Notes - Jiang et al. 2022

Core Idea : the high frequency directional wave spectrum contains enough local wind information that you can invert it for wind speed and wind direction, but the wind it sees is not instantaneous anemometer wind, it is closer to a 40 minute lagged, wave integrated wind.

Implications: independent evidence shows that high-frequency directional wave spectra encode local wind forcing, especially in the 0.2–0.5 Hz band, which motivates treating tail energy, spreading, and alignment as physically meaningful air–sea coupling diagnostics.

Central Question

Can we estimate near surface wind speed and wind direction using only buoy wave spectra?

Importance Compact wave buoys measure waves well but often do not directly measure winds. This paper builds a DNN that maps spectral information $(E,{\alpha }_1,{\alpha }_2,r_1,r_2)(f)\to (U_{10},{\theta }_w)$

Physical premise the high frequency part of the spectrum is wind-coupled, so it carries information about the wind forcing that generated it

Previous work used $E(f)\sim \alpha u_{\ast }g{f}^{-4}$ -> a more physically explicit approach baed on the equilibrium range Voermans: derive wind from assumed tail physics.
Jiang: let the network learn the empirical tail-to-wind mapping.

Data

Buoys They use NDBC buoy data from 101 buoys over 2014–2018, giving more than 1.6 million records. The wave spectra span about $0.02\le f\le 0.485\mathrm{Hz}$ with 47 frequency bins.

Setup $5\text{ Fourier coefficients}\times 47\text{ frequency bins}=235\text{ inputs}$ So for each time, the DNN sees a vectorized version of: $E(f),{\alpha }_1(f),{\alpha }_2(f),r_1(f),r_2(f)$

Model structure

$235\to 64\to 64\to 1$

They train two separate networks: ${\text{DNN}}_{U10}:{\mathbf{x}}_{wave}\to U_{10}$ ${\text{DNN}}_{dir}:{\mathbf{x}}_{wave}\to {\theta }_w$

Main result

Wave spectra estimate winds surprisingly well

For instantaneous collocation, the DNN gets about ${\mathrm{RMSE}}_{U10}\sim 1.3\mathrm{m}{\mathrm{s}}^{-1}$ and wind direction RMSE around: $\sim 16^{\circ }$ for winds stronger than $7\mathrm{m}{\mathrm{s}}^{-1}$

They shift the wind measurements backward in time and ask: what past wind is best represented by the current wave spectrum? The best performance occurs when the wave spectrum is used to estimate wind about 40 minutes earlier Then the error improves to about: ${\mathrm{RMSE}}_{U10}\sim 1.1\mathrm{m}{\mathrm{s}}^{-1}$ ${\mathrm{RMSE}}_{dir}\sim 14^{\circ }\text{for }{U}_{10}>7\mathrm{m}{\mathrm{s}}^{-1}$

**The wave spectrum is not an instantaneous wind sensor. It is an integrated memory of recent forcing.

Figures

1. the wave-to-wind inversion works best where the observed frequency spectrum is not strongly contaminated by current-induced Doppler shifting, coastal complexity, or sensor problems.
2. At low wind, wave energy is weak and geophysical noise/current effects matter more. At high wind, the air–sea system becomes nonlinear, and the NDBC anemometer correction to U_{10} may itself become questionable.Panels d–f do the same for wind direction. Direction is poor at low wind because the wind-sea direction is less strongly defined. It improves quickly as wind speed increases.
3. The RMSE minimum occurs near 40–50 min for wind speed and 40–60 min for wind direction.
4. If the DNN is usually unbiased and suddenly has persistent bias at one buoy, the issue may be sensor failure or corrupted buoy data. This matters for me because a wave-derived wind estimate could be used as a consistency check on Spotter/buoy/model collocation.
5. wind speed lives strongly in the high-frequency energy level, while wind direction lives strongly in the high-frequency directional coefficients. For wind speed, the key frequencies are above about $f>0.1\mathrm{Hz}$ with particularly strong importance near: $f\sim 0.2\mathrm{Hz}$ For wind direction, the most important region is higher: $f>0.25\mathrm{Hz}$ with strongest importance near: $f\sim 0.35-0.4\mathrm{Hz}$

What the DNN learns

The sensitivity tests show that For $U_{10}$, the dominant input is: $E(f)$ especially high-frequency E. For wind direction, the dominant input is: ${\alpha }_1(f)$ especially high-frequency ${\alpha }_1$ But $r_1,$ ${\alpha }_2$, and $E$ also help. This is important because directional spreading is not irrelevant. It contains secondary wind information.

