How the forecast is built

Not a single wave height. The model takes the whole directional spectrum arriving at the coast: every swell train, with its own height, period and angle. Two swells of the same size do entirely different things to a break depending on their period and where they come from.

Swell

CDIP's Monitoring and Prediction (MOP) system, run by Scripps. Nearshore nodes roughly every 300 feet along the California coast, carrying full directional spectra rather than a summary height and period.

Seafloor

NOAA NCEI CUDEM, 1/9 arc-second (2022–2023), resampled to a 10-foot grid covering about half a mile in every direction from each break.

Tide

NOAA CO-OPS. Observed where a station reports it, predicted beyond.

Wind and weather

Open-Meteo, hourly, including sea-surface temperature.

Seasonal context

25 years of CDIP MOP hindcast (2000–2025), binned by week of year, so a given week can be read against its own history.

Ground truth

Sentinel-2 satellite imagery, used to check the model, not to drive it.

A parabolic mild-slope solver marches that spectrum shoreward across the real seabed, computing refraction, shoaling, diffraction and finally breaking. It is deterministic physics, not a statistical fit, so the same inputs always produce the same wave field.

Waves break when height reaches a fixed fraction of depth (H ≥ 0.78·d, the McCowan depth limit). Breaking is solved on the combined sea rather than train by train: superposed peaks can break where no single train would, and energy-summing already-saturated trains would inflate surf-zone height past what the depth allows.

Which train organises the break is decided by which one breaks biggest, not which is tallest offshore. Breaking height follows energy flux, so a 2 ft 17-second groundswell out-breaks a 3 ft 6-second windswell. Choosing on offshore height alone gets this backwards.

Where swell has to wrap a point over miles before it arrives (a bay, an inlet, anything tucked behind a headland), a local grid can only bend the last few hundred yards, so the swell would show up far too oblique. Those spots get a separately traced wrap from the offshore node to the break. At Bolinas that is the difference between a swell arriving 64° off shore-normal and one that has actually turned toward the beach.

Satellite imagery gives an independent read on what the ocean actually did. For a chosen day we find a cloud-free Sentinel-2 pass over a spot and pull the real swell crest direction out of the image with a 2-D FFT. The crest lines are usually invisible to the eye, but they are periodic, and the transform separates them from the noise.

That observed direction goes next to our prediction for the same hour, and three things get judged: whether the approach angle matches, whether the model reproduces the wrap the satellite shows around the point, and whether the break sits where the whitewater actually is.

The extraction reports its own confidence and refuses to return a direction when the signal is too weak rather than inventing one. Where it is confident, it has recovered a swell direction to within 3° of the nearest offshore buoy.

Horizon

7 days, then a lower-confidence extended outlook to 10 days.

Resolution

3-hourly to 72 hours, 6-hourly beyond.

Confidence

Every step carries high, medium or low, degrading with lead time. The extended outlook is always low.

Ceiling

Seven days is a hard limit, not a choice. It is how far the underlying wave models run. Beyond that is climatology, and it is labelled as such.

This forecast was generated Aug 26, 3:44 AM PDT.

Each of these is a real limitation, stated because a forecast that hides its error bars is worth less than one that doesn’t.