Comparison of two data assimilation algorithms for shallow water flows

Issam S. Strub, Julie Percelay, Olli-Pekka Tossavainen, Alexandre M. BayenNetworks Heterog. Media 2009 · DOI

Abstract

This article presents the comparison of two algorithms for data assimilation of two dimensional shallow water flows. The first algorithm is based on a linearization of the model equations and a quadratic programming (QP) formulation of the problem. The second algorithm uses Ensemble Kalman Filtering (EnKF) applied to the non-linear two dimensional shallow water equations. The two methods are implemented on a scenario in which boundary conditions and Lagrangian measurements are available. The performance of the methods is evaluated using twin experiments with experimentally measured bathymetry data and boundary conditions from a river located in the Sacramento Delta. The sensitivity of the algorithms to the number of drifters, low or high discharge and time sampling frequency is studied.

Introduction

Accurately monitoring river hydraulics is a cornerstone of effective water management, influencing everything from drinking water supplies to irrigation and flood control. While 2D shallow water equations provide a robust mathematical framework for modeling these flows, real-world applications often suffer from model drift caused by simplified assumptions or evolving environmental conditions.

To maintain accuracy, researchers use "data assimilation"—a process that continuously synchronizes mathematical models with real-world sensor data. Traditionally, this relied on fixed (Eulerian) sensors. However, the rise of low-cost, GPS-enabled Lagrangian drifters offers a more flexible alternative. These sensors move with the flow, providing rich spatial data, but they introduce significant mathematical complexity because their locations are constantly changing relative to the model's fixed computational grid.

Methodology

This study evaluates two distinct mathematical frameworks for integrating Lagrangian drifter data into 2D shallow water models:

1. Quadratic Programming (QP) / Variational Approach

The QP method simplifies the problem by linearizing the complex flow equations. This transformation allows the assimilation to be treated as a massive optimization task. The goal is to identify the initial state of the river that best minimizes the discrepancy between the model and the observed drifter paths over a specific observation window (typically one hour). By using the CPLEX solver, this method is highly efficient, capable of converging on a solution in mere seconds.

2. Ensemble Kalman Filtering (EnKF) / Sequential Approach

In contrast, the EnKF method preserves the full, non-linear physics of the river flow. It utilizes the professional hydrodynamic software TELEMAC 2D to run an "ensemble" of approximately 100 simultaneous simulations, each with slightly varied initial conditions. As new GPS coordinates arrive, the filter uses the statistical spread of this ensemble to "nudge" the model toward reality in real-time. While physically more rigorous, this approach is computationally demanding.

Experimental Setup

The researchers conducted "twin experiments" to benchmark these algorithms. In this setup, a high-resolution simulation serves as the "ground truth" to generate synthetic drifter data. This data is then fed into the test algorithms to see how closely they can reconstruct the original flow. To ensure practical relevance, the simulation environment was modeled after a 930-meter section of the Sacramento River, utilizing:

  • Bathymetry: Real-world riverbed depth data from the U.S. Geological Survey (USGS).
  • Boundary Conditions: Discharge and elevation data from the Delta Simulation Model II (DSM2), covering both high-flow (March 2006) and low-flow (June 2006) scenarios.

Key Insights and Comparative Analysis

Computational Efficiency vs. Physical Fidelity

The most stark difference between the two methods is speed. The QP method proved to be a "speed demon," reaching convergence in roughly 5 seconds on a standard desktop. Conversely, the EnKF required between 6 to 8 hours to process a single hour of river time. This suggests that while EnKF is more physically accurate, it is currently impractical for real-time emergency response without significant parallelization.

Deployment Strategies

The algorithms respond differently to how drifters are deployed. The QP method performs optimally when all drifters are released simultaneously at the start of the window, providing a strong anchor for the optimization. The EnKF, being a sequential filter, performs better when drifters are released in staggered intervals, providing a steady stream of updates to the model.

Sensitivity to Flow and Sampling

The QP method is highly sensitive to sampling frequency and flow velocity. If a drifter moves too quickly and "skips" over a grid cell between GPS pings, the linearized model struggles to maintain a lock on its path. This was particularly evident in high-discharge scenarios (e.g., snowmelt runoff), where QP errors increased to approximately 25%. The EnKF is much more robust in these scenarios, as its statistical nature allows it to bridge gaps in data more effectively.

Results and Limitations

Both methods demonstrated strong performance in standard conditions, with relative Root Mean Square (RMS) errors for velocity generally falling between 5% and 15%. However, both algorithms faced challenges near river banks. Because drifters naturally follow the main channel, the algorithms lack data to correct velocity near the shores, leading to higher relative errors in those specific zones.

Furthermore, while increasing the number of drifters (from 8 to 24) significantly improved the QP results, the EnKF showed diminishing returns, suggesting that a smaller number of well-placed sensors may be sufficient for sequential filtering.

Conclusion

This research highlights a clear trade-off in data assimilation for river hydraulics. The QP method offers a rapid, low-cost solution suitable for scenarios where computational resources are limited and flow is relatively stable. The EnKF provides a more robust, non-linear solution capable of handling complex flow dynamics, albeit at a much higher computational price. Future work will focus on optimizing sensor placement and extending these frameworks to include salinity and sediment transport models.

Summary updated

BibTeX
@article{strub2009comparison,
  author = {Issam S. Strub and Julie Percelay and Olli-Pekka Tossavainen and Alexandre M. Bayen},
  title = {{Comparison of two data assimilation algorithms for shallow water flows}},
  journal = {Networks Heterog. Media},
  year = {2009},
  doi = {10.3934/nhm.2009.4.409},
}
Figure