⚡ electrical engineering

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

This paper introduces SIMBA, a novel bidirectional deep learning framework that jointly performs atmospheric profile retrieval and hyperspectral infrared radiance simulation for FY-4A GIIRS data, utilizing cycle-consistency constraints and a bidirectional Mamba architecture to outperform existing methods for numerical weather prediction applications.

Original authors: Jingdong Shen, Fu Wang*, Qifeng Lu, Hao Huang, Chunqiang Wu, Chi Yang, Xiaofang Liu

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jingdong Shen, Fu Wang*, Qifeng Lu, Hao Huang, Chunqiang Wu, Chi Yang, Xiaofang Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Earth's atmosphere as a giant, multi-layered cake. Meteorologists need to know exactly how much heat (temperature) and moisture (humidity) are in each layer of this cake to predict the weather accurately.

For decades, scientists have used satellites to look at this cake from space. However, the satellite doesn't see the cake layers directly; it only sees the "glow" (infrared light) coming off the top. The challenge is figuring out what the cake looks like inside just by looking at the glow. This is called retrieval.

Traditionally, scientists have used complex physics equations (like a very strict recipe book) to guess the cake layers from the glow. But these recipes are slow and computationally heavy. Recently, computers have started using "deep learning" (AI) to learn this recipe faster. However, most AI models only learn one way: Glow $\rightarrow$ Cake. They guess the layers, but they don't check if their guess actually produces the original glow they started with.

The New Solution: SIMBA

The authors of this paper created a new AI framework called SIMBA. Think of SIMBA as a two-way translator that acts like a strict editor.

Instead of just guessing the cake layers from the glow, SIMBA does two things at once:

The Translator (Retrieval): It looks at the satellite's glow and guesses the temperature and humidity layers of the atmosphere.
The Editor (Forward Simulation): It takes that guess and tries to "re-generate" the glow. It asks, "If the atmosphere really looked like my guess, what would the satellite see?"

The "Cycle-Consistency" Trick

Here is the clever part: SIMBA forces these two steps to check each other.

If the AI guesses the layers, but when it tries to re-generate the glow, the result looks nothing like the original satellite image, the AI knows it made a mistake.
It then goes back and fixes its guess.

The paper calls this a "cycle-consistency constraint." Imagine you are trying to translate a sentence from English to French and back to English. If the final English sentence doesn't make sense, you know your French translation was wrong. SIMBA uses this loop to ensure its guesses are physically consistent with reality.

The "Mamba" Brain

To handle the atmosphere, which has 101 different vertical layers (like 101 floors in a skyscraper), the AI needs to understand how the top floor affects the bottom floor.

Old AI models (like LSTMs) sometimes get tired or confused when looking at long sequences of data.
SIMBA uses a new type of AI brain called Mamba. Think of Mamba as a super-efficient librarian who can instantly remember the relationship between the top floor and the bottom floor of the skyscraper without getting overwhelmed. This allows SIMBA to understand the "vertical dependencies" of the weather very well.

What Did They Find?

The researchers tested SIMBA using data from China's FY-4A satellite (which has a special instrument called GIIRS that sees infrared light) and compared it against other AI models.

Better Guesses: SIMBA was more accurate at guessing the temperature and humidity of the atmosphere than the other models.
Better Consistency: Because of the "two-way" check, the guesses SIMBA made were more consistent with the actual satellite light.
Cloudy vs. Clear: The model was trained on cloudy days (which are messy and hard to predict) but worked surprisingly well on clear days too, without needing to be retrained.
Efficiency: It is fast and doesn't require massive amounts of computer memory compared to some other complex models.

The Bottom Line

The paper claims that SIMBA is a better tool for turning satellite light into weather data because it doesn't just guess; it double-checks its own work by simulating the result. This makes the data more reliable for weather forecasting systems.

Important Note: The authors are careful to say that while SIMBA is a powerful "differentiable" tool (meaning it can be used in advanced math calculations for weather models), it is not a perfect replacement for the traditional physics equations yet. It is a data-driven assistant that needs further testing against real-world ground measurements (like weather balloons) before it can be fully trusted for daily weather forecasting operations.

Technical Summary: SIMBA – A Bidirectional Retrieval–Forward Simulation Framework

Problem Statement
Hyperspectral infrared observations, such as those from the Geostationary Interferometric Infrared Sounder (GIIRS) on the FY-4A satellite, provide critical vertical structure information for Numerical Weather Prediction (NWP). However, existing deep learning approaches for atmospheric profile retrieval predominantly focus on one-way mapping from radiances to atmospheric states. These methods often neglect the reverse process (forward simulation from profiles to radiances) and fail to explicitly enforce consistency between the atmospheric state space and the radiance observation space. Furthermore, traditional models struggle to capture the long-range vertical dependencies inherent in thermodynamic profiles, and the physical coherence of learned mappings is often insufficient for rigorous data assimilation applications.

Methodology
The authors propose SIMBA (Simulation-constrained Bidirectional Mamba framework), a unified architecture designed to jointly perform atmospheric profile retrieval and radiance reconstruction. The framework is built upon the following core components:

Bidirectional Architecture: SIMBA consists of two coupled branches:
1. Retrieval Branch: Encodes Long-Wave (LW) and Medium-Wave (MW) radiance observations (along with auxiliary variables) to estimate atmospheric temperature and specific humidity profiles.
2. Forward Simulation Branch: Encodes atmospheric profiles to reconstruct LW and MW radiances.
Cycle-Consistency Constraint: The two branches are linked via a cycle-consistency mechanism. Retrieved profiles must be able to reconstruct the original input radiances, and reconstructed radiances from reference profiles must recover the corresponding atmospheric states. This enforces consistency in both state and observation spaces.
Bidirectional Mamba Backbone: To address the challenge of modeling vertical dependencies across 101 pressure levels, the framework employs a Bidirectional Mamba state-space module. This architecture captures long-range dependencies with linear computational complexity, outperforming traditional RNNs and Transformers in efficiency and vertical sequence modeling.
FiLM Conditioning: A Feature-wise Linear Modulation (FiLM) mechanism injects radiative and geometric information into the profile decoder, dynamically adjusting the hidden states based on the specific observation characteristics.
Training Data: The model is trained on collocated FY-4A/GIIRS radiance data and ERA5 reanalysis profiles. The dataset includes both cloudy-sky (158,649 samples) and clear-sky (112,151 samples) conditions, with the model trained on cloudy data and evaluated on both.

Key Contributions

Unified Framework: Introduction of a bidirectional retrieval–forward simulation framework that jointly links atmospheric state space and radiance observation space for FY-4A/GIIRS hyperspectral data.
Cycle-Consistency Mechanism: Implementation of a cycle-consistency constraint to couple retrieval and forward branches, improving observation-space consistency and the stability of the radiance–profile mapping.
Vertical Dependency Modeling: Integration of a Bidirectional Mamba module to effectively model long-range vertical dependencies in thermodynamic profiles, demonstrating superior performance in both retrieval and reconstruction tasks.
Differentiable Forward Modeling: Provision of a data-driven, differentiable forward-modeling basis that allows for potential future Jacobian-related analysis and sensitivity studies, while acknowledging the necessity of rigorous Radiative Transfer Model (RTM) validation.

Experimental Results
SIMBA was evaluated against bidirectional variants of CNN, MLP, Transformer, and LSTM (BiCNN, BiMLP, BiTransformer, BiLSTM) under cloudy and clear-sky conditions.

Retrieval Performance: Under cloudy-sky conditions, SIMBA achieved the lowest errors for temperature (RMSE: 1.0032 K, MAE: 0.7131 K) and specific humidity (RMSE: 0.5684, MAE: 0.2530), outperforming the best baseline (BiLSTM) by approximately 3.0% and 1.5% respectively. It also maintained superior performance on clear-sky samples without retraining.
Reconstruction Performance: SIMBA demonstrated the best consistency in radiance reconstruction for both LW and MW bands, achieving the lowest RMSE and highest $R^2$ values. The model showed particularly stable performance across different spectral channels, including those sensitive to the middle and upper troposphere.
Ablation Studies: Comparisons with single-direction variants (Retrieval-Mamba and Forward-Mamba) confirmed that the bidirectional closed-loop design significantly improves both retrieval accuracy and radiance reconstruction consistency, rather than gains being solely attributable to the Mamba backbone.
Efficiency: SIMBA operates with 1.13M trainable parameters and an inference time of 0.157 ms per sample on an NVIDIA RTX 4090, offering a balance of accuracy and computational efficiency.

Significance and Limitations
The paper posits that SIMBA offers a more physically consistent relationship between hyperspectral observations and atmospheric states, aligning better with the needs of NWP data assimilation. The "validate-by-reconstruction" strategy provides a methodological reference for ensuring geophysical estimates remain consistent with their corresponding observations.

However, the authors maintain a modest stance regarding the framework's current operational readiness:

Not a Replacement for RTMs: SIMBA is presented as a data-driven differentiable forward operator, not a complete replacement for physical RTMs. Residual errors remain non-negligible.
Validation Scope: The reported accuracy reflects consistency with ERA5 reanalysis, not independent validation against in situ radiosonde observations. The spatial scale mismatch between GIIRS footprints and ERA5 grids may introduce representativeness errors.
Physical Interpretability: While the framework is differentiable, the resulting channel sensitivities (Jacobians) currently contain high-frequency noise, limiting their immediate physical interpretability.
Future Work: The authors identify the need for independent validation using radiosondes, separate modeling for clear and cloudy skies, and quantitative comparisons with physical RTMs (e.g., RTTOV) before the framework can be considered for operational assimilation.