Towards detailed and interpretable hybrid modeling of continental-scale bird migration

Fiona Lippert, Bart Kranstauber, Patrick Forré, E. Emiel van Loon

TL;DR

Researchers have developed FluxRGNN+, a hybrid model that combines fluid dynamics with deep learning to forecast continental-scale bird migration. By decoupling the computational grid from sparse radar locations and adding physics-informed velocity constraints, the model achieves higher-resolution predictions and significantly better generalization to regions without radar coverage.

What Problem Does This Paper Address?

Every year, billions of birds migrate across continents, impacting biodiversity and posing risks to aviation and wind energy infrastructure. To mitigate human-wildlife conflicts, we need accurate, near-term forecasts.

Currently, we rely on weather radar networks (like NEXRAD in the US) to track these movements. However, radars are spatially sparse, leaving massive "blind spots" between stations. Existing forecasting models are often either purely data-driven "black boxes" that ignore the physical process of movement, or they are tied to coarse, irregular grids defined by the radar locations themselves. This makes it difficult to generate detailed, ecologically interpretable maps of where birds are taking off, landing, and flying in the gaps between radars.

Methods

The authors build upon FluxRGNN, a hybrid architecture that uses the continuity equation (a fundamental principle of fluid dynamics) to ensure that bird biomass is conserved—meaning birds don't just appear or disappear; they must flow from one area to another.

They introduce two major innovations in FluxRGNN+:

  1. Grid Decoupling: The model no longer uses irregular cells centered on radars. Instead, it operates on a standardized, high-resolution hexagonal grid (Uber’s H3 index). They implement mapping functions ($f_{R \to C}$ and $f_{C \to R}$) to translate sparse radar observations into this dense grid and back again.
  2. Velocity-Based Flux Parameterization: Instead of letting the neural network learn abstract "flow rates," they force it to explicitly predict bird velocities. These velocities are then used in a standard numerical scheme (an "upwind scheme") to calculate movement.
Figure 1. Map of the NEXRAD weather radar network, with black circles indicating measurement areas Am, and the hexagonal tes- sellation on which movements are modeled. During encoding, fR→C maps sparse radar measurements to cell space (bottom left), in which the forecast is generated based on within-cell source/sink terms si and cell-to-cell fluxes Fj→i (bottom center). Finally, fC→R maps cell-level predictions back to measurement space (bottom right).

A tunable loss parameter ($\lambda$) allows researchers to balance the trade-off between raw density accuracy and the ecological realism of the predicted flight paths.

Datasets

  • NEXRAD: Data from 143 weather radars across the contiguous United States, providing bird density and velocity measurements.
  • ERA5: Atmospheric reanalysis data (wind, temperature, pressure, etc.) used as environmental predictors.
  • NLCD: Land cover data to incorporate habitat types into the model's decision-making process.
  • Timeframe: The study focused on the autumn migration season (August 1 to November 15) from 2013 to 2021.

Key Insights and Ideas

  • Physics as a Regularizer: Even without direct supervision on bird velocities ($\lambda=0$), the underlying movement model acts as a powerful constraint. This allows the model to maintain spatial consistency, leading to much more robust predictions in "blind spots" compared to standard machine learning models like XGBoost.
  • The Interpretability Pareto Front: There is an inherent trade-off between predicting bird density perfectly and predicting realistic flight directions. By increasing velocity supervision, the model produces movement patterns that align much better with biological reality, which is crucial for ecological research.
  • Extrapolation is the Killer App: Because the model understands flow, it can "fill in the blanks" between radars far more effectively than local models that only look at weather conditions at a single point.

Key Results

  • High Performance on a Finer Grid: FluxRGNN+ matches the predictive accuracy of the original model while providing much more detailed spatial patterns. It achieved an RMSE of ~31.6 birds/km².
  • Superior Recall: The model significantly outperformed baseline models in identifying rare, high-intensity migration events—the "peak nights" that are most important for conservation and safety.
  • Robust Extrapolation: In spatial cross-validation (where radars are hidden during training), FluxRGNN+'s error only increased by ~9.7%, whereas the error for the data-driven XGBoost baseline jumped by over 26%.
Spatial cross-validation shows that FluxRGNN+ (right three pairs) maintains much lower error and higher precision/recall on unseen "test" radars compared to XGBoost (left pair).

Limitations and Future Work

  • Mapping Simplicity: The current radar-to-cell mapping uses a simple weighted average. The authors suggest that using Graph Neural Networks (GNNs) for this step could further reduce information loss.
  • Uncertainty Quantification: The model does not yet account for measurement errors or model uncertainty. Future versions could integrate data assimilation techniques common in meteorology to provide probabilistic forecasts.
  • The Accuracy-Interpretability Trade-off: Users must decide whether they prioritize the most accurate density numbers or the most realistic movement vectors, as maximizing one slightly degrades the other.

Summary updated

Abstract

Hybrid modeling aims to augment traditional theory-driven models with machine learning components that learn unknown parameters, sub-models or correction terms from data. In this work, we build on FluxRGNN, a recently developed hybrid model of continental-scale bird migration, which combines a movement model inspired by fluid dynamics with recurrent neural networks that capture the complex decision-making processes of birds. While FluxRGNN has been shown to successfully predict key migration patterns, its spatial resolution is constrained by the typically sparse observations obtained from weather radars. Additionally, its trainable components lack explicit incentives to adequately predict take-off and landing events. Both aspects limit our ability to interpret model results ecologically. To address this, we propose two major modifications that allow for more detailed predictions on any desired tessellation while providing control over the interpretability of model components. In experiments on the U.S. weather radar network, the enhanced model effectively leverages the underlying movement model, resulting in strong extrapolation capabilities to unobserved locations.
BibTeX
@misc{lippert2024towards,
  author = {Fiona Lippert and Bart Kranstauber and Patrick Forré and E. Emiel van Loon},
  title = {{Towards detailed and interpretable hybrid modeling of continental-scale bird migration}},
  year = {2024},
  eprint = {2407.10259},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.10259},
}
Figure