SCI-VIS: Lightweight Visibility Tagging for Coastal Camera Monitoring

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

Observations and measurements made from coastal cameras are conditional on the frame being interpretable. Sea fog, glare, rain and condensation on the housing, haze and low light each degrade the imagery, and the shoreline positions, wave measurements and counts derived from a degraded frame are corrupted or spurious. This paper evaluates a per-frame visibility classifier, based on a compact convolutional network, that assigns each frame one or more of six conditions (Clear, Glare, Fog, Rain/Blur, Hazy, Dark). It is trained and evaluated on 33,569 manually labelled frames from 757 cameras, labelled under a written protocol. On a 20% test split from cameras in training, primary-label accuracy is 93.1% and exact multi-rank accuracy 84.1%, against 67.9% for the majority class; per-class F1 ranges from 0.96 for Clear to 0.79 for Hazy. Under a camera-disjoint cross-validation, in which every camera is scored by a classifier that has not seen it, primary-label accuracy is 81.0%, and Fog, Hazy and Dark are recognised in fewer than half of their frames. Withholding one condition at a time from half of the cameras that have it shows the contribution of a camera’s own labelled examples: recall of the withheld condition is 77% for Rain/Blur, 64% for Fog, 45% for Glare, 19% for Dark and 16% for Hazy, against 86 to 99% for the full classifier on the same frames. Reliable recognition of a camera’s hazy and dark segments therefore requires labelled examples of those conditions from that camera, which sets what a network has to label and retrain on as it grows.