Rip current detection in nearshore areas through UAV video analysis using optical flow and almost local-isometric embedding techniques on the sphere

Rip currents cause most surf-beach lifeguard rescues and most beach drowning incidents, yet automated detection relies on costly instrumentation that scales poorly across the many beaches that lack lifeguards. We propose an end-to-end pipeline producing pixel-level rip-current likelihood maps directly from raw unmanned aerial vehicle (UAV) video without manual ground-control points — to our knowledge the first to do so from uncalibrated drone video, prior automated approaches operating at video-clip granularity or requiring surveyed camera installations. The pipeline composes (i) dense optical-flow estimation with five classical algorithm variants; (ii) a sphere-aware projection mapping pixel coordinates to local-isometric ground coordinates, with a tilt-aware Jacobian correction of per-pixel velocity; (iii) a CNN six-class scene parser (sky/sea/land/foam/structure/void), trained on frames verified by dense pixel-matching to contain no extraction from either evaluation video; (iv) a per-pixel temporal accumulator flagging pixels whose flow direction agrees with the locally-estimated offshore direction within ±90∘ — the physical boundary of seaward-directed flow, shown robust in an 18-scene sensitivity sweep; and (v) post-processing by Gaussian smoothing (σ=25 px) and a normalized cross-algorithm ensemble. We evaluate on two UAV-captured rip-current videos at Palm Beach and Haulover Park, Florida, annotated by consensus among three coastal-science domain experts. At the physically-neutral T=90∘ default the framework attains precision-recall AUC 0.636 at Palm Beach and 0.660 at Haulover; per-scene selection — of the angular tolerance, ensemble membership, mixing weight, sphere tilt, and which modules are engaged — against the same expert annotation used for evaluation raises these to 0.7066 and 0.7040, reported as an optimistic upper bound rather than an expected-performance estimate. We release the source code, the trained sea-mask refiner, a YOLOv8m-seg deep-learning baseline, and the evaluation scripts. Evaluation is limited to these two expert-annotated scenes, and the method requires visible surface tracers (foam, dye, or sediment fronts) advected by the current.