MambaRain: Multi-Scale Mamba-Attention Framework for 0-3 Hour Precipitation Nowcasting
MambaRain: Multi-Scale Mamba-Attention Framework for 0-3 Hour Precipitation Nowcasting
Authors: Chunlei Shi, Cui Wu, Xiang Xu, Hao Li, Ni Fan, Xue Han, Yongchao Feng, Yufeng Zhu, Boyu Liu, Zengliang Zang, Hongbin Wang, Yanlan Yang, Dan Niu Date: 2026-05-14 Paper ID: openalex:2605.14606
Summary
MambaRain is a hybrid encoder-decoder framework designed to overcome performance degradation in precipitation nowcasting beyond 90 minutes. By integrating Mamba’s linear-complexity state space mechanisms for global temporal modeling with self-attention for explicit spatial dependency capture, the model effectively addresses the limitations of purely sequential or attention-based architectures. A custom spectral loss formulation further improves forecast sharpness, resulting in superior performance over current deterministic models in the 2-3 hour horizon.
Key Contributions
- Introduces MambaRain, a hybrid architecture combining selective state space models (Mamba) for temporal dynamics and self-attention for spatial correlation modeling.
- Enables effective 0-3 hour precipitation nowcasting, extending reliable prediction horizons beyond the traditional 2-hour window.
- Proposes a spectral loss formulation to minimize blurring artifacts in chaotic precipitation fields, preserving fine-scale motion information.
Open Questions & Future Work
Key Concepts
- mamba-attention-hybrid-spatiotemporal-modeling: An architectural pattern that integrates selective state-space models for global temporal dynamics with self-attention for local spatial feature extraction in spatiotemporal tasks.
Archivist Review
I approved the hybrid architecture pattern rather than the specific model name to ensure reusability. I also approved the uncertainty bottleneck in nowcasting as it is a fundamental challenge for the transition from deterministic to operational-grade weather AI. The spectral loss was rejected as it is a common signal processing technique applied in a paper-local manner.
Approved Concepts
- Mamba-Attention Hybrid Spatiotemporal Modeling: Provides a concrete structural blueprint for combining the temporal efficiency of selective state-space models with the spatial fidelity of self-attention.
Approved Open Questions
- Probabilistic Nowcasting Uncertainty Bottleneck: Deterministic point estimates in chaotic systems often miss the multi-modal nature of extreme weather, making uncertainty quantification a vital next step for operational adoption.
Rejected Candidates
- [concept] MambaRain (
mambarain) - paper_local: Specific architecture names are typically paper-local; the underlying hybrid approach is more broadly reusable.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.14606
- paper_id
- 2605.14606
- paper_source
- openalex
- domain
- nlp
- tags
- nlp
- architectures
-
- datasets
-
- concept_slugs
- mamba-attention-hybrid-spatiotemporal-modeling
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-17T06:09:07Z