SCI-VIS: Lightweight Visibility Tagging for Coastal Camera Monitoring

Michael Reinwald , Ketron Mitchell-Wynne, Zane D'Amico, Mercedes Stringer, Benjamin Freeston

Automated assessment of imaging conditions is a necessary pre-processing step for any computer vision pipeline deployed on outdoor cameras subject to atmospheric variability. We describe SCI-VIS, a lightweight per-frame visibility classifier deployed within the Surfline Coastal Intelligence (SCI) pipeline. The Surfline coastal-camera network spans ~1,200 cameras worldwide. SCI-VIS tags each incoming frame with one or more concurrent conditions: Clear, Glare, Fog, Rain/Blur, Hazy, and Dark. The classifier combines a compact handcrafted image-feature representation with a gradient-boosted tree classifier, deployed CPU-only at the per-rewind cadence of the SCI archive. A multi-rank label scheme allows each frame to carry multiple simultaneous condition tags. The model is trained on 12,435 frames from a 15,543-frame manually-labelled dataset drawn from 595 cameras (a subset of the full network). On the 3,108-frame held-out test set it achieves 92.9% primary-label accuracy and 87.3% exact multi-rank match accuracy, with strong performance on the dominant conditions (F1 = 0.96 for Clear, 0.90 for Glare, 0.87 for Fog) and weaker recall on the rarest conditions (F1 = 0.46 for Hazy); because the train/test split is at the frame level, these figures are in-distribution estimates over the operational camera distribution the model is deployed against. Confusion analysis shows that the residual error is biased toward the majority Clear class - a direct consequence of class imbalance (74% Clear) - so misclassifications fall on borderline degraded frames and are addressable through class-weighted training or threshold adjustment. The model requires no GPU and is invoked once per rewind (the fixed-duration video segment forming the atomic unit of the camera archive), making the ~1.8 s per-frame CPU cost operationally negligible.