That supports the idea that diagnostics like: $R_{hf}=\frac{E_{0.2-0.5}}{E_{total}}$ and high-frequency alignment/spreading are not arbitrary. They are in the same frequency band where the wind signal is strongest.

Applicability

This paper supports your conceptual argument that: the high-frequency tail is a dynamically active, wind-coupled part of the spectrum, not merely a residual tail.

In my language:

• $E_{hf}$ carries wind-speed / stress-related information.
• ${\alpha }_{hf}$ carries wind-direction / alignment information.
• $r_{hf}$ or directional spreading carries information about organization of the wind sea.
• The high-frequency tail has memory and may lag the wind forcing.
• Currents and Doppler shifting can seriously complicate frequency-space interpretation.

That last point is huge. Jiang basically gives a citation-backed reason why frequency-bin-based high-frequency diagnostics may become distorted when currents are strong: ${\omega }_{obs}=\omega (k)+\mathbf{k}\cdot {\mathbf{U}}_c$ So the same physical wavenumber spectrum can appear shifted in frequency space.

Limitations

1. Training data are mostly ordinary NDBC conditions, not tropical cyclone extremes.
2. High winds are underrepresented, so the model underestimates large $U_{10}$.
3. The target wind is anemometer-derived $U_{10}$, which itself may be questionable in extreme seas.
4. The model uses frequency spectra, so strong currents and Doppler shifting can corrupt the learned mapping.
5. It retrieves wind, not stress. It does not directly solve the drag/saturation question.

independent evidence shows that high-frequency directional wave spectra encode local wind forcing, especially in the 0.2–0.5 Hz band, which motivates treating tail energy, spreading, and alignment as physically meaningful air–sea coupling diagnostics.

My story

Prior work has shown that high-frequency directional wave spectra can retrieve near-surface wind speed and direction with scatterometer-like accuracy. Sensitivity tests indicate that wind speed information is concentrated primarily in high-frequency spectral energy, while wind direction information is concentrated in high-frequency directional coefficients. This supports the use of high-frequency tail energy, alignment, and spreading as diagnostics of wind–wave coupling in tropical cyclone conditions.

Within tropical cyclones, do different high-frequency tail organizations occur at similar bulk sea states, and could those differences affect wave-supported stress and drag?

Implications

The mapping is relatively low dimensional -> adding more neurons does not improve performance. This implies that the mapping $S(f,\theta )\to U_{10}$ is actually fairly simple.

Most of the 235 inputs are redundant : The actual physical state space appears much smaller than the initial inputs

The model is basically learning a weighted spectral moment: When they block certain frequencies:

• low frequencies matter very little
• frequencies near 0.2 Hz matter enormously for wind speed
• frequencies near 0.35-0.4 Hz matter enormously for direction It is an automated feature-discovery paper, the DNN is identifying where the wind information lives.

Directional spreading contains real information : when they remove $r_1$ performance gets worse. Two spectra can have similar energy levels but different directional organization. If the DNN can improve wind retrieval using spreading information, does that imply spreading is tied to air-sea momentum transfer? This paper strongly suggests spreading is not merely decorative information.

The network discovers wave age without being told : The DNN learns latent versions of those variables. (my variables) -> similar to manifold work findings

The 40-minute lag is probably the most physically important hidden layer : The spectrum is a memory system. The network can only retrieve the wind because the waves have stored information about recent forcing.

Frequency-space diagnostics can be misleading in currents

**If directional organization were irrelevant, the network would ignore it. ->suggests the atmosphere “cares” about more than bulk energy alone.

if jake is trying to : Estimate how Doppler contamination affects derived quantities while retaining compatibility with existing spectral products -> his approach may be convenient

I am asking : What is the actual high-frequency tail organization? How much energy truly exists between 0.2-0.5 Hz? How does the tail relate to wind stress? -> the new grid is highly important because my diagnostics are frequency dependent