Not just how fast waves oscillate.
Which way do they travel?
Two sea states can share exactly the same frequency spectrum and spectral wave height, yet look completely different across the water. Spread the variance across directions, then split it into crossing swell and wind-sea systems. Compare three representations of one two-dimensional spectrum.
A · Frequency–direction variance density S(f, θ)
Horizontal: frequency (Hz). Vertical: propagation direction θ in degrees, plotted on a 0° to 360° compass-style axis, clockwise from north. 360° is the same direction as 0°, so a lobe close to north can wrap across the top and bottom edges of the plot. Colour represents S(f, θ) (m²/Hz/rad); the colour scale stays fixed at 0–32 m²/(Hz·rad) across scenarios. Markers identify the mean directions of the two modelled systems.
B · Directional rose (integrated over frequency)
Each sector spans exactly 5°. Its radius is the angular marginal G(θ) = ∫S(f, θ)df in m²/rad, on a fixed 0–1.5 scale. Radius, not wedge area, encodes density. Thin coloured arrows identify the prescribed system means; they do not measure density.
C · The unchanged frequency marginal S(f)
Blue = swell partition; coral = wind-sea partition; dark curve = sum. Integrating S(f, θ) over all directions recovers this identical frequency spectrum in every preset.
Illustrative two-partition spectrum: Gaussian frequency shapes (discretely normalized), wrapped Gaussian direction shapes (discretely normalized), 0.005 Hz × 5° grid. Angle is the direction of travel on a compass-style bearing axis (0° north, 90° east, 180° south, 270° west), not the nautical “coming from” convention. Radians are used in integration and in the S(f, θ) units.
A cell contributes Δm₀ ≈ S(f, θ) × 0.005 Hz × (5π/180 rad). Sum across direction to obtain the frequency marginal S(f) in m²/Hz; sum across frequency to obtain the angular marginal G(θ) in m²/rad. Integrating either marginal gives 0.48 m². Spreading the same variance over more angles lowers the density peaks without changing Hm0.
Three views, three different questions
A frequency–direction heatmap reveals which periods travel in which directions. Two separate patches can indicate crossing swell and wind sea, even when the total variance is unchanged.
A directional rose collapses all frequencies into one angular distribution. It helps show whether waves propagate toward one dominant direction or several—but it cannot tell which direction belongs to the longer-period system.
A frequency marginal collapses directions instead. It reports variance at each frequency but cannot tell whether the corresponding waves travel together or cross at different angles.
A single-point surface-elevation time series cannot uniquely resolve directions. Directional instruments combine additional spatial, slope, velocity or other information. Directional wave modelling likewise tracks variance or action across frequency and direction, subject to physical assumptions and source terms.
Try a thought experiment
- Choose Aligned, then Broad. What stays identical on the frequency-marginal plot?
- Choose Crossing seas and increase the separation. Which plot best identifies the two distinct wave systems?
- Increase σ. Why does the peak colour weaken even though the total spectral variance and Hm0 are fixed?
What is a spectral partition?
A partition groups a spectral system by its frequency and direction. Here swell and wind sea are prescribed Gaussian frequency shapes, with variances 0.30 and 0.18 m². Those names identify teaching components rather than an automatic classification: assigning observed wind sea normally also requires wind information. Overlapping systems are not always uniquely separable from spectral shape alone.
From bearings to the plot, and from “toward” to “from”
A compass bearing θ increases clockwise from north. Canvas coordinates are x = cₓ + r sin θ and y = cᵧ − r cos θ. The corresponding Cartesian angle measured counterclockwise from east is α = 90° − θ, modulo 360°. A coming-from bearing is (θ + 180°) modulo 360°: waves travelling toward 320° come from 140°.
The heatmap has 72 cells covering [0°, 360°), centred at 2.5°, 7.5°, …, 357.5°. Its bottom 360° edge joins its top 0° edge; there is no extra 360° cell. A lobe crossing north appears split only because a circular coordinate has been laid flat. The rose joins the same two edges naturally.
Directional density is not directional variance
G(θ) is variance per radian; its integral is elevation variance in m². It is not the statistical variance of the angles themselves. Bearings are circular: 359° and 1° are close neighbours, so ordinary arithmetic averages and variances can mislead. The width σ here sets a wrapped-Gaussian shape. The model normalizes the sampled density so ΣDjΔθ = 1.
A density per degree is a different numerical quantity: Sdegree = Sradianπ/180. Using a per-radian density with 5 instead of 5π/180 in an integral would overestimate variance by 180/π.
Why one elevation sensor cannot identify direction
At one location, waves with the same frequency can produce the same elevation history regardless of bearing. A directional buoy also measures horizontal motion or slopes; arrays measure spatial phase differences. Cross-spectra between such channels constrain directional moments. Reconstructing a full directional distribution still involves assumptions, finite angular resolution and possible ambiguities, especially for crossing or reflected systems.
Go deeper: how the directional spectrum works
Conservation of variance when spreading over direction
Each partition p has its own frequency distribution Sp(f) and normalized directional spreading Dp(θ). Because ∫Dpdθ = 1, spreading changes the location of variance in direction, not its total. Integrating over frequency and direction gives m0 = 0.48 m² in every case.
Directional bins, units and coordinate conventions
Here Δf = 0.005 Hz and Δθ = 5π/180 rad. Therefore S(f, θ) has units m²/(Hz·rad). Plotted angular labels are in degrees for accessibility; a radian conversion is needed to integrate densities expressed per radian. In real applications document whether a directional angle indicates travel toward or coming from. This chapter uses travel toward, with degrees increasing clockwise from north, while the density and integrals still use radians internally.
What this model does and does not predict
The two partition shapes are idealized and prescribed rather than inferred from a wave buoy. The plots do not include geographical propagation, finite-depth dispersion, nonlinear interaction, refraction, diffraction, currents or source/sink terms. A genuine directional spectral wave model solves an energy or action balance and evolves the spectrum through space and time; this is a controlled visualization of spectral representation alone.