SIMBA: ABidirectional Retrieval Forward Simulation Framework for Modeling FY-4A GIIRS Hyperspectral Infrared Radiances Toward NWP Applications

Abstract: Hyperspectral infrared observations are an important data source for numerical weather prediction (NWP) because they provide rich information on the vertical structure of atmospheric temperature and humidity. However, most existing deep learning methods mainly focus on one-way retrieval from radiances to atmospheric profiles, while the reverse radiance simulation process and the consistency between atmospheric state space and radiance observation space are insufficiently considered. In this study, we propose SIMBA, a unified bidirectional retrieval-forward simulation framework for FY-4A GIIRS hyperspectral infrared radiance modeling toward NWP applications. The framework jointly performs atmospheric profile retrieval and radiance reconstruction, introduces a cycle-consistency constraint to strengthen the coupling between the two processes, and employs a bidirectional Mamba state-space module to capture long-range dependencies along pressure levels. Using collocated FY-4A GIIRS observations and ERA5 reanalysis data, the proposed method is evaluated for temperature retrieval, specific humidity retrieval, long-wave radiance reconstruction, and medium-wave radiance reconstruction. Experimental results show that SIMBA outperforms several representative deep learning baselines across both retrieval and reconstruction tasks, while ablation experiments confirm the contribution of the bidirectional design and cycle-consistency mechanism. These results demonstrate that the proposed framework is effective for joint atmospheric profile retrieval and hyperspectral infrared radiance modeling, and suggest potential for future Jacobian-related analysis and NWP-oriented extensions.

• The paper introduces SIMBA, a cycle-consistent framework that jointly retrieves atmospheric profiles and simulates radiances for improved data assimilation.
• It employs bidirectional Mamba state-space layers and FiLM conditioning to accurately model LW and MW channels, ensuring robust vertical dependency capture.
• The approach outperforms traditional methods with RMSE reductions up to 7.2%, demonstrating stability and enhanced precision in both clear- and cloudy-sky conditions.

Bidirectional Retrieval-Forward Simulation Framework for FY-4A GIIRS Hyperspectral Infrared Radiance Modeling

Introduction

The accurate assimilation of hyperspectral infrared radiances from geostationary sensors is fundamental for enhancing numerical weather prediction (NWP). The FY-4A Geostationary Interferometric Infrared Sounder (GIIRS) provides high-resolution, temporally continuous measurements critical for retrieving atmospheric thermodynamic profiles. Traditional retrieval approaches typically employ deep learning architectures for unidirectional mapping—from radiance observations to atmospheric profiles—without explicitly modeling the reverse radiance simulation or enforcing observation-state consistency. This paper introduces SIMBA, a cycle-consistent and bidirectional retrieval-forward simulation framework, to address these deficiencies and align retrieval architectures with the forward operator requirements of data assimilation systems in operational NWP (2606.19943).

Methodological Framework

SIMBA is architected as a unified bidirectional modeling system, integrating retrieval (radiance-to-profile) and forward simulation (profile-to-radiance) branches, each constructed with stacked bidirectional Mamba state-space layers for long-range vertical dependency capture. The cycle-consistency constraint intrinsically links the two branches, enforcing both state-space and observation-space consistency during optimization. The radiance encoder leverages spectrum-specific multilayer perceptron (MLP) blocks for LW and MW channels, fused through FiLM conditioning to modulate profile generation. Auxiliary variables (e.g., pressure-level embeddings, geometric factors) are incorporated to maintain radiative transfer context-awareness.

The loss function is composed of profile retrieval, radiance reconstruction, and cycle-consistency terms, supporting joint gradient propagation across the closed-loop architecture. Training and evaluation utilize collocated GIIRS radiance observations and ERA5 reanalysis profiles, resulting in a substantial dataset of cloudy-sky (158,649 samples) and clear-sky (112,151 samples) cases spanning the full vertical and spectral extent of GIIRS.

Comparative and Ablation Results

SIMBA is systematically benchmarked against Bidirectional CNN, MLP, Transformer, and LSTM baselines, each augmented for bidirectional processing. Across all metrics (RMSE, MAE, $R^2$) and both cloudy- and clear-sky conditions, SIMBA demonstrates superior performance—retrieval RMSE reductions of 3.0–7.2% for temperature and 1.5–1.05% for specific humidity relative to the strongest baselines. In radiance reconstruction, SIMBA achieves lowest LW and MW errors and maintains channel-wise stability, especially in absorption-sensitive spectral regions and challenging vertical layers. Ablation experiments further isolate the contribution of SIMBA's cycle-consistency: closed-loop coupling yields 2–3.5% improvements over single-directional Mamba dependencies.

The architecture also exhibits high computational efficiency: inference times and memory footprints are competitive or superior to baselines, enabling feasible operational deployment. Channel-selection strategies—based on Jacobian and cumulative influence coefficients—prioritize high-sensitivity spectral bands for temperature and humidity, optimizing input informativeness and reducing redundancy.

Theoretical and Practical Implications

The explicit bidirectionality and cycle-consistency offer improved physical coherence for integration into variational data assimilation systems, where forward operators and reliable sensitivity estimates are required. The differentiable forward model branch enables direct Jacobian and sensitivity analysis, supporting channel impact diagnostics and future assimilation extensions. However, while SIMBA's differentiable operator mimics RTM function under evaluated conditions, it does not guarantee physical equivalence or valid sensitivity diagnostics without further in situ and RTM-based validation.

Residual retrieval and reconstruction errors remain—especially due to reliance on ERA5, whose spatial-scale mismatch with GIIRS introduces representativeness errors. Independent radiosonde validation and cross-seasonal/regional robustness assessments are necessary for operational generalization. Physical interpretability remains limited: perturbation sensitivity analysis exposes high-frequency noise in learned channel sensitivities, emphasizing the need for methodological advancements in noise suppression and interpretability.

Future Directions

Future research will focus on independent validation against radiosonde networks, separate modeling by sky conditions, and seasonal/interannual variability extension. Quantitative comparisons of SIMBA-generated Jacobians with RTM (e.g., RTTOV) outputs and interpretability improvements via regularization or post-hoc analysis are imperative. The framework's applicability to NWP assimilation will depend on proven robustness across operationally relevant scenarios and demonstrable consistency in both state and observation spaces.

Conclusion

SIMBA establishes a formal bidirectional retrieval-forward simulation paradigm for hyperspectral radiance-profile modeling, optimized for GIIRS and NWP-oriented applications (2606.19943). The cycle-consistent closed-loop architecture, leveraging bidirectional Mamba blocks, enhances retrieval accuracy, radiance reconstruction stability, and vertical/channel-wise modeling fidelity compared to standard deep learning baselines. While strongly aligned with data assimilation requirements, further validation and interpretability developments are required before operational adoption. This framework sets a new methodological foundation for physically coherent, joint retrieval-forward modeling in satellite hyperspectral remote sensing and assimilation.