QG-LaDA: Reconstructing Quasi-Geostrophic Streamfunction from Lagrangian Drifters

Q G-L A D A:
Quasi-Geostrophic Lagrangian Assimilation of Drifter Abservations

MSc Thesis · University of Calgary · 2025

Ahmad Droobi1

1 Schulich School of Engineering, University of Calgary
Advisor: Mustafa Mohamad

Poster

QG-LaDA live lab screenshot: truth, WHERE EnKF, and hybrid streamfunction panels with drifters
Identical-twin lab. Truth, WHERE EnKF, and hybrid EnKF–PF reconstructions of $\psi$ from 40 drifters.

Video

Animated comparison of truth, WHERE, and hybrid streamfunction
Truth vs WHERE vs hybrid. Black markers are Lagrangian tracers.

TL;DR

Abstract

This work contributes to the understanding of high-dimensional fluid flows, focusing on quasi-geostrophic systems, by reconstructing turbulent flow fields from a partial, noisy, time-sequential set of tracer measurements. Lagrangian drifters stream local velocity rather than a gridded streamfunction; the inverse problem is to recover $\psi$ from those tracks. A spectral quasi-geostrophic twin generates truth and a forecast ensemble. Helmholtz inversion maps vorticity to streamfunction. A localized stochastic Ensemble Kalman Filter (WHERE) assimilates the noisy velocities, with an optional hybrid EnKF–particle-filter branch. The public lab is a runnable companion to the University of Calgary MSc thesis Data-driven filtering techniques for turbulent flow models (a Lagrangian data assimilation approach) (Droobi, 2025).

Method

1. Initialization

A spectral barotropic QG model on a periodic $[0, 2\pi]^2$ $\beta$-plane produces the truth and the forecast ensemble. Forty drifters sit on a jittered lattice and observe noisy velocities ($\sigma_o = 0.04$). Streamfunction is recovered from vorticity by Helmholtz inversion.

$$\partial_t q + J(\psi, q) + \beta \partial_x \psi = F - d q - \nu (-\nabla^2)^p q$$ $$q = \nabla^2 \psi - \mu \psi, \qquad \hat{\psi} = -\hat{q}/(\kappa^2 + \mu)$$

2. Localized analysis

WHERE is a stochastic EnKF on vorticity. Predicted velocities at the truth tracer sites form the observation operator. Gaspari–Cohn localization ($L_{\mathrm{loc}} = 1.55$) tapers the sample covariances; multiplicative inflation $\alpha = 1.08$ keeps the ensemble from collapsing. Analysis interval is $0.35$ time units on the committed run ($N_e = 24$).

3. Hybrid refinement

After the EnKF update, members can be reweighted by the Gaussian velocity likelihood and resampled when the effective sample size drops below $0.55\,N_e$. A $1.5\%$ jitter restores sample covariance after cloning. On this QG twin the hybrid matches large-scale correlation but does not beat WHERE on RMSE.

Results

Committed identical-twin Python run, $N = 32$, $N_e = 24$, $40$ tracers, $t_{\mathrm{end}} = 28$. Not a production forecast score and not thesis Table 10.1 (different resolution and horizon).

FilterXCORRMSENote
WHERE EnKF0.9640.256clears XCOR target > 0.9
Hybrid EnKF–PF0.9450.367smoother; worse small-scale RMSE
Free run (no DA)0.3160.960phase lost within a few eddy turnovers
WHERE versus free-run streamfunction
WHERE holds the gyre. The free run loses phase.

BibTeX

@mastersthesis{droobi2025lada,
  title   = {Data-Driven Filtering Techniques for Turbulent Flow Models
             (A Lagrangian Data Assimilation Approach)},
  author  = {Droobi, Ahmad},
  school  = {University of Calgary},
  year    = {2025},
  doi     = {10.11575/PRISM/49934},
  url     = {https://hdl.handle.net/1880/122341}
}