[Paper Review] Fuxi-DA: A Generalized Deep Learning Data Assimilation Framework for Assimilating Satellite Observations

Xiaoze Xu, Xiuyu Sun+4 more|arXiv (Cornell University)|Apr 12, 2024

Meteorological Phenomena and Simulations4 citations

TL;DR

Fuxi-DA is a generalized deep learning-based data assimilation framework that efficiently integrates satellite radiance observations from Fengyun-4B's AGRI instrument into numerical weather prediction. By leveraging a U-Net-like architecture with multi-scale feature fusion and a multi-time-step loss, it reduces analysis errors and improves 6–60 hour forecasts, outperforming baseline models in both global and regional metrics.

ABSTRACT

Data assimilation (DA), as an indispensable component within contemporary Numerical Weather Prediction (NWP) systems, plays a crucial role in generating the analysis that significantly impacts forecast performance. Nevertheless, the development of an efficient DA system poses significant challenges, particularly in establishing intricate relationships between the background data and the vast amount of multi-source observation data within limited time windows in operational settings. To address these challenges, researchers design complex pre-processing methods for each observation type, leveraging approximate modeling and the power of super-computing clusters to expedite solutions. The emergence of deep learning (DL) models has been a game-changer, offering unified multi-modal modeling, enhanced nonlinear representation capabilities, and superior parallelization. These advantages have spurred efforts to integrate DL models into various domains of weather modeling. Remarkably, DL models have shown promise in matching, even surpassing, the forecast accuracy of leading operational NWP models worldwide. This success motivates the exploration of DL-based DA frameworks tailored for weather forecasting models. In this study, we introduces FuxiDA, a generalized DL-based DA framework for assimilating satellite observations. By assimilating data from Advanced Geosynchronous Radiation Imager (AGRI) aboard Fengyun-4B, FuXi-DA consistently mitigates analysis errors and significantly improves forecast performance. Furthermore, through a series of single-observation experiments, Fuxi-DA has been validated against established atmospheric physics, demonstrating its consistency and reliability.

Motivation & Objective

• To develop a generalized deep learning framework for assimilating diverse satellite observations into numerical weather prediction systems.
• To overcome the limitations of traditional variational data assimilation methods, including high computational cost, complex observation operator design, and limited use of cloudy-sky observations.
• To improve initial condition accuracy by integrating AGRI radiance data from Fengyun-4B, which are often underutilized in operational systems.
• To validate the framework’s physical consistency and reliability through single-observation experiments and comparison with ERA5 reanalysis.
• To demonstrate that deep learning-based assimilation can match or exceed the performance of leading operational NWP models in medium-range forecasting.

Proposed method

• Fuxi-DA employs a U-Net-like encoder-decoder architecture with skip connections, using 3×3 convolutional layers, layer normalization, and SiLU activation functions.
• The model uses a multi-scale feature fusion mechanism, where features from down-sampling and up-sampling stages are concatenated to preserve spatial and channel-wise information.
• A pixel-shuffle layer enables feature upscaling with an upsampling factor of 2 in each up-sampling stage.
• The training uses a multi-time-step loss function combining analysis error (t=0) and forecast error (t=1 to t=10), with latitude-weighted L1 loss to account for spherical geometry.
• The loss function is optimized via AdamW with a warmup and cosine annealing learning rate schedule over 6000 iterations on 4 A100 GPUs.
• The model is trained on ERA5 reanalysis as ground truth, with satellite observations used as input to reconstruct accurate atmospheric state fields.

Figure 1: The comparison between the Fuxi-DA and the variational data assimilation.

Experimental results

Research questions

• RQ1 Can a deep learning-based data assimilation framework effectively reduce analysis errors when assimilating AGRI radiance observations from Fengyun-4B?
• RQ2 How does Fuxi-DA compare to baseline models (e.g., FuXi and a correction-only model) in terms of forecast accuracy across 6–60 hour ranges?
• RQ3 Does Fuxi-DA maintain physical consistency when assimilating single observations, as validated against atmospheric physics?
• RQ4 To what extent does the inclusion of satellite data improve the representation of atmospheric states, especially in cloudy and precipitating conditions?
• RQ5 Can the multi-time-step loss strategy enhance long-range forecast performance by jointly optimizing analysis and forecast fields?

Key findings

• Fuxi-DA significantly reduces latitude-weighted RMSE in both analysis and 6–60 hour forecasts compared to the control experiment (EXP_CTRL), with improvements observed across all 10 forecast time steps.
• The assimilation experiment (EXP_ASSI) shows a normalized RMSE difference of -12.3% compared to EXP_CTRL, indicating a 12.3% relative reduction in error across the 6–60 hour forecast window.
• In regional evaluation over East Asia, Fuxi-DA reduced RMSE by 14.1% compared to the control model, demonstrating enhanced performance in dynamically active regions.
• Single-observation experiments confirm that Fuxi-DA produces physically consistent responses, validating its reliability and interpretability.
• The correction model (EXP_CORR) trained without satellite data shows higher baseline errors, confirming that Fuxi-DA’s improvements stem from effective assimilation, not just model correction.
• The multi-time-step loss function contributes to improved forecast skill by jointly minimizing analysis and forecast errors, as evidenced by consistent error reduction across all forecast horizons.

Figure 2: Comparison of analysis after a single assimilation. The time series of the globally-averaged latitude-weighted RMSE (a, b, c, d) and the time series of the normalized RMSE difference (e, f, g, h) of EXP_CTRL (black lines), EXP_CORR (blue lines) and EXP_ASSI (red lines), and the spatial dis