HorizonUNet: A Structure-Aware Spatiotemporal Mamba Architecture for Extreme Precipitation Nowcasting

Abstract

Determining precipitation nowcasting in severe weather conditions is still a difficult problem because of the complicated and anisotropic nature of convective storm clouds and the drawbacks associated with the current state-of-the-art deep learning algorithms. Existing solutions face a trade-off between temporal error accumulation and spatial averaging with respect to ConvRNN-based architectures, the $O\left(N^{2}\right)$ computational complexity of transformers, and high temporal latency and instabilities of generative models. This paper introduces HorizonUNet, a spatiotemporal architecture combining the linear-time efficiency of Mamba2 State Space Models with a hierarchical encoder-decoder design. A Mamba2D block with independent horizontal and vertical scans, fused through a learnable gating mechanism, enables structure-aware modeling of meteorological patterns. The proposed model jointly processes multimodal SEVIR data, specifically radar VIL and IR069, to produce t+30 and t+60 forecasts simultaneously in a single pass, using a physics-informed loss. In SEVIR 2018 dataset, the model achieves results of $\mathbf{P O D}_{\text {heavy }}=0.70$ at $\mathbf{t + 6 0}$ and $\mathbf{C S I}_{\text {heavy }}=0.20$, with a structural location error $\mathbf{S A L}_{\mathbf{L}}=0.056$.