This AI Framework for Satellite Data Doesn't Just Predict Weather — It Reconstructs What It Missed

Hyperspectral infrared observations from satellites like China's FY-4A GIIRS are critical for numerical weather prediction (NWP) because they capture vertical temperature and humidity profiles. Most existing deep learning approaches only perform one-way retrieval from radiances to atmospheric states, ignoring the reverse simulation process and consistency between the two spaces. A new paper on arXiv introduces SIMBA (State-space Integrated Modeling with Bidirectional Alignment), a unified framework that does both: it retrieves atmospheric profiles from radiances and simultaneously reconstructs radiances from those profiles. The novelty lies in a cycle-consistency loss that enforces coupling between the two directions, and a bidirectional Mamba state-space module that captures long-range dependencies across pressure levels.

Using collocated FY-4A GIIRS observations and ERA5 reanalysis data, SIMBA was evaluated on four tasks: temperature retrieval, specific humidity retrieval, long-wave radiance reconstruction, and medium-wave radiance reconstruction. It consistently outperformed several deep learning baselines (e.g., Transformer-based models) in both retrieval accuracy and reconstruction quality. Ablation studies confirmed that the bidirectional design and cycle-consistency mechanism each contribute significantly to performance. The authors note that SIMBA's structure is amenable to future Jacobian analysis and direct integration into NWP data assimilation systems, potentially enabling more accurate and physically consistent use of satellite data for weather forecasting.