approach that remains accurate in the nighttime-infrared quadrants, and its mIoU varies by less than 0.04 across confidence thresholds 0.15–0.35 at the reported prompt and stays within 0.48–0.66 across four prompt wordings. Reference-mask reliability was independently confirmed by a second annotator (inter-annotator IoU 0.71, Dice 0.80). Inference costs ~46 ms per frame on a consumer laptop GPU (not an edge platform), supporting further evaluation of multi-camera processing. These results establish text-prompted foundation models as a practical, training-free candidate for drainage surveillance across day/night and rain/no-rain conditions—promising for operational evaluation, though not yet operationally validated—and delineate the specific conditions under which they fail